📚 Full Archive

Artificial intelligence (AI), particularly large language models (LLMs), is increasingly applied to radiographic interpretation in healthcare. In dentistry, radiographic imaging is essential for diagnosis and treatment planning, yet remains subject to variability and human error. AI may enhance diagnostic accuracy and consistency.

To evaluate and compare the diagnostic accuracy, consistency, and interpretive performance of multimodal AI models-ChatGPT, Grok, and MANUS-with expert radiologists in dental radiograph interpretation.

A total of 120 anonymised radiographs (40 OPGs, 40 periapical, 40 CT slices) were selected from validated academic sources. Two board-certified oral and maxillofacial radiologists established gold standard diagnoses. Each image was independently assessed by the three AI models under standardised prompting. Diagnostic accuracy and intra-model consistency were analysed using descriptive statistics, Cohen's kappa, McNemar's test, and logistic regression.

In the first assessment, MANUS and ChatGPT achieved 92.5% accuracy (111/120), while Grok reached 88.3% (106/120). Performance improved in the second round: MANUS 95.0%, ChatGPT 93.3%, and Grok 90.8%, compared with 96.7% for radiologists. ChatGPT showed the highest reproducibility (κ = 0.937), whereas MANUS demonstrated the highest overall accuracy. Strong agreement was observed between ChatGPT and MANUS, with greater variability in Grok. No significant systematic bias was detected between AI outputs and radiologist benchmarks.

The evaluated LLMs demonstrated diagnostic performance comparable to expert radiologists. MANUS excelled in accuracy and ChatGPT in reproducibility, supporting their potential as adjunct tools in dental radiology, while maintaining the need for expert oversight.Clinical trial number: Not applicable.

Large language models (LLMs) show promise in automatically detecting errors in radiology reports, but their performance remains insufficiently validated in large-scale, real-world clinical datasets.

This study aimed to systematically evaluate the performance of LLMs in detecting and correcting errors in Chinese radiology reports derived from authentic clinical data.

A large-scale dataset of 4480 Chinese radiology reports with modification records containing real clinical practice-generated errors was retrospectively collected between January 2023 and June 2024 at a single institution. After exclusions, 1363 reports containing 1551 errors were included. The dataset covers various anatomical parts of the body from different imaging modalities and was randomly divided into a test set (n=1263) and an internal validation set (n=100). Additionally, 100 error-free reports were added to the internal validation set. An additional 200 English-language reports from the Medical Information Mart for Intensive Care (MIMIC-III) were used for external validation. Eight human readers and 8 widely adopted LLMs, enhanced by prompt engineering, were tasked with error detection. Overall and subgroup detection performance and reading time were evaluated. Correction suggestions from the 2 best-performing LLMs were reviewed by a senior radiologist.

On the test set, DeepSeek-R1 achieved the highest overall detection rate at 89% (95% CI 87%-90%), significantly better than the other 7 models (P=.001-.007). On the internal validation set, DeepSeek-R1 and Claude-3.5-Sonnet achieved detection rates of 83% (100/120; 95% CI 76%-89%) and 80% (96/120; 95% CI 72%-86%), respectively. DeepSeek-R1 showed performance comparable to radiologists (83%, 95% CI 76%-89% vs 80%, 95% CI 72%-86% for junior radiologists and 78%, 95% CI 70%-85% for senior radiologists; P=.39 and P=.19, respectively) and significantly better performance than that of nonradiologists and nonphysicians (83%, 95% CI 76%-89% vs 66%, 95% CI 57%-74% and 38%, 95% CI 30%-47%; P<.001, respectively). DeepSeek-R1 showed a false-positive rate comparable to radiologists (DeepSeek-R1 vs senior radiologists and junior radiologists, 3% vs 0% and 1%; P=.25 and P=.61, respectively) and a significantly lower rate than nonradiologists and nonphysicians (3% vs 13% and 17%; P=.02 and P=.002, respectively). On the external validation set, DeepSeek-R1 and Claude-3.5-Sonnet achieved detection rates of 94% (95% CI 89%-97%) and 93% (95% CI 88%-97%), respectively. The correction accuracy of DeepSeek-R1 and Claude-3.5-Sonnet was 95% and 91%, respectively.

Enhanced LLMs, particularly DeepSeek-R1, demonstrated robust performance in error detection and correction within real-world Chinese radiology reports, supporting their clinical use for automated quality assurance and integration into workflows to improve reporting accuracy and efficiency.

Information extraction (IE) from clinical texts has advanced rapidly with recent advances in natural language processing, particularly the advent of large language models (LLMs). However, inconsistent and incomplete reporting of methodologies limits reproducibility, comparability, and clinical translation. We aimed to develop a consensus-based reporting guideline tailored to clinical IE studies.

We developed the Clinical Information Extraction Reporting Guideline (CINEX) through a multi-phase process. The initiative was prospectively registered on the EQUATOR Network as a reporting guideline under development, and a detailed Delphi study protocol was published in advance. A scoping review informed an initial set of items, which was refined through a 3-round electronic Delphi study with 20 international experts, followed by a final consensus meeting. Items were iteratively refined based on predefined inclusion criteria and expert feedback.

The final CINEX guideline comprises 29 checklist items grouped into 5 domains: information model, architecture, data, annotation, and outcomes. The 3 eDelphi rounds included 20, 15, and 12 experts, respectively. Two items were added after round one. Consensus for inclusion was reached for a 21 and additional 7 items after rounds 2 and 3, respectively.

CINEX provides a structured framework to improve transparency, reproducibility, and interpretability in clinical IE research. By standardizing reporting of key methodological components, such as data provenance, annotation processes, and evaluation strategies, it facilitates meaningful comparison across studies and supports safer clinical implementation.

CINEX complements existing AI reporting standards by addressing domain-specific challenges in clinical IE.

Radiology reports are written for clinicians, leaving the majority of patients unable to understand their own imaging findings. Large language models (LLMs) offer a means to simplify reports for patient-facing use; however, most implementations rely on cloud-based platforms that raise data privacy and regulatory concerns. Arabic-speaking populations - over 400 million people worldwide - remain substantially underserved in medical AI research, with no prospectively evaluated AI Arabic radiology report simplification system previously reported.

This study aimed to conduct a preliminary prospective feasibility and safety evaluation of patient-perceived understandability outcomes and radiologist-assessed clinical safety of an on-site, institutionally governed AI system that generates simplified radiology reports in plain English and Arabic for Arabic-speaking outpatients.

In this prospective single-center observational study conducted at King Faisal Specialist Hospital and Research Centre - Jeddah (KFSHRC-J) in January 2026, 98 adult outpatients (mean age 50.0 ± 14.9 years; 50 men, 48 women; IRB #2251488) reviewed three versions of their own radiology report in randomized order: a traditional radiologist report, an AI-simplified English report, and an AI-simplified Arabic report. The system used a two-stage pipeline: Qwen3-14B-FP8 with structured prompt engineering for English simplification (Stage 1), and a fine-tuned Hala-1.2B model for Arabic translation (Stage 2; 8,319 training samples). Patient-perceived understandability was measured on a 5-point Likert scale (1 = Very Easy; 5 = Very Difficult). Three board-certified radiologists independently assessed clinical safety using a standardized seven-domain rubric; inter-rater agreement was quantified using ICC(2,k).

AI-simplified Arabic reports produced a median perceived understandability score of 1 (IQR 1-2) versus 5 (IQR 3-5) for traditional reports (p < 0.001; rank-biserial r = 0.97), with 94.9% (93/98) of patients preferring the Arabic simplified version. Safety evaluation demonstrated 96.9% (95/98) of reports safe for patient release, with a hallucination rate of 3.1% (n = 3; 1 unsafe, 2 safe by consensus). Inter-rater reliability was moderate by conventional thresholds (ICC 0.48-0.55) due to a ceiling effect from uniformly high safety scores (mean fidelity 4.8/5.0; 90.8% highest rating; SD = 0.3), with 91.2% absolute agreement on binary safety classification and Fleiss' kappa = 0.71 for safety.

In this preliminary prospective feasibility evaluation, a locally deployed, clinician- and AI team-built system significantly improved patient-perceived understandability of radiology reports for Arabic-speaking patients, with a radiologist-assessed safety profile that is promising but requires further validation before broad clinical deployment. The hallucination rate of 3.1% (1.0% unsafe) is lower than published pooled benchmarks, though direct comparisons are limited by methodological heterogeneity across studies (pooled 7.2% across 38 simplification studies; 1.12% for fine-tuned models). To our knowledge, this is among the first prospectively evaluated AI systems for Arabic radiology report simplification, highlighting the potential of on-site AI deployment as a safe, privacy-preserving approach for multilingual radiology communication, warranting further investigation of its contribution to health equity.

Photon -Counting CT (PCCT) provides higher resolution, reduced dose, and valuable spectral data, but generates a large number of images per study, underlining radiologist and Picture Archiving and Communication Systems (PACS) capacity. While no guidelines defining when PCCT should be preferred over conventional Energy Integrating Detector (EID) CT in neuroradiology, and manual decision-making being impractical, Natural Language Processing (NLP)-based large language models (LLMs) may help automate routing of CT requests to the most appropriate CT technology.

We conducted a retrospective study using Spanish-language neuroradiology CT requests retrieved from the Radiology Information System (RIS) (January 2012-October 2025). A random sample of 800 requests was independently labeled by two neuroradiologists as Basic Protocol (BP); feasible on EID-CT or basic PCCT) or Advanced Protocol (AP); requiring full-advanced PCCT), considering the gold standard an expert consensus (Cohen's k = 0.74). For discriminative modeling, transformer classifiers (Spanish bidirectional encoder representations from transformers (BERT), biomedical RoBERTa, Llama-3.1-8B, and Mistral-7B) were fine-tuned for binary BP/AP classification and evaluated with stratified five-fold cross-validation. For generative modeling, ChatGPT 5.2 and Gemini 3 were applied in a zero-shot setting using structured prompting and post-processed into binary outputs. Performance was assessed using accuracy, precision, recall, specificity, F1-score, and AUC.

Discriminative transformers achieved the best performance for the classification, with RoBERTa leading (accuracy 0.925, precision 0.906, recall 0.967, F1 0.935, AUC 0.920) and the lowest total error, without a significant difference versus Spanish BERT. In contrast, RoBERTa outperformed Llama, Mistral, Gemini, and ChatGPT with statistically significant improvements, while the generative LLMs yielded the poorest overall performance.

Discriminative transformer models, particularly RoBERTa, enabled highly accurate automatic identification of CT requests requiring AP to be performed at PCCT, substantially outperforming general-purpose generative LLMs. These findings support this exploratory proof-of-concept study of AI-driven eligibility routing to optimize PCCT utilization and streamline neuroradiology workflows.

Large language models (LLMs) show promise for converting complex radiology reports into patient-centric language, but inherent output instability may limit clinical application.

To quantitatively assess the translational accuracy, error rates, and instability of various LLMs when generating patient-centric radiology reports, and evaluate demographic influences on report readability.

This retrospective study evaluated 320 de-identified radiology reports processed by three LLMs using a two-stage (baseline and optimized) prompt engineering strategy. Two senior radiologists evaluated medical accuracy, completeness, and recommendation suitability. Readability was evaluated by 16 non-medical participants stratified by age and education.

Professional radiological evaluation revealed that all tested models exhibited inherent instability, omitted information, and tended to generate risk-averse, generalized clinical recommendations. To address these limitations, optimized structured prompts significantly reduced model output variance and improved translational accuracy, with particularly prominent effects observed in DeepSeek-R1 and ChatGPT-4.0. Overall, large language models significantly enhanced the readability of radiology reports (P < 0.05), with DeepSeek-R1 achieving the best performance. However, patients' self-reported comprehension of the reports was affected by demographic characteristics.

Large language models can effectively improve the readability of radiology reports, yet all such models inherently suffer from output instability and information omission. Optimized structured prompting can substantially reduce the variability of model outputs and improve the accuracy of medical text translation. Nevertheless, LLMs should currently be strictly confined to human-supervised auxiliary tools rather than applied as standalone clinical solutions.

To evaluate text-only versus vision-enabled performance of late-2025 large language models (LLMs) on the Japan Diagnostic Radiology Board Examination (JDRBE) and compare model performance with newly board-certified radiologists.

Image-based questions from the JDRBE 2021 and 2023–2025 were collected, and ground truth answers were determined by expert consensus. Four commercial multimodal LLMs were evaluated: Gemini 2.5 Pro (March 2025, baseline), Gemini 3 Pro, GPT-5.1, and Claude Opus 4.5 (all released in November 2025). Each question was answered with image input (“vision”) and without images (“text-only”). For the JDRBE 2025, subjective legitimacy of responses was independently rated by two radiologists using a five-point Likert scale, and low-rated responses were further analyzed by error type. Additional analyses on the JDRBE 2025 subset included image-shuffling and multi-run variability assessment (five runs). Model accuracies were also compared with those of five newly board-certified radiologists who passed the JDRBE 2025.

Gemini 3 Pro achieved the highest accuracy among all models, scoring 85.3% (279/327) in the vision condition and significantly outperforming its text-only accuracy (74.3%, P < 0.001). Gemini 2.5 Pro and Claude Opus 4.5 also improved with image input, whereas GPT-5.1 did not. For the JDRBE 2025, Gemini 3 Pro in the vision condition received the highest legitimacy ratings, and its accuracy (88%) was above the range observed in a reference group of five newly board-certified radiologists (65%–83%), but hallucination was still the most common error type. Image-shuffling analysis using the 2025 subset showed no performance gain in all models, supporting reliance on visual input. Multi-run variability analysis showed high agreement across runs.

Among late-2025 commercial LLMs, Gemini 3 Pro demonstrated board-level performance on the JDRBE through direct medical image interpretation. SECONDARY ABSTRACT: The performance of vision-enabled large language models on the Japan Diagnostic Radiology Board Examination was evaluated. Among the models released in November 2025, Gemini 3 Pro demonstrated significant capabilities in direct medical image interpretation, achieving accuracy above that of a reference group of five newly board-certified radiologists. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s11604-026-01983-x.

Poor cardiac MR image quality can prompt repeat examinations and hinder clinical decision-making.

To evaluate whether pre-imaging clinical information, extracted using a large language model (LLM), is independently associated with cardiac MR image quality. STUDY TYPE: Retrospective. POPULATION: 1006 adults undergoing clinical cardiac MR examinations. FIELD STRENGTH/SEQUENCE: 1.5 T and 3 T scanners with cine, black blood, MR angiogram, or late gadolinium enhancement protocols. ASSESSMENT: Image quality was categorized per study as excellent, slightly limited, severely limited, or nondiagnostic using institutional reporting conventions finalized by radiologists and cardiologists. A HIPAA-compliant LLM assigned image quality labels based on radiology reports through an iteratively refined prompt, with reliability confirmed by two radiologists. Labels were binarized as Good (excellent and slightly limited) versus Poor (severely limited and nondiagnostic). A repeat-imaging-adjusted image quality label was used in a sensitivity analysis. Pre-imaging clinical and patient variables were extracted from electronic health records. Associations between variables and image quality labels were investigated. STATISTICAL TESTS: Cohen's kappa (κ) for label agreement. Chi-square and t-tests for univariate analysis. Variance inflation factor (VIF) and multivariable logistic regression. Significance level: p < 0.05.

Binarized image quality labels showed substantial agreement with interpreters' assessments for both the primary dataset (κ = 0.689) and the repeat-adjusted dataset (κ = 0.879). There was no significant multicollinearity (VIF = 1.01-1.39). Cognitive and communication impairment (OR 1.81, 95% CI [1.30-2.54], p < 0.001) and respiratory issues (1.57 [1.14-2.17], p = 0.006) were independently associated with poor image quality. These associations remained significant in the repeat-adjusted sensitivity analysis (cognitive and communication impairment (OR 1.75, 95% CI [1.27-2.44], p < 0.001) and respiratory compromise (OR 1.37, 95% CI [1.04-1.82], p = 0.027)). Other clinical variables were not independently associated after adjustment. DATA

Cognitive/communication impairment and respiratory compromise were independently associated with poor cardiac MR image quality. TECHNICAL EFFICACY: Stage 2. This study examined why some cardiac magnetic resonance scans have poor image quality. Poor‐quality scans can limit diagnosis and may require repeat imaging. It analyzed adult examinations and used artificial intelligence (AI) to extract clinical information from the medical records available before the scan. Cognitive or communication difficulties and breathing‐related problems were associated with worse image quality. These findings might suggest that these patients may benefit from additional preparation before scanning. More broadly, this study shows that information in the medical record can be efficiently extracted using AI to support large‐scale clinical research.

To develop an ontology-driven framework that standardizes heterogeneous local radiology procedure names by decomposing them into semantic components and aligning them to LOINC/RSNA Radiology Playbook codes using constrained large language model (LLM)-based parsing and selection.

Radiology procedure names from two tertiary hospitals in Korea were parsed into semantic components using LLM prompting with retrieval-augmented generation, aligned with the LOINC/RSNA Radiology Playbook (version 2.80). Ontology-based similarity scoring quantify correspondence between parsed components and Playbook candidates' attributes, and retrieve the Top 10 candidates, followed by LLM-based selection within this candidate set. Performance was evaluated against direct Playbook code name-based mapping and conventional similarity metrics using a radiologist-curated gold reference.

A total of 3,326 local procedure names were analyzed. Ontology-based mapping approach substantially outperformed direct Playbook code name mapping across all evaluation metrics. At the candidate retrieval stage, ontology-based attribute matching achieved recall@5 of up to 0.78 (internal) and 0.89 (external), compared with 0.51 and 0.47 for direct mapping. After LLM-based selection, the ontology-based approach achieved a final selection recall@1 of up to 0.70 (internal) and 0.81 (external), exceeding direct mapping (0.48 and 0.50) more than 20 percentage points in both settings (p < 0.001).

Decomposing procedure names into ontology-grounded semantic components enables robust handling of heterogeneous local terminology, while constraining LLM reasoning to structured selection tasks mitigates hallucination and preserves semantic fidelity. Ontology-driven knowledge encoding provides a scalable and reliable approach to standardizing radiology procedure names, supporting cross-institutional interoperability and secondary data use of imaging research.

To evaluate whether large language models (LLMs) can generate accurate, clinically valid, and usable letters to appeal insurance denials for radiology procedures.

This pilot study generated insurance appeal letters for a simulated clinical scenario. Four LLMs (Claude 3.5, Nova Pro, Llama-3.1-70B, ChatGPT-4o) were used with zero-shot, few-shot, and retrieval-augmented generation (RAG) techniques. Four board-certified interventional radiologists, blinded to model and technique, scored letters for content (accuracy, personalization, references), grammar and structure (readability, tone, persuasiveness), and usability (estimated editing time, usefulness as a template). References were verified for accuracy, and outputs were carefully examined for hallucinations. Statistical analyses included ANOVA, Chi-square, and Fleiss' Kappa for interrater reliability.

Mean content and grammar scores were 3.9 ± 0.95 and 4.3 ± 0.9 (out of 5), with no significant differences by model or technique (p >.05). Reviewer agreement was poor (Fleiss' Kappa -0.18 for content, -0.085 for grammar). Hallucinations were flagged by reviewers in 16/48 assessments, significantly more often with the online model (ChatGPT-4o: 58% vs offline 25%; p =.03). Of 44 references, 80% from the offline models were fabricated compared with 29% from ChatGPT-4o (p <.001). Estimated editing time was less than 10 min in 71% of responses, and the reviewers felt the letters would be useful as templates in 73% of cases.

LLM-generated appeal letters for insurance denials were generally well received, with high usability and adequate quality. However, fabricated references and hallucinations remain prevalent, necessitating careful human review before clinical use.

Recent years have seen a rapid development of artificial intelligence (AI) tools to enhance radiologists' workflow, but most have focused on image acquisition and interpretation. Generating radiology reports is a critical element of practice but remains a source of inefficiency and cognitive load.

To investigate the performance of an AI based tool integrated into dictation software to automate generation of report impressions and compare it with radiologist generated impressions.

One hundred consecutive abdominal radiology reports were retrospectively selected in January 2024 from a single center. An AI-generated impression (GI) was created for each report and compared with the original radiologist impression (RI). Ten subspecialty abdominal radiologists evaluated the blinded, randomized pairs. Each impression was evaluated on a 5-point Likert scale for coherence, comprehensiveness, and factual consistency, and an overall preference for the impression was recorded. Cumulative link mixed models and binomial logistic regression were used where appropriate.

GI was preferred in 38% of reports (95% CI 32-45%), RI in 45% (95% CI 39-50%), with no preference in 17% of instances. Mixed-effects logistic regression demonstrated that GI was rated as equivalent or preferred to RI (odds ratio 1.34, 95% CI 1.004-1.788, p = 0.02). Radiologists rated GI equivalent or superior to RI in 79% of cases for coherence, 66% for comprehensiveness, and 77% for factual consistency.

The AI generated impressions were clinically acceptable and rated equivalent or superior to radiologist-generated impressions in the majority of cases across key quality metrics.

Artificial intelligence (AI) is increasingly proposed as a solution to improve efficiency in radiology and nuclear medicine, particularly in the context of workforce shortages. However, adoption of AI-based clinical decision support systems (AI-CDSS) remains slow, due to limited model transparency. Explainable AI (XAI) may improve clinician acceptance by supporting oversight and trust. This study evaluated the impact of different XAI explanation types on radiologists' willingness to adopt AI systems.

Ten nuclear medicine radiologists from eight UK institutions performed lung cancer TNM staging using whole-body PET/CT scans supported by a simulated AI-CDSS. Three explanation approaches were assessed: input feature attribution, high-level concept explanations and global algorithmic transparency. Adoption likelihood and explanation usefulness were rated using Likert scales and analysed with nonparametric sign tests. Semi-structured interviews were additionally analysed through thematic evaluation supported by large language model-assisted coding with human verification.

All explanation approaches significantly increased radiologists' willingness to adopt the AI system compared to a black-box model (p < 0.05). Explanations were consistently considered useful in enabling participants to confirm or challenge AI staging recommendations (p < 0.001). Qualitative findings highlighted the importance of clinical relevance, error detection and decision support value. A trade-off between explanation depth and usability was identified as a key factor influencing preferences.

Incorporating XAI into nuclear medicine CDSS enhances radiologists' acceptance and provides clinically meaningful information for oversight of AI recommendations, in accordance with the EU AI Act. These findings support the role of XAI in facilitating integration of AI tools into diagnostic workflows.

Large language models (LLMs) are becoming increasingly popular in medicine and neurosurgery. Because LLMs are not trained in specific subspecialties or diagnoses, a better understanding of the implications, effectiveness, and use of LLMs by users and clinicians is necessary. To better understand LLM's effectiveness in neurosurgery and aneurysms, we compared community and physician feedback on ChatGPT 4o and Gemini 1.5 Flash responses to frequently asked questions regarding brain aneurysms.

External surveys were made available on the Brain Aneurysm Foundation page for patients and families to complete and internal surveys were distributed and completed by physicians in the department of neurosurgery at Boston Children's Hospital.

In the community survey assessing response usefulness and helpfulness, ChatGPT and Gemini provided different response quality despite similar AI sentiment. Clarity of procedure explanation (p = 0.04), discussion of alternative procedures (p = 0.02), and discussion of procedure risks (p = 0.01) were different. The physician survey, assessing response accuracy, safety, and helpfulness, also found differences in multiple domains. Importantly, differences were found in consistency with current medical knowledge and practice guidelines (p = 0.001), omittance of key points (p < 0.001), and amount of clinically relevant detail included (p < 0.001).

LLMs had variable performance across several key domains, consistent with previous research. Despite the apparent advantages of ChatGPT, physician feedback highlighted the continued need for information oversight. Interestingly, community participants consistently found LLM responses to be better than physician ones, while physicians found LLM responses to be similar or somewhat worse than the one they would have provided.

To develop and externally validate integrated models combining three-dimensional (3D) volumetric segmentation and large language model (LLM)-based slice-wise classification for improved detection of clinically significant Prostate Cancer (csPCa) on multiparametric Magnetic Resonance Imaging (mpMRI).

This retrospective multi-institutional study included 5050 patients (3896 for model development and 1154 for external validation) who underwent mpMRI. A 3D V-Net was trained for voxel-wise csPCa segmentation, generating three patient-level metrics: positive volume, positive slice count, and positive slice rate. A four-billion-parameter MedGemma-IT LLM was fine-tuned for slice-level classification, producing positive slice count and rate metrics. The optimal V-Net and LLM metrics were integrated into two combined models: Combined Model 1 (logistic regression of V-Net and LLM positive slice rates) and Combined Model 2 (zero-adjusted Gamma regression of V-Net positive volume and LLM positive slice rate). Performance was evaluated in the external validation cohort using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), decision curve analysis (DCA), net reclassification improvement (NRI), and integrated discrimination improvement (IDI).

In external validation, Combined Model 1 achieved an AUROC of 0.900 (95% CI: 0.882-0.918) and Combined Model 2 achieved 0.885 (95% CI: 0.865-0.905), significantly outperforming all individual V-Net and LLM metrics (ΔAUROC: 0.060-0.145; all P < 0.001). Both combined models demonstrated significant improvements in NRI and IDI for most comparisons (P < 0.05 for most) and provided superior standardized net benefit across clinically relevant risk thresholds (0.3-0.9) on DCA.

Integrating 3D volumetric segmentation with LLM-based slice-wise classification significantly improves csPCa detection accuracy and clinical utility compared to single-modality approaches, demonstrating promising generalizability in an independent, multi-institutional external validation cohort with substantial scanner and protocol heterogeneity.

To evaluate whether a custom agentic artificial intelligence (AI) pipeline can overcome the limitations of general-purpose large language model tools, when compared with a generic commercial tool (Google NotebookLM [NBLM]), for generating podcast-style summaries of radiology research articles.

Twenty-two PDF-format original research articles published in the April 2025 issue of Radiology were processed using our Programmable, Phoneme-Aware PDF-to-Podcast Pipeline (P5) and NBLM to generate 44 audio episodes. P5 utilizes a multi-agent workflow for script generation, quality assurance, pronunciation enhancement, and audio synthesis. Four radiologists from a pool of 25 (7 generalists and 18 specialists) were randomly assigned to evaluate each blinded audio episode, yielding 176 total evaluations. The primary outcomes were the number of hallucinations (factual errors) per episode and the percentage of hallucination-free episodes. Secondary outcomes included the number of inappropriate statements, mispronunciations, and flow disruptions; the composite quality score (Quality Assessment of Educational Podcasts [QAEP]); the key results coverage score; and overall listener preference. Data were analyzed using generalized linear mixed models.

The P5 method produced significantly fewer hallucinations per episode compared with NBLM (mean, 0.32 vs. 0.93; P < 0.001) and a higher proportion of hallucination-free episodes (71.6% [63/88] vs. 56.8% [50/88]; P = 0.013), consistently across generalists and specialists. P5 demonstrated significantly fewer mispronunciations (mean, 0.11 vs. 1.62; P < 0.001) and flow disruptions (mean, 0.19 vs. 1.06; P < 0.001) per episode, as well as higher mean QAEP composite scores (4.62 vs. 4.28; P < 0.001), compared with NBLM, consistently across generalists and specialists. Raters preferred P5 over NBLM in 72.7% (64/88) of comparisons (P = 0.003).

Our custom agentic AI pipeline generated podcast-style summaries of radiology research articles with significantly higher quality and greater listener preference than the generic commercial tool.

Radiologists often employ diagnostic certainty phrases (DCPs) to convey levels of confidence in imaging interpretations. Prior research in adult radiology demonstrated wide variability in DCP usage, potentially complicating communication with clinicians and patients. Little is known about these practices in pediatric radiology. We aimed to characterize DCP use among pediatric radiologists in a large academic institution.

We retrospectively analyzed radiology reports from a freestanding pediatric hospital system between October 2023 and September 2024. From 309,751 pediatric imaging reports meeting inclusion criteria, 10,000 were randomly sampled to identify DCPs using a natural language processing pipeline (GPT-4o, OpenAI, San Francisco, California). After exclusions and manual review, 122 unique DCPs were identified. We then counted the occurrence of these 122 DCPs across the remaining 309,751 radiology report first impressions. Usage rates were evaluated across patient demographics, imaging modalities, clinical settings, and radiologist seniority. Statistical significance was assessed using two-sided one-sample t tests with significance assigned to values P ≤ .001.

Of the 309,751 analyzable reports, 20.9% contained at least one DCP in the first impression. The mean DCP frequency was 0.27 per report. MRI (0.54, P < .001), patients aged 1 to 5 years (0.34, P < .001), and emergency cases (0.31, P < .001) showed significantly higher usage. Midcareer and early-career radiologists employed more DCPs (0.30, P < .001) than senior radiologists (0.22, P < .001). The most common phrases included "likely," "could represent," and "cannot rule out."

DCPs are frequently used in pediatric radiology, with variability influenced by modality, patient age, clinical context, and radiologist seniority. Such variability may compound existing uncertainty of the clinical diagnostic process and possibly indicate an opportunity for standardized uncertainty reporting approaches.

Nomograms predicting the likelihood of sentinel lymph node (SLN) metastasis in early-stage breast cancer can aid surgical decision-making but are underused due to the burden of variable review and input. This study evaluated whether OpenAI's large language models (LLMs) can extract information required for nomogram use and reproduce the estimated rate of SLN metastasis using the Memorial Sloan Kettering and MD Anderson nomograms.

The study analyzed the de-identified radiology and pathology notes for 20 patients. Three prompts were tested: (1) o1 prompted to generate SLN metastasis estimates without nomograms, (2) o1 prompted to use the nomogram from online calculators with and without serial corrections, and (3) GPT-4o prompted to use chain-of-thought reasoning from nomogram variables and the corresponding point values from a pictorial nomogram. Artificial intelligence (AI) estimates of SLN metastasis rates were compared with a physician-expert's manual use of nomograms.

OpenAI o1 captured all clinical variables in 65% of cases without serial correction (94-95% of individual variables) and 80% of cases with serial correction (96-97% of individual variables), with tumor size most frequently misidentified. Agreement between LLM-generated and physician-calculated risk prediction was low (exact matches in 0-10% of cases, near agreement in 20-45% of cases), indicating moderate rater reliability (intraclass correlation coefficient [ICC], 0.57-0.62). The optimized GPT-4o prompt demonstrated greater agreement (exact matches in 25% and near agreement in 90% of cases) and reliability (ICC, 0.95).

Large language models can reliably extract nomogram inputs from clinical notes but require task-specific prompt engineering for accurate SLN metastasis risk estimation. Currently, automating nomogram-based risk estimators with AI may not justify the significant resources required for optimization.

MRI is important for cardiac disease evaluation, but accurate diagnosis remains challenging in less experienced centers. Although large language models (LLMs) have shown promise in medical imaging diagnosis, their application in cardiac MRI is limited. HYPOTHESIS: LLMs may be effective in achieving cardiac MRI diagnosis based on standardized descriptions. STUDY TYPE: Retrospective. POPULATION: A total of 203 hypertrophic cardiomyopathy, 186 dilated cardiomyopathy, 46 hypertensive heart disease, 198 ischemic cardiomyopathy, 38 constrictive pericarditis, 45 cardiac amyloidosis, 91 myocarditis, and 144 normal controls. FIELD STRENGTH/SEQUENCES: Balanced steady-state free-precession, short tau inversion recovery, and breath-hold inversion-recovery segmented gradient-echo sequences at 3.0 T. ASSESSMENT: Clinical and cardiac MRI information from each subject was converted into standardized descriptions and input into Generative Pre-trained Transformer-4.5 (GPT-4.5), GPT-4 Omni (GPT-4o), Deepseek-V3, and Deepseek-R1 LLMs. Cardiac MRI information included LV function, wall thickness and motion, and abnormalities on T2WI, perfusion and late gadolinium enhancement sequences. Each model was asked to generate an imaging diagnosis. In addition, a medical student (8 months experience) and three radiologists (junior, mid-level and senior: with 3, 6, and 10 years' experience, respectively) provided diagnoses based on cardiac MRI images and clinical information. STATISTIC TESTS: Frequency-weighted sensitivity and specificity were calculated. The diagnostic performances of the LLMs and human readers were compared using the McNemar test with Bonferroni correction. A p value < 0.05 was considered significant.

All LLMs showed excellent frequency-weighted specificity (0.973-0.983). The frequency-weighted sensitivities of all LLMs were not significantly different from that of the junior radiologist, were significantly higher than that of the medical student, and significantly inferior to those of the senior radiologist (GPT-4.5: 0.863, GPT-4o: 0.821, Deepseek-V3: 0.843, and Deepseek-R1: 0.851 vs. junior radiologist: 0.850, all adjusted p = 1.000; vs. medical student: 0.731, all adjusted p < 0.001; vs. senior radiologist: 0.942, all adjusted p < 0.001). Additionally, the mid-level radiologist achieved a frequency-weighted sensitivity of 0.895, outperforming all LLMs except GPT-4.5. DATA

LLMs may generate accurate diagnoses from standardized cardiac MRI descriptions, potentially benefiting less experienced physicians. TECHNICAL EFFICACY: Stage 5. Reading cardiac MRI scans can be difficult, especially for doctors who have less clinical experience. New artificial intelligence tools called large language models (LLMs) may help support this process. In our study, we found that several LLMs were able to make diagnostic judgments at a level similar to junior radiologists. This suggests that LLMs could be used as supportive tools to provide an initial interpretation of cardiac MRI examinations. Such assistance may help improve diagnostic efficiency, reduce workload, and promote more consistent early evaluation, particularly in settings where experienced cardiac imaging specialists are not readily available.

Large language models (LLMs) that can process both images and text are increasingly being used in radiology. This study aimed to evaluate the performance of LLMs including GPT-4 Omni (GPT-4o), Claude-3.5-Sonnet (Claude), and Gemini 1.5 Pro (Gemini) in differentiating benign and malignant nodules in liver US cases and compare it with that of human readers.

Four hundred liver US cases with pathologically confirmed liver nodules visible on B-mode US from January 2020 to November 2024 were randomly selected in this retrospective study. They were divided into a development set (n = 100) and a test set (n = 300). Five prompt groups for LLMs including US image [I-only], image description [D-only], image and description [I+D], image and liver US e-textbook [I+T], and image and medical history [I+H] were evaluated to identify the optimal input in development set. In test set, accuracy of LLMs in differentiating benign and malignant liver nodules was compared with that of human readers using McNemar's test.

In development set, the prompt group I+H for all LLMs exhibited the highest diagnostic accuracy in differentiating benign and malignant liver nodules, being considering as the optimal input (taking GPT-4o as an example, with I-only, 57.0% [as reference]; D-only, 62.0%, p = 0.55; I+D, 62.0%, p = 0.54; I+T, 62.0%, p = 0.36; I+H, 77.0%, p = 0.01). In test set, LLMs with I+H outperformed junior group and showed similar accuracy to senior group (Junior, 70.0% [as reference1]; Senior, 78.3% [as reference2]; GPT-4o, 83.3%, P1 < .001, P2 = .10; Claude, 77.0%, p1 = 0.04, p2 = 0.72; Gemini, 75.3%, p1 = 0.14, p2 = 0.36).

Large language models with US image and medical history inputs achieved accuracy comparable to senior radiologists and superior to junior radiologists in differentiating benign and malignant liver nodules.

To evaluate large language models (LLMs) against supervised baselines for fine-grained, lesion-level detection of incidentalomas requiring follow-up, addressing the limitations of current document-level classification systems.

We utilized a dataset of 400 annotated radiology reports containing 1623 verified lesion findings. We compared two supervised transformer-based encoders (BioClinicalModernBERT, ModernBERT) against four generative LLM configurations (Llama 3.1-8B, Fine-tuned Llama 3.1-8b, GPT-4o, GPT-OSS-20B). We introduced a novel inference strategy using lesion-tagged inputs and anatomy-aware prompting to ground model reasoning. Performance was evaluated using class-specific F1-scores.

The anatomy-informed GPT-OSS-20B model achieved the highest performance, yielding an incidentaloma-positive macro-F1 of 0.79. This surpassed all supervised baselines (maximum macro-F1: 0.70) and closely matched the inter-annotator agreement of 0.76. Explicit anatomical grounding yielded statistically significant performance gains across GPT-based models (p<0.05), while a majority-vote ensemble of the top systems further improved the macro-F1 to 0.90. Error analysis revealed that anatomy-aware LLMs demonstrated superior contextual reasoning in distinguishing actionable findings from benign lesions.

Generative LLMs, when enhanced with structured lesion tagging and anatomical context, significantly outperform traditional supervised encoders and achieve performance comparable to human experts. This approach offers a reliable, interpretable pathway for automated incidental finding surveillance in radiology workflows.

This study aims to establish a retrospective, single-centre, feasibility-oriented benchmark for medical imaging quality control(QC) and to evaluate the potential of multiple large language models for chest X-ray radiograph(CXR) technical QC and CT report consistency assessment, based on a relatively small, radiologist-annotated dataset derived from routine clinical practice.

This retrospective, single-centre study included 161 CXRs and 219 structured CT reports from routine clinical practice. Twelve labels were used for CXR QC, including eleven radiologist-defined error categories and one error-free label, while nine labels were used for CT report evaluation, including eight inconsistency categories and one error-free label. All cases were annotated using a radiologist consensus reference standard. Multiple large language models(LLMs) and multimodal large language models(MLLMs) were evaluated using Micro-F1 and Macro-F1 for CXR QC and expert-based Micro-F1 for CT report QC.

For CXR QC, Gemini 2.0 Flash showed the strongest performance, achieving robust category-level generalization, while GPT-4o and Qwen2.5-VL-72B-Instruct demonstrated more balanced but weaker performance. In CT report QC, DeepSeek-R1 achieved the highest recall (62.23%) and the best overall performance. Across models, protocol-report mismatches and metric inconsistencies were the most common error types.

This study presents an initial, feasibility-oriented multimodal benchmark for medical imaging quality control, showing that LLM's performance is highly task- and modality-dependent. Given the limited sample size, single-centre design, and Chinese-language scope, the findings support future multicentral validation and workflow-integrated evaluation, rather than immediate clinical deployment.

To evaluate the feasibility and limitations of real-world, text-only inference of PI-RADS v2.1 categories from prostate MRI reports using large language models, with lesion-level and zone-aware analysis.

This single-center retrospective study included 1,205 lesion-level entries from 1,118 patients derived from semi-structured prostate MRI reports after removal of all explicit PI-RADS elements. ChatGPT-4o was prompted to assign numeric PI-RADS categories based solely on report text. Agreement with radiologist-assigned reference categories was assessed using exact agreement, Cohen's κ, and class-wise metrics. Analyses were performed overall, by zone (peripheral vs transition), and using collapsed risk strata (1-2/3/4-5). Discordant cases were reviewed to identify error mechanisms and severity. Human interobserver agreement, intra-model reproducibility, temporal stability, and a paired model-version sensitivity analysis comparing ChatGPT-4o with GPT-5.2 were also evaluated.

Overall exact agreement was 72.9% (κ = 0.538; macro-F1 = 61.2%), with a systematic tendency toward overcalling. Agreement was higher in the peripheral zone than in the transition zone (κ = 0.476 vs 0.077, reference PI-RADS 3-5). PI-RADS 3 showed the lowest precision and recall, with frequent bidirectional misclassification. Collapsing categories improved agreement (κ = 0.610). Incorrect diffusion-weighted imaging subscores were the most common error mechanism, with zone-specific differences. Clinically high-impact downgrades of PI-RADS 4-5 to 1-2 were rare (1.6%). Human interobserver agreement was excellent (κ = 0.916-0.967). GPT-5.2 outperformed ChatGPT-4o in paired analyses but produced invalid outputs in a minority of cases.

Text-only large language models can infer radiologist-assigned PI-RADS v2.1 categories from real-world prostate MRI reports with moderate agreement, but performance is zone dependent and limited around PI-RADS 3, particularly in the transition zone. These models are best suited as supervised tools for quality control rather than autonomous decision-making.

To develop and externally validate ultrasound nomograms combining BI-RADS features and quantitative morphometric characteristics, and to compare their performance with expert radiologists and large language models in biopsy recommendation and malignancy prediction for breast lesions.

In this multi-center, multi-national study, 1747 women with breast lesions underwent ultrasound across three centers in Iran and Turkey. A total of 10 BIRADS and 26 morphological features were extracted from each lesion. Three nomograms based on BI-RADS, morphometric, and both feature sets were constructed. Three radiologists (one senior, two general) and two ChatGPTs including ChatGPT-o3 and o4-mini-high interpreted de-identified breast lesion images. Diagnostic performance for biopsy recommendation and malignancy prediction was assessed across all cohorts.

According to the pooled results, although the difference between the fused nomogram and the BI-RADS version was not statistically significant, the fused version consistently outperformed all models in biopsy recommendation and malignancy prediction (AUCs of 0.901 and 0.853, respectively) compared to BI-RADS nomogram (AUCs of 0.898 and 0.834), morphometric nomogram (AUCs of 0.825 and 0.708), radiologist1 (AUCs of 0.820 and 0.729), radiologist2 (AUCs of 0.605 and 0.719), radiologist3 (AUCs of 0.728 and 0.699), ChatGPT-o3 (AUCs of 0.729 and 0.689), and o4-mini-high (AUCs of 0.713 and 0.695).

The proposed BI-RADS-morphometric nomogram outperforms standalone nomogram models, LLMs, and radiologists in guiding biopsy decisions and predicting malignancy. The proposed novel fused nomogram has the potential to reduce unnecessary biopsies and enhance personalized decision-making in breast imaging.

To evaluate how a custom web application integrating clinician-edited large language model (LLM)-generated summaries, clickable definitions, and artificial intelligence (AI)-generated videos affects radiology report comprehension, feature preferences, and overall sentiment toward AI-assisted report summaries.

This prospective study recruited participants between May and July 2025 at a hospital-based outpatient imaging floor before their scheduled examinations at a tertiary university hospital. Following exam completion and report publication, patient-friendly AI summaries were generated and reviewed by a radiologist for accuracy. Participants were then shown a web application containing their own de-identified, AI-augmented reports featuring clinician-edited LLM-generated summaries with clickable terms and AI videos. Participants were surveyed on comprehension, feature usefulness, and attitudes toward LLM summaries.

Participants (n=101, 40 male/61 female, racially diverse) ranged from 20 to 82 (mean 58±15) years old. Overall comprehension improved significantly (median pre: 4.00, post: 5.00, p<0.001), with 47.52% (n=48) identifying LLM-summaries as most helpful. However, LLM-summaries required manual clinician edits (average per summary: 24.75 words removed; 0.13 words added, lexical similarity = 84.63%; semantic similarity = 98.25%). When asked if they were comfortable with LLM-summaries without clinician edits, most participants reported being only Somewhat comfortable (27.72%) or Very uncomfortable (25.74%).

This prospective study demonstrates that interactive, LLM-driven applications can significantly improve self-reported patient comprehension of complex radiology reports, emphasizing their potential to enhance patient-centered communication. However, patients had reservations about clinician-edited LLM-generated summaries, indicating that successful integration is contingent on professional oversight - an added workload that may limit scalable real-world implementation.

Accurate documentation of distant recurrence sites in breast cancer is essential for evaluating treatment effectiveness and outcomes research. However, such information is embedded in unstructured clinical notes, making manual abstraction labor-intensive. Large language models (LLMs) offer a scalable solution for extracting complex information from heterogeneous clinical narratives; however, generic LLMs often lack the specialized clinical reasoning needed for accurate interpretation of oncologic documentation. This study aims to develop an efficient LLM-based framework to automatically extract distant recurrence sites from free-text documentation. MATERIALS &

We used clinical notes, pathology and radiology reports from recurrent breast cancer patients at Mayo Clinic (n = 766) for model development and evaluated generalizability on internal hold-out samples (n = 112) and an external Stanford Medicine cohort (n = 110). For cross-disease domain adaptation, we further validated on prostate cancer patients (n = 49). Our proposed framework employs BioLinkBERT, a pretrained language model (PLM) backbone, with weak supervision and an epoch-wise entropy optimization to address limited labeled data and class imbalance across recurrence sites. The fine-tuned model was compared against state-of-the-art models, including Llama2-7B, Llama-3-8B and MedAlpaca, using precision, recall, and F1-score.

The fine-tuned model outperformed generic and domain-specific LLM baselines, with notable gains in identifying multi-site distant recurrence. In-domain validation showed consistent F1-score improvement (average 0.78), particularly for rare recurrence sites. The model also demonstrated strong performance on the external Stanford cohort and on prostate cancer, achieving F1-score of 0.83 and 0.93, respectively.

This study presents an efficient, weakly supervised LLM framework that accurately extracts metastatic recurrence sites, reducing reliance on manual chart review. The results demonstrate that relatively small LLMs, optimized with domain-aware weak supervision, can outperform larger models for complex oncologic information extraction. The model is released as a platform-independent Docker image to support seamless cancer registry integration.

This study evaluates the performance of ChatGPT, a large language model (LLM), in selecting appropriate imaging modalities for breast imaging scenarios using the American College of Radiology (ACR) Appropriateness Criteria (AC). We aim to compare the agreement of ChatGPT with the ACR AC to that of breast radiologists at a single institution in selecting appropriate imaging modalities. METHODS/MATERIALS: The study utilized ten randomly selected clinical variants from the ACR AC breast imaging category. Outputs were obtained from ChatGPT-3.5, ChatGPT-4, and ChatGPT-4o using the versions available in July 2024. The ChatGPT versions and four breast radiologists rated the appropriateness of 81 imaging decisions on a scale from 1 to 9. For each imaging option within a clinical scenario, the ratings provided by the radiologists and four independent samplings of ChatGPT's responses were aggregated. Agreement between ratings from ChatGPT, radiologists, and the ACR AC was analyzed using generalized estimating equations (GEEs) and Bland-Altman plots to assess consistency and bias.

Radiologists had the lowest overall mean bias (0.2438) relative to the ACR (p = 0.489). All versions of ChatGPT had larger mean biases that were significant (GPT-4o: 2.463; GPT-4: 1.7623, GPT-3.5: 2.4691, all p<0.001). All had a slope bias (p < 0.001), but radiologists had the smallest slope bias. In summary, radiologists were closer to the ACR AC and were oftentimes as variable or even less variable as a group than the same ChatGPT version at the same time.

ChatGPT shows promise as an AI tool for imaging decision-making, but current versions lack the accuracy, consistency, and reproducibility demonstrated by experienced radiologists. The study underscores the importance for human oversight in clinical applications and the need for further development to improve ChatGPT's and other LLMs' reliability and alignment with established guidelines.

To address limited quantitative data on sustainable use of large language models (LLMs) in radiology, we quantified the resource footprint of LLMs for labelling CT pulmonary embolism reports and assessed how a hybrid rule-based-LLM workflow changes time, cost and carbon emissions compared with manual labelling.

In this single-centre retrospective study, 2923 structured CT reports were labelled using four workflows: a rule-based extractor (RBE), an LLM-only pipeline using 18 open-weight and four proprietary models, a hybrid RBE-LLM pipeline that routed RBE failures to an LLM, and full manual labelling by radiologists. Ground truth was based on radiologist adjudication. For each LLM, we measured per-report latency, estimated CO2 emissions and cost. Radiologists recorded the labelling time per report.

Manual labelling required 32.8 h for 2923 reports (40.4 s/report; €0.42/report) with 95.0% accuracy (95% CI: 93.7-96.2). LLM-only pipelines were less accurate (85.1%; 95% CI: 84.9-85.5) but reduced labelling time to 12.4 h and cost to €2.60 (both p < 0.001). Hybrid RBE-LLM workflows yielded the highest accuracy (98.5%) and lowest resource use: across 22 models, switching from LLM-only to hybrid reduced time (6.7 to 0.97 h), cost (€1.19 to €0.17), and CO2 (0.82 to 0.12 kg; all p < 0.001).

LLM-only labelling reduced labour time and direct costs compared with manual annotation. A hybrid RBE-LLM pipeline that forwards rule-based failures to an LLM concentrated compute where needed and markedly decreased time, cost and emissions, supporting targeted deployment of LLMs for sustainable data-annotation workflows in radiology. CRITICAL RELEVANCE: By quantifying time, cost and carbon emissions of manual, rule-based, LLM and hybrid report labelling, this study identifies sustainable workflows for deploying LLMs in routine radiology reporting. KEY POINTS: Manual expert labelling of CT pulmonary embolism reports is time-intensive and costly. Mid-sized LLM configurations provide favourable trade-offs between performance and resource use. Hybrid rule-based-LLM workflows sustain accuracy while reducing resource demands.

To compare open-weight and proprietary large language models (LLMs), a rule-based extractor (RBE) and radiologists for labelling pulmonary embolism CT reports, and to test whether a hybrid RBE-LLM workflow improves labelling performance.

This single-centre retrospective study included structured CT reports from October 2021 to March 2025. Three labelling pipelines were evaluated: an RBE; a model-agnostic LLM extractor (18 open-weight, four GPT-4 variants); and a hybrid pipeline routing only RBE failures to an LLM. Ground truth was defined at the report-text level by deterministic schema matching for initially RBE-valid fields and blinded adjudication of RBE-invalid fields by two attending radiologists. Eight radiologists provided a human baseline. Outcomes included F1 scores, accuracy, LLM-based salvage of RBE failures, and labelling time.

In total, 2,923 reports from 2,923 patients (mean age 66 ± 17 years; 1,465 women) were included. Falcon3-10b and GPT-4.1-mini achieved similar item-level performance (F1 0.98 [95% CI, 0.97-0.98] for both; p = 0.70) and both exceeded the RBE (F1 0.81 [95% CI, 0.80-0.82]; p < 0.001). Salvage of RBE failures was comparable between open-weight and proprietary models (88.1% vs 91.9%; p = 0.12). The hybrid RBE-LLM workflow achieved 99.8% accuracy and F1 0.99 (0.98-0.99), exceeding both the RBE and pooled radiologists (F1 0.92 [95% CI, 0.90-0.93]; all p < 0.001).

Schema-constrained open-weight and proprietary LLMs exceeded rule-based extraction and, at the upper end of performance, matched a pooled radiologist label-transfer baseline. A rules-first, targeted LLM workflow enabled near-perfect extraction from finalised structured pulmonary embolism CT reports. RELEVANCE STATEMENT: A rules-first LLM workflow can automate high-fidelity extraction of structured CT findings from finalised radiology reports, enabling scalable, auditable, and more consistent cohort curation for clinical research, registries, and quality improvement. KEY POINTS: A hybrid rules-first workflow combining a rule-based extractor (RBE) with targeted large language model (LLM) salvage achieved the highest overall performance for labelling of pulmonary embolism CT reports (F1, 0.99; accuracy, 99.8%). The top standalone open-weight and proprietary LLMs (Falcon3-10b and GPT-4.1-mini) both exceeded the RBE and, at the upper end of performance, matched a pooled radiologist label-transfer baseline. The hybrid workflow reduced cohort-curation time from 32.2 h for radiologists to 1.0 h while reducing LLM calls by 85.6%, because the LLM was only triggered for rule-failed fields.

To develop and evaluate a glass-box, agentic-style radiology pipeline that separates perception from reasoning for auditable multiclass diagnosis on cine cardiac magnetic resonance imaging (MRI), and to quantify accuracy, robustness across decoding temperatures, and fidelity/safety of generated narrative explanations.

Using the labeled Automated Cardiac Diagnosis Challenge training cohort (n = 100; five diagnostic classes), cine bSSFP images were segmented at end-diastole (ED) and end-systole (ES) with a pretrained nnU-Net, and 17 clinically interpretable biomarkers were extracted. A large language model (LLM) (GPT-OSS-120B) queried prompts under three different prompt strategies (V1-V3) with majority-vote self-consistency after a stratified split into prompt development (n = 20) and independent evaluation (n = 80). Temperatures (T = 0.1, 1.0, and 2.0) were tested for stability. A decoupled narrative module generated radiologist-style reports. Narratives underwent radiologist audit for numeric fidelity and clinical safety. Machine learning algorithms [Random Forest, Support Vector Machine (SVM), Logistic Regression, Decision Tree] were trained on the same biomarker set for benchmarking.

Automated segmentation showed high agreement with reference masks [Dice at ED: right ventricle [RV] cavity 0.984 ± 0.004, left ventricle (LV)] myocardium 0.965 ± 0.009, LV cavity 0.989 ± 0.003; ES: RV cavity 0.979 ± 0.013, LV myocardium 0.975 ± 0.009, LV cavity 0.985 ± 0.005). The hierarchical veto-logic strategy (V3) achieved an accuracy of 0.925 (95% confidence interval: 0.863-0.975) and a macro-F1 of 0.924, remaining stable across temperatures, outperforming V2 (accuracy 0.787-0.800) and V1 (0.562-0.600). Reproducibility was highest for V3 at T = 0.1 (Fleiss' kappa: 0.969) with a low failure rate (0.83%). Narrative generation produced 97.5% valid reports with 100% numeric fidelity and audited safety ≥ 97.5%. Performance was comparable to supervised models (Random Forest accuracy 0.938; SVM/Logistic Regression accuracy 0.925).

In this single-dataset internal evaluation, a glass-box workflow combining automated segmentation-derived biomarkers with an LLM enables robust multiclass cardiac MRI diagnosis while producing numerically faithful, safety-audited narratives, supporting auditability and governance for radiology artificial intelligence (AI). External multicenter validation is needed to confirm generalizability. CLINICAL

A glass-box, biomarker-driven agentic-style workflow enables auditable cine cardiac MRI classification with numerically grounded explanations, addressing interpretability and stability barriers that limit translation of radiology AI into routine practice.

Generative artificial intelligence (GenAI) is enhancing virtual patient simulations in health care education by enabling dynamic, adaptive interactions, reshaping how clinical skills are taught. A synthesis of the current evidence is needed to guide implementation and future research, given the pace of technological advancement.

This systematic review aims to synthesize empirical research on the design, implementation, and educational impact of GenAI-supported virtual patients in health care education.

A systematic search was conducted across 5 databases (CINAHL, Medline, Embase, Scopus, and Web of Science) from their inception to March 19, 2026. Reference lists of included studies and relevant systematic reviews were also screened. Peer-reviewed studies in English that evaluated GenAI-supported virtual patients using quantitative or mixed methods were included. Two reviewers independently screened studies and extracted data. Study quality and risk of bias were assessed critically using JBI (Joanna Briggs Institute) checklists, with disagreements resolved by consensus.

A total of 15 studies met the inclusion criteria (total participants N=645), spanning health care disciplines, including nursing, medicine, pharmacy, radiography, and medical first-responder training. The virtual patients varied in design; input modalities included text (9 studies), voice (5 studies), or hybrid (1 study); output was text (9 studies), speech (5 studies), or both (1 study); 6 studies used 3D-embodied avatars, while 9 used nonembodied interfaces. A total of 13 studies used OpenAI GPT models (eg, ChatGPT), 1 used a fine-tuned model from a different provider, and 1 evaluated multiple model families (Claude, GPT, and open-source). Further, 6 studies used controlled experimental designs, including 3 randomized controlled trials (RCTs); the remainder were cross-sectional or prepost evaluations. Primary outcomes included user perceptions (14 studies), communication skills (4 studies), clinical reasoning (3 studies), and performance (7 studies). In controlled comparisons, GenAI-supported virtual patients consistently improved outcomes relative to control conditions: for example, enhanced clinical decision-making (RCT, n=21), ophthalmology history-taking skills (RCT, n=26), and medical history-taking performance (crossover RCT, n=20). The evidence base is characterized by brief intervention durations, a predominant reliance on single-session interactions, and a general lack of underpinning educational theory. No meta-analysis was performed due to the limited number of studies and significant heterogeneity in designs, interventions, and outcome measures.

The evidence supports the feasibility and acceptability of GenAI-supported virtual patients, with positive learner perceptions and promising outcomes for skills development. However, critical limitations remain in emotional-behavioral complexity, simulation adaptability, and research design rigor (eg, limited use of control groups and validated instruments). The review offers educators, instructional designers, and policymakers actionable insights for integrating dynamic, artificial intelligence-driven simulations while identifying crucial gaps-such as the need for theoretical grounding, longitudinal studies, and standardized design protocols-that must be addressed for safe and effective implementation.

To develop a privacy-preserving method for structuring free-text mammography reports using a locally fine-tuned, open-source large language model (LLM).

In this multicenter study, 7161 unstructured mammography reports were collected from three institutions. The open-source Llama-3 model was fine-tuned via supervised learning using pseudo-labels from the commercial Qwen-Max model with low-rank adaptation. All labels were pseudo-labels generated by the commercial Qwen-Max model rather than human annotations. Structured outputs followed a BI-RADS-oriented nested JSON schema. Performance was evaluated across 23 features using Precision, Recall, and F1-score. Structural integrity was assessed using the JSON format accuracy (JFA) and field integrity accuracy (FIA) metrics. Statistical comparisons were performed using the paired Wilcoxon signed-rank test and Cohen's d effect size.

A total of 7161 reports were retrospectively obtained from three institutions and analyzed. The fine-tuned model achieved strong performance at epoch 10 (Precision 0.942, Recall 0.929, F1-score 0.932), with JFA and FIA reaching 0.964 and 1.000, respectively, showing significant gains over the base model (p < 0.05, Cohen's d > 0.8). While slightly below Qwen-Max overall, the model exhibited moderate yet statistically significant differences (p < 0.05; 0.5 < Cohen's d < 0.8), particularly in the "special signs" category (F1 = 0.737 vs 0.947).

This method effectively converts mammography reports into structured data using a locally fine-tuned, open-source LLM. Although there is a slight performance trade-off, it improves privacy and can be deployed locally. Its accuracy, clinical relevance, and compliance make it a practical solution for medical institutions. KEY POINTS: Question Free-text mammography reports lack standardization, making structured extraction difficult, while existing solutions often compromise privacy, adaptability, or require costly commercial tools. Findings Our fine-tuned LLaMA-3 model achieved high extraction accuracy (F1-score: 0.932) and complete structural integrity (FIA: 1.000) within a fully local deployment pipeline. Clinical relevance This method enables standardized and privacy-preserving mammography reporting without disruption to the clinical workflow, supporting safer AI integration, enhanced data quality, and compliance with regulations such as the General Data Protection Regulation (GDPR).

This study evaluates seven open-source Large Language Models (LLMs) in summarizing radiology reports of acute ischemic stroke patients treated with mechanical thrombectomy and predicting angiography-based outcome measures relevant to post-thrombectomy reperfusion.

2000 mechanical thrombectomy reports (findings and summarizing impression section as gold standard) were split into training set (N = 1900) for model fine-tuning and test set (N = 100). A two-step evaluation was performed: (1) Quantitative analyses of seven LLMs with metrics ROUGE-1, -2, -L, METEOR, BERTScore (F1) and BLEU comparing LLM-generated summaries against gold-standard impressions. (2) Qualitative manual evaluation of the four best-performing models by two radiologists, assessing correctness and completeness across key parameters: outcome-relevant scores, vessel information, occlusion side, number of passes, relevant additional information, hallucinations, and grammar quality. Statistical significance was assessed via a two-tailed, four-sample χ² test, followed by post hoc pairwise χ² comparisons.

BioMistral-7b scored highest across most quantitative metrics (ROUGE-1: 0.47, ROUGE-2: 0.30, ROUGE-L: 0.43, METEOR: 0.46, BERTScore (F1): 0.82). Manual evaluation revealed gemma-2-9b most frequently documented pass counts (56 out of 100 cases (56%); p < 0.02 vs. Llama-3.1-8b/mistral-7b-instruct), while mistral-7b-instruct described them most often correctly (29 out of 38 mentioned passes (76.32%); p < 0.02 vs. BioMistral-7b and p < 0.01 vs. gemma-2-9b). All four manually evaluated LLMs performed moderately well in predicting "Thrombolysis-In-Cerebral-Ischemia (TICI)" Score (correctness rate ranging from 66 to 71%; p = 0.89).

All four manually evaluated LLMs effectively summarized thrombectomy reports and demonstrated moderate accuracy predicting TICI scores. Their integration into radiology workflows could enhance efficiency, warranting further clinical validation. KEY POINTS: Question Specifically fine-tuned Large Language Models (LLMs) can improve radiology workflow by automatically summarizing thrombectomy reports and inferring angiographic classifications from textual descriptions. Findings Fine-tuned LLMs achieve similar performance in summarizing thrombectomy reports, with each model performing best in specific categories and showing moderate accuracy in correct "Thrombolysis-In-Cerebral-Ischemia (TICI)" Score prediction (66-71%). Clinical relevance Integrating fine-tuned LLMs into radiology workflows may accelerate decision-making and improve patient outcomes by automatically summarizing reports and assessing recanalization success, while future work should enhance contextual understanding, address ambiguous inputs, and limit hallucinations.

To evaluate the association of patient characteristics, community-level social determinants of health, and cyst risk categories with completion of follow-up recommendations for incidental Pancreatic Cystic Lesions (PCLs).

We retrospectively identified consecutive patients (2013-2023) whose MRI radiology reports described PCLs. A fine-tuned LLaMA-3.1 8B Instruct large language model was used to extract PCL features. Lesions were classified using the 2017 ACR white paper: Category 1 (low risk), Category 2 (worrisome features), or Category 3 (high-risk stigmata). We recorded demographics and follow-up imaging or endoscopic ultrasound dates. Community-level factors were characterized by the 2020 CDC Social Vulnerability Index (SVI), stratified into quartiles. The primary outcome, "inappropriate follow-up," combined late and no follow-up. Multivariable binomial regression was applied to evaluate associations with inappropriate follow-up.

In 7,745 patients (mean age 66.3 years; 4,796 women), 92.9% (7,198/7,745) of cysts were Category 1, 6.4% (498/7,745) were Category 2, and 0.6% (49/7,745) were Category 3. Only 36.3% of patients completed appropriate follow-up, 12.1% were late, and 51.6% were lost to follow-up. Inappropriate follow-up was high in every cyst category: 64.2% in Category 1, 59.4% in Category 2 and 49.0% in Category 3. In multivariable analysis, non-English primary language (RR 1.08; 95% CI, 1.02-1.14) and residing in more vulnerable communities of the 3rd quartiles of the socioeconomic Social Vulnerability Index subcategory (RR 1.07; 95% CI, 1.02-1.12) were associated with inappropriate follow-up. Higher age-adjusted Charlson Comorbidity Index (CCI ≥ 4) (RR .84; 95% CI, .79-.88), CCI 2-3 (RR .84; 95% CI, .79-.88), and higher-risk cysts in patients under 65 years of age (RR .76; 95% CI, .65-.89) were associated with completed follow-up.

Follow-up completion for incidental PCLs was low. Factors most consistently associated with follow-up completion were language barriers, residence in socioeconomically vulnerable communities, age-adjusted CCI and higher-risk features among those under 65 years.

To evaluate the performance of large language models (LLMs) in predicting molecular types of adult-type diffuse gliomas according to the 2021 WHO classification using MRI radiology reports.

This retrospective study included 2169 patients diagnosed with adult-type diffuse gliomas (294 oligodendrogliomas, 295 IDH-mutant astrocytomas, and 1580 IDH-wildtype glioblastomas) between July 2005 and March 2024 from four hospitals in Asia and Europe. Seven proprietary and open-source LLMs were assessed: GPT-4o-mini, GPT-4.1-mini, Llama 3.1 8B, Llama 3.1 70B, Qwen2.5 7B, Deepseek-r1 8B, and Mistal 7B. The performance of LLMs in classifying molecular types was compared based on the provision of relevant knowledge of glioma imaging findings (knowledge-based vs. naïve prompt). The impact of radiologists' subspecialization in neuro-oncology, report quality, and reporting language on LLMs' performance was also evaluated.

LLMs achieved significantly higher (naïve vs. knowledge-based; GPT-4o-mini, 77.0% vs. 79.1%, p < 0.001; Qwen2.5 7B, 75.9% vs. 79.5%, p < 0.001; Deepseek-r1 8B, 66.0% vs. 73.2%, p < 0.001) or comparable accuracy (GPT-4.1-mini, 78.7% vs. 78.6%; Llama 3.1 70B, 78.0% vs. 78.1%; Mistral 7B, 58.4% vs. 57.4%) using knowledge-based prompt compared to naïve prompt, except for Llama 3.1 8B (65.4% vs. 44.6%, p < 0.001). Differences in accuracy were more pronounced in smaller-sized LLMs. Additionally, the accuracy was significantly higher with reports by neuro-oncology specialists and high-quality reports in all LLMs (p < 0.001).

LLMs may provide preoperative information on the tumor types of adult-type diffuse gliomas from MRI reports by providing relevant knowledge in the prompt. Informative and descriptive reports could further enhance LLMs' performance. KEY POINTS: Question Our study aimed to evaluate large language models' (LLMs) ability to efficiently predict molecular types of adult-type diffuse gliomas according to the 2021 WHO classification. Findings Larger models generally showed better accuracy and were less sensitive to domain-specific knowledge. Their performance improved when using high-quality, longer reports or reports by neuro-oncology specialists. Clinical relevance These findings highlight the potential role of LLMs in predicting glioma molecular types, underscoring the importance of informative and descriptive reports in enhancing their performance.

Radiology reports are essential for clinical decision-making and must meet standards of clarity, completeness, and diagnostic accuracy. However, quality assessment is often subjective, time-consuming, and dependent on expert reviewers. Large language models (LLMs) offer a promising alternative for automating this process. This study evaluates whether LLMs can objectively and consistently assess the formal and diagnostic quality of neuroradiology reports.

We analysed 277 neuroradiology reports, originally authored by 10 radiologists and subsequently evaluated by a different radiologist to assess formal and diagnostic quality. Reports were annotated using six quality criteria: diagnostic discrepancy, report structure, clarity, completeness, grading/staging system use, and recommendation of additional tests. Three locally deployed LLMs (LLaMA 3.2, DeepSeek R1:7B, and Gemma 3:4B) were evaluated for their ability to classify reports according to these criteria.

LLMs showed strong performance in recognising positive categories, with DeepSeek achieving the highest accuracy in report structure (70.40%) and test recommendations (68.59%). However, performance declined significantly in detecting issues such as report completeness or diagnostic discrepancies, reflected in low F1-scores for those categories.

While current models have limitations in identifying subtle errors or complex clinical nuances, LLMs may assist in automating aspects of radiology report quality control. With further refinement, they could help improve consistency and support clinical decision-making.

Ambiguous or incomplete documentation is a recurrent bottleneck in radiation oncology workflows, leading to inefficiencies in communication and potential treatment delays. Large language models (LLMs) pose a solution to addressing these ambiguities without added burden to clinical staff. We aim to assess the effectiveness of Meta's open-source Llama 3.3 model in using physician consultation notes to isolate and classify anatomical treatment sites and create helpful extractive summaries for each patient.

Semi-structured interviews with five radiation therapists revealed that CT simulation orders lack the necessary details to acquire the appropriate image. A retrospective cohort of 100 patient notes was used for iterative prompt engineering. The final model was evaluated on an independent test cohort of 52 patient notes. The LLM's accuracy in identifying the treatment site was benchmarked against two human observers (a medical physicist and a physician) as well as the final delivered treatment plan (ground truth). The helpfulness and accuracy of the AI-generated summaries were also rated by both observers on a 5-point Likert scale.

Llama 3.3 achieved a weighted accuracy of 94.2% [95%CI: 89.4%-98.1%] when compared to sites isolated by either observer. When compared to the sites isolated from the retrospectively delivered plans, the model reached a weighted accuracy of 92.3% [95% CI: 87.5%-97.1%]. The model classified the anatomical sites with a weighted accuracy of 96.2% [95%CI: 87.0% -98.9%]. The AI-generated summaries were highly rated by both observers (Observer 1: 4.96 [95%CI: 4.87-5.00] and Observer 2: 4.58 [95% CI: 4.38-4.73]).

This pilot study provides foundational evidence that LLMs can classify data with high accuracy, achieve benchmarks comparable to human experts when isolating anatomical treatment sites, and produce clinically helpful summaries. Our results suggest that LLMs can be effectively integrated to streamline complex radiotherapy workflows in the clinic.

Automated structuring of radiology reports is essential for data utilization and the development of medical artificial intelligence models. However, manual annotation by experts is labor-intensive, and processing real clinical data through commercial large language models (LLMs) presents significant privacy risks. These challenges are particularly pronounced for non-English languages like Japanese, where specialized medical corpora are scarce. While synthetic data generation offers a potential privacy-preserving alternative, its effectiveness in capturing complex clinical nuances-such as negation and contextual dependencies-to train robust classification models without any real-world training data has not been fully established.

This study aimed to develop a context-aware sentence classification model for Japanese radiology reports using an entirely synthetic training pipeline, thereby eliminating reliance on real-world clinical data during the development phase. Furthermore, we sought to evaluate the generalizability of this approach by validating the model's performance on diverse, multi-institutional, real-world reports.

Japanese radiology reports (n=3104) were generated using GPT-4.1 and automatically annotated at the sentence level into 4 categories (background, positive finding, negative finding, and continuation) using GPT-4.1-mini. The synthetic data were partitioned into training (n=2670), validation (n=334), and test (n=100) sets. We fine-tuned several models, including lightweight local LLMs (Qwen3 and Llama 3.2 series) using low-rank adaptation and Japanese text classification models (Bidirectional Encoder Representations from Transformers [BERT]-base Japanese v3, Japanese Medical Robustly Optimized BERT Pretraining Approach [JMedRoBERTa]-base, and ModernBERT-Ja-130M). External validation was performed using 280 real-world reports (3477 sentences) from 7 institutions in the Japan Medical Image Database, with ground-truth labels established by board-certified radiologists. Evaluation metrics included accuracy, macro-averaged F1 (macro F1) score, and positive predictive value for positive findings (PPV_1).

All models achieved high performance on the synthetic test set (accuracy: 0.938-0.951; macro F1-score: 0.924-0.940). Overall performance declined on the external validation dataset (accuracy: 0.783-0.813; macro F1-score: 0.761-0.790), reflecting distributional differences between synthetic and real-world reports; however, PPV_1 remained stable and high across datasets (eg, 0.957 on the synthetic test set vs 0.952 on the external validation dataset for Qwen3 [4B]). Parsing errors occurred in LLM-based approaches (19-260 sentences, 0.55%-7.48% in the external dataset).

This study demonstrates the feasibility of developing context-aware sentence classification models for Japanese radiology reports using a training pipeline based entirely on synthetic data. The stability of PPV_1 indicates that the models successfully captured the essential clinical terminology and linguistic patterns required to identify positive findings in real-world reports, despite the observed performance degradation during external validation. This approach substantially reduces manual annotation requirements and privacy risks, providing a scalable foundation for constructing structured radiology datasets to support the development of clinically relevant medical artificial intelligence models.

Incomplete clinical details on magnetic resonance imaging (MRI) examination requests (MERs) can lead to suboptimal protocol selection. An institutional secure large language model (sLLM) with access to manually retrieved salient data from the electronic medical record (EMR) may improve request completeness and protocol accuracy across multiple MRI subspecialties.

The objective of this study was to compare clinician MERs with sLLM-augmented MERs for information quality and to evaluate the protocoling accuracy of the sLLM versus board-certified radiologists across body, musculoskeletal, and neuroradiology MRI.

This retrospective study included 608 random outpatient MRI examinations performed between September 2023 and July 2024 (body 206, musculoskeletal 203, neuroradiology 199). The cohort comprised 528 patients (mean 51.2 years, SD 19.2; range 4-93; n=279, 52.8% women, n=249, 47.2% men). MERs without EMR access were excluded. A privately hosted Anthropic Claude 3.5 model (temperature 0) augmented each MER with manually retrieved salient EMR data and, via rule-based parsing, mapped the extracted elements onto predefined institutional criteria to recommend region or coverage and contrast use. Two experienced radiologists established a consensus reference standard. Two board-certified general radiologists (Rad 3 and Rad 4) and the sLLM were compared with this standard. Clinical information quality was graded using the Reason-for-Exam Imaging Reporting and Data System (RI-RADS). Interrater reliability was quantified with Gwet AC1. Paired accuracies were compared with the McNemar test to determine whether there was a statistically significant difference.

Interreader agreement for RI-RADS was almost perfect for sLLM-augmented MERs (AC1 0.97, 95% CI 0.94-0.99) and moderate for clinician MERs (AC1 0.43, 95% CI 0.34-0.52). Limited or deficient clinical information (RI-RADS C/D) fell to 0% to 0.7% (0/608 to 4/608) with sLLM augmentation vs 4.1% to 20.4% (25/608 to 124/608) for clinician MERs. Overall protocol accuracy was 93.1% (566/608; 95% CI 89.6-96.6) for the sLLM, 91.4% (556/608; 95% CI 87.6-95.3) for Rad 3, and 92.1% (560/608; 95% CI 88.4-95.8) for Rad 4 (sLLM vs Rad 3 P=.23 vs Rad 4 P=.40). Region or coverage accuracy was similar (sLLM: 579/608, 95.2%; Rad 3: 585/608, 96.2%; Rad 4: 573/608, 94.2%; P=.46 and P=.36). Contrast decisions were more accurate using the sLLM at 94.4% (574/608; 95% CI 91.3-97.5) vs Rad 3 at 92.1% (560/608; 95% CI 88.4-95.8; P=.027) and were not significantly different to Rad 4 at 92.9% (565/608; 95% CI 89.4-96.4; P=.16). Subspecialty analyses showed similar patterns, with the sLLM outperforming Rad 4 for musculoskeletal MRI contrast decisions (96.6% vs 91.1%; P=.006) and matching readers elsewhere. Manual review indicated that sLLM improvements arose from EMR details not listed on the MER (infection/inflammation, tumor history, prior surgery). No clinically significant hallucinations were identified in a manual review of discordant cases.

Across body, musculoskeletal, and neuroradiology MRI, sLLM-augmented examination requests improved clinical context and enhanced contrast selection while demonstrating accuracy comparable to general radiologists for region or coverage. Integrating sLLMs into routine vetting workflows may reduce manual workload in protocol selection for more efficient, standardized protocoling.

Staging gynecological malignancies is a complex process, and radiologists should be familiar with the evolution of FIGO staging criteria. Large Language Models (LLMs) offer potential to support radiologists by automating classification tasks from free-text MRI reports.

We conducted a retrospective study using two curated datasets of pelvic MRI reports from patients with cervical (n = 261, FIGO 2018) and endometrial cancer (n = 555, FIGO 2023). A general-purpose LLM (Cohere Command-A) was evaluated under three prompting strategies (zero-shot, guided, and chain-of-thought [CoT]), using exact stage accuracy, an ordinal FIGO distance metric, and the rate of severe errors. The Cohere Command-A model was chosen for its long-context reasoning, instruction-following capabilities, reproducible fixed version, and secure handling of sensitive clinical data. While alternative LLMs (eg, GPT-4o, Gemini, Llama-3, DeepSeek) could offer complementary insights, access, resources, and compliance constraints limited broader comparisons.

For cervical cancer, CoT prompting achieved the highest accuracy (80.5%) and the lowest FIGO distance, with 23 severe misclassifications (≥2-stage deviation), outperforming guided and zero-shot prompting. For endometrial cancer, all strategies performed appropriately, with CoT again yielding the best results (accuracy, 90.6%) and the lowest number of severe misclassifications (37 cases), compared with guided and zero-shot prompting. In a small subset of cases with no agreement between any prompting strategy and the reference label, manual review showed that only a minority presented potentially suboptimal annotations, suggesting that CoT-based predictions may also help flag doubtful reports.

The LLMs used demonstrated strong performance in automatically assigning FIGO stages for cervical and endometrial cancers from MRI reports. Their integration could reduce workload and improve consistency in staging. Further validation is needed before clinical implementation.

Automated extraction of information from cardiac reports would benefit both clinical reporting and research. Large language models (LLMs) hold promise for such automation, but their clinical performance and practical implementation across various computational environments remain unclear. This study aims to evaluate the feasibility and performance of LLM-based classification of echocardiogram and invasive coronary angiography reports, using real-world clinical data across local, high-performance computing and cloud-based platforms. METHODS AND

The angiography and echocardiography reports of 1000 patients, admitted with acute coronary syndrome, were labelled for multiple key diagnostic elements, including left ventricular function (LVF), culprit vessel, and acute occlusions. Report classification models were developed using LLMs via (i) prompt-based and (ii) fine-tuning approaches. Performance was assessed across different model types and compute infrastructures, with attention to class imbalance, ambiguous label annotations, and implementation costs. Large language models demonstrated strong performance in extracting structured diagnostic information from cardiac reports. Cloud-based models (such as GPT-4o) achieved the highest accuracy (0.87 for culprit vessel and 1.0 for LVF) and generalizability, but also smaller models run on a local high-performance cluster achieved reasonable accuracy, especially for less complex tasks (0.634 for culprit vessel and 0.984 for LVF). Classification was feasible with minimal pre-processing, enabling potential integration into electronic health record systems or research pipelines. Class imbalance, reflective of real-world prevalence, had a greater impact on fine-tuning approaches.

Large language models can reliably classify structured cardiology reports across diverse computed infrastructures. Their accuracy and adaptability support their use in clinical and research settings, particularly for scalable report structuring and dataset generation.

To evaluate the application value of three ChatGPT versions and Gemini in pathology report simplification tasks for prostate cancer.

This retrospective study assessed GPT-3.5, GPT-4.0, GPT-4o, and Gemini on pathology reports from 228 prostate cancer patients across two institutions. Data were split into internal (center 1, n = 171) and external (center 2, n = 57) cohorts. Using specific prompts, models generated simplified texts. The evaluation of outputs included three main dimensions: (1) human scoring by patients, clinicians, and pathologists; (2) readability scores; and (3) BERT-based semantic similarity scores. Statistical comparisons employed paired t -tests or Wilcoxon signed-rank tests. Statistical consistency between raters was assessed using squared weighted kappa, intraclass correlation coefficient(3,1), and percent agreement, with 95% confidence intervals calculated for all metrics.

GPT-4o (Few-Shot) achieved the highest accuracy and comprehensiveness scores from pathologists, while Gemini demonstrated the best understandability. Patient and clinician understandability ratings were consistently high across models. Mean Reading Grade Level scores varied between internal and external datasets, with GPT-4o Few-Shot performing best overall. BERT-based semantic similarity scores demonstrated distinct trends across models, reflecting differences in text simplification strategies.

LLMs adopt distinct trade-off strategies between simplifying pathology reports and preserving their structure and logic, influenced by prompt design and textual style. Their application shows potential to enhance patient comprehension and clinical communication. Future work should focus on domain-specific fine-tuning to ensure safe and reliable clinical integration.

To develop and validate a fine-tuned DeepSeek R1-based Radiology Examination Chatbot (REC) to optimize clinical interaction between cancer patients and radiology healthcare providers (RHPs), thereby effectively providing emotional support for cancer patients and reducing burnout among RHPs. DESIGN, SETTING, AND

Audio recordings of multi-scenarios (appointment triage (AT), pre-examination preparation (PP), radiology clinic services (RCS)) were collected from the radiology departments of three tertiary hospitals (n = 36,511 min). This study conducts two independent randomized controlled sub-trials for distinct patient groups: Sub-trial 1 evaluates AT/PP participants (1,424 patients, 1:1 randomized to RHP + REC or RHP), while Sub-trial 2 assesses RCS participants (638 patients, 1:1 randomized to the same groups). Due to differing patient populations, the sub-trials were designed and implemented separately. INTERVENTION: The REC was fine-tuned using domain-specific dialogue data (80% for training) and scenario-specific prompts, with GPT-o1 as a comparative benchmark. Sub-trials randomized patients to RHP + REC or RHP groups. MAIN OUTCOME AND MEASURES: The primary outcome included dialogue quality (empathy, frustration, emotional regulation, factuality, integrity, and satisfaction), while the secondary outcomes comprised burnout (exhaustion, depersonalization, and personal achievement) and image quality (CT/MRI), all assessed via Likert scales and statistical tests.

RHP + REC group demonstrated superior dialogue quality in AT (factuality: 4.12 ± 0.86 vs. 3.39 ± 1.21, P < 0.001) and PP (satisfaction: 3.73 ± 0.11 vs. 3.19 ± 0.18, P < 0.001), with reduced burnout (exhaustion: 1.85 ± 0.91 vs. 2.40 ± 1.22, P < 0.01). CT image quality improved significantly (4.35 ± 0.51 vs. 4.00 ± 0.52, P < 0.01), and similar results were achieved in MRI examinations (4.12 ± 0.51 vs. 3.79 ± 0.58, P = 0.02). However, REC underperformed in empathy and emotional regulation during emotionally complex RCS (3.88 ± 0.67 vs. 4.42 ± 0.53, P = 0.002; 3.87 ± 0.19 vs. 4.12 ± 0.27, P = 0.004). Ablation studies confirmed the necessity of fine-tuning and scenario-specific prompts for performance. CONCLUSION AND RELEVANCE: Overall, the DeepSeek R1-based REC synergistically enhanced multi-scenarios clinical interaction, provided emotional support for cancer patients, and reduced RHPs' burnout, offering a scalable solution to optimize radiology workflows. TRIAL REGISTRATION: Chinese Clinical Trial Registry Registration number: (ChiCTR2500102740|| http://www.chictr.org.cn/ ), Registration Date: 2025-05-26. KEY POINTS: QUESTION: Can a DeepSeek R1 fine-tuned chatbot (REC) optimize clinical interactions, provide emotional support to patients, and reduce burnout among radiology healthcare providers (RHPs) across multiple scenarios in radiology cancer care (appointment triage (AT), pre-examination preparation (PP), radiology clinic services (RCS))?

REC significantly improved dialogue quality (factuality, integrity, satisfaction) and imaging quality (CT/MRI scans) in both AT and PP scenarios, while simultaneously reducing occupational burnout among RHPs. In emotionally complex scenario like the RCS, although REC outperformed humans in factuality and integrity, its empathy capabilities and emotional regulation lagged behind humans. Ablation experiments confirmed that domain-specific fine-tuning and scenario-adapted prompt templates were crucial to REC’s performance gains. MEANING: REC functions as an efficient workflow aid in radiology departments—enhancing efficiency for structured tasks (scheduling, procedural guidance), alleviating clinical burden, and delivering standardized patient support for cancer cases. Its limitations in emotionally sensitive contexts suggest future integration should combine human providers’ emotional intelligence or incorporate multimodal technologies (e.g., facial expression/tone analysis) for model optimization. This study establishes a benchmark for contextual validation of medical LLMs, advancing safe AI implementation in cancer care.

To compare the quality and clinical usefulness of large language model (LLM)-generated lumbar spine magnetic resonance imaging (MRI) reports with radiologist-written ones and assess whether medical professionals can distinguish between them.

This retrospective observational single-center study was approved by the local ethics committee. A total of 125 lumbar spine MRI reports (104 human-written, 21 LLM-generated using ChatGPT-4o) were anonymized, randomized, and blindly evaluated by five medical professionals (one board-certified radiologist, two radiology residents, one general practitioner, one orthopedic surgeon), all with basic familiarity with LLM. Each report was scored on a five-point Likert scale for clinical relevance, clarity, completeness, diagnostic accuracy, and intelligibility, whereas general practitioner and orthopedic surgeon evaluated intelligibility only. Evaluators also classified each report as AI-generated or human-written. Accuracy was defined as the proportion of correctly classified reports in distinguishing LLM-generated from radiologist-written texts. Mann-Whitney U or Student's t-tests were used.

Radiologists' reports consistently received higher median scores across all domains (p < 0.001). No differences were found in the description of the imaging technique (p > 0.175). No clinically false statements were identified in the LLM-generated reports. Identification accuracy varied widely among evaluators: Board-certified radiologist achieved 88.0% accuracy (sensitivity 66.7%, specificity 92.3%), Resident 1 65.6% (14.3%, 76.0%), Resident 2 94.4% (66.7%, 100%), orthopedic surgeon 78.4% (90.5%, 76.0%) and general practitioner 65.6% (81.0%, 62.5%).

Radiologist-written lumbar spine MRI reports outperform LLM-generated reports in quality and structure. However, some AI-generated reports were indistinguishable from human ones, particularly for non-specialized readers. LLMs may support radiologists in structured reporting and improve workflow efficiency, while maintaining diagnostic reliability. RELEVANCE STATEMENT: Large language models can draft lumbar spine MRI reports, but currently lack the quality and consistency of radiologist reports. With radiologist supervision, large language models may improve reporting efficiency while preserving diagnostic reliability and supporting clinical decision-making. KEY POINTS: LLM-generated reports are clinically coherent and stylistically comparable to those written by expert radiologists. Radiologist-written reports scored significantly higher for clinical relevance, findings, and structure. LLM-generated reports were sometimes misclassified as human-written by clinicians.

Accurate esophageal cancer staging relies on 18F fluorodeoxyglucose positron emission tomography (18F FDG-PET), but its interpretation is complex and time-intensive. This diagnostic burden is exacerbated by significant workforce shortages in both radiology and surgery, thus necessitating automated support systems. The emergence of advanced large language models (LLMs) has raised expectations for their potential to fulfill this role in complex medical tasks.

We evaluated the diagnostic accuracy of LLMs for staging esophageal cancer using 18F FDG-PET images, with a focus on their ability to assess lymph nodes (LNs; clinical N [cN]) and distant metastases (clinical M [cM]) for automated radiology reporting.

This retrospective study included 120 consecutive adult patients who were diagnosed with esophageal squamous cell carcinoma and underwent 18F FDG-PET/computed tomography at Tohoku University Hospital between January 2019 and December 2021. Patients with prior treatment, nonsquamous cell carcinoma histology, or blood glucose levels ≥200 mg/dL were excluded. Frontal maximum-intensity projection positron emission tomography images were extracted, standardized, and analyzed along with information regarding the tumor location. Six LLMs (GPT-5, GPT-4.5, GPT-4.1, OpenAI-o3, -o1, and GPT-4 Turbo) and 4 blinded human evaluators (a nuclear medicine specialist, a gastrointestinal surgeon, and 2 radiology residents) assessed the presence of thoracic and abdominal LN metastases on a region-level basis and determined cN and cM staging on a patient-level basis. The model analyses were performed using the application programming interface in a zero-shot setting. Radiology reports served as the reference standard. Diagnostic agreement and accuracy were evaluated using Cohen κ and the Cochran Q test. Additionally, to account for the class imbalance in the dataset, the Matthews Correlation Coefficient was calculated as a robust metric for binary classification performance. Post hoc McNemar tests were performed with Bonferroni correction; statistical significance for pairwise comparisons was set at P<.0083 (adjusted from P<.05) using JMP Pro (version 18.0; SAS Institute Inc).

The average accuracy was 41/120 (34%) to 94/120 (78%) for LLMs and 72/120 (60%) to 102/120 (85%) for physicians, with significantly higher accuracy for physicians (P<.05) in the thoracic LN, abdominal LN, and cN stages. Interrater reliability was slight to fair for LLMs (κ: -0.07 to 0.25) and fair to substantial for physicians (κ: 0.27 to 0.74). Matthews Correlation Coefficient scores were consistently higher for physicians (0.28 to 0.75) than for LLMs (-0.07 to 0.32). Among the LLMs, GPT-5 demonstrated the highest overall accuracy, with newer LLMs showing improved diagnostic accuracy when compared with previous models in identifying abdominal LN metastases and cM staging, though they showed weaker consistency for cN staging. For example, in thoracic LN detection, GPT-5 achieved 76/120 (63%) accuracy, whereas other LLMs achieved 72/120 (60%) or lower accuracy.

Although current LLMs have not yet reached physician-level accuracy in comprehensive staging, recent models show promise in assisting with specific diagnostic tasks.

Large language models (LLMs) show promise in healthcare but are constrained by hallucinations, static knowledge, and limited domain specificity. Fine-tuning (FT) and retrieval-augmented generation (RAG) offer complementary solutions, with FT embedding domain reasoning and RAG enabling dynamic, up-to-date knowledge access. Hybrid FT + RAG frameworks have been proposed to improve factual accuracy and clinical reliability. This scoping review synthesizes current evidence on such hybrids in healthcare AI. (2)

The search across PubMed, IEEE Xplore, Google Scholar, and Embase identified studies implementing explicit FT + RAG hybrids in healthcare or biomedical tasks. Eligible studies reported empirical evaluations of LLM performance or behavior. Data were extracted on base models, FT strategies, RAG architectures, applications, and performance outcomes. (3)

Seven studies met inclusion criteria. FT + RAG systems consistently outperformed FT-only or RAG-only approaches across QA, clinical summarization, report generation, and decision support tasks. Parameter-efficient FT methods (e.g., LoRA) were common, while RAG implementations varied (dense, hybrid, hierarchical, multimodal, federated). Reported benefits included improved accuracy, reduced hallucination, and greater clinician preference and feasibility in protected settings. (4)

FT + RAG frameworks represent a promising direction for clinically grounded healthcare AI, combining domain-specific reasoning with transparent, up-to-date retrieval. Future work should prioritize standardized evaluation, workflow integration, and governance to enable safe deployment.

To evaluate the performance of large language models (LLMs), including retrieval-augmented generation (RAG)-based approaches, in extracting components and management recommendations from structured coronary computed tomography angiography (CCTA) reports according to the Coronary Artery Disease Reporting and Data System (CAD-RADS 2.0).

A total of 320 fully structured CCTA reports were analyzed using LLM. Closed-source standard ChatGPT‑5, NotebookLM (RAG-based model), and a RAG-adapted ChatGPT‑5 model (ChatGPT-5-RAG) were used. Each model extracted the CAD-RADS category, plaque burden, presence of high-risk plaque (HRP), other modifiers, full score, and management recommendations in accordance with the CAD-RADS 2.0 guidelines. We compared LLM outputs with reference standards determined by two expert cardiovascular radiologists.

ChatGPT-5-RAG showed the highest accuracy for CAD-RADS classification (0.959, 95% CI: 0.932-0.976), plaque burden (0.912, 95% CI: 0.876-0.939), HRP detection (0.988, 95% CI: 0.968-0.995), other modifiers (0.950, 95% CI: 0.920-0.969), and full score (0.828, 95% CI: 0.783-0.866). Closed-source ChatGPT‑5 showed the weakest performance across all components. Significant statistical differences were found among the three models (p < 0.001). Management recommendations were qualitatively rated on a three-point Likert scale; although agreement between models was low, ChatGPT-5-RAG and NotebookLM performed almost perfectly (median 3 points).

This study demonstrates that RAG-enhanced LLMs significantly improve accuracy and reliability in extracting CAD-RADS 2.0 components and generating clinical management recommendations. The findings highlight the potential of RAG-based LLMs as innovative, explainable tools for automated and standardized CCTA reporting in clinical radiology workflows. ZUSAMMENFASSUNG: ZIEL: Ziel der vorliegenden Arbeit war es, die Leistungsfähigkeit von Large-Language-Modellen (LLM) bei der Herausfilterung von Komponenten und Therapieempfehlungen aus strukturierten Befunden von Computertomographie-Angiographien der Koronarien (CCTA) gemäß Coronary Artery Disease Reporting and Data System (CAD-RADS 2.0) zu untersuchen, einschließlich Ansätzen auf der Grundlage von durch Abfrage von Informationen verbesserter Generierung von Texten („retrieval-augmented generation“, RAG). MATERIAL UND METHODE: Es wurden 320 vollständig strukturierte CCTA-Befunde mittels LLM ausgewertet. Dafür wurden Closed-Source-Standard-ChatGPT‑5, NotebookLM (RAG-basiertes Modell) und ein RAG-adaptiertes ChatGPT-5-Modell (ChatGPT-5-RAG) eingesetzt. Jedes Modell extrahierte die CAD-RADS-Kategorie, Plaquelast, das Vorliegen von Hochrisikoplaques (HRP), sonstige Modifikatoren, den vollständigen Score und die Therapieempfehlungen in Übereinstimmung mit den CAD-RADS‑2.0‑Leitlinien. Die LLM-Ergebnisse wurden mit Referenzstandards verglichen, die durch 2 Fachärzte für den Bereich kardiovaskuläre Radiologie festgelegt worden waren. ERGEBNISSE: ChatGPT-5-RAG wies die größte Genauigkeit auf bei der CAD-RADS-Klassifizierung (0,959; 95%-Konfidenzintervall, 95%-KI: 0,932–0,976), Plaquelast (0,912; 95%-KI: 0,876–0,939), HRP-Erkennung (0,988; 95%-KI: 0,968–0,995), sonstigen Modifikatoren (0,950; 95%-KI: 0,920–0,969) und vollständigem Score (0,828; 95%-KI: 0,783–0,866). Für Closed-Source-ChatGPT‑5 zeigte sich die schwächste Leistung hinsichtlich aller Komponenten. Zwischen den 3 Modellen wurden signifikante statistische Unterschiede festgestellt (p < 0,001). Die Therapieempfehlungen wurden qualitativ auf einer 3‑Punkt-Likert-Skala bewertet; auch wenn die Übereinstimmung zwischen den Modellen gering war, war die Leistung von ChatGPT-5-RAG und NotebookLM fast perfekt (im Mittel 3 Punkte). SCHLUSSFOLGERUNG: In der vorliegenden Studie wurde gezeigt, dass mittels RAG verbesserte LLM die Genauigkeit und Verlässlichkeit bei der Herausfilterung von CAD-RADS‑2.0‑Komponenten und der Generierung klinischer Therapieempfehlungen signifikant verbessern. Diese Ergebnisse unterstreichen das Potenzial von RAG-basierten LLM als innovative, erklärbare Instrumente für die automatisierte und standardisierte CCTA-Befunderstellung in den Arbeitsabläufen der klinischen Radiologie.

Radiology impressions guide clinical care. Large Language Models (LLMs)-drafted impressions can drift into generic, off-style text. Retrieval-augmented generation (RAG) enables context-aware few-shot prompting during inference.

This retrospective IRB-approved study included 11,998 CT pulmonary angiography (CTPA) reports. We built a retrieval bank from 11,399 reports and reserved 599 reports for testing. GPT-4o and LLaMA 3.1-70B generated impressions from the "findings" section using three setups: zero-shot, fixed random few-shot, and dynamic retrieval-selected few-shot (top-k semantic matches; k = 3/5/10). We ran temperatures 0, 0.7, 1. We scored outputs against the original impressions with ROUGE and BERTScore F1, report mean scores with 95% confidence intervals, and tested for statistical significance using Wilcoxon signed-rank test.

Dynamic retrieval-based few-shot prompting outperformed zero-shot and fixed few-shot prompting across all configurations (all p < 0.05). The highest scores were observed at temperature 0 and k = 10. ROUGE-1 F1 increased to 0.44-0.47 for GPT-4o and 0.37-0.50 for LLaMA, versus 0.35-0.37 and 0.25-0.37, respectively, in zero-shot prompting. Lower temperature and larger k were associated with higher similarity scores.

Dynamic, case-matched retrieval improved alignment of LLM-generated CTPA impressions with reference impressions on automated text-similarity metrics. Scores remained moderate, and radiologists' verification is still required before clinical deployment.

Breast ultrasound (US) imaging is essential for early breast cancer detection, yet generating diagnostic reports is labor-intensive, particularly when incorporating multimodal elastography.

This study presents a novel framework that combines multimodal large language models and deep learning to generate semiautomated breast US reports. This framework bridges the gap between manual and fully automated workflows by integrating radiologist annotations with advanced image classification and structured report compilation. A total of 2,119 elastography images and 60 annotated patient cases were retrospectively collected from two US machines.

The system demonstrated robust performance in elastography classification, achieving areas under the receiver operating characteristic curve of 0.92, 0.91, and 0.88 for shear-wave, strain, and Doppler images, respectively. In the evaluated dataset, the report generation module correctly identified all suspicious masses across both US machines, achieving 100% sensitivity in lesion detection, with an average report generation time of 31 s per patient using the GE Healthcare machine and 36 s using the Supersonic Image machine.

The proposed framework enables accurate, efficient, and device-adaptable breast US report generation by combining multimodal DL and prompt-based LLM inference. It significantly reduces radiologist workload and demonstrates potential for scalable deployment in real-world clinical workflows.