LLMs have demonstrated significant potential in Medical Report Generation (MRG), yet their development requires large amounts of medical image-report pairs, which are commonly scattered across multiple centers. Centralizing these data is exceptionally challenging due to privacy regulations, thereby impeding model development and broader adoption of LLM-driven MRG models. To address this challenge, we present FedMRG, the first framework that leverages Federated Learning (FL) to enable privacy-preserving, multi-center development of LLM-driven MRG models, specifically designed to overcome the critical challenge of communication-efficient LLM training under multi-modal data heterogeneity. To start with, our framework tackles the fundamental challenge of communication overhead in FL-LLM tuning by employing low-rank factorization to efficiently decompose parameter updates, significantly reducing gradient transmission costs and making LLM-driven MRG feasible in bandwidth-constrained FL settings. Furthermore, we observed the dual heterogeneity in MRG under the FL scenario: varying image characteristics across medical centers, as well as diverse reporting styles and terminology preferences. To address this, we further enhance FedMRG with (1) client-aware contrastive learning in the MRG encoder, coupled with diagnosis-driven prompts, which capture both globally generalizable and locally distinctive features while maintaining diagnostic accuracy; and (2) a dual-adapter mutual boosting mechanism in the MRG decoder that harmonizes generic and specialized adapters to address variations in reporting styles and terminology. Through extensive evaluation of our established FL-MRG benchmark, we demonstrate the generalizability and adaptability of FedMRG, underscoring its potential in harnessing multi-center data and generating clinically accurate reports while maintaining communication efficiency.

Medical report generation from imaging data remains a challenging task in clinical practice. While large language models (LLMs) show great promise in addressing this challenge, their effective integration with medical imaging data still deserves in-depth exploration. In this paper, we present MRG-LLM, a novel multimodal large language model (MLLM) that combines a frozen LLM with a learnable visual encoder and introduces a dynamic prompt customization mechanism. Our key innovation lies in generating instance-specific prompts tailored to individual medical images through conditional affine transformations derived from visual features. We propose two implementations: prompt-wise and promptbook-wise customization, enabling precise and targeted report generation. Extensive experiments on IU X-ray and MIMIC-CXR datasets demonstrate that MRG-LLM achieves state-of-the-art performance in medical report generation. Our code will be made publicly available.

<i>"Just Accepted" papers have undergone full peer review and have been accepted for publication in <i>Radiology: Artificial Intelligence</i>. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content.</i> Purpose To develop and evaluate an artificial intelligence (AI) system for generating breast ultrasound (BUS) reports. Materials and Methods This retrospective study included 104,364 cases from three hospitals (January 2020-December 2022). The AI system was trained on 82,896 cases, validated on 10,385 cases, and tested on an internal set (10,383 cases) and two external sets (300 and 400 cases). Under blind review, three senior radiologists (> 10 years of experience) evaluated AI-generated reports and those written by one midlevel radiologist (7 years of experience), as well as reports from three junior radiologists (2-3 years of experience) with and without AI assistance. The primary outcomes included the acceptance rates of Breast Imaging Reporting and Data System (BI-RADS) categories and lesion characteristics. Statistical analysis included one-sided and two-sided McNemar tests for non-inferiority and significance testing. Results In external test set 1 (300 cases), the midlevel radiologist and AI system achieved BI-RADS acceptance rates of 95.00% [285/300] versus 92.33% [277/300] (<i>P</i> < .001; non-inferiority test with a prespecified margin of 10%). In external test set 2 (400 cases), three junior radiologists had BI-RADS acceptance rates of 87.00% [348/400] versus 90.75% [363/400] (<i>P</i> = .06), 86.50% [346/400] versus 92.00% [368/400] ( <i>P</i> = .007), and 84.75% [339/400] versus 90.25% [361/400] (<i>P</i> = .02) with and without AI assistance, respectively. Conclusion The AI system performed comparably to a midlevel radiologist and aided junior radiologists in BI-RADS classification. ©RSNA, 2025.

Different medical imaging modalities capture diagnostic information at varying spatial resolutions, from coarse global patterns to fine-grained localized structures. However, most existing vision-language frameworks in the medical domain apply a uniform strategy for local feature extraction, overlooking the modality-specific demands. In this work, we present MedMoE, a modular and extensible vision-language processing framework that dynamically adapts visual representation based on the diagnostic context. MedMoE incorporates a Mixture-of-Experts (MoE) module conditioned on the report type, which routes multi-scale image features through specialized expert branches trained to capture modality-specific visual semantics. These experts operate over feature pyramids derived from a Swin Transformer backbone, enabling spatially adaptive attention to clinically relevant regions. This framework produces localized visual representations aligned with textual descriptions, without requiring modality-specific supervision at inference. Empirical results on diverse medical benchmarks demonstrate that MedMoE improves alignment and retrieval performance across imaging modalities, underscoring the value of modality-specialized visual representations in clinical vision-language systems.

Diagnostic imaging interpretation involves distilling multimodal clinical information into text form, a task well-suited to augmentation by generative artificial intelligence (AI). However, to our knowledge, impacts of AI-based draft radiological reporting remain unstudied in clinical settings. To prospectively evaluate the association of radiologist use of a workflow-integrated generative model capable of providing draft radiological reports for plain radiographs across a tertiary health care system with documentation efficiency, the clinical accuracy and textual quality of final radiologist reports, and the model's potential for detecting unexpected, clinically significant pneumothorax. This prospective cohort study was conducted from November 15, 2023, to April 24, 2024, at a tertiary care academic health system. The association between use of the generative model and radiologist documentation efficiency was evaluated for radiographs documented with model assistance compared with a baseline set of radiographs without model use, matched by study type (chest or nonchest). Peer review was performed on model-assisted interpretations. Flagging of pneumothorax requiring intervention was performed on radiographs prospectively. The primary outcomes were association of use of the generative model with radiologist documentation efficiency, assessed by difference in documentation time with and without model use using a linear mixed-effects model; for peer review of model-assisted reports, the difference in Likert-scale ratings using a cumulative-link mixed model; and for flagging pneumothorax requiring intervention, sensitivity and specificity. A total of 23 960 radiographs (11 980 each with and without model use) were used to analyze documentation efficiency. Interpretations with model assistance (mean [SE], 159.8 [27.0] seconds) were faster than the baseline set of those without (mean [SE], 189.2 [36.2] seconds) (P = .02), representing a 15.5% documentation efficiency increase. Peer review of 800 studies showed no difference in clinical accuracy (χ2 = 0.68; P = .41) or textual quality (χ2 = 3.62; P = .06) between model-assisted interpretations and nonmodel interpretations. Moreover, the model flagged studies containing a clinically significant, unexpected pneumothorax with a sensitivity of 72.7% and specificity of 99.9% among 97 651 studies screened. In this prospective cohort study of clinical use of a generative model for draft radiological reporting, model use was associated with improved radiologist documentation efficiency while maintaining clinical quality and demonstrated potential to detect studies containing a pneumothorax requiring immediate intervention. This study suggests the potential for radiologist and generative AI collaboration to improve clinical care delivery.

To evaluate the feasibility of generative pre-trained transformer-4 (GPT-4) in generating structured reports (SRs) from mixed-language (English and Korean) narrative-style CT reports for pancreatic ductal adenocarcinoma (PDAC) and to assess its accuracy in categorizing PDCA resectability. This retrospective study included consecutive free-text reports of pancreas-protocol CT for staging PDAC, from two institutions, written in English or Korean from January 2021 to December 2023. Both the GPT-4 Turbo and GPT-4o models were provided prompts along with the free-text reports via an application programming interface and tasked with generating SRs and categorizing tumor resectability according to the National Comprehensive Cancer Network guidelines version 2.2024. Prompts were optimized using the GPT-4 Turbo model and 50 reports from Institution B. The performances of the GPT-4 Turbo and GPT-4o models in the two tasks were evaluated using 115 reports from Institution A. Results were compared with a reference standard that was manually derived by an abdominal radiologist. Each report was consecutively processed three times, with the most frequent response selected as the final output. Error analysis was guided by the decision rationale provided by the models. Of the 115 narrative reports tested, 96 (83.5%) contained both English and Korean. For SR generation, GPT-4 Turbo and GPT-4o demonstrated comparable accuracies (92.3% [1592/1725] and 92.2% [1590/1725], respectively; <i>P</i> = 0.923). In the resectability categorization, GPT-4 Turbo showed higher accuracy than GPT-4o (81.7% [94/115] vs. 67.0% [77/115], respectively; <i>P</i> = 0.002). In the error analysis of GPT-4 Turbo, the SR generation error rate was 7.7% (133/1725 items), which was primarily attributed to inaccurate data extraction (54.1% [72/133]). The resectability categorization error rate was 18.3% (21/115), with the main cause being violation of the resectability criteria (61.9% [13/21]). Both GPT-4 Turbo and GPT-4o demonstrated acceptable accuracy in generating NCCN-based SRs on PDACs from mixed-language narrative reports. However, oversight by human radiologists is essential for determining resectability based on CT findings.

Recent advancements in multimodal Large Language Models (LLMs) have significantly enhanced the automation of medical image analysis, particularly in generating radiology reports from chest X-rays (CXR). However, these models still suffer from hallucinations and clinically significant errors, limiting their reliability in real-world applications. In this study, we propose Look & Mark (L&M), a novel grounding fixation strategy that integrates radiologist eye fixations (Look) and bounding box annotations (Mark) into the LLM prompting framework. Unlike conventional fine-tuning, L&M leverages in-context learning to achieve substantial performance gains without retraining. When evaluated across multiple domain-specific and general-purpose models, L&M demonstrates significant gains, including a 1.2% improvement in overall metrics (A.AVG) for CXR-LLaVA compared to baseline prompting and a remarkable 9.2% boost for LLaVA-Med. General-purpose models also benefit from L&M combined with in-context learning, with LLaVA-OV achieving an 87.3% clinical average performance (C.AVG)-the highest among all models, even surpassing those explicitly trained for CXR report generation. Expert evaluations further confirm that L&M reduces clinically significant errors (by 0.43 average errors per report), such as false predictions and omissions, enhancing both accuracy and reliability. These findings highlight L&M's potential as a scalable and efficient solution for AI-assisted radiology, paving the way for improved diagnostic workflows in low-resource clinical settings.

The automated generation of radiology reports from chest X-ray images holds significant promise in enhancing diagnostic workflows while preserving patient privacy. Traditional centralized approaches often require sensitive data transfer, posing privacy concerns. To address this, the study proposes a Multimodal Federated Learning framework for chest X-ray report generation using the IU-Xray dataset. The system utilizes a Vision Transformer (ViT) as the encoder and GPT-2 as the report generator, enabling decentralized training without sharing raw data. Three Federated Learning (FL) aggregation strategies: FedAvg, Krum Aggregation and a novel Loss-aware Federated Averaging (L-FedAvg) were evaluated. Among these, Krum Aggregation demonstrated superior performance across lexical and semantic evaluation metrics such as ROUGE, BLEU, BERTScore and RaTEScore. The results show that FL can match or surpass centralized models in generating clinically relevant and semantically rich radiology reports. This lightweight and privacy-preserving framework paves the way for collaborative medical AI development without compromising data confidentiality.

The task of generating medical reports automatically is of paramount importance in modern healthcare, offering a substantial reduction in the workload of radiologists and accelerating the processes of clinical diagnosis and treatment. Current challenges include handling limited sample sizes and interpreting intricate multi-modal and multi-view medical data. In order to improve the accuracy and efficiency for radiologists, we conducted this investigation. This study aims to present a novel methodology for medical report generation that leverages Multi-View Contrastive Learning (MVCL) applied to MRI data, combined with a Symptom Consultant (SC) for extracting medical insights, to improve the quality and efficiency of automated medical report generation. We introduce an advanced MVCL framework that maximizes the potential of multi-view MRI data to enhance visual feature extraction. Alongside, the SC component is employed to distill critical medical insights from symptom descriptions. These components are integrated within a transformer decoder architecture, which is then applied to the Deep Wrist dataset for model training and evaluation. Our experimental analysis on the Deep Wrist dataset reveals that our proposed integration of MVCL and SC significantly outperforms the baseline model in terms of accuracy and relevance of the generated medical reports. The results indicate that our approach is particularly effective in capturing and utilizing the complex information inherent in multi-modal and multi-view medical datasets. The combination of MVCL and SC constitutes a powerful approach to medical report generation, addressing the existing challenges in the field. The demonstrated superiority of our model over traditional methods holds promise for substantial improvements in clinical diagnosis and automated report generation, indicating a significant stride forward in medical technology.

Medical image captioning is a challenging task that requires generating clinically accurate and semantically meaningful descriptions of radiology images. While recent vision-language models (VLMs) such as BLIP, BLIP2, Gemini and ViT-GPT2 show strong performance on natural image datasets, they often produce generic or imprecise captions when applied to specialized medical domains. In this project, we explore the effectiveness of fine-tuning the BLIP model on the ROCO dataset for improved radiology captioning. We compare the fine-tuned BLIP against its zero-shot version, BLIP-2 base, BLIP-2 Instruct and a ViT-GPT2 transformer baseline. Our results demonstrate that domain-specific fine-tuning on BLIP significantly improves performance across both quantitative and qualitative evaluation metrics. We also visualize decoder cross-attention maps to assess interpretability and conduct an ablation study to evaluate the contributions of encoder-only and decoder-only fine-tuning. Our findings highlight the importance of targeted adaptation for medical applications and suggest that decoder-only fine-tuning (encoder-frozen) offers a strong performance baseline with 5% lower training time than full fine-tuning, while full model fine-tuning still yields the best results overall.