Existing radiology report generation (RRG) studies mostly adopt autoregressive (AR) approaches to produce textual descriptions token-by-token for specific clinical radiographs, where they are susceptible to error propagation problems if irrelevant contents are half-way generated, leading to potential ill-presenting of precise diagnoses, especially when there exist complicated abnormalities in radiographs. Although the non-AR paradigm, e.g., diffusion model, provides an alternative solution to tackle the problem from AR by generating all contents in parallel, the mechanism of using Gaussian noise in existing diffusion models still has significant room to improve when such models are used in particular circumstances, i.e., providing proper guidance in controlling noises in the diffusive process to ensure precise report generation. In this paper, we propose to conduct RRG with diffusion networks by controlling the noise with task-specific features, which leverages irrelevant visual and textual information as noise rather than the stochastic Gaussian noise, and allows the diffusion networks to filter particular information through iterative denoising, thus performing a precise and controlled report generation process. Experiments on IU X-Ray and MIMIC-CXR demonstrate the superiority of our approach compared to strong baselines and state-of-the-art solutions. Human evaluation and noise type analysis show that comprehensive noise control greatly helps diffusion networks to refine the generation of global and local report contents.
29 August 2025 in Journal of Biomedical Informatics
Introduce the task of wound care multimodal multilingual visual question answering, provide baseline performances, and identify areas of future study. A dataset of wound care multimodal multilingual visual question answering (VQA) was created using consumer health questions asked online. Practicing US medical doctors were tasked with providing metadata and expert responses labels. Several instruct-enabled, multilingual visual question answering models (GPT-4o, Gemini-1.5-Pro, and Qwen-VL) were tested to benchmark performances. Finally, automatic evaluations were tested against domain expert response ratings. A multilingual dataset of 477 wound care cases, 768 responses, 748 images, 3k structured data labels, 1362 translation instances, and 10k judgements was constructed (https://osf.io/xsj5u/). Metadata scores ranged from 0.32-0.78 accuracy depending on classification type; response generation performances 0.06 BLEU, 0.66 BERTScore, 0.45 ROUGE-L in English and 0.12 BLEU, 0.69 BERTScore, and 0.50 ROUGE-L in Chinese. We construct and explore the tasks of multimodal, multilingual VQA. We hope the work here can inspire further research in wound care metadata classification, VQA response generation, and open response automatic evaluation.
31 August 2026 in The Canadian Journal of Psychiatry
Background: We evaluated whether summaries and large language models (LLMs) preserve predictive performance for inpatient violence risk and assessed performance across demographics. Methods: We conducted a retrospective study of inpatient encounters at Harborview Medical Center between March 2021 and September 2023. Cases were inpatient behavioural violence events, and each was matched to 10 control encounters. For each encounter, we used clinical notes from the 72 hours before the event for cases and a matched window for controls. We compared Clinical-Longformer trained on original notes with pipelines that first generated summaries and then classified risk. Summaries were either general or guided by clinician-defined risk entities. We also tested Llama-3.1–8B and MedGemma-27B in zero-shot (ZS) and fine-tuned settings. Primary endpoints were positive-class precision, recall, and F1. We also reported overall and subgroup area under the receiver operating characteristic curve (AUROC) with bootstrap confidence intervals for sex, race, ethnicity, age, and mental health flag. Results: Clinical-Longformer on original notes achieved the strongest performance (F1 = 0.776, AUROC = 0.959). Replacing original notes with general or entity-guided summaries reduced performance (F1 = 0.465 and 0.510). Llama-3.1–8B improved with sequence-classification fine-tuning on original notes but remained below Clinical-Longformer (F1 = 0.701; AUROC = 0.933), while ZS prompting performed poorly. MedGemma-27B also remained inferior, with its best performance from sequence-classification fine-tuning on original notes (F1 = 0.612, AUROC = 0.947). Subgroup AUROCs were high across sex, race, ethnicity, and age, with lower performance among patients with a documented mental health flag. Conclusion: Direct classification of original notes remained the most reliable strategy. Summaries of clinical notes may compress or alter cues needed for discrimination, and off-the-shelf prompting was inadequate as a stand-alone predictor. A practical path is to anchor risk scoring in long-context discriminative models and use LLMs for auditable, clinician-facing summaries or rationales. Prospective and external validation are needed before clinical deployment.