Jean-Benoit Delbrouck

Sabri Eyuboglu, Maya Varma, Khaled Saab et al. · stanford

Machine learning models that achieve high overall accuracy often make systematic errors on important subsets (or slices) of data. Identifying underperforming slices is particularly challenging when working with high-dimensional inputs (e.g. images, audio), where important slices are often unlabeled. In order to address this issue, recent studies have proposed automated slice discovery methods (SDMs), which leverage learned model representations to mine input data for slices on which a model performs poorly. To be useful to a practitioner, these methods must identify slices that are both underperforming and coherent (i.e. united by a human-understandable concept). However, no quantitative evaluation framework currently exists for rigorously assessing SDMs with respect to these criteria. Additionally, prior qualitative evaluations have shown that SDMs often identify slices that are incoherent. In this work, we address these challenges by first designing a principled evaluation framework that enables a quantitative comparison of SDMs across 1,235 slice discovery settings in three input domains (natural images, medical images, and time-series data). Then, motivated by the recent development of powerful cross-modal representation learning approaches, we present Domino, an SDM that leverages cross-modal embeddings and a novel error-aware mixture model to discover and describe coherent slices. We find that Domino accurately identifies 36% of the 1,235 slices in our framework - a 12 percentage point improvement over prior methods. Further, Domino is the first SDM that can provide natural language descriptions of identified slices, correctly generating the exact name of the slice in 35% of settings.

Yunhe Gao, Yabin Zhang, Chong Wang et al. · stanford

Foundation models have transformed vision and language by learning general-purpose representations from large-scale unlabeled data, yet 3D medical imaging lacks analogous approaches. Existing self-supervised methods rely on low-level reconstruction or contrastive objectives that fail to capture the anatomical semantics critical for medical image analysis, limiting transfer to downstream tasks. We present MASS (MAsk-guided Self-Supervised learning), which treats in-context segmentation as the pretext task for learning general-purpose medical imaging representations. MASS's key insight is that automatically generated class-agnostic masks provide sufficient structural supervision for learning semantically rich representations. By training on thousands of diverse mask proposals spanning anatomical structures and pathological findings, MASS learns what semantically defines medical structures: the holistic combination of appearance, shape, spatial context, and anatomical relationships. We demonstrate effectiveness across data regimes: from small-scale pretraining on individual datasets (20-200 scans) to large-scale multi-modal pretraining on 5K CT, MRI, and PET volumes, all without annotations. MASS demonstrates: (i) few-shot segmentation on novel structures, (ii) matching full supervision with only 20-40\% labeled data while outperforming self-supervised baselines by over 20 in Dice score in low-data regimes, and (iii) frozen-encoder classification on unseen pathologies that matches full supervised training with thousands of samples. Mask-guided self-supervised pretraining captures broadly generalizable knowledge, opening a path toward 3D medical imaging foundation models without expert annotations. Code is available: https://github.com/Stanford-AIMI/MASS.

Justin Xu, Zhihong Chen, Andrew Johnston et al. · oxford, stanford

Recent developments in natural language generation have tremendous implications for healthcare. For instance, state-of-the-art systems could automate the generation of sections in clinical reports to alleviate physician workload and streamline hospital documentation. To explore these applications, we present a shared task consisting of two subtasks: (1) Radiology Report Generation (RRG24) and (2) Discharge Summary Generation ("Discharge Me!"). RRG24 involves generating the 'Findings' and 'Impression' sections of radiology reports given chest X-rays. "Discharge Me!" involves generating the 'Brief Hospital Course' and 'Discharge Instructions' sections of discharge summaries for patients admitted through the emergency department. "Discharge Me!" submissions were subsequently reviewed by a team of clinicians. Both tasks emphasize the goal of reducing clinician burnout and repetitive workloads by generating documentation. We received 201 submissions from across 8 teams for RRG24, and 211 submissions from across 16 teams for "Discharge Me!".

Dave Van Veen, Cara Van Uden, Louis Blankemeier et al.

Analyzing vast textual data and summarizing key information from electronic health records imposes a substantial burden on how clinicians allocate their time. Although large language models (LLMs) have shown promise in natural language processing (NLP), their effectiveness on a diverse range of clinical summarization tasks remains unproven. In this study, we apply adaptation methods to eight LLMs, spanning four distinct clinical summarization tasks: radiology reports, patient questions, progress notes, and doctor-patient dialogue. Quantitative assessments with syntactic, semantic, and conceptual NLP metrics reveal trade-offs between models and adaptation methods. A clinical reader study with ten physicians evaluates summary completeness, correctness, and conciseness; in a majority of cases, summaries from our best adapted LLMs are either equivalent (45%) or superior (36%) compared to summaries from medical experts. The ensuing safety analysis highlights challenges faced by both LLMs and medical experts, as we connect errors to potential medical harm and categorize types of fabricated information. Our research provides evidence of LLMs outperforming medical experts in clinical text summarization across multiple tasks. This suggests that integrating LLMs into clinical workflows could alleviate documentation burden, allowing clinicians to focus more on patient care.

Benjamin Warner, Ratna Sagari Grandhi, Max Kieffer et al.

Evaluating large language models (LLMs) for medical applications remains challenging due to benchmark saturation, limited data accessibility, and insufficient coverage of relevant tasks. Existing suites have either saturated, heavily depend on restricted datasets, or lack comprehensive model coverage. We introduce Medmarks, a fully open-source evaluation suite with 30 benchmarks spanning question answering, information extraction, medical calculations, and open-ended clinical reasoning. We perform a systematic evaluation of 61 models across 71 configurations using verifiable metrics and LLM-as-a-Judge. Our results show that frontier reasoning models (Gemini 3 Pro Preview, GPT-5.1, & GPT-5.2) achieve the highest performance across both benchmarks, most frontier proprietary models are significantly more token efficient than open-weight alternatives, medically fine-tuned models outperform their generalist counterparts, and that models are susceptible to answer-order bias (particularly smaller models and Grok 4). A subset of our evals (Medmarks-T) can be directly used as reinforcement learning environments to post-train LLMs for medical reasoning. Code is available at https://github.com/MedARC-AI/Medmarks

Maya Varma, Jean-Benoit Delbrouck, Sarah Hooper et al. · stanford

Vision-language models (VLMs), such as CLIP and ALIGN, are generally trained on datasets consisting of image-caption pairs obtained from the web. However, real-world multimodal datasets, such as healthcare data, are significantly more complex: each image (e.g. X-ray) is often paired with text (e.g. physician report) that describes many distinct attributes occurring in fine-grained regions of the image. We refer to these samples as exhibiting high pairwise complexity, since each image-text pair can be decomposed into a large number of region-attribute pairings. The extent to which VLMs can capture fine-grained relationships between image regions and textual attributes when trained on such data has not been previously evaluated. The first key contribution of this work is to demonstrate through systematic evaluations that as the pairwise complexity of the training dataset increases, standard VLMs struggle to learn region-attribute relationships, exhibiting performance degradations of up to 37% on retrieval tasks. In order to address this issue, we introduce ViLLA as our second key contribution. ViLLA, which is trained to capture fine-grained region-attribute relationships from complex datasets, involves two components: (a) a lightweight, self-supervised mapping model to decompose image-text samples into region-attribute pairs, and (b) a contrastive VLM to learn representations from generated region-attribute pairs. We demonstrate with experiments across four domains (synthetic, product, medical, and natural images) that ViLLA outperforms comparable VLMs on fine-grained reasoning tasks, such as zero-shot object detection (up to 3.6 AP50 points on COCO and 0.6 mAP points on LVIS) and retrieval (up to 14.2 R-Precision points).

Yabin Zhang, Maya Varma, Yunhe Gao et al.

Out-of-distribution (OOD) detection aims to identify samples that deviate from in-distribution (ID). One popular pipeline addresses this by introducing negative labels distant from ID classes and detecting OOD based on their distance to these labels. However, such labels may present poor activation on OOD samples, failing to capture the OOD characteristics. To address this, we propose \underline{T}est-time \underline{A}ctivated \underline{N}egative \underline{L}abels (TANL) by dynamically evaluating activation levels across the corpus dataset and mining candidate labels with high activation responses during the testing process. Specifically, TANL identifies high-confidence test images online and accumulates their assignment probabilities over the corpus to construct a label activation metric. Such a metric leverages historical test samples to adaptively align with the test distribution, enabling the selection of distribution-adaptive activated negative labels. By further exploring the activation information within the current testing batch, we introduce a more fine-grained, batch-adaptive variant. To fully utilize label activation knowledge, we propose an activation-aware score function that emphasizes negative labels with stronger activations, boosting performance and enhancing its robustness to the label number. Our TANL is training-free, test-efficient, and grounded in theoretical justification. Experiments on diverse backbones and wide task settings validate its effectiveness. Notably, on the large-scale ImageNet benchmark, TANL significantly reduces the FPR95 from 17.5\% to 9.8\%. Codes are available at \href{https://github.com/YBZh/OpenOOD-VLM}{YBZh/OpenOOD-VLM}.

Pierre Chambon, Christian Bluethgen, Jean-Benoit Delbrouck et al.

Multimodal models trained on large natural image-text pair datasets have exhibited astounding abilities in generating high-quality images. Medical imaging data is fundamentally different to natural images, and the language used to succinctly capture relevant details in medical data uses a different, narrow but semantically rich, domain-specific vocabulary. Not surprisingly, multi-modal models trained on natural image-text pairs do not tend to generalize well to the medical domain. Developing generative imaging models faithfully representing medical concepts while providing compositional diversity could mitigate the existing paucity of high-quality, annotated medical imaging datasets. In this work, we develop a strategy to overcome the large natural-medical distributional shift by adapting a pre-trained latent diffusion model on a corpus of publicly available chest x-rays (CXR) and their corresponding radiology (text) reports. We investigate the model's ability to generate high-fidelity, diverse synthetic CXR conditioned on text prompts. We assess the model outputs quantitatively using image quality metrics, and evaluate image quality and text-image alignment by human domain experts. We present evidence that the resulting model (RoentGen) is able to create visually convincing, diverse synthetic CXR images, and that the output can be controlled to a new extent by using free-form text prompts including radiology-specific language. Fine-tuning this model on a fixed training set and using it as a data augmentation method, we measure a 5% improvement of a classifier trained jointly on synthetic and real images, and a 3% improvement when trained on a larger but purely synthetic training set. Finally, we observe that this fine-tuning distills in-domain knowledge in the text-encoder and can improve its representation capabilities of certain diseases like pneumothorax by 25%.

Maya Varma, Jean-Benoit Delbrouck, Sophie Ostmeier et al.

Vision-language models (VLMs) have made great strides in addressing temporal understanding tasks, which involve characterizing visual changes across a sequence of images. However, recent works have suggested that when making predictions, VLMs may rely on static feature biases, such as background or object features, rather than dynamic visual changes. Static feature biases are a type of shortcut and can contribute to systematic prediction errors on downstream tasks; as a result, identifying and characterizing error-inducing static feature biases is critical prior to real-world model deployment. In this work, we introduce TRoVe, an automated approach for discovering error-inducing static feature biases learned by temporal VLMs. Given a trained VLM and an annotated validation dataset associated with a downstream classification task, TRoVe extracts candidate static features from the dataset and scores each feature by (i) the effect of the feature on classification errors as well as (ii) the extent to which the VLM relies on the feature when making predictions. In order to quantitatively evaluate TRoVe, we introduce an evaluation framework consisting of 101 trained temporal VLMs paired with ground-truth annotations for learned static feature biases. We use this framework to demonstrate that TRoVe can accurately identify error-inducing static feature biases in VLMs, achieving a 28.6% improvement over the closest baseline. Finally, we apply TRoVe to 7 off-the-shelf VLMs and 2 temporal understanding tasks, surfacing previously-unknown static feature biases and demonstrating that knowledge of learned biases can aid in improving model performance at test time. Our code is available at https://github.com/Stanford-AIMI/TRoVe.

Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen et al.

Neural image-to-text radiology report generation systems offer the potential to improve radiology reporting by reducing the repetitive process of report drafting and identifying possible medical errors. These systems have achieved promising performance as measured by widely used NLG metrics such as BLEU and CIDEr. However, the current systems face important limitations. First, they present an increased complexity in architecture that offers only marginal improvements on NLG metrics. Secondly, these systems that achieve high performance on these metrics are not always factually complete or consistent due to both inadequate training and evaluation. Recent studies have shown the systems can be substantially improved by using new methods encouraging 1) the generation of domain entities consistent with the reference and 2) describing these entities in inferentially consistent ways. So far, these methods rely on weakly-supervised approaches (rule-based) and named entity recognition systems that are not specific to the chest X-ray domain. To overcome this limitation, we propose a new method, the RadGraph reward, to further improve the factual completeness and correctness of generated radiology reports. More precisely, we leverage the RadGraph dataset containing annotated chest X-ray reports with entities and relations between entities. On two open radiology report datasets, our system substantially improves the scores up to 14.2% and 25.3% on metrics evaluating the factual correctness and completeness of reports.

Zhihong Chen, Maya Varma, Xiang Wan et al.

Radiology report summarization (RRS) is a growing area of research. Given the Findings section of a radiology report, the goal is to generate a summary (called an Impression section) that highlights the key observations and conclusions of the radiology study. However, RRS currently faces essential limitations.First, many prior studies conduct experiments on private datasets, preventing reproduction of results and fair comparisons across different systems and solutions. Second, most prior approaches are evaluated solely on chest X-rays. To address these limitations, we propose a dataset (MIMIC-RRS) involving three new modalities and seven new anatomies based on the MIMIC-III and MIMIC-CXR datasets. We then conduct extensive experiments to evaluate the performance of models both within and across modality-anatomy pairs in MIMIC-RRS. In addition, we evaluate their clinical efficacy via RadGraph, a factual correctness metric.

Yabin Zhang, Chong Wang, Yunhe Gao et al.

Chest X-rays (CXRs) are among the most frequently performed imaging examinations worldwide, yet rising imaging volumes increase radiologist workload and the risk of diagnostic errors. Although artificial intelligence (AI) systems have shown promise for CXR interpretation, most generate only final predictions, without making explicit how visual evidence is translated into radiographic findings and diagnostic predictions. We present CheXOne, a reasoning-enabled vision-language model for CXR interpretation. CheXOne jointly generates diagnostic predictions and explicit, clinically grounded reasoning traces that connect visual evidence, radiographic findings, and these predictions. The model is trained on 14.7 million instruction and reasoning samples curated from 30 public datasets spanning 36 CXR interpretation tasks, using a two-stage framework that combines instruction tuning with reinforcement learning to improve reasoning quality. We evaluate CheXOne in zero-shot settings across visual question answering, report generation, visual grounding and reasoning assessment, covering 17 evaluation settings. CheXOne outperforms existing medical and general-domain foundation models and achieves strong performance on independent public benchmarks. A clinical reader study demonstrates that CheXOne-drafted reports are comparable to or better than resident-written reports in 55% of cases, while effectively addressing clinical indications and enhancing both report writing and CXR interpretation efficiency. Further analyses involving radiologists reveal that the generated reasoning traces show high clinical factuality and provide causal support for the final predictions, offering a plausible explanation for the performance gains. These results suggest that explicit reasoning can improve model performance, interpretability and clinical utility in AI-assisted CXR interpretation.

Mélanie Roschewitz, Kenneth Styppa, Yitian Tao et al.

Vision-language models (VLM) have markedly advanced AI-driven interpretation and reporting of complex medical imaging, such as computed tomography (CT). Yet, existing methods largely relegate clinicians to passive observers of final outputs, offering no interpretable reasoning trace for them to inspect, validate, or refine. To address this, we introduce RadAgent, a tool-using AI agent that generates CT reports through a stepwise and interpretable process. Each resulting report is accompanied by a fully inspectable trace of intermediate decisions and tool interactions, allowing clinicians to examine how the reported findings are derived. In our experiments, we observe that RadAgent improves Chest CT report generation over its 3D VLM counterpart, CT-Chat, across three dimensions. Clinical accuracy improves by 6.0 points (36.4% relative) in macro-F1 and 5.4 points (19.6% relative) in micro-F1. Robustness under adversarial conditions improves by 24.7 points (41.9% relative). Furthermore, RadAgent achieves 37.0% in faithfulness, a new capability entirely absent in its 3D VLM counterpart. By structuring the interpretation of chest CT as an explicit, tool-augmented and iterative reasoning trace, RadAgent brings us closer toward transparent and reliable AI for radiology.

Chong Wang, Yabin Zhang, Yunhe Gao et al.

Foundation models for medical imaging are typically pretrained on increasingly large datasets, following a "scale-at-all-costs" paradigm. However, this strategy faces two critical challenges: large-scale medical datasets often contain substantial redundancy and severe class imbalance that bias representation learning toward over-represented patterns, and indiscriminate training regardless of heterogeneity in data quality incurs considerable computational inefficiency. Here we demonstrate that active, principled data curation during pretraining can serve as a viable, cost-effective alternative to brute-force dataset enlargement. We introduce CheXficient, a chest X-ray (CXR) foundation model that selectively prioritizes informative training samples. CheXficient is pretrained on only 22.7% of 1,235,004 paired CXR images and reports while consuming under 27.3% of the total compute budget, yet achieving comparable or superior performance to its full-data counterpart and other large-scale pretrained models. We assess CheXficient across 20 individual benchmarks spanning 5 task types, including non-adapted off-the-shelf evaluations (zero-shot findings classification and crossmodal retrieval) and adapted downstream tasks (disease prediction, semantic segmentation, and radiology report generation). Further analyses show that CheXficient systematically prioritizes under-represented training samples, improving generalizability on long-tailed or rare conditions. Overall, our work offers practical insights into the data and computation demands for efficient pretraining and downstream adaptation of medical vision-language foundation models.

Eva Prakash, Maayane Attias, Pierre Chambon et al.

Objective: To enhance automated de-identification of radiology reports by scaling transformer-based models through extensive training datasets and benchmarking performance against commercial cloud vendor systems for protected health information (PHI) detection. Materials and Methods: In this retrospective study, we built upon a state-of-the-art, transformer-based, PHI de-identification pipeline by fine-tuning on two large annotated radiology corpora from Stanford University, encompassing chest X-ray, chest CT, abdomen/pelvis CT, and brain MR reports and introducing an additional PHI category (AGE) into the architecture. Model performance was evaluated on test sets from Stanford and the University of Pennsylvania (Penn) for token-level PHI detection. We further assessed (1) the stability of synthetic PHI generation using a "hide-in-plain-sight" method and (2) performance against commercial systems. Precision, recall, and F1 scores were computed across all PHI categories. Results: Our model achieved overall F1 scores of 0.973 on the Penn dataset and 0.996 on the Stanford dataset, outperforming or maintaining the previous state-of-the-art model performance. Synthetic PHI evaluation showed consistent detectability (overall F1: 0.959 [0.958-0.960]) across 50 independently de-identified Penn datasets. Our model outperformed all vendor systems on synthetic Penn reports (overall F1: 0.960 vs. 0.632-0.754). Discussion: Large-scale, multimodal training improved cross-institutional generalization and robustness. Synthetic PHI generation preserved data utility while ensuring privacy. Conclusion: A transformer-based de-identification model trained on diverse radiology datasets outperforms prior academic and commercial systems in PHI detection and establishes a new benchmark for secure clinical text processing.

Justin Xu, Xi Zhang, Javid Abderezaei et al.

We introduce RadEval, a unified, open-source framework for evaluating radiology texts. RadEval consolidates a diverse range of metrics, from classic n-gram overlap (BLEU, ROUGE) and contextual measures (BERTScore) to clinical concept-based scores (F1CheXbert, F1RadGraph, RaTEScore, SRR-BERT, TemporalEntityF1) and advanced LLM-based evaluators (GREEN). We refine and standardize implementations, extend GREEN to support multiple imaging modalities with a more lightweight model, and pretrain a domain-specific radiology encoder, demonstrating strong zero-shot retrieval performance. We also release a richly annotated expert dataset with over 450 clinically significant error labels and show how different metrics correlate with radiologist judgment. Finally, RadEval provides statistical testing tools and baseline model evaluations across multiple publicly available datasets, facilitating reproducibility and robust benchmarking in radiology report generation.

Eva Prakash, Yunhe Gao, Chong Wang et al.

Chest radiograph interpretation requires temporal reasoning over prior and current studies, yet most vision-language models are trained on static image-report pairs and lack explicit supervision for modeling longitudinal change. We introduce CheXTemporal, a dataset for temporally grounded reasoning in chest radiography consisting of paired prior-current chest X-rays (CXR) with finding-level temporal and spatial annotations. The dataset includes a five-class progression taxonomy (new, worse, stable, improved, resolved), localized spatial supervision of pathology, explicit spatial-temporal alignment across paired studies, and multi-source coverage for cross-domain evaluation. We additionally construct a 280K-pair silver dataset with automatically derived temporal and anatomical supervision for large-scale evaluation under weaker supervision. Using these resources, we evaluate multiple state-of-the-art vision-language CXR models on grounding and progression-classification tasks in a zero-shot setting. Across both gold and silver evaluations, current models exhibit consistent limitations in spatial grounding, fine-grained temporal reasoning, and robustness under distribution shift. In particular, models perform substantially better on salient progression categories such as worse than on temporally subtle states such as stable and resolved, suggesting limited modeling of longitudinal disease evolution in chest radiography.

Johannes Moll, Markus Graf, Tristan Lemke et al.

Vision-language models (VLMs) often produce chain-of-thought (CoT) explanations that sound plausible yet fail to reflect the underlying decision process, undermining trust in high-stakes clinical use. Existing evaluations rarely catch this misalignment, prioritizing answer accuracy or adherence to formats. We present a clinically grounded framework for chest X-ray visual question answering (VQA) that probes CoT faithfulness via controlled text and image modifications across three axes: clinical fidelity, causal attribution, and confidence calibration. In a reader study (n=4), evaluator-radiologist correlations fall within the observed inter-radiologist range for all axes, with strong alignment for attribution (Kendall's $τ_b=0.670$), moderate alignment for fidelity ($τ_b=0.387$), and weak alignment for confidence tone ($τ_b=0.091$), which we report with caution. Benchmarking six VLMs shows that answer accuracy and explanation quality can be decoupled, acknowledging injected cues does not ensure grounding, and text cues shift explanations more than visual cues. While some open-source models match final answer accuracy, proprietary models score higher on attribution (25.0% vs. 1.4%) and often on fidelity (36.1% vs. 31.7%), highlighting deployment risks and the need to evaluate beyond final answer accuracy.

Stefania L. Moroianu, Christian Bluethgen, Pierre Chambon et al. · stanford

Achieving robust performance and fairness across diverse patient populations remains a challenge in developing clinically deployable deep learning models for diagnostic imaging. Synthetic data generation has emerged as a promising strategy to address limitations in dataset scale and diversity. We introduce RoentGen-v2, a text-to-image diffusion model for chest radiographs that enables fine-grained control over both radiographic findings and patient demographic attributes, including sex, age, and race/ethnicity. RoentGen-v2 is the first model to generate clinically plausible images with demographic conditioning, facilitating the creation of a large, demographically balanced synthetic dataset comprising over 565,000 images. We use this large synthetic dataset to evaluate optimal training pipelines for downstream disease classification models. In contrast to prior work that combines real and synthetic data naively, we propose an improved training strategy that leverages synthetic data for supervised pretraining, followed by fine-tuning on real data. Through extensive evaluation on over 137,000 chest radiographs from five institutions, we demonstrate that synthetic pretraining consistently improves model performance, generalization to out-of-distribution settings, and fairness across demographic subgroups. Across datasets, synthetic pretraining led to a 6.5% accuracy increase in the performance of downstream classification models, compared to a modest 2.7% increase when naively combining real and synthetic data. We observe this performance improvement simultaneously with the reduction of the underdiagnosis fairness gap by 19.3%. These results highlight the potential of synthetic imaging to advance equitable and generalizable medical deep learning under real-world data constraints. We open source our code, trained models, and synthetic dataset at https://github.com/StanfordMIMI/RoentGen-v2 .

Johannes Moll, Louisa Fay, Asfandyar Azhar et al.

Radiology reports are critical for clinical decision-making but often lack a standardized format, limiting both human interpretability and machine learning (ML) applications. While large language models (LLMs) have shown strong capabilities in reformatting clinical text, their high computational requirements, lack of transparency, and data privacy concerns hinder practical deployment. To address these challenges, we explore lightweight encoder-decoder models (<300M parameters)-specifically T5 and BERT2BERT-for structuring radiology reports from the MIMIC-CXR and CheXpert Plus datasets. We benchmark these models against eight open-source LLMs (1B-70B), adapted using prefix prompting, in-context learning (ICL), and low-rank adaptation (LoRA) finetuning. Our best-performing lightweight model outperforms all LLMs adapted using prompt-based techniques on a human-annotated test set. While some LoRA-finetuned LLMs achieve modest gains over the lightweight model on the Findings section (BLEU 6.4%, ROUGE-L 4.8%, BERTScore 3.6%, F1-RadGraph 1.1%, GREEN 3.6%, and F1-SRR-BERT 4.3%), these improvements come at the cost of substantially greater computational resources. For example, LLaMA-3-70B incurred more than 400 times the inference time, cost, and carbon emissions compared to the lightweight model. These results underscore the potential of lightweight, task-specific models as sustainable and privacy-preserving solutions for structuring clinical text in resource-constrained healthcare settings.

Louis Blankemeier, Ashwin Kumar, Joseph Paul Cohen et al.

The large volume of abdominal computed tomography (CT) scans coupled with the shortage of radiologists have intensified the need for automated medical image analysis tools. Previous state-of-the-art approaches for automated analysis leverage vision-language models (VLMs) that jointly model images and radiology reports. However, current medical VLMs are generally limited to 2D images and short reports. Here to overcome these shortcomings for abdominal CT interpretation, we introduce Merlin, a 3D VLM that learns from volumetric CT scans, electronic health record data and radiology reports. This approach is enabled by a multistage pretraining framework that does not require additional manual annotations. We trained Merlin using a high-quality clinical dataset of paired CT scans (>6 million images from 15,331 CT scans), diagnosis codes (>1.8 million codes) and radiology reports (>6 million tokens). We comprehensively evaluated Merlin on 6 task types and 752 individual tasks that covered diagnostic, prognostic and quality-related tasks. The non-adapted (off-the-shelf) tasks included zero-shot classification of findings (30 findings), phenotype classification (692 phenotypes) and zero-shot cross-modal retrieval (image-to-findings and image-to-impression). The model-adapted tasks included 5-year chronic disease prediction (6 diseases), radiology report generation and 3D semantic segmentation (20 organs). We validated Merlin at scale, with internal testing on 5,137 CT scans and external testing on 44,098 CT scans from 3 independent sites and 2 public datasets. The results demonstrated high generalization across institutions and anatomies. Merlin outperformed 2D VLMs, CT foundation models and off-the-shelf radiology models. We also release our trained models, code, and dataset, available at: https://github.com/StanfordMIMI/Merlin.

Sophie Ostmeier, Justin Xu, Zhihong Chen et al.

Evaluating radiology reports is a challenging problem as factual correctness is extremely important due to the need for accurate medical communication about medical images. Existing automatic evaluation metrics either suffer from failing to consider factual correctness (e.g., BLEU and ROUGE) or are limited in their interpretability (e.g., F1CheXpert and F1RadGraph). In this paper, we introduce GREEN (Generative Radiology Report Evaluation and Error Notation), a radiology report generation metric that leverages the natural language understanding of language models to identify and explain clinically significant errors in candidate reports, both quantitatively and qualitatively. Compared to current metrics, GREEN offers: 1) a score aligned with expert preferences, 2) human interpretable explanations of clinically significant errors, enabling feedback loops with end-users, and 3) a lightweight open-source method that reaches the performance of commercial counterparts. We validate our GREEN metric by comparing it to GPT-4, as well as to error counts of 6 experts and preferences of 2 experts. Our method demonstrates not only higher correlation with expert error counts, but simultaneously higher alignment with expert preferences when compared to previous approaches.

Jean-Benoit Delbrouck, Noé Tits, Mathilde Brousmiche et al.

Understanding expressed sentiment and emotions are two crucial factors in human multimodal language. This paper describes a Transformer-based joint-encoding (TBJE) for the task of Emotion Recognition and Sentiment Analysis. In addition to use the Transformer architecture, our approach relies on a modular co-attention and a glimpse layer to jointly encode one or more modalities. The proposed solution has also been submitted to the ACL20: Second Grand-Challenge on Multimodal Language to be evaluated on the CMU-MOSEI dataset. The code to replicate the presented experiments is open-source: https://github.com/jbdel/MOSEI_UMONS.

Zhihong Chen, Maya Varma, Justin Xu et al. · mila, oxford

Over 1.4 billion chest X-rays (CXRs) are performed annually due to their cost-effectiveness as an initial diagnostic test. This scale of radiological studies provides a significant opportunity to streamline CXR interpretation and documentation. While foundation models are a promising solution, the lack of publicly available large-scale datasets and benchmarks inhibits their iterative development and real-world evaluation. To overcome these challenges, we constructed a large-scale dataset (CheXinstruct), which we utilized to train a vision-language foundation model (CheXagent). We systematically demonstrated competitive performance across eight distinct task types on our novel evaluation benchmark (CheXbench). Beyond technical validation, we assessed the real-world utility of CheXagent in directly drafting radiology reports. Our clinical assessment with eight radiologists revealed a 36% time saving for residents using CheXagent-drafted reports, while attending radiologists showed no significant time difference editing resident-drafted or CheXagent-drafted reports. The CheXagent-drafted reports improved the writing efficiency of both radiology residents and attending radiologists in 81% and 61% of cases, respectively, without loss of quality. Overall, we demonstrate that CheXagent can effectively perform a variety of CXR interpretation tasks and holds potential to assist radiologists in routine clinical workflows.

Maya Varma, Jean-Benoit Delbrouck, Zhihong Chen et al.

Fine-tuned vision-language models (VLMs) often capture spurious correlations between image features and textual attributes, resulting in degraded zero-shot performance at test time. Existing approaches for addressing spurious correlations (i) primarily operate at the global image-level rather than intervening directly on fine-grained image features and (ii) are predominantly designed for unimodal settings. In this work, we present RaVL, which takes a fine-grained perspective on VLM robustness by discovering and mitigating spurious correlations using local image features rather than operating at the global image level. Given a fine-tuned VLM, RaVL first discovers spurious correlations by leveraging a region-level clustering approach to identify precise image features contributing to zero-shot classification errors. Then, RaVL mitigates the identified spurious correlation with a novel region-aware loss function that enables the VLM to focus on relevant regions and ignore spurious relationships during fine-tuning. We evaluate RaVL on 654 VLMs with various model architectures, data domains, and learned spurious correlations. Our results show that RaVL accurately discovers (191% improvement over the closest baseline) and mitigates (8.2% improvement on worst-group image classification accuracy) spurious correlations. Qualitative evaluations on general-domain and medical-domain VLMs confirm our findings.

Jean-Benoit Delbrouck, Justin Xu, Johannes Moll et al. · oxford, stanford

Automated radiology report generation from chest X-ray (CXR) images has the potential to improve clinical efficiency and reduce radiologists' workload. However, most datasets, including the publicly available MIMIC-CXR and CheXpert Plus, consist entirely of free-form reports, which are inherently variable and unstructured. This variability poses challenges for both generation and evaluation: existing models struggle to produce consistent, clinically meaningful reports, and standard evaluation metrics fail to capture the nuances of radiological interpretation. To address this, we introduce Structured Radiology Report Generation (SRRG), a new task that reformulates free-text radiology reports into a standardized format, ensuring clarity, consistency, and structured clinical reporting. We create a novel dataset by restructuring reports using large language models (LLMs) following strict structured reporting desiderata. Additionally, we introduce SRR-BERT, a fine-grained disease classification model trained on 55 labels, enabling more precise and clinically informed evaluation of structured reports. To assess report quality, we propose F1-SRR-BERT, a metric that leverages SRR-BERT's hierarchical disease taxonomy to bridge the gap between free-text variability and structured clinical reporting. We validate our dataset through a reader study conducted by five board-certified radiologists and extensive benchmarking experiments.

Melanie Rieff, Maya Varma, Ossian Rabow et al.

Multimodal in-context learning (ICL) remains underexplored despite significant potential for domains such as medicine. Clinicians routinely encounter diverse, specialized tasks requiring adaptation from limited examples, such as drawing insights from a few relevant prior cases or considering a constrained set of differential diagnoses. While multimodal large language models (MLLMs) have shown advances in medical visual question answering (VQA), their ability to learn multimodal tasks from context is largely unknown. We introduce SMMILE, the first expert-driven multimodal ICL benchmark for medical tasks. Eleven medical experts curated problems, each including a multimodal query and multimodal in-context examples as task demonstrations. SMMILE encompasses 111 problems (517 question-image-answer triplets) covering 6 medical specialties and 13 imaging modalities. We further introduce SMMILE++, an augmented variant with 1038 permuted problems. A comprehensive evaluation of 15 MLLMs demonstrates that most models exhibit moderate to poor multimodal ICL ability in medical tasks. In open-ended evaluations, ICL contributes only an 8% average improvement over zero-shot on SMMILE and 9.4% on SMMILE++. We observe a susceptibility for irrelevant in-context examples: even a single noisy or irrelevant example can degrade performance by up to 9.5%. Moreover, we observe that MLLMs are affected by a recency bias, where placing the most relevant example last can lead to substantial performance improvements of up to 71%. Our findings highlight critical limitations and biases in current MLLMs when learning multimodal medical tasks from context. SMMILE is available at https://smmile-benchmark.github.io.

Alois Thomas, Maya Varma, Jean-Benoit Delbrouck et al.

Automating radiology report generation with Large Vision-Language Models (LVLMs) holds great potential, yet these models often produce clinically critical hallucinations, posing serious risks. Existing hallucination detection methods frequently lack the necessary sentence-level granularity or robust generalization across different LVLM generators. We introduce a novel approach: a sentence-level Process Reward Model (PRM) adapted for this vision-language task. Our PRM predicts the factual correctness of each generated sentence, conditioned on clinical context and preceding text. When fine-tuned on MIMIC-CXR with weakly-supervised labels, a lightweight 0.5B-parameter PRM outperforms existing verification techniques, demonstrating, for instance, relative improvements of 7.5% in Matthews Correlation Coefficient and 1.8% in AUROC over strong white-box baselines on outputs from one LVLM. Unlike methods reliant on internal model states, our PRM demonstrates strong generalization to an unseen LVLM. We further show its practical utility: PRM scores effectively filter low-quality reports, improving F1-CheXbert scores by 4.5% (when discarding the worst 10% of reports). Moreover, when guiding a novel weighted best-of-N selection process on the MIMIC-CXR test set, our PRM show relative improvements in clinical metrics of 7.4% for F1-CheXbert and 0.6% for BERTScore. These results demonstrate that a lightweight, context-aware PRM provides a model-agnostic safety layer for clinical LVLMs without access to internal activations

Dave Van Veen, Cara Van Uden, Maayane Attias et al.

We systematically investigate lightweight strategies to adapt large language models (LLMs) for the task of radiology report summarization (RRS). Specifically, we focus on domain adaptation via pretraining (on natural language, biomedical text, or clinical text) and via discrete prompting or parameter-efficient fine-tuning. Our results consistently achieve best performance by maximally adapting to the task via pretraining on clinical text and fine-tuning on RRS examples. Importantly, this method fine-tunes a mere 0.32% of parameters throughout the model, in contrast to end-to-end fine-tuning (100% of parameters). Additionally, we study the effect of in-context examples and out-of-distribution (OOD) training before concluding with a radiologist reader study and qualitative analysis. Our findings highlight the importance of domain adaptation in RRS and provide valuable insights toward developing effective natural language processing solutions for clinical tasks.

Jean-Benoit Delbrouck, Noé Tits, Stéphane Dupont

This paper aims to bring a new lightweight yet powerful solution for the task of Emotion Recognition and Sentiment Analysis. Our motivation is to propose two architectures based on Transformers and modulation that combine the linguistic and acoustic inputs from a wide range of datasets to challenge, and sometimes surpass, the state-of-the-art in the field. To demonstrate the efficiency of our models, we carefully evaluate their performances on the IEMOCAP, MOSI, MOSEI and MELD dataset. The experiments can be directly replicated and the code is fully open for future researches.

Jean-Benoit Delbrouck, Bastien Vanderplaetse, Stéphane Dupont

Recently, generative adversarial networks (GAN) have gathered a lot of interest. Their efficiency in generating unseen samples of high quality, especially images, has improved over the years. In the field of Natural Language Generation (NLG), the use of the adversarial setting to generate meaningful sentences has shown to be difficult for two reasons: the lack of existing architectures to produce realistic sentences and the lack of evaluation tools. In this paper, we propose an adversarial architecture related to the conditional GAN (cGAN) that generates sentences according to a given image (also called image captioning). This attempt is the first that uses no pre-training or reinforcement methods. We also explain why our experiment settings can be safely evaluated and interpreted for further works.

Jean-Benoit Delbrouck, Antoine Maiorca, Nathan Hubens et al.

As new data-sets for real-world visual reasoning and compositional question answering are emerging, it might be needed to use the visual feature extraction as a end-to-end process during training. This small contribution aims to suggest new ideas to improve the visual processing of traditional convolutional network for visual question answering (VQA). In this paper, we propose to modulate by a linguistic input a CNN augmented with self-attention. We show encouraging relative improvements for future research in this direction.

Jean-Benoit Delbrouck, Stéphane Dupont

Even with the growing interest in problems at the intersection of Computer Vision and Natural Language, grounding (i.e. identifying) the components of a structured description in an image still remains a challenging task. This contribution aims to propose a model which learns grounding by reconstructing the visual features for the Multi-modal translation task. Previous works have partially investigated standard approaches such as regression methods to approximate the reconstruction of a visual input. In this paper, we propose a different and novel approach which learns grounding by adversarial feedback. To do so, we modulate our network following the recent promising adversarial architectures and evaluate how the adversarial response from a visual reconstruction as an auxiliary task helps the model in its learning. We report the highest scores in term of BLEU and METEOR metrics on the different datasets.

Jean-Benoit Delbrouck, Stéphane Dupont

When searching for an object humans navigate through a scene using semantic information and spatial relationships. We look for an object using our knowledge of its attributes and relationships with other objects to infer the probable location. In this paper, we propose to tackle the visual navigation problem using rich semantic representations of the observed scene and object-oriented targets to train an agent. We show that both allows the agent to generalize to new targets and unseen scene in a short amount of training time.

Jean-Benoit Delbrouck, Stéphane Dupont

Neural Image Captioning (NIC) or neural caption generation has attracted a lot of attention over the last few years. Describing an image with a natural language has been an emerging challenge in both fields of computer vision and language processing. Therefore a lot of research has focused on driving this task forward with new creative ideas. So far, the goal has been to maximize scores on automated metric and to do so, one has to come up with a plurality of new modules and techniques. Once these add up, the models become complex and resource-hungry. In this paper, we take a small step backwards in order to study an architecture with interesting trade-off between performance and computational complexity. To do so, we tackle every component of a neural captioning model and propose one or more solution that lightens the model overall. Our ideas are inspired by two related tasks: Multimodal and Monomodal Neural Machine Translation.

Jean-Benoit Delbrouck, Stéphane Dupont

This paper describes the UMONS solution for the Multimodal Machine Translation Task presented at the third conference on machine translation (WMT18). We explore a novel architecture, called deepGRU, based on recent findings in the related task of Neural Image Captioning (NIC). The models presented in the following sections lead to the best METEOR translation score for both constrained (English, image) -> German and (English, image) -> French sub-tasks.

Jean-Benoit Delbrouck

This paper describes the UMONS solution for the OMG-Emotion Challenge. We explore a context-dependent architecture where the arousal and valence of an utterance are predicted according to its surrounding context (i.e. the preceding and following utterances of the video). We report an improvement when taking into account context for both unimodal and multimodal predictions.

Jean-Benoit Delbrouck, Stéphane Dupont

We propose a new and fully end-to-end approach for multimodal translation where the source text encoder modulates the entire visual input processing using conditional batch normalization, in order to compute the most informative image features for our task. Additionally, we propose a new attention mechanism derived from this original idea, where the attention model for the visual input is conditioned on the source text encoder representations. In the paper, we detail our models as well as the image analysis pipeline. Finally, we report experimental results. They are, as far as we know, the new state of the art on three different test sets.

Jean-Benoit Delbrouck, Stéphane Dupont, Omar Seddati

In Multimodal Neural Machine Translation (MNMT), a neural model generates a translated sentence that describes an image, given the image itself and one source descriptions in English. This is considered as the multimodal image caption translation task. The images are processed with Convolutional Neural Network (CNN) to extract visual features exploitable by the translation model. So far, the CNNs used are pre-trained on object detection and localization task. We hypothesize that richer architecture, such as dense captioning models, may be more suitable for MNMT and could lead to improved translations. We extend this intuition to the word-embeddings, where we compute both linguistic and visual representation for our corpus vocabulary. We combine and compare different confi

Jean-Benoit Delbrouck, Stéphane Dupont

In state-of-the-art Neural Machine Translation (NMT), an attention mechanism is used during decoding to enhance the translation. At every step, the decoder uses this mechanism to focus on different parts of the source sentence to gather the most useful information before outputting its target word. Recently, the effectiveness of the attention mechanism has also been explored for multimodal tasks, where it becomes possible to focus both on sentence parts and image regions that they describe. In this paper, we compare several attention mechanism on the multimodal translation task (English, image to German) and evaluate the ability of the model to make use of images to improve translation. We surpass state-of-the-art scores on the Multi30k data set, we nevertheless identify and report different misbehavior of the machine while translating.

Jean-Benoit Delbrouck, Stephane Dupont

In state-of-the-art Neural Machine Translation, an attention mechanism is used during decoding to enhance the translation. At every step, the decoder uses this mechanism to focus on different parts of the source sentence to gather the most useful information before outputting its target word. Recently, the effectiveness of the attention mechanism has also been explored for multimodal tasks, where it becomes possible to focus both on sentence parts and image regions. Approaches to pool two modalities usually include element-wise product, sum or concatenation. In this paper, we evaluate the more advanced Multimodal Compact Bilinear pooling method, which takes the outer product of two vectors to combine the attention features for the two modalities. This has been previously investigated for visual question answering. We try out this approach for multimodal image caption translation and show improvements compared to basic combination methods.