Arjun K. Manrai

Thomas Buckley, James A. Diao, Pranav Rajpurkar et al.

Multimodal foundation models have shown compelling but conflicting performance in medical image interpretation. However, the mechanisms by which these models integrate and prioritize different data modalities, including images and text, remain poorly understood. Here, using a diverse collection of 1014 multimodal medical cases, we evaluate the unimodal and multimodal image interpretation abilities of proprietary (GPT-4, Gemini Pro 1.0) and open-source (Llama-3.2-90B, LLaVA-Med-v1.5) multimodal foundational models with and without the use of text descriptions. Across all models, image predictions were largely driven by exploiting text, with accuracy increasing monotonically with the amount of informative text. By contrast, human performance on medical image interpretation did not improve with informative text. Exploitation of text is a double-edged sword; we show that even mild suggestions of an incorrect diagnosis in text diminishes image-based classification, reducing performance dramatically in cases the model could previously answer with images alone. Finally, we conducted a physician evaluation of model performance on long-form medical cases, finding that the provision of images either reduced or had no effect on model performance when text is already highly informative. Our results suggest that multimodal AI models may be useful in medical diagnostic reasoning but that their accuracy is largely driven, for better and worse, by their exploitation of text.

Luke W. Sagers, James A. Diao, Matthew Groh et al.

Dermatological classification algorithms developed without sufficiently diverse training data may generalize poorly across populations. While intentional data collection and annotation offer the best means for improving representation, new computational approaches for generating training data may also aid in mitigating the effects of sampling bias. In this paper, we show that DALL$\cdot$E 2, a large-scale text-to-image diffusion model, can produce photorealistic images of skin disease across skin types. Using the Fitzpatrick 17k dataset as a benchmark, we demonstrate that augmenting training data with DALL$\cdot$E 2-generated synthetic images improves classification of skin disease overall and especially for underrepresented groups.

Luke W. Sagers, James A. Diao, Luke Melas-Kyriazi et al.

While hundreds of artificial intelligence (AI) algorithms are now approved or cleared by the US Food and Drugs Administration (FDA), many studies have shown inconsistent generalization or latent bias, particularly for underrepresented populations. Some have proposed that generative AI could reduce the need for real data, but its utility in model development remains unclear. Skin disease serves as a useful case study in synthetic image generation due to the diversity of disease appearance, particularly across the protected attribute of skin tone. Here we show that latent diffusion models can scalably generate images of skin disease and that augmenting model training with these data improves performance in data-limited settings. These performance gains saturate at synthetic-to-real image ratios above 10:1 and are substantially smaller than the gains obtained from adding real images. As part of our analysis, we generate and analyze a new dataset of 458,920 synthetic images produced using several generation strategies. Our results suggest that synthetic data could serve as a force-multiplier for model development, but the collection of diverse real-world data remains the most important step to improve medical AI algorithms.

Aashna P. Shah, Michelle M. Li, Yash Lal et al.

Blood-based biomarkers underpin clinical diagnosis and management, yet their interpretation relies largely on fixed population reference intervals that ignore stable, intra-patient variability. As such, population-based interpretation can mask meaningful deviation from an individual's baseline, risking delayed disease detection. To remedy this, there have been increasing efforts to personalize blood biomarker interpretation using individual testing histories. However, these methods may overfit to sparse data, inflating false-positive rates and unnecessary follow-up, and can also unwittingly include unrecognized or subclinical disease. Here, we leverage nearly 2 billion longitudinal laboratory measurements from over 1.6 million individuals across North America, the Middle East, and East Asia, to show that while laboratory values are highly individual, purely personalized intervals routinely overfit, classifying up to 68% of measurements as abnormal, without corresponding associations with adverse clinical outcomes. We then introduce NORMA, a conditional transformer-based framework that generates reference intervals by conditioning on both a patient's history and population-level data about "normal" variation. NORMA-derived intervals achieve higher precision for predicting outcomes, including mortality, acute kidney injury, and chronic disease. These findings caution against over-personalization in laboratory medicine and demonstrate that anchoring individual trajectories to population-level priors outperforms either approach alone. To promote transparency, we publicly release the model, code, and an interactive user interface for accessible, individualized laboratory interpretation.

Luke Melas-Kyriazi, Arjun K. Manrai

Unsupervised domain adaptation is a promising technique for semantic segmentation and other computer vision tasks for which large-scale data annotation is costly and time-consuming. In semantic segmentation, it is attractive to train models on annotated images from a simulated (source) domain and deploy them on real (target) domains. In this work, we present a novel framework for unsupervised domain adaptation based on the notion of target-domain consistency training. Intuitively, our work is based on the idea that in order to perform well on the target domain, a model's output should be consistent with respect to small perturbations of inputs in the target domain. Specifically, we introduce a new loss term to enforce pixelwise consistency between the model's predictions on a target image and a perturbed version of the same image. In comparison to popular adversarial adaptation methods, our approach is simpler, easier to implement, and more memory-efficient during training. Experiments and extensive ablation studies demonstrate that our simple approach achieves remarkably strong results on two challenging synthetic-to-real benchmarks, GTA5-to-Cityscapes and SYNTHIA-to-Cityscapes. Code is available at: https://github.com/lukemelas/pixmatch

Peter G. Brodeur, Thomas A. Buckley, Zahir Kanjee et al.

A seminal paper published by Ledley and Lusted in 1959 introduced complex clinical diagnostic reasoning cases as the gold standard for the evaluation of expert medical computing systems, a standard that has held ever since. Here, we report the results of a physician evaluation of a large language model (LLM) on challenging clinical cases against a baseline of hundreds of physicians. We conduct five experiments to measure clinical reasoning across differential diagnosis generation, display of diagnostic reasoning, triage differential diagnosis, probabilistic reasoning, and management reasoning, all adjudicated by physician experts with validated psychometrics. We then report a real-world study comparing human expert and AI second opinions in randomly-selected patients in the emergency room of a major tertiary academic medical center in Boston, MA. We compared LLMs and board-certified physicians at three predefined diagnostic touchpoints: triage in the emergency room, initial evaluation by a physician, and admission to the hospital or intensive care unit. In all experiments--both vignettes and emergency room second opinions--the LLM displayed superhuman diagnostic and reasoning abilities, as well as continued improvement from prior generations of AI clinical decision support. Our study suggests that LLMs have achieved superhuman performance on general medical diagnostic and management reasoning, fulfilling the vision put forth by Ledley and Lusted, and motivating the urgent need for prospective trials.

Thomas A. Buckley, Kian R. Weihrauch, Katherine Latham et al.

Despite being widely used to support clinical care, general-purpose large multimodal models (LMMs) have generally shown poor or inconclusive performance in medical image interpretation, particularly in pathology, where gigapixel images are used. However, prior studies have used either low-resolution thumbnails or random patches, which likely underestimated model performance. Here, we ask whether LMMs can be adapted to reason coherently and accurately in the evaluation of such images. In this study, we introduce Gigapixel Image Agent for Navigating Tissue (GIANT), the first framework that allows LMMs to iteratively navigate whole-slide images (WSIs) like a pathologist. Accompanying GIANT, we release MultiPathQA, a new benchmark, which comprises 934 WSI-level questions, encompassing five clinically-relevant tasks ranging from cancer diagnosis to open-ended reasoning. MultiPathQA also includes 128 questions, authored by two professional pathologists, requiring direct slide interpretation. Using MultiPathQA, we show that our simple agentic system substantially outperforms conventional patch- and thumbnail-based baselines, approaching or surpassing the performance of specialized models trained on millions of images. For example, on pathologist-authored questions, GPT-5 with GIANT achieves 62.5% accuracy, outperforming specialist pathology models such as TITAN (43.8%) and SlideChat (37.5%). Our findings reveal the strengths and limitations of current foundation models and ground future development of LMMs for expert reasoning in pathology.

Thomas A. Buckley, Riccardo Conci, Peter G. Brodeur et al.

BACKGROUND: For over a century, the New England Journal of Medicine Clinicopathological Conferences (CPCs) have tested the reasoning of expert physicians and, recently, artificial intelligence (AI). However, prior AI evaluations have focused on final diagnoses without addressing the multifaceted reasoning and presentation skills required of expert discussants. METHODS: Using 7102 CPCs (1923-2025) and 1021 Image Challenges (2006-2025), we conducted extensive physician annotation and automated processing to create CPC-Bench, a physician-validated benchmark spanning 10 text-based and multimodal tasks, against which we evaluated leading large language models (LLMs). Then, we developed "Dr. CaBot," an AI discussant designed to produce written and slide-based video presentations using only the case presentation, modeling the role of the human expert in these cases. RESULTS: When challenged with 377 contemporary CPCs, o3 (OpenAI) ranked the final diagnosis first in 60% of cases and within the top ten in 84% of cases, outperforming a 20-physician baseline; next-test selection accuracy reached 98%. Event-level physician annotations quantified AI diagnostic accuracy per unit of information. Performance was lower on literature search and image tasks; o3 and Gemini 2.5 Pro (Google) achieved 67% accuracy on image challenges. In blinded comparisons of CaBot vs. human expert-generated text, physicians misclassified the source of the differential in 46 of 62 (74%) of trials, and scored CaBot more favorably across quality dimensions. To promote research, we are releasing CaBot and CPC-Bench. CONCLUSIONS: LLMs exceed physician performance on complex text-based differential diagnosis and convincingly emulate expert medical presentations, but image interpretation and literature retrieval remain weaker. CPC-Bench and CaBot may enable transparent and continued tracking of progress in medical AI.