Robbie Holland

Ashwin Kumar, Robbie Holland, Corey Barrett et al.

Recent medical multimodal foundation models are built as multimodal LLMs (MLLMs) by connecting a CLIP-pretrained vision encoder to an LLM using LLaVA-style finetuning. This two-stage, decoupled approach introduces a projection layer that can distort visual features. This is especially concerning in medical imaging where subtle cues are essential for accurate diagnoses. In contrast, early-fusion generative approaches such as Chameleon eliminate the projection bottleneck by processing image and text tokens within a single unified sequence, enabling joint representation learning that leverages the inductive priors of language models. We present CheXmix, a unified early-fusion generative model trained on a large corpus of chest X-rays paired with radiology reports. We expand on Chameleon's autoregressive framework by introducing a two-stage multimodal generative pretraining strategy that combines the representational strengths of masked autoencoders with MLLMs. The resulting models are highly flexible, supporting both discriminative and generative tasks at both coarse and fine-grained scales. Our approach outperforms well-established generative models across all masking ratios by 6.0% and surpasses CheXagent by 8.6% on AUROC at high image masking ratios on the CheXpert classification task. We further inpaint images over 51.0% better than text-only generative models and outperform CheXagent by 45% on the GREEN metric for radiology report generation. These results demonstrate that CheXmix captures fine-grained information across a broad spectrum of chest X-ray tasks. Our code is at: https://github.com/StanfordMIMI/CheXmix.

Robbie Holland, Oliver Leingang, Hrvoje Bogunović et al.

Deep learning has potential to automate screening, monitoring and grading of disease in medical images. Pretraining with contrastive learning enables models to extract robust and generalisable features from natural image datasets, facilitating label-efficient downstream image analysis. However, the direct application of conventional contrastive methods to medical datasets introduces two domain-specific issues. Firstly, several image transformations which have been shown to be crucial for effective contrastive learning do not translate from the natural image to the medical image domain. Secondly, the assumption made by conventional methods, that any two images are dissimilar, is systematically misleading in medical datasets depicting the same anatomy and disease. This is exacerbated in longitudinal image datasets that repeatedly image the same patient cohort to monitor their disease progression over time. In this paper we tackle these issues by extending conventional contrastive frameworks with a novel metadata-enhanced strategy. Our approach employs widely available patient metadata to approximate the true set of inter-image contrastive relationships. To this end we employ records for patient identity, eye position (i.e. left or right) and time series information. In experiments using two large longitudinal datasets containing 170,427 retinal OCT images of 7,912 patients with age-related macular degeneration (AMD), we evaluate the utility of using metadata to incorporate the temporal dynamics of disease progression into pretraining. Our metadata-enhanced approach outperforms both standard contrastive methods and a retinal image foundation model in five out of six image-level downstream tasks related to AMD. Due to its modularity, our method can be quickly and cost-effectively tested to establish the potential benefits of including available metadata in contrastive pretraining.

Martin J. Menten, Johannes C. Paetzold, Veronika A. Zimmer et al.

The skeleton of a digital image is a compact representation of its topology, geometry, and scale. It has utility in many computer vision applications, such as image description, segmentation, and registration. However, skeletonization has only seen limited use in contemporary deep learning solutions. Most existing skeletonization algorithms are not differentiable, making it impossible to integrate them with gradient-based optimization. Compatible algorithms based on morphological operations and neural networks have been proposed, but their results often deviate from the geometry and topology of the true medial axis. This work introduces the first three-dimensional skeletonization algorithm that is both compatible with gradient-based optimization and preserves an object's topology. Our method is exclusively based on matrix additions and multiplications, convolutional operations, basic non-linear functions, and sampling from a uniform probability distribution, allowing it to be easily implemented in any major deep learning library. In benchmarking experiments, we prove the advantages of our skeletonization algorithm compared to non-differentiable, morphological, and neural-network-based baselines. Finally, we demonstrate the utility of our algorithm by integrating it with two medical image processing applications that use gradient-based optimization: deep-learning-based blood vessel segmentation, and multimodal registration of the mandible in computed tomography and magnetic resonance images.

Robbie Holland, Oliver Leingang, Christopher Holmes et al.

Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Current grading systems based on imaging biomarkers only coarsely group disease stages into broad categories and are unable to predict future disease progression. It is widely believed that this is due to their focus on a single point in time, disregarding the dynamic nature of the disease. In this work, we present the first method to automatically discover biomarkers that capture temporal dynamics of disease progression. Our method represents patient time series as trajectories in a latent feature space built with contrastive learning. Then, individual trajectories are partitioned into atomic sub-sequences that encode transitions between disease states. These are clustered using a newly introduced distance metric. In quantitative experiments we found our method yields temporal biomarkers that are predictive of conversion to late AMD. Furthermore, these clusters were highly interpretable to ophthalmologists who confirmed that many of the clusters represent dynamics that have previously been linked to the progression of AMD, even though they are currently not included in any clinical grading system.

Robbie Holland, Thomas R. P. Taylor, Christopher Holmes et al.

Clinicians spend a significant amount of time reviewing medical images and transcribing their findings regarding patient diagnosis, referral and treatment in text form. Vision-language models (VLMs), which automatically interpret images and summarize their findings as text, have enormous potential to alleviate clinical workloads and increase patient access to high-quality medical care. While foundational models have stirred considerable interest in the medical community, it is unclear whether their general capabilities translate to real-world clinical utility. In this work, we demonstrate that OpenAI's ChatGPT-4o model, in addition to two foundation VLMs designed for medical use, markedly underperform compared to practicing ophthalmologists on specialist tasks crucial to the care of patients with age-related macular degeneration (AMD). To address this, we initially identified the essential capabilities required for image-based clinical decision-making, and then developed a curriculum to selectively train VLMs in these skills. The resulting model, RetinaVLM, can be instructed to write reports that significantly outperform those written by leading foundation medical VLMs and ChatGPT-4o in disease staging (F1 score of 0.63 vs. 0.33) and patient referral (0.67 vs. 0.50), and approaches the diagnostic performance of junior ophthalmologists (who achieve 0.77 and 0.78 on the respective tasks). Furthermore, in a single-blind reader study two senior ophthalmologists with up to 32 years of experience found RetinaVLM's reports were found to be substantially more accurate than those by ChatGPT-4o (64.3% vs. 14.3%). These results reinforce that our curriculum-based approach provides a blueprint towards specializing foundation medical VLMs for real-world clinical tasks.

Philipp Wesp, Robbie Holland, Vasiliki Sideri-Lampretsa et al.

Vision foundation models (FMs) achieve state-of-the-art performance in medical imaging. However, they encode information in abstract latent representations that clinicians cannot interrogate or verify. The goal of this study is to investigate Sparse Autoencoders (SAEs) for replacing opaque FM image representations with human-interpretable, sparse features. We train SAEs on embeddings from BiomedParse (biomedical) and DINOv3 (general-purpose) using 909,873 CT and MRI 2D image slices from the TotalSegmentator dataset. We find that learned sparse features: (a) reconstruct original embeddings with high fidelity (R2 up to 0.941) and recover up to 87.8% of downstream performance using only 10 features (99.4% dimensionality reduction), (b) preserve semantic fidelity in image retrieval tasks, (c) correspond to specific concepts that can be expressed in language using large language model (LLM)-based auto-interpretation. (d) bridge clinical language and abstract latent representations in zero-shot language-driven image retrieval. Our work indicates SAEs are a promising pathway towards interpretable, concept-driven medical vision systems. Code repository: https://github.com/pwesp/sail.

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.

Vasiliki Sideri-Lampretsa, Sophie Starck, Robbie Holland et al.

Reusing the early layers of cohort-trained INRs as initialization for new signals has been shown to accelerate and improve signal fitting, yet it remains unclear which layers of the shared encoder learn transferable representations and what those representations encode. We address both questions for two standard backbones, SIREN and Fourier-feature MLPs (FFMLP). First, sweeping the freeze depth across the shared encoder at test time, we find that the optimum coincides with the layer of highest weight stable rank. Moreover, freezing at this depth matches or improves on the standard fine-tuning recipe across all our experiments. Second, identifying which layer transfers does not characterize what that layer encodes. To address this we adopt sparse autoencoders (SAEs), the dominant tool in mechanistic interpretability, and present the first SAE decomposition of INR activations into sparse dictionary atoms. Interestingly, SIREN and FFMLP achieve comparable cohort-fitting quality, but learn qualitatively different dictionaries. Cohort SIREN's atoms are localized, tiling the coordinate plane such that each atom fires in a confined region independent of cohort content. Cohort FFMLP's atoms are image-spanning, tracing the contours of memorized cohort signals. Single-atom ablations confirm causal use of these dictionaries: a single FFMLP atom out of 4096 can drop PSNR by up to 10.6 dB across the image, while SIREN ablations remain confined to where the atom fires. Together, these results give the first mechanistic account of what transfers in cohort-trained INRs and turn their activations into inspectable dictionary atoms. These tools open a path towards characterizing what INRs encode and towards architectures designed for generalization rather than memorization.

Robbie Holland, Rebecca Kaye, Ahmed M. Hagag et al.

Diseases are currently managed by grading systems, where patients are stratified by grading systems into stages that indicate patient risk and guide clinical management. However, these broad categories typically lack prognostic value, and proposals for new biomarkers are currently limited to anecdotal observations. In this work, we introduce a deep-learning-based biomarker proposal system for the purpose of accelerating biomarker discovery in age-related macular degeneration (AMD). It works by first training a neural network using self-supervised contrastive learning to discover, without any clinical annotations, features relating to both known and unknown AMD biomarkers present in 46,496 retinal optical coherence tomography (OCT) images. To interpret the discovered biomarkers, we partition the images into 30 subsets, termed clusters, that contain similar features. We then conduct two parallel 1.5-hour semi-structured interviews with two independent teams of retinal specialists that describe each cluster in clinical language. Overall, both teams independently identified clearly distinct characteristics in 27 of 30 clusters, of which 23 were related to AMD. Seven were recognised as known biomarkers already used in established grading systems and 16 depicted biomarker combinations or subtypes that are either not yet used in grading systems, were only recently proposed, or were unknown. Clusters separated incomplete from complete retinal atrophy, intraretinal from subretinal fluid and thick from thin choroids, and in simulation outperformed clinically-used grading systems in prognostic value. Overall, contrastive learning enabled the automatic proposal of AMD biomarkers that go beyond the set used by clinically established grading systems. Ultimately, we envision that equipping clinicians with discovery-oriented deep-learning tools can accelerate discovery of novel prognostic biomarkers.

Chengzhi Shen, Martin J. Menten, Hrvoje Bogunović et al.

Analyzing temporal developments is crucial for the accurate prognosis of many medical conditions. Temporal changes that occur over short time scales are key to assessing the health of physiological functions, such as the cardiac cycle. Moreover, tracking longer term developments that occur over months or years in evolving processes, such as age-related macular degeneration (AMD), is essential for accurate prognosis. Despite the importance of both short and long term analysis to clinical decision making, they remain understudied in medical deep learning. State of the art methods for spatiotemporal representation learning, developed for short natural videos, prioritize the detection of temporal constants rather than temporal developments. Moreover, they do not account for varying time intervals between acquisitions, which are essential for contextualizing observed changes. To address these issues, we propose two approaches. First, we combine clip-level contrastive learning with a novel temporal embedding to adapt to irregular time series. Second, we propose masking and predicting latent frame representations of the temporal sequence. Our two approaches outperform all prior methods on temporally-dependent tasks including cardiac output estimation and three prognostic AMD tasks. Overall, this enables the automated analysis of temporal patterns which are typically overlooked in applications of deep learning to medicine.

Daniel Scholz, Ayhan Can Erdur, Robbie Holland et al.

Magnetic resonance imaging (MRI) is an invaluable tool for clinical and research applications. Yet, variations in scanners and acquisition parameters cause inconsistencies in image contrast, hindering data comparability and reproducibility across datasets and clinical studies. Existing scanner harmonization methods, designed to address this challenge, face limitations, such as requiring traveling subjects or struggling to generalize to unseen domains. We propose a novel approach using a conditioned diffusion autoencoder with a contrastive loss and domain-agnostic contrast augmentation to harmonize MR images across scanners while preserving subject-specific anatomy. Our method enables brain MRI synthesis from a single reference image. It outperforms baseline techniques, achieving a +7% PSNR improvement on a traveling subjects dataset and +18% improvement on age regression in unseen. Our model provides robust, effective harmonization of brain MRIs to target scanners without requiring fine-tuning. This advancement promises to enhance comparability, reproducibility, and generalizability in multi-site and longitudinal clinical studies, ultimately contributing to improved healthcare outcomes.

Alina F. Dima, Suprosanna Shit, Huaqi Qiu et al.

Vessels are complex structures in the body that have been studied extensively in multiple representations. While voxelization is the most common of them, meshes and parametric models are critical in various applications due to their desirable properties. However, these representations are typically extracted through segmentations and used disjointly from each other. We propose a framework that joins the three representations under differentiable transformations. By leveraging differentiable voxelization, we automatically extract a parametric shape model of the vessels through shape-to-segmentation fitting, where we learn shape parameters from segmentations without the explicit need for ground-truth shape parameters. The vessel is parametrized as centerlines and radii using cubic B-splines, ensuring smoothness and continuity by construction. Meshes are differentiably extracted from the learned shape parameters, resulting in high-fidelity meshes that can be manipulated post-fit. Our method can accurately capture the geometry of complex vessels, as demonstrated by the volumetric fits in experiments on aortas, aneurysms, and brain vessels.

Hongwei Bran Li, Fernando Navarro, Ivan Ezhov et al.

Uncertainty in medical image segmentation tasks, especially inter-rater variability, arising from differences in interpretations and annotations by various experts, presents a significant challenge in achieving consistent and reliable image segmentation. This variability not only reflects the inherent complexity and subjective nature of medical image interpretation but also directly impacts the development and evaluation of automated segmentation algorithms. Accurately modeling and quantifying this variability is essential for enhancing the robustness and clinical applicability of these algorithms. We report the set-up and summarize the benchmark results of the Quantification of Uncertainties in Biomedical Image Quantification Challenge (QUBIQ), which was organized in conjunction with International Conferences on Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2020 and 2021. The challenge focuses on the uncertainty quantification of medical image segmentation which considers the omnipresence of inter-rater variability in imaging datasets. The large collection of images with multi-rater annotations features various modalities such as MRI and CT; various organs such as the brain, prostate, kidney, and pancreas; and different image dimensions 2D-vs-3D. A total of 24 teams submitted different solutions to the problem, combining various baseline models, Bayesian neural networks, and ensemble model techniques. The obtained results indicate the importance of the ensemble models, as well as the need for further research to develop efficient 3D methods for uncertainty quantification methods in 3D segmentation tasks.