Teaching AI to See Cancer: An Interview with Supriya Vijay

Pathology sits at the center of almost every cancer diagnosis, yet the infrastructure supporting it is under severe strain. The digital pathology market reached $1.31 billion in 2024 and is projected to hit $3.86 billion by 2033, a trajectory driven by a hard clinical reality: demand for pathology interpretation is growing faster than the workforce trained to provide it. For AI to help close any part of that gap, the models doing the work have to be trained on data that reflects what pathologists actually encounter.

Supriya Vijay is a Senior Software Engineer with over a decade of experience in machine learning and artificial intelligence. She is a co-author of the widely cited paper "Domain-Specific Optimization and Diverse Evaluation of Self-Supervised Models for Histopathology" and a named inventor on a pending U.S. patent application for self-supervised training of machine-learned image processing models for histopathology. She was a core technical contributor to research developing foundation models for analyzing microscopic tissue images, which have since been made publicly available as open research tools.

We spoke with Supriya about the challenge of adapting self-supervised learning to one of medicine's most demanding visual domains.

Pathology images are not like the images most AI systems are trained on. What makes them so different from an engineering perspective?

Histopathology images are fundamentally different from the kinds of images most computer vision systems are built to handle. A single whole-slide image can be 100,000 by 100,000 pixels or larger. The structures that pathologists examine - cell shapes, the spatial arrangement of nuclei, the density of the surrounding connective tissue - exist at multiple scales simultaneously. A model trained on images of everyday objects is not learning specifics that transfer meaningfully to this domain. The visual grammar is different.

The harder problem is that tissue appearance varies enormously across tissue types, staining protocols, and preparation methods - something well documented in the computational pathology literature. A slide prepared in one lab can have a completely different color profile than the same tissue prepared in another. Models trained on data from one institution often fail when deployed at a different one. That distribution shift is one of the central technical problems in the field, and it is why building a foundation model for pathology requires training strategies and augmentations engineered specifically for this domain.

Traditional AI development relies on massive amounts of hand-labeled data, which is a major bottleneck in medicine. How do foundation models change that equation?

The global density of pathologists is roughly 12.5 per million people, about 4 times below the North American level. Having those experts manually annotate massive training datasets is expensive, and the supply of people who can do it is limited. Self-supervised learning changes the equation - the model learns useful representations from unlabeled data by solving its own internal tasks, without requiring a pathologist to label every slide. The goal is to build a foundation model that learns rich, general-purpose representations across a wide range of tissue types and magnifications, which can then dramatically reduce the data, compute, and expertise needed to develop downstream tools.

Building a robust foundation model requires pre-training on a massive corpus of pathology images, optimizing the training strategy specifically for this domain, and evaluating the resulting representations across a highly diverse set of downstream tasks. That diversity of evaluation is essential - a model that performs well on one tissue type or one staining modality can still fail on others, so the evaluation suite must be structured to surface those failures. Today, high-quality pre-trained pathology models are increasingly available as open research tools. In the era of large language models and multimodal AI, these representations are especially critical - they serve as the foundational visual building blocks that allow language models to reason over massive gigapixel tissue slides alongside clinical text.

In self-supervised learning, the model has to learn what makes two images similar without human labels. How do you define what the model should pay attention to versus what it should ignore?

I think the important part is deciding what counts as a meaningful training signal when the ground truth is something only a trained pathologist can interpret. Self-supervised methods like SimCLR and Masked Siamese Networks learn by comparing different augmented views of the same image patch. The model is trained to recognize that a cropped, color-shifted, and blurred version of a tissue patch still represents the same underlying structure. But the augmentations that work well for natural images do not necessarily make sense for pathology. What counts as noise and what counts as signal are fundamentally different in this domain.

Getting that right requires careful thinking about how patches are sampled, how augmentations are designed, and which variations the model should learn to look past versus which it should preserve. That is where close collaboration with pathologists becomes essential. Staining variation, for instance, reflects differences in lab preparation, not in the underlying biology. Cellular density variation, on the other hand, carries information the model needs to retain. Those are competing pressures, and the pre-training design has to resolve them explicitly.

When engineers drill down from general computer vision into healthcare AI, what are the non-obvious traps they run into?

The most persistent challenge was getting the data partitioning right. In standard image benchmarks, you can split by image. In pathology, you have to split by patient or case, because there are often multiple slides from the same patient, and slides from the same patient can look nearly identical. If you split by slide instead of by patient, you risk data leakage. It is a well-known pitfall, but it introduces real complexity. You have to ensure that no patient's data appears in both the pre-training set and any downstream evaluation set, even across different tasks. Maintaining that separation across a benchmark spanning many tissue types and multiple data sources is nontrivial.

The other challenge is color. In standard AI training, you might randomly tweak an image's brightness or hue so the model learns to ignore changes in lighting. But in pathology, the color variation that matters does not come from lighting - it comes from differences in the chemical dyes used to stain tissue slides across different labs. Domain-specific stain augmentation methods, like RandStainNA, which simulates realistic variation in staining across different lab protocols, have proven to be among the most consistently beneficial improvements in this space. But the augmentation strategy has to be calibrated carefully: too aggressive and it destroys color information that pathologists say carries real diagnostic value in certain tissue types; too conservative and the model latches onto staining differences that are artifacts of preparation rather than biology. Finding that balance is iterative and requires clinical input throughout. That kind of domain expertise is not something you can substitute with more compute.

You have also served as a reviewer for both specialized health AI workshops and major machine learning conferences like ICML. What has reviewing taught you about the state of evaluation in the field?

Reviewing for the Structured Data for Health workshop is valuable because it enables you to read work from adjacent areas of health AI - clinical time series, EHR modeling, wearable data. The problems around data quality, distribution shift, and evaluation validity look structurally similar to what you would face in medical imaging. The specifics are different, but the failure modes are often the same: models that look good on benchmark splits but break on out-of-distribution cases, evaluations too narrow to surface the failures that would matter in deployment, training sets that over-represent well-resourced institutions.

That exposure reinforces the importance of rigorous evaluation design. It is easy to report a strong number on a single dataset and declare success. The harder and more informative thing is to test whether a learned representation holds up across tissue types, cancer subtypes, and preparation methods. That instinct came directly from seeing how often single-dataset evaluations fail to generalize.

Reviewing for ICML 2026 is a different kind of challenge. At the main conference, submissions span the full breadth of machine learning, including areas well outside my direct research. You have to assess rigor, novelty, and significance across work that might be in optimization theory, generative modeling, or reinforcement learning. That requires being honest about where your expertise ends and being careful not to over-weight what you happen to find interesting.

What reviewing at that scale has taught me is how much the quality of evaluation design varies even among strong submissions. The methods sections are usually tight. The evaluation sections are where the gaps show up.I look for those patterns in my own work now in a way I did not before.

The market for AI in pathology is expanding rapidly. As the field matures over the next few years, what do you see as the most critical unsolved engineering problems?

The hardest open problem is generalization across institutions. This is especially true as the field moves toward multimodal AI, pairing large language models with vision systems. An LLM can only reason effectively about an image if the underlying visual representation is robust. It is relatively tractable to get a model that works well on data from institutions that resemble your training set. It is much harder to build models that hold up when deployed in a setting with different scanners, different staining protocols, and different slide preparation practices. That gap between benchmark performance and real-world performance is where I believe the most important work still needs to happen. 

The second challenge is workflow integration. You can build the most accurate model in the world, but if it requires pathologists to jump through extra software screens or does not integrate cleanly into existing workflows, it will not be adopted. Hospitals are dealing with fragmented IT infrastructure and massive data silos. Moving AI from a standalone research sandbox into an enterprise-grade platform is a massive, underappreciated engineering hurdle.

The final piece is evaluation infrastructure. It has become increasingly clear that we need evaluation datasets reflecting the actual diversity of pathology practice Building those datasets requires collaboration with clinical partners who have access to diverse, real-world tissue samples and the expertise to annotate them correctly. That is slow, careful work, and it does not attract the same attention as building a new model architecture. But it is foundational work. The AI in pathology market is projected to reach $347.4 million by 2030 at a 26.5% annual growth rate. If that investment funds new architectures but not the evaluation infrastructure to test them honestly, the field will keep producing work that looks stronger on benchmarks than it performs in practice. The research investment has to follow the hard problems, not just the publishable ones.