Not Medium member,[click here for access]
Google docs version click[here] Artificial intelligence is rapidly transforming digital pathology. In recent years, large vision foundation models have demonstrated impressive performance across a wide range of computational pathology tasks, from tissue classification and tumor detection to survival prediction and biomarker discovery. By learning rich visual representations from millions of image patches, these models promise to reduce the need for extensive manual annotation while providing a common foundation for numerous downstream applications. It is therefore tempting to believe that, as in natural language processing, a single powerful foundation model could eventually serve as a universal backbone for all pathology tasks.
However, pathology is fundamentally different from the natural images on which many of these models were originally developed. Every organ possesses its own unique tissue architecture, cellular organization, and disease progression patterns. The microscopic appearance of a healthy colon bears little resemblance to that of normal breast tissue, and the biological features that distinguish malignant lesions in the skin are entirely different from those used to diagnose gastric or prostate cancer. These differences are not merely cosmetic — they reflect distinct developmental biology, tissue function, and tumor microenvironments that pathologists rely on every day to make clinical decisions.
This raises an important question: Can a single foundation model truly learn the intricate morphological language of every organ equally well? Or does training one model across many organs inevitably require compromises, forcing it to learn broadly useful features at the expense of subtle, organ-specific characteristics that are often the most clinically informative?
I believe this question deserves far more attention than it currently receives. While the field has understandably focused on building increasingly larger and more general foundation models, relatively little effort has been devoted to exploring the opposite direction: developing models that specialize in the unique biological context of individual organs. Just as medicine is organized into specialties because expertise in one organ system does not automatically translate to another, perhaps artificial intelligence in pathology should also embrace specialization.
In this article, I explore the idea of organ-specific foundation models — self-supervised models trained exclusively on histopathology images from a single organ. Rather than attempting to learn a universal representation of all tissues, these models could focus on mastering the morphological and biological patterns that matter most within their domain. If successful, they may provide more meaningful embeddings for tissue classification, prognosis prediction, treatment response modeling, and ultimately, more trustworthy AI systems for clinical pathology.
At first glance, a histopathology image may appear to be little more than a colorful mosaic of cells stained with hematoxylin and eosin. To a trained pathologist, however, every tissue tells a unique biological story. The arrangement of cells, the surrounding stroma, the organization of glands, the distribution of immune cells, and even subtle variations in nuclear morphology all reflect the normal function of an organ and the diseases that affect it. These microscopic patterns are the language of pathology, and like any language, each organ has its own vocabulary and grammar.
Consider the breast and the colon. Breast tissue is composed of branching ducts and lobules embedded within a fibrous and adipose-rich stroma. Diagnosing breast cancer often requires evaluating architectural distortion, stromal reactions, hormone receptor-associated morphology, and the interaction between tumor cells and the surrounding microenvironment. In contrast, the colon is organized into densely packed crypts lined by epithelial cells that continuously renew themselves. Pathologists examining colorectal tissue focus on features such as crypt architecture, gland formation, mucin production, invasion through the bowel wall, and inflammatory changes. Although both tissues may develop adenocarcinomas, the biological processes driving disease — and the visual cues used to recognize them — are fundamentally different.
These differences extend far beyond normal anatomy. Every organ develops cancer through distinct molecular pathways, interacts with a unique immune microenvironment, and exhibits its own spectrum of histological subtypes. Consequently, the features that are most informative for predicting patient prognosis or treatment response in one organ may be irrelevant in another. A model trained to recognize biologically meaningful patterns in breast cancer cannot be expected to automatically understand the glandular complexity of colorectal cancer or the layered architecture of skin without learning an entirely different set of visual representations.
This presents a challenge for today’s foundation models. Most are designed to learn representations that generalize across many tissues and diseases. While this broad training undoubtedly improves versatility, it may also encourage models to prioritize visual features that are common across organs while overlooking the subtle, tissue-specific characteristics that pathologists rely on in daily practice. In other words, generalization may come at the cost of specialization.
Human medicine offers an interesting analogy. We do not expect a cardiologist to possess the same level of expertise in dermatopathology as a specialist who has spent years studying skin diseases. Both are highly trained physicians, yet each develops deep knowledge within a specific biological domain. Perhaps artificial intelligence should follow a similar trajectory. Instead of relying on a single universal model to understand every tissue in the human body, we may achieve better clinical performance by developing models that become true experts in the morphology of individual organs.
If organ-specific foundation models have the potential to learn richer and more biologically meaningful representations, an obvious question follows: why haven’t we built them already? The answer lies in one of the greatest challenges facing computational pathology today — the scarcity of expertly annotated data. Modern pathology departments generate millions of whole-slide images every year, creating an unprecedented repository of digital tissue specimens. At first glance, this abundance of data appears to be an ideal resource for training artificial intelligence models. However, the vast majority of these images are unlabeled. Unlike natural image datasets, where an object can often be annotated by non-experts, histopathology requires years of specialized medical training. Identifying subtle patterns such as tumor grade, lymphovascular invasion, perineural invasion, or the boundaries between benign and malignant tissue often demands careful review by experienced pathologists. Producing high-quality annotations at this level is not only time-consuming but also expensive, making large, expertly labeled datasets exceptionally difficult to assemble.
The problem becomes even more pronounced when developing organ-specific models. A single institution may possess hundreds of thousands of breast pathology images but only a small fraction with comprehensive diagnostic labels or long-term clinical outcomes. The same limitation exists for colorectal, gastric, prostate, and skin pathology datasets. As a result, researchers often face a paradox: an abundance of digital slides but a scarcity of usable labels. This imbalance has become one of the primary obstacles preventing AI systems from reaching their full potential in clinical pathology.
Traditional supervised learning depends heavily on these annotations. The performance of a model is fundamentally constrained by the quantity and quality of the labeled examples it receives. Collecting enough expert annotations to train robust deep learning models across multiple organs, disease subtypes, and clinical endpoints is rarely feasible, particularly for rare cancers or uncommon pathological features. Simply scaling supervised learning is therefore unlikely to solve this problem.
Instead of asking pathologists to label millions of images, perhaps we should ask a different question: Can AI learn the visual language of tissue before anyone tells it what the tissue represents? This simple shift in perspective has led to one of the most exciting developments in modern computer vision — self-supervised learning. Rather than relying on manually assigned labels, self-supervised models learn directly from the intrinsic structure of the data itself, opening the possibility of leveraging the enormous collections of unlabeled whole-slide images that already exist in pathology archives around the world.
For decades, the success of deep learning has depended on one critical ingredient: labeled data. Whether the goal is to classify tissue types, detect tumors, or predict patient outcomes, conventional supervised learning assumes that every training image comes with a corresponding ground-truth label. In computational pathology, this assumption quickly becomes impractical. While hospitals routinely generate vast numbers of whole-slide images, only a small fraction receive the detailed annotations required to train modern deep learning models. As a result, an enormous amount of potentially valuable biological information remains untapped. Self-supervised learning (SSL) offers a fundamentally different approach. Instead of learning from human-provided labels, SSL allows a model to generate its own learning signal directly from the data. During training, the model is presented with multiple augmented views of the same tissue image — for example, different crops, rotations, or color variations — and learns to recognize that these images represent the same underlying biological structure. By solving carefully designed pretext tasks, the model gradually learns to distinguish meaningful morphological patterns from superficial variations introduced during image acquisition. Remarkably, it accomplishes this without ever being told whether the tissue is benign or malignant, or even which organ it originated from.
A useful analogy is how pathology residents are trained. Before they can confidently diagnose disease, they spend months or even years examining thousands of slides, becoming familiar with the normal architecture of different tissues and the subtle ways those tissues change in disease. They first learn to see before they learn to diagnose. Self-supervised learning follows a remarkably similar philosophy. Rather than memorizing labels from the outset, the model first develops a visual understanding of tissue morphology, cellular organization, and spatial relationships. Once these representations have been learned, only a relatively small number of labeled examples are needed to adapt the model to downstream tasks such as tissue classification, prognosis prediction, or treatment response prediction.
This shift in learning strategy is particularly well suited to digital pathology because it transforms one of the field’s greatest limitations into one of its greatest strengths. Instead of viewing the millions of unlabeled whole-slide images stored in pathology archives as unusable data, self-supervised learning treats them as an invaluable source of biological knowledge. Every archived slide becomes an opportunity for the model to learn how tissues are organized, how cells interact within their microenvironment, and how morphology varies across patients and disease states.
For organ-specific foundation models, this capability is especially compelling. Rather than requiring exhaustive annotations for every organ, researchers can leverage the abundance of unlabeled slides available for breast, colorectal, gastric, skin, or prostate pathology to learn rich, organ-specific representations. The labeled datasets that pathologists painstakingly create can then be reserved for fine-tuning models on clinically relevant tasks, dramatically reducing the annotation burden while allowing the model to develop a much deeper understanding of tissue biology. In essence, self-supervised learning changes the question from “How can we label enough data to train a good model?” to “How can we enable the model to learn as much as possible before a single label is provided?” For a field where unlabeled data is abundant but expert annotations are scarce, this shift represents not merely an incremental improvement, but a fundamentally new way of building artificial intelligence for computational pathology.
If self-supervised learning is the idea that a model can learn from unlabeled data, the next question is how it actually accomplishes this. Over the past few years, researchers have developed several self-supervised learning algorithms, but two approaches have emerged as particularly influential for learning visual representations: DINOv2 and Barlow Twins. Although they differ in their training strategies, both share the same objective — to help a neural network understand the underlying structure of images without relying on human-provided labels. Imagine showing a pathologist two slightly different photographs of the same tissue section. One image might be zoomed in, another rotated, or perhaps the staining appears slightly darker. Despite these differences, the pathologist immediately recognizes that both images represent the same biological specimen. Self-supervised learning aims to teach neural networks this same ability: to focus on meaningful biological structures while ignoring superficial variations introduced during image acquisition or data preprocessing.
DINOv2 achieves this through a teacher-student learning framework. During training, different augmented views of the same tissue image are presented to two versions of the model. The teacher network produces a high-level representation of the tissue, while the student network learns to generate a similar representation from another augmented view. Over thousands of training iterations, the student gradually learns to recognize the essential morphological features that remain consistent regardless of image orientation, color variation, or cropping. Rather than memorizing pixels, it begins to encode biologically meaningful concepts such as glandular architecture, nuclear organization, stromal composition, and tissue structure.
Barlow Twins approaches the same challenge from a different perspective. Instead of relying on a teacher-student framework, it learns by comparing the representations generated from two augmented versions of the same image. The model is encouraged to make these representations highly similar while ensuring that each feature captures distinct information rather than repeating the same patterns. In other words, it strives to produce embeddings that are both consistent and informative. This encourages the network to learn diverse morphological characteristics without becoming overly dependent on any single visual feature.
Although their learning mechanisms differ, the outcome is remarkably similar. Both methods produce rich feature representations — often referred to as embeddings — that summarize the complex morphology of tissue images in a way that downstream models can readily use. These embeddings become the foundation for tasks such as tissue classification, prognosis prediction, biomarker discovery, and treatment response prediction.
For computational pathology, the choice between DINOv2 and Barlow Twins is not simply a question of which algorithm achieves the highest benchmark score. It is a question of **which learning strategy captures the biological language of tissue most effectively**. If organ-specific foundation models are to become a reality, understanding how different self-supervised objectives influence the quality, robustness, and interpretability of these learned representations will be just as important as improving predictive accuracy itself
If self-supervised learning enables AI to learn from millions of unlabeled pathology images, an intriguing question naturally emerges: **what if we stopped trying to build one foundation model for every organ and instead built a dedicated foundation model for each one?**
The current generation of pathology foundation models is largely trained on diverse collections of whole-slide images spanning multiple organs and disease types. This broad training strategy aims to produce versatile models that perform reasonably well across a wide range of downstream tasks. While this generalization is undoubtedly valuable, it also requires the model to learn representations that are applicable across highly heterogeneous tissues. In doing so, the model may prioritize visual patterns that are common across organs while underrepresenting the subtle morphological features that define individual tissue types.
An alternative approach is to embrace specialization. Imagine training a self-supervised model exclusively on breast pathology slides, allowing it to spend millions of training iterations learning the intricate architecture of mammary ducts, stromal remodeling, immune infiltration, and the diverse morphological patterns associated with benign and malignant breast disease. A separate model could then be trained exclusively on colorectal tissue, developing a deep understanding of crypt organization, glandular differentiation, mucin production, and the progression from adenoma to carcinoma. Rather than becoming a generalist, each model would become an expert in the biological language of a single organ.
This idea mirrors the evolution of modern medicine itself. Physicians begin with a broad medical education, but meaningful expertise is achieved through specialization. A breast pathologist develops a level of morphological intuition that differs fundamentally from that of a gastrointestinal or dermatopathologist because each field presents unique diagnostic challenges and biological complexities. Artificial intelligence may benefit from following the same principle. Instead of expecting one model to master every tissue in the human body, we could develop specialized foundation models that acquire a much deeper understanding of individual organ systems.
Such models could provide richer embeddings for a wide range of downstream applications, including tissue classification, biomarker discovery, prognosis prediction, and treatment response prediction. More importantly, because their representations are learned from a single biological domain, they may better capture the subtle structural and cellular relationships that are most relevant for clinical decision-making. This could improve not only predictive performance but also the interpretability and trustworthiness of AI systems used in routine pathology practice.
Of course, this idea raises important scientific questions. Do organ-specific models consistently outperform general-purpose foundation models? How much training data is required before specialization becomes advantageous? Which self-supervised learning objectives produce the most biologically meaningful embeddings? And can specialized models transfer knowledge to related tasks within the same organ? These questions remain largely unexplored, but answering them could reshape how the next generation of AI systems for computational pathology is designed.
Ultimately, I do not see organ-specific foundation models as competitors to general foundation models. Rather, I see them as complementary tools within a broader AI ecosystem. General-purpose models provide versatility and broad applicability, while organ-specific models offer the depth of understanding needed for complex, clinically meaningful applications. Finding the right balance between these two approaches may be one of the most important challenges — and opportunities — in the future of computational pathology.
Building an organ-specific foundation model is conceptually straightforward, although implementing one at scale requires careful engineering and close collaboration between computational scientists and pathologists. The process begins with assembling a large collection of whole-slide images from a single organ — for example, breast, colorectal, gastric, or skin pathology. Importantly, these images do not need detailed annotations. Instead, they only need to represent the diversity of tissue morphology encountered in routine clinical practice, including normal tissue, precancerous lesions, malignant tumors, and the wide range of biological variation observed across patients.
Once the dataset has been curated, each whole-slide image is divided into thousands of smaller image patches, or tiles, which serve as the basic training units for the model. These tiles are then used to train a self-supervised learning algorithm such as DINOv2 or Barlow Twins. During this stage, the model repeatedly observes different augmented views of the same tissue tile, gradually learning to recognize invariant biological structures while ignoring superficial variations caused by staining, image acquisition, or preprocessing. Over millions of training iterations, the model begins to develop a rich internal representation of the organ’s unique cellular architecture, tissue organization, and morphological patterns — all without requiring a single manual annotation.
After this self-supervised pretraining phase, the learned representations can be adapted to a variety of downstream clinical tasks using relatively small labeled datasets. Rather than training a new model from scratch for every application, researchers can simply fine-tune the pretrained foundation model to classify tissue subtypes, predict patient prognosis, estimate treatment response, or identify clinically relevant biomarkers. Because the model has already learned the visual language of the organ, far fewer annotated examples are needed to achieve strong performance, significantly reducing the burden on pathologists.
Evaluating these models extends beyond measuring predictive accuracy. An important question is whether the learned embeddings genuinely capture biologically meaningful information. Visualization techniques such as UMAP or t-SNE can reveal whether similar tissue types naturally cluster together in the learned feature space, while explainability methods such as attention maps and class activation mapping (CAM) can help identify the tissue regions driving the model’s predictions. Ultimately, these computational findings must be interpreted alongside expert pathologists to ensure that the model is learning clinically plausible features rather than relying on spurious correlations.
Although this workflow focuses on a single organ, the framework is inherently scalable. The same strategy could be applied independently to breast, colorectal, gastric, prostate, lung, skin, or other tissue types, creating a family of specialized foundation models rather than relying on a single universal model. Over time, these specialized models could form a comprehensive ecosystem of AI tools, each designed to understand the unique biology of its respective organ while sharing a common self-supervised learning framework.
The true value of organ-specific foundation models will not be measured by benchmark scores alone, but by their ability to address real challenges faced by pathologists and clinicians. Every day, pathology laboratories process an enormous number of tissue specimens, while the demand for expert interpretation continues to grow. At the same time, diseases such as cancer are becoming increasingly complex, requiring the integration of histopathology, molecular biomarkers, and clinical information to guide treatment decisions. Artificial intelligence has the potential to support this process — not by replacing pathologists, but by providing consistent, quantitative analyses that augment human expertise.
One of the most immediate applications of organ-specific foundation models is improving routine diagnostic workflows. A model trained specifically on breast pathology, for example, could rapidly distinguish benign from malignant lesions, identify regions suspicious for invasive disease, or prioritize slides that require urgent review. Similarly, a colorectal-specific model could assist in detecting subtle glandular abnormalities or quantifying histological features associated with disease progression. By automating repetitive and time-consuming tasks, these systems could allow pathologists to focus their attention on diagnostically challenging cases where clinical judgment is most valuable.
Beyond diagnosis, organ-specific embeddings may unlock entirely new opportunities for precision medicine. Because these models learn rich representations of tissue morphology, they could serve as a foundation for predicting patient prognosis, identifying biomarkers associated with therapeutic response, or discovering previously unrecognized relationships between tissue architecture and disease biology. Rather than functioning solely as image classifiers, foundation models could become computational tools for understanding how microscopic tissue organization reflects the underlying molecular and cellular mechanisms of disease.
Another important advantage is interpretability. Trust is essential for the adoption of AI in healthcare, and clinicians need to understand why a model reaches a particular conclusion. Organ-specific models, trained within a narrower biological domain, may produce representations that are easier to relate to established histopathological concepts. Combined with explainability techniques such as attention maps and feature attribution methods, these models could highlight the tissue regions and morphological patterns that influenced a prediction, allowing pathologists to evaluate whether the model’s reasoning is biologically plausible rather than treating its output as a black box.
Ultimately, the goal is not to create artificial intelligence that replaces pathologists, but to develop systems that amplify their expertise. Just as radiologists now routinely work alongside computer-aided detection tools, future pathologists may work with organ-specific foundation models that provide quantitative insights, prioritize diagnostically relevant regions, and integrate complex imaging data into clinically meaningful predictions. If successful, these models could help transform digital pathology from a discipline focused primarily on image interpretation into one that also enables biological discovery, risk prediction, and truly personalized patient care.
Over the past decade, computational pathology has evolved from developing algorithms for narrow image analysis tasks to building foundation models capable of supporting a wide range of clinical applications. This transition has been driven by remarkable advances in deep learning, the increasing availability of digital pathology data, and the emergence of self-supervised learning. Yet, as these models become larger and more general, it is worth asking whether scaling alone is the optimal path forward. In a field where every organ possesses distinct biological architecture, disease mechanisms, and diagnostic criteria, specialization may prove to be just as important as generalization.
Organ-specific foundation models represent one possible direction for the next generation of computational pathology. Rather than attempting to learn a universal representation of every tissue in the human body, these models could develop a deeper understanding of the unique morphology and biology of individual organs. Whether this approach ultimately outperforms general-purpose foundation models remains an open scientific question, but it is one that deserves careful investigation. Understanding the trade-offs between broad generalization and biological specialization may fundamentally shape how future AI systems are designed for medicine.
Perhaps the most exciting aspect of this idea is that it shifts the conversation beyond simply improving predictive accuracy. The ultimate goal is not to build models that classify images more efficiently, but to develop computational systems that help us better understand disease. If self-supervised learning can uncover representations that reflect meaningful tissue organization, cellular interactions, and disease progression, these models may become valuable tools not only for diagnosis but also for biomarker discovery, prognosis prediction, and hypothesis generation in biomedical research.
As digital pathology continues to expand, the availability of large-scale whole-slide image repositories will only increase, creating unprecedented opportunities to train increasingly sophisticated models. At the same time, close collaboration between computer scientists, pathologists, molecular biologists, and clinicians will remain essential to ensure that these models learn biologically meaningful features and address clinically relevant problems. Advances in artificial intelligence alone will not transform pathology; meaningful progress will come from combining computational innovation with deep biological and clinical expertise.
Whether the future belongs to universal foundation models, organ-specific models, or a hybrid ecosystem of both remains to be seen. However, one thing is becoming increasingly clear: the next generation of AI for pathology will not be defined solely by larger datasets or more powerful algorithms, but by how well these technologies capture the complexity of human biology and translate that understanding into better patient care. Exploring organ-specific foundation models is one step toward that vision, and it may help redefine how we think about artificial intelligence in computational pathology
Organ-Specific Embedding Models for Computational Pathology was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.