Pathology -- the microscopic examination of tissue to diagnose cancer -- remains the gold standard for determining tumor type, grade, and molecular features that guide treatment. Yet the field faces a perfect storm of pressures: a global shortage of pathologists, rapidly expanding diagnostic complexity, and an ever-growing workload of cancer cases.
Modern pathology demands far more than traditional morphological assessment. Pathologists must now interpret complex immunohistochemistry staining patterns, molecular mutation assays, genomic profiling results, and their integration with imaging data -- all while managing increasing caseloads. These converging pressures create an urgent need for intelligent tools to augment pathological practice.
The solution has been enabled by the widespread adoption of whole slide imaging (WSI) -- technology that converts glass microscope slides into high-resolution digital files. WSI digitization creates the essential data substrate for AI algorithms, unlocking the potential for large-scale computational analysis of previously inaccessible pathological archives.
This review traces AI's evolution in cancer pathology from its earliest task-specific algorithms through to today's powerful foundation models and conversational AI copilots, examining both the remarkable capabilities these systems offer and the significant challenges that must be overcome before they can be safely deployed in clinical practice.
The evolution of AI in pathology has proceeded through three phases. The first era (pre-2020s) used task-specific convolutional neural networks trained for one narrow job -- detecting mitoses, grading Gleason patterns in prostate cancer (achieving accuracy ~0.70 versus ~0.61 for average pathologists), or classifying lung cancer subtypes. These systems excelled in their specific task but required separate models for different problems, extensive pixel-level annotations, and often failed to generalize beyond their training institution.
The critical limitations were scalability barriers (one prostate cancer model required annotating over 112 million image patches), data silos (models trained on TCGA data often failed on community hospital slides), and architectural constraints (systems built for biopsies could not analyze resection specimens without retraining). These challenges motivated the shift to a fundamentally different AI philosophy.
Foundation models represent this paradigm shift: large neural networks pre-trained on massive, diverse datasets using self-supervised learning -- learning from the data itself without requiring expert-labeled annotations. Inspired by the success of ChatGPT in natural language, pathology foundation models learn general representations of tissue appearance that can be rapidly adapted (fine-tuned) to any downstream diagnostic task with minimal additional labeled data.
Key enabling advances included Meta's DINOv2 framework (demonstrating that training data diversity matters more than volume), Vision Transformer architectures (which replaced CNNs for capturing long-range dependencies across gigapixel slides), and GPU clusters enabling training on over 100 million image tiles. Together these advances precipitated an explosion of powerful pathology-specific foundation models.
UNI (from Mahmood Lab, 2024) was trained on over 100 million image tiles from more than 100,000 diagnostic whole slides spanning 20 major tissue types, using the DINOv2 self-supervised algorithm with a Vision Transformer (ViT-Large) backbone. It demonstrated strong performance across cancer subtyping (lung, kidney, brain) and metastasis detection tasks. A newer version, UNI2, extends these capabilities further. UNI is openly available for academic research.
CONCH (also from Mahmood Lab) pioneered multimodal pre-training in pathology, training on 1.17 million histopathology image-caption pairs using both visual and language objectives. This enables CONCH to understand the relationship between tissue appearance and clinical descriptions, supporting tasks like cross-modal retrieval (finding images matching a text description) and zero-shot classification. It was evaluated on 14 diverse benchmarks across TCGA and other external datasets.
GigaPath (Microsoft, 2024) used a novel two-stage training strategy: first learning tile-level features with DINOv2, then integrating slide-level context using LongNet, a transformer architecture designed to handle the extremely long sequences created by dividing a gigapixel slide into thousands of tiles. Trained on 171,189 whole slides from over 30,000 patients, GigaPath achieved state-of-the-art performance on 25 of 26 benchmark tasks including cancer subtyping and mutation prediction.
mSTAR is notable for its tri-modal training on whole slide images, pathology reports, and gene expression (RNA-seq) data simultaneously -- the most comprehensive data integration to date. It was evaluated on 97 tasks across 15 categories. Atlas from Mayo Clinic is trained on the largest dataset of any pathology foundation model and introduces RudolfV, an architecture optimized specifically for pathological tissue analysis. Both models demonstrate that increasing data diversity and incorporating molecular information leads to more powerful general representations.
Beyond passive analysis tools, a new generation of AI copilots combines the visual capabilities of foundation models with large language models (LLMs) to create interactive conversational interfaces for pathology. Rather than just producing a classification score, these systems can engage in dialogue -- accepting questions in natural language and providing detailed explanations of their reasoning.
PathChat integrates UNI's visual encoder with a language model, creating a multimodal system capable of answering clinical questions about pathology images. In evaluations on multiple-choice pathology questions, PathChat achieved 78.1% accuracy using images alone and improved to 89.5% when clinical context was also provided. This demonstrates that combining visual and clinical information yields substantially better performance than either alone.
Other copilots include PathAsst, SmartPath, and SlideChat, each exploring different architectures for integrating visual and language understanding in pathology. SmartPath introduced the concept of autonomous task planning -- where the AI system can break complex diagnostic questions into subtasks and execute them sequentially without constant human direction, approaching more agentic behavior.
These copilots move AI from a tool that produces outputs to a collaborative partner that can explain its reasoning, answer follow-up questions, and assist with report generation. This conversational capability is expected to be crucial for clinical adoption, as it allows pathologists to interrogate AI conclusions and build appropriate trust in the system's recommendations.
Despite impressive benchmark performance, foundation models face a validation crisis: they often perform poorly in zero-shot testing (applying a model to a new task without any task-specific training). One study found that models excelling on familiar benchmark tasks showed dramatically reduced performance when applied to rare diseases or atypical cancer presentations not well represented in their training data. Published performance metrics are largely derived from the same datasets used for development.
The black-box problem is particularly severe in pathology. A foundation model may correctly classify a tumor but be unable to explain which morphological features drove its decision in a way a pathologist can verify. Without interpretability, clinicians cannot identify when a model is making errors for the wrong reasons -- such as classifying a tissue type based on staining artifacts rather than true biological features. This undermines trust and creates regulatory barriers.
Generative AI components, particularly in copilots, introduce the risk of hallucinations -- generating plausible-sounding but factually incorrect information. In pathology, where diagnostic errors can directly harm patients, a model confidently providing an incorrect diagnosis or citing a nonexistent clinical guideline is a serious safety concern. Robust evaluation protocols, human oversight requirements, and uncertainty quantification are essential safeguards.
Additional challenges include algorithmic bias (models may perform worse on tissue samples from underrepresented patient populations), data heterogeneity (variations in staining protocols and scanner types across institutions degrade performance), and the lack of standardized regulatory frameworks for validating and approving AI systems that continuously learn and update.
Realizing the transformative potential of foundation models and copilots in pathology requires solving both technical and governance challenges in parallel. On the technical side, priority areas include developing robust external validation frameworks using truly independent datasets, creating standardized benchmarks that test generalization across institutions, and advancing explainable AI approaches that make model reasoning transparent and auditable.
Novel regulatory frameworks are needed that address the unique characteristics of AI systems in pathology: their ability to continuously improve with new data, their dependence on the quality and diversity of training sets, and their interaction with human oversight. Current medical device regulations were designed for static tools and require adaptation for AI systems that may behave differently as they receive new data.
Despite these challenges, the trajectory is clear. Foundation models have already demonstrated capabilities that task-specific AI could not achieve: few-shot learning, zero-shot transfer across cancer types, and multimodal integration of imaging with molecular and clinical data. As validation frameworks mature and interpretability improves, these systems hold genuine promise to address the global pathologist shortage, reduce diagnostic variability, accelerate cancer discovery, and ultimately improve patient outcomes.