The histopathological slide has been the diagnostic gold standard in medicine for over a century. A thin section of tissue, stained and mounted, examined under a microscope by a trained pathologist. The judgment is expert, time-intensive, and subject to the inevitable variability of human perception. Artificial intelligence is not replacing that judgment — it is beginning to augment, systematize, and in some cases exceed it.
Digital pathology — the digitization of histological slides into whole-slide images (WSIs) — has created the substrate for computational analysis at scale. What was once a glass slide examined sequentially is now a gigapixel image that algorithms can process in full, identifying regions of interest, quantifying cellular features, and correlating morphological patterns with molecular and clinical outcomes.
What deep learning sees
Deep learning models applied to histopathological images have demonstrated remarkable performance across a range of tasks. A comprehensive review published in The Lancet Digital Health in 2025 summarizes the current state: on haematoxylin and eosin-stained slides, algorithms have shown accuracy comparable to subspecialist pathologists in tumor detection, classification, grading, and subtyping across multiple cancer types. Beyond replicating human assessment, these models can detect morphological signatures predictive of molecular alterations — identifying, for instance, microsatellite instability or HER2 status directly from H&E images, without additional molecular testing.
Wang et al. (2024), publishing in Nature, presented a pathology foundation model trained on large-scale WSI data that demonstrated generalization across cancer diagnosis and prognosis prediction tasks. Foundation models — large pre-trained architectures adapted for downstream tasks — are rapidly becoming the dominant paradigm in computational pathology, following the same trajectory that transformed natural language processing.
The challenge of clinical translation
Despite these advances, the gap between research performance and clinical deployment remains substantial. A systematic review published in Computational and Structural Biotechnology Journal (2025) identifies the key obstacles: computational scaling for gigapixel images, the scarcity of well-annotated training data, interpretability gaps that limit pathologist trust, and domain shifts across institutions and staining protocols. A model trained on slides from one institution may fail on slides from another.
Regulatory approval adds another layer. As of mid-2025, only two AI-based tools have received FDA authorization for use in histopathology — Paige Prostate and Galen Second Read — despite more than 1,200 AI-enabled medical devices receiving clearance across other medical domains.
Quantitative morphometry and fractal approaches
Alongside deep learning, classical quantitative morphometry continues to contribute to computational pathology. Fractal dimension analysis of vascular and tissue structures provides substrate-neutral descriptors that capture geometric complexity independent of staining variation — an advantage in settings where deep learning models may overfit to specific preparation protocols. In our own work in pathology, the intersection of fractal-based image analysis and machine learning has shown methodological promise for characterizing tissue microenvironment organization in ways that neither approach reaches alone.
References
Wang, X., Zhao, J., Marostica, E., et al. (2024). A pathology foundation model for cancer diagnosis and prognosis prediction. Nature, 634, 970–978.
Marra, A., Morganti, S., Pareja, F., et al. (2025). Artificial intelligence entering the pathology arena in oncology: current applications and future perspectives. Annals of Oncology, 36, 712–725.
Anonymous authors (2025). Application of artificial intelligence and digital tools in cancer pathology. The Lancet Digital Health. DOI: 10.1016/S2589-7500(25)00115-3
Anonymous authors (2024). Machine learning methods for histopathological image analysis: Updates in 2024. Computational and Structural Biotechnology Journal. DOI: 10.1016/j.csbj.2024.12.001
— Daniela Di Marco, Operation Knowledge
Leave a Reply