The gap between AI's promise in cancer diagnostics and what is actually cleared for clinical use is far wider than most assume. Despite years of headline-grabbing research, the regulatory landscape for AI-assisted cancer pathology remains sparse—a finding with real consequences for patients, oncologists, and the institutions investing in digital pathology infrastructure.
A comprehensive review published in Frontiers in Digital Health mapped the full universe of FDA-approved and CE-marked AI tools for whole-slide image cancer histopathology—the digital scanning and automated analysis of tissue biopsies. Researchers identified only four FDA-cleared solutions, all covering a narrow band of cancer applications. The review systematically examined how these approved products were developed and validated, comparing them against state-of-the-art research pipelines that have not crossed the regulatory threshold. Key variables analyzed included learning modalities (supervised versus self-supervised), processing architectures (including transformer-based models), statistical validation strategies, and biomarker coupling approaches. Approved products were found to integrate relatively well into existing pathologist workflows, but the gap between research-only AI and clinically deployable AI remains substantial.
This assessment carries notable implications for the field. AI cancer pathology research has accelerated dramatically since 2020, with large foundation models and multi-modal transformers demonstrating performance rivaling trained pathologists on certain tasks. Yet the regulatory bottleneck reflects genuine scientific challenges: prospective validation at scale, performance across demographically diverse tissue samples, and reproducibility across scanner hardware from different manufacturers remain underaddressed in many published studies. The review's emphasis on patient safety as a central organizing concern—rather than benchmark accuracy—signals a maturing discipline. The finding that biomarker assays may be paired with emerging targeted therapies is particularly forward-looking, but it underscores how clinical adoption requires not just algorithmic sophistication but tightly controlled diagnostic-therapeutic linkage. This review is confirmatory of a widely suspected reality: AI histopathology is powerful in research contexts but still nascent in practice.