A major review says computers are getting good at reading cancer slides and spotting tell‑tale molecular clues—yet most tools still need stronger proof before routine use.
Artificial intelligence (AI) is moving from lab benches into pathology labs, where doctors diagnose cancer by examining tissue on microscope slides. A comprehensive review from an international oncology working group finds AI can help detect and classify tumors, estimate prognosis, and even infer molecular features from standard stained slides—promising faster, more consistent reads and wider access to high‑quality diagnostics. But the authors stress that, so far, no prognostic or predictive AI biomarker has the top‑tier clinical evidence required for broad adoption.
What’s new
The team mapped the field using a systematic review and update: 68 studies on tumor detection/classification, 143 on molecular biomarker assessment, and 92 on AI models that predict outcomes like relapse or survival. A Venn diagram in Figure 2 (page 4) shows how these domains overlap across hundreds of published studies.
How the tech works
Deep‑learning systems are trained on “whole‑slide images” (WSIs) of tissue to recognize patterns humans may miss. As illustrated in Figure 1 (page 3), today’s models range from supervised tools that learn from labeled examples to “foundation models” trained on massive image collections and adapted to many tasks. These systems can quantify tumor features, immune cells, and more—turning pixel patterns into usable clinical signals.
What AI can already do—encouraging signs
- Detect & classify cancer: In challenges and large datasets, top algorithms matched or exceeded pathologists working under time pressure when finding metastases or grading tumors (for example, in prostate cancer).
- Read biomarkers from routine slides: Models are learning to estimate markers such as ER/PR/HER2 in breast cancer; quantify tumor‑infiltrating lymphocytes; and prescreen for microsatellite instability (MSI)/mismatch‑repair deficiency, which can guide immunotherapy—potentially reducing costly confirmatory testing.
- Hint at the genome: Several studies show AI can predict common mutations (e.g., TP53) or copy‑number changes directly from H&E slides—opening a door to faster triage for full molecular testing.
- Regulatory beachheads exist: The FDA authorized Paige Prostate Detect in 2021 to assist prostate biopsy reads, reporting an average 70% reduction in false‑negative diagnoses; in the EU, CE‑marked tools such as RlapsRisk BC (breast cancer relapse risk) and MSIntuit CRC (MSI prescreening) are approved, though real‑world use remains limited.
Mind the gaps
Despite rapid progress, the review highlights key hurdles: pre‑analytical variability (how tissue is processed and scanned), data sharing and bias, overfitting to local data, explainability, and uncertain regulatory pathways and reimbursement. Critically, no AI biomarker yet meets “level IA/IB” evidence (the standard typically required for clinical decision‑making), and most studies are retrospective. Prospective trials and standardized evaluation are needed before AI results can reliably change care.
What’s next
Huge “foundation” and generalist medical AI models trained on hundreds of thousands to more than a million slides are arriving, built to handle multiple tasks and combine pathology with radiology, genomics, and clinical data. Early studies suggest this multimodal approach can improve risk stratification and treatment prediction—but it will require larger, diverse datasets and transparent reporting to earn clinicians’ trust.
Main takeaway
AI isn’t replacing pathologists—it’s becoming a powerful second set of eyes. In the near term, patients may benefit as hospitals digitize slides and adopt validated tools that prescreen cases, speed up reads, and flag who might need extra testing. For AI to shape individual treatment choices, stronger, prospective evidence—and clear rules of the road—must come first.
Source: Marra A., Morganti S., Pareja F., et al. “Artificial intelligence entering the pathology arena in oncology: current applications and future perspectives,” Annals of Oncology (2025).
Editor’s note: This article is for information only and is not a substitute for professional medical advice.