Are H&E-based computational models transforming molecular pathology diagnostics in cancer?

Background: Advances in computational pathology using deep learning have enabled the prediction of molecular biomarkers and therapy response directly from hematoxylin and eosin (H&E) whole slide images (WSIs), offering a potential complement to molecular assays. Methods: Following PRISMA guidelines, we searched Embase, Web of Science, and PubMed (01/2020–04/2025) for studies applying H&E-based WSI computational analysis to predict prognosis or treatment response in solid tumors, with molecular pathology comparators. Data extraction and quality assessment with a 7-item QUADAS-2/PROBAST-adapted tool was performed. Results: Of 322 records screened, 64 met inclusion. Many studies focused on breast, colorectal, lung, gastric, or prostate cancer. Targets included Microsatellite Instability, mutations (e.g., EGFR, TP53), protein biomarkers (e.g., PD-L1, HER2), and gene expression profiles. Convolutional neural networks were most common, alongside Vision Transformers and foundation models. Median dataset size for model development was 509 cases (range: 13–12,592); 64% used independent external validation. AUROCs ranged from 0.58 to 0.97. Clinical readiness varied: 41% were proof-of-concept only, while 30% had multi-cohort validation. Median quality score was 71.4%, with frequent gaps in cohort representativeness and method reporting. Conclusions: H&E-based WSI computational analysis can accurately predict diverse biomarkers and therapy-relevant endpoints, sometimes nearing molecular assay performance. Dataset quality, external validation, and reporting variability currently limit clinical adoption; standardized benchmarking and multi-institutional validation are needed for integration into precision oncology workflows.