Answers are grounded in this paper and cite their sources.
Background: Convolutional neural networks (CNNs) are increasingly applied to haematoxylin and eosin (H&E)-stained colorectal cancer (CRC) histopathology, but it remains unclear how far reported performance reflects readiness for clinical use. Methods: Following PRISMA 2020, PubMed and Scopus were searched to 30 August 2026, and two reviewers independently selected studies; 102 studies were included. Each was classified by principal task, level of analysis, and validation strategy, and risk of bias was assessed with an adapted QUADAS-2 tool. Results: Most studies addressed binary tumour-versus-benign classification (53.9%) or multiclass tissue classification (26.5%); dysplasia grading, segmentation, and molecular prediction were less common. Over half evaluated isolated patches, 62.7% relied on public datasets, and only 21.6% tested a fixed model on an independent external cohort. Only 6 (5.9%) studies were at low risk of bias. Near-perfect accuracies came mainly from randomly partitioned benchmarks, performance declined on independent cohorts, and no study reported a prospective trial. Conclusions: This systematic review provides a task- and validation-stratified appraisal of CNNs in CRC histopathology, showing that the main barrier to clinical adoption is no longer algorithmic accuracy but the quality of validation evidence. Patient-level data separation, multicentre external validation, slide-level evaluation, and prospective evaluation of pathologist–AI collaboration are priorities.
No takes yet. Share an insight, caveat, or question.
Antzoulas et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: