Computational pathology in precision oncology: evidence, validation, and clinical translation

Histopathology is central to cancer diagnosis, but conventional interpretation is limited by interobserver variability and insufficient quantitative integration. Computational pathology applies machine learning to digitized pathology images to support diagnosis and grading, molecular biomarker inference, prognostic assessment, and treatment-response evaluation. This review examines clinically relevant approaches, criteria for evaluating evidence maturity, and requirements for clinical translation. We distinguish supervised learning, weakly supervised multiple-instance learning, foundation models, and multimodal or spatial systems, and summarize representative studies by clinical task, endpoint, cohort, validation strategy, performance metric, and translational stage. Although foundation and multimodal models enhance representation learning and data efficiency in selected settings, they do not eliminate risks related to domain shift, data leakage, calibration error, bias, or workflow mismatch. Responsible clinical adoption requires representative multicenter datasets, patient-level splitting, task-appropriate metrics, uncertainty and abstention strategies, pathologist-in-the-loop integration, regulatory oversight, and post-deployment monitoring. Computational pathology should therefore be regarded as an emerging decision-support platform whose clinical value must be established prospectively rather than inferred from retrospective discrimination alone.