Weakly Supervised Learning: Toward Robust and Scalable Cancer Pathology Artificial Intelligence

Artificial intelligence models are increasingly being developed for cancer diagnostics, leveraging digital pathology slides alongside clinical and molecular data to characterize disease features at the patient level and support faster diagnosis and treatment planning. Current model-training approaches rely on detailed annotations of disease features in training data, leaving the vast reservoir of routine clinical data underutilized. Weakly supervised learning (WSL) is a machine learning framework that trains models using coarse or incomplete labels for medical images, rather than requiring detailed annotations specifying the precise locations of target features. This viewpoint argues that WSL should be more widely adopted for developing future oncological pathology models because it reduces dependence on detailed manual annotation. This can substantially expand the pool of data available for model development. When combined with reliable label extraction, appropriate model design, and rigorous validation, such expansion may improve performance across diverse patient populations and facilitate scalable model development. Since WSL can learn associations between image- or patient-level labels and tissue representations of the pathological target without requiring exhaustive spatial annotations, it may help identify candidate morphological features associated with cancer phenotypes and therapeutic outcomes. Such model-derived associations require independent validation before they can be interpreted as clinically useful indicators. A further advantage is that certain WSL frameworks, particularly multiple instance learning approaches employing appropriate instance-selection strategies, can focus model decision-making on diagnostically informative tissue regions. By reducing reliance on irrelevant (noisy) or misleading (spurious) features, these approaches may improve model robustness. With rigorous external validation, WSL could accelerate the development of scalable computational pathology systems with potential clinical utility in cancer healthcare.