Difficulties and Recommendations for AI-Based Prediction of Prostate Cancer Aggressiveness in Digital Pathology

Prostate cancer is among the most common cancers in men with around 1.4 million new cases each year world-wide. A vital part in the diagnosis of prostate cancer is the evaluation of its severity using biopsies and histopathology. Recent progress in artificial intelligence-based image analysis has led to a flurry of algorithms for the automated analysis of prostate cancer histopathological data focusing on the detection of cancerous areas, the grading of cancer severity, and patient outcome. Some of these approaches have reached human expert-level performance and digital models trained directly on patient outcomes might surpass human performance in the future. Although these results hold great promise for the future usage of digital pathology in clinical settings, several bottlenecks remain to be addressed. Especially the robustness, reliability and trustworthiness of predictions must be guaranteed across a wide range of variation in protocols and instrumentation. While human experts are relatively robust to technical and biological variation in biopsies, artificial intelligence-based systems tend to struggle with differences in staining intensity, color, scanner type, and image resolution, impeding the clinical usage of digital models. In this work we highlight salient problems and minimal requirements of computational pathology for future use in clinical settings, while focusing on prostate cancer as a use case. In particular, we highlight data and model problems and solutions that include data variability, dataset size, and data annotations, as well as model robustness to data heterogeneity, model prediction confidence, and the explainability of model decisions. While model and data requirements for successful computational pathology in clinics will be highlighted, legal, ethical, and deployment requirements will not be addressed in this review. In summary, we provide a short overview of the field, salient problems, and potential solutions to harvest the full potential of digital pathology for prostate cancer in clinical practice.