Shedding light on the black box of a neural network used to detect prostate cancer in whole slide images by occlusion-based explainability

Diagnostic histopathology faces increasing demands due to aging populations and expanding healthcare programs. Semi-automated diagnostic systems employing deep learning methods are one approach to alleviate this pressure. The learning models for histopathology are inherently complex and opaque from the user's perspective. Hence different methods have been developed to interpret their behavior. However, relatively limited attention has been devoted to the connection between interpretation methods and the knowledge of experienced pathologists. The main contribution of this paper is a method for comparing morphological patterns used by expert pathologists to detect cancer with the patterns identified as important for inference of learning models. Given the patch-based nature of processing large-scale histopathological imaging, we have been able to show statistically that the VGG16 model could utilize all the structures that are observable by the pathologist, given the patch size and scan resolution. The results show that the neural network approach to recognizing prostatic cancer is similar to that of a pathologist at medium optical resolution. The saliency maps identified several prevailing histomorphological features characterizing carcinoma, e.g., single-layered epithelium, small lumina, and hyperchromatic nuclei with halo. A convincing finding was the recognition of their mimickers in non-neoplastic tissue. The method can also identify differences, i.e., standard patterns not used by the learning models and new patterns not yet used by pathologists. Saliency maps provide added value for automated digital pathology to analyze and fine-tune deep learning systems and improve trust in computer-based decisions.