Deep Visualisation-Based Interpretable Analysis of Digital Pathology Images for Colorectal Cancer

Colorectal cancer is one of the most widespread cancers in western countries. As with most forms of cancers, its diagnosis can only be made by performing a histopathological examination from tissue collected in the suspicious area. This process is time-consuming for the pathologist and consequently for the patient, whereas recent advances in computer vision and machine learning have shown that models can match the performance of human experts in classification and segmentation challenges. However the medical and ethical responsibilities that come with the use of artificial intelligence in healthcare demand that predictions are interpretable, and not just be the output of a black-box model. This paper demonstrates how visualisation of learned features play an important role in making the decision-making process transparent and helps to justify alignment with clinical features of colorectal cancer.