A whole-slide is greater than the sum of its...patches

Muscular-invasive bladder cancer (MIBC) is a common formof cancer which can necessitate complex treatment decisions.Different methods involving machine learning have been developed with the goal of improving and making MIBC diagnosis more specific, and thus limiting the amount of invasive testing needed for MIBC patients. A particularly fruitful direction of research involves the use of tissue images and the application of deep learning. In order to deal with extremelylarge whole slide images (WSIs), the state of the art methodsapproach the problem by using a patch-based convolutionalneural network which takes small patches (often 256 × 256pixels) of WSIs as input and provides a classification of cancerous or not-cancerous as output. Patch-to-slide classification is then often achieved by classifying a WSI as cancerous if and only if the majority of its patches are classified as cancerous. In this work we compare different approaches to the integration of local, patch based decisions, as a means of arriving at a robust global, WSI based classification. Our results suggest that an absolute, positive patch count based decisionmaking, with an appropriately learnt threshold, achieves the best results.