Visual interpretability for patch-based classification of breast cancer histopathology images

Mara Graziani, Vincent Andrearczyk, Henning Müller

Abstract: Decision support for digital histopathology has increased in an important way thanks to very good results using deep learning techniques in the past few years but neural networks have largely remained black boxes. Visualization methods address the interpretability of neural networks. If applied to histopathology images, they can improve the trust of pathologists in automated support tools for cancer diagnosis. We perform model interpretation and decision explanation of our binary patch-based breast cancer classifier. Experimental results show that morphological features of the nuclei influence network decisions in an important way.

Author Affiliation: University of Applied Sciences Western Switzerland (HES-SO), University of Geneva (UNIGE), Geneva, Switzerland

Keywords: histopathology, convolutional network, network interpretability, visualization