In this paper, we address the question of interpretability in the context of whole slide images (WSI) classification with the formalization of the design of WSI classification architectures and propose a piece-wise interpretability approach, relying on gradient-based methods, feature visualization and multiple instance learning context. After training two WSI classification architectures on Camelyon-16 WSI dataset, highlighting discriminative features learned, and validating our approach with pathologists, we propose a novel manner of computing interpretability slide-level heat-maps, based on the extracted features, that improves tile-level classification performances.

We measure the improvement using the tile-level AUC that we called Localization AUC, and show an improvement of more than 0.2. We also validate our results with a RemOve And Retrain (ROAR) measure. Then, after studying the impact of the number of features used for heat-map computation, we propose a corrective approach, relying on activation colocalization of selected features, that improves the performances and the stability of our proposed method.

View full text

Authors: 

Antoine Pirovano 1,2,*, Hippolyte Heuberger 1, Sylvain Berlemont 1, SaÏd Ladjal 2 and Isabelle Bloch 2,3




1 Keen Eye, 75012 Paris, France; hippolyte.heuberger@keeneye.ai (H.H.); sylvain.berlemont@keeneye.ai (S.B.)

2 LTCI, Télécom Paris, Institut Polytechnique de Paris, 91120 Palaiseau, France; said.ladjal@telecom-paris.fr (S.L.); isabelle.bloch@telecom-paris.fr (I.B.)

3 Centre National de la Recherche Scientifique, Laboratoire d’Informatique de Paris 6, Sorbonne Université, 75005 Paris, France

* Correspondence: antoine.pirovano@keeneye.ai