Résumé

Visualization methods for Convolutional Neural Net­works (CNNs) are spreading within the medical com­munity to obtain explainable AI (XAI). The sole quali­tative assessment of the explanations is subject to a risk of confirmation bias. This paper proposes a methodol­ogy for the quantitative evaluation of common visual­ization approaches for histopathology images, i.e. Class Activation Mapping and Local-Interpretable Model­Agnostic Explanations. In our evaluation, we propose to assess four main points, namely the alignment with clinical factors, the agreement between XAI methods, the consistency and repeatability of the explanations. To do so, we compare the intersection over union of multi­ple visualizations of the CNN attention with the seman­tic annotation of functionally different nuclei types. The experimental results do not show stronger attributions to the multiple nuclei types than those of a randomly ini­tialized CNN. The visualizations hardly agree on salient areas and LIME outputs have particularly unstable re­peatability and consistency. The qualitative evaluation alone is thus not sufficient to establish the appropriate­ness and reliability of the visualization tools. The code is available on GitHub at bit.ly/2K4BHKz.

Détails

Actions

PDF