A Framework for Evaluating Model Trustworthiness in Classification of Very High Resolution Histopathology Images

Abstract

In computer vision, one approach to explaining a deep learning model’s decision is to show regions of visual evidence upon which the model makes a decision. Typically, this evidence is represented in the form of a saliency map which conveys how much an image region is contributing to the model’s decision. For a model to be trustworthy, it is expected that this saliency region should provide relevant information. In this work, we use model "trustworthiness" or "rationale" to describe how much relevant information the model is using to determine the image class. For medical images, this information connects to biological relevance. For very high resolution histopathology image applications, such as gigapixel whole-slide image classification, where patch-based multiple-instance based learning approach is taken to determine the patch label, this biological relevance has to be determined both at the patch and the image level. In this work, we present a novel patch-based model trustworthiness evaluation framework for very high resolution histopathology images. Our trustworthiness framework takes two approaches: spatial overlap based and feature based evaluation. For the overlap based approach, we check overlap with the annotation provided with the database to see if they have biological relevance, since for tumor positive patches only high probability regions from within the annotated regions are likely to be relevant. For feature based approach, we train an interpretability model using the sub-patches of the training set, extract features and cluster them. Then based on the distance from these clusters we determine if there is any biological rationale behind the prediction. Finally, we propose four patch-level and four image-level rationale metrics that evaluate the biological relevance of the information used by the classifier to decide on the patch class. Our experiment using the CAMELYON16 dataset shows the efficacy of this approach for model trustworthiness evaluation and explainability.

Ähnliche Arbeiten