A Comparative Study of Explainability Methods for Whole Slide Classification of Lymph Node Metastases using Vision Transformers

Abstract

Recent advancements in deep learning (DL), such as transformer networks, have shown promise in enhancing the performance of medical image analysis. In pathology, automated whole slide imaging (WSI) has transformed clinical workflows by streamlining routine tasks and diagnostic and prognostic support. However, the lack of transparency of DL models, often described as “black boxes”, poses a significant barrier to their clinical adoption. This necessitates the use of explainable AI methods (xAI) to clarify the decision-making processes of the models. Heatmaps can provide clinicians visual representations that highlight areas of interest or concern for the prediction of the specific model. Generating them from deep neural networks, especially from vision transformers, is non-trivial, as typically their self-attention mechanisms can lead to overconfident artifacts. The aim of this work is to evaluate current xAI methods for transformer models in order to assess which yields the best heatmaps in the histopathological context. Our study undertakes a comparative analysis for classifying a publicly available dataset comprising of N=400 WSIs of lymph node metastases of breast cancer patients. Our findings indicate that heatmaps calculated from Attention Rollout and Integrated Gradients are limited by artifacts and in quantitative performance. In contrast, removal-based methods like RISE and ViT-Shapley exhibit better qualitative attribution maps, showing better results in the well-known interpretability metrics for insertion and deletion. In addition, ViT-Shapley shows faster runtime and the most promising, reliable and practical heatmaps. Incorporating the heatmaps generated from approximate Shapley values in pathology reports could help to integrate xAI in the clinical workflow and increase trust in a scalable manner.