Evaluation of Post-hoc Interpretability Methods in Breast Cancer Histopathological Image Classification

Methods for post-hoc interpretability are essential for understanding neural network results. Recent years have seen the emergence of numerous post-hoc techniques, but their application to certain tasks, such as histopathological image classification for breast cancer, can produce varied and unpredictable outcomes. Frameworks for quantitative assessment are essential for evaluating each method’s effectiveness. The implementation of post-hoc interpretability methodologies is however hampered by the shortcomings of current frameworks, particularly in high-risk industries. In this study, the performance levels of several common post-hoc interpretability methods are systematically evaluated and compared in the context of histopathological image classification for breast cancer. The study is based on six post-hoc interpretability methods, 3 datasets, and 3 deep neural network models, compared via a RemOve And Retrain (ROAR) approach. The results show that Shapley value sampling obtains the best overall performance in the context of the chosen breast cancer histopathological image datasets.