Diagnosis and Interpretation of Breast Cancer Using Explainable Artificial Intelligence

Abstract

Breast cancer is one of the leading causes of death among women and timely intervention is the key to curb it. This has necessitated Information Technology (IT) researchers and professionals to continually create models that can help in early detection of breast cancer, the area of interpretation has, however, not been explored. This has motivated this research to visually interpret breast cancer diagnosis to Pathologists and even layman who wishes to know. In this research the BreakHis dataset from Kaggle Challenge was used. A ResNet50 model (adopted in this research) was trained, using deep learning in order to classify the breast tumor as either malignant or benign. The result obtained from testing the model was 96.84% which outperformed results achieved by other researchers who used the same deep learning methodology. The classification of breast cancer diagnosis from histopathological images were later interpreted using eXplainable Artificial Intelligence) (AI) techniques like Integrated Gradient (IG), GradientShap (GS) and Occlusion, which gave reasons why a particular histopathological image was considered as Benign or Malignant. Comparing these three techniques, Occlusion was found to have more predictive results based on visualization and time of execution. This research did not only classify histopathological images as either benign or malignant but also gave reasons for its results unlike other earlier studies.