Comparative Evaluation of Explainable AI Techniques for Histopathology-Based Cancer Detection Using Deep Learning

Abstract

The analysis of histopathological images is an important method of diagnosing cancer. Deep learning models, such as convolutional neural networks and transformer-based models have demonstrated great potential in automated cancer detection. However, they are black-box and cannot be understood in a clinical context. In this paper, a comparative study is conducted to analyze the explainable artificial intelligence (XAI) approaches for cancer detection based on histopathology images on a hybrid model based on deep learning networks: ResNet-18 and Vision Transformer (ViT). The hybrid model is trained on histopathology images and three explainability mechanisms are used, including Grad-CAM, Grad-CAM++, and Score-CAM to visualize regions of decisions made. The created heatmaps are evaluated according to the accuracy of localization, noise minimization, and interpretability. Experimental results indicate that Grad-CAM++ provides better localization and Score-CAM generates more smooth and noisy explanations. This comparative analysis enhances transparency of models and helps to make more credible medical decisions.