Gradient-based Interpretation on Convolutional Neural Network for Classification of Pathological Images

Digital pathological images have been developing very fast with the recent progress in imaging hardware. It provides tons of imaging data for diagnostic purposes in medical applications. Data-driven algorithms, especially convolutional neural networks (CNN), have been successfully applied in various image analysis tasks. However, neural network models are regarded as black box model and they are supposed to be more interpretable and transparent for real-life applications, especially in the field of medical image analysis. In this work, we trained a CNN model on digital pathological images from lymph nodes sections to classify metastatic tissue, achieving F1 score of over 86%. Furthermore, we investigate the possible ways to understand how the model makes decisions using gradient-based method.