Visualizing Salient Network Activations in Convolutional Neural Networks for Medical Image Modality Classification

Abstract

Convolutional neural network (CNN) has become the architecture of choice for visual recognition tasks. However, these models are perceived as black boxes since there is a lack of understanding of their learned behavior from the underlying task of interest. This lack of transparency is a drawback since poorly understood model behavior could adversely impact subsequent decision-making. Researchers use novel machine learning (ML) tools to classify the medical imaging modalities. However, it is poorly understood how these algorithms discriminate the modalities and if there are implicit opportunities for improving visual information access applications in computational biomedicine. In this study, we visualize the learned weights and salient network activations in a CNN based Deep Learning (DL) model to determine the image characteristics that lend themselves for improved classification with a goal of developing informed clinical question-answering systems. To support our analysis we cross-validate model performance to reduce bias and generalization errors and perform statistical analyses to assess performance differences.