Visualizing abnormalities in chest radiographs through salient network activations in Deep Learning

This study aims to visualize salient network activations in a customized Convolutional Neural Network (CNN) based Deep Learning (DL) model, applied to the challenge of chest X-ray (CXR) screening. Computer-aided detection (CAD) software using machine learning (ML) approaches have been developed for analyzing CXRs for abnormalities with an aim to reduce delays in resource-constrained settings. However, field experts often need to know how these techniques arrive at a decision. In this study, we visualize the task-specific features and salient network activations in a customized DL model towards understanding the learned parameters, model behavior and optimizing its architecture and hyper-parameters for improved learning. The performance of the customized model is evaluated against the pre-trained DL models. It is found that the proposed model precisely localizes the abnormalities, aiding in improved abnormality screening.