Deep Learning for Grading Cardiomegaly Severity in Chest X-Rays: An Investigation

Abstract

This study investigates using deep convolutional neural networks for automatic detection of cardiomegaly in digital chest X-rays (CXRs). First, we employed and fine-tuned several deep CNN architectures to detect presence of cardiomegaly in CXRs. Next, we introduce a CXR-based pre-trained model where we first fully train an architecture with a very large CXR dataset, and fine-tune the system with cardiomegaly CXRs. Finally, we investigate the correlation between softmax probability of an architecture and the severity of the disease. We used two publicly available datasets, NLM-Indiana Collection and NIH-CXR datasets. Based on our preliminary results (i) data-driven approach produces better results than prior rulebased approaches developed for cardiomegaly detection, (ii) our preliminary experiment with alternative pre-trained model is promising, and (iii) the system is more confident if severity increases.