A Study on the Deep Learning Models for Lung Disease Detection from Chest X-Ray Images

Chest X-ray (CXR) imaging is a popular, inexpensive, and non-invasive diagnostic method for examining lung diseases. However, even radiologists often have lower accuracy with CXR data than more expensive methods like CT or MRI, due to lower precision of CXR images. Here, we present a new hybrid deep learning model of image features obtained through a ResNet50 model and clinical data taken through a Random Forest (RF) classifier. The image processing pipeline involved scaling, background/foreground noise removal, feature augmentation, and data cleaning for the clinical data, to present robust input to both deep learning model and RF classifier. The deep learning image features obtained from the ResNet50 model were fused with clinical data obtained from the RF classifier to complete a final classification. Our results compared model levels of accuracy, precision, recall, and F1-score against modern CNN and RNN baselines. Our experiments on a curated and classified CXR data set demonstrated that our ResNet50+RF hybrid model achieved the best performance, with 92% accuracy, 95.9% precision, 96% recall, and a 96% F1-score; the CNN model achieved an accuracy of 91%. RNN performance was lower overall, with an accuracy of 89%. These conclusions suggest the potential of extracting deep features from an image, while also performing ensemble learning on clinical data to improve classification accuracy with a scalable model for real-world clinical decision support systems.