Comparison of Deep Learning Approaches in Classification of the Chest X-Ray

Abstract

Chest X-Ray is a radiological examination that is commonly used in clinical practice and is easy to access. Deep convolutional neural networks (DCNN) are used to make the computer-aided diagnosis (CAD) of diseases on chest radiography. Deep convolutional neural networks help the radiologist to diagnose better. In this study, the ChestX-Ray14 data set was examined to understand different modern deep learning networks in diagnosing chest diseases. X-Ray image quality was improved by applying a three-step process including crop, histogram equalization and contrast-limited adaptive histogram equalization to the data sets. For training and validation purposes, images in the dataset were applied to the model with and without preprocessing. It was determined that the processed datasets provided more accurate results than the original images.  In this study, AlexNet, ResNet50, and GoogLeNet deep learning architectures were used to determine the presence of chest disease. The performances of these models, which generally classify normal and abnormal results from chest radiographs, were analyzed using preprocessed ChestX-Ray14 datasets and comparative evaluations were made. The most accurate was the ResNet architecture, where we used the preprocessed datasets to detect abnormalities, with 91.46% accuracy and 0.9584 area under curve (AUC) results.