Effect of Chest X-Ray Contrast Image Enhancement on Pneumonia Detection using Convolutional Neural Networks

Abstract

The problem in pneumonia detection using a chest X-ray is the differences in the image quality due to the difference in the image acquisition procedure. To overcome the issue, the image enhancement technique is used in the preprocessing step. The purpose is to examine the effect of three image enhancement techniques to detect pneumonia, i.e. histogram equalization (HE), contrast limited adaptive HE and exposure fusion framework. These enhanced images are used as input images in pneumonia detection using VGG16 convolutional neural network architecture. In total, 3,151 chest X-ray images are used. The best performance is achieved by the exposure fusion framework image enhancement technique. The combination of exposure fusion framework and VGG16 give the training loss and accuracy of 0.2113 and 0.9451, and validation and accuracy loss of 0.6034 and 0.8670. Deeper analysis shows that the exposure fusion framework not only stretches the image intensity but also keeps the shape of the histogram remains. This technique will minimize the information loss in the enhanced image during the enhancement process.