Understanding CNN to Automatically Diagnose Rheumatoid Arthritis using Hand Radiographs

Abstract

Rheumatoid arthritis may be described as a chronic inflammatory disorder which affects the joints by damaging the body's tissue. Therefore, the identification and detection of rheumatoid arthritis by hand, especially during its development or pre-diagnostic phases, requires an effective system analysis. Hand Radiographs is one such cheaper and popular method used by doctors.  This system is designed to detect rheumatoid arthritis of the hand using image processing techniques and a neural network of convolution. The system comprises of two main phases. The image processing phase is the first stage in which images are processed using image processing. These techniques include pre-processing, image segmentation and feature extraction using gabor filter. The second phase the extracted features being used as inputs for the neural convolution network, which classifies the hand images as normal or abnormal (arthritic). Classification is carried out based on the CNN, which involves the training of the network with normal and abnormal hand images. Data entails 190 images of hand radio graphs. The system trained on 133 images and tested on 57 images with module accuracy 94.46%. The sensitivity of the network is 0.95, specificity 0.82.