Progressive Faster Residual Convolutional Neural Network for Improving Osteoarthritis of the Temporomandibular Joint Detection

Abstract

Osteoarthritis of the Temporomandibular Joint (TMJ-OA) is a chronic condition that affects the TMJ and is characterized by the progressive degeneration of the internal surfaces of the joint. Several deep learning models were adopted for identifying the TMJ-OA from the panoramic dental X-ray scans. Amongst, an Optimized Generative Adversarial Network (OGAN) with Faster Residual Convolutional Neural Network (FRCNN) produces more synthetic images to train the FRCNN for recognizing TMJ-OA cases. But, its accuracy was comparatively low while recognizing Region-of-Interest (RoI) from the panoramic scans that have analogous objects. Hence in this paper, an OGAN with a Progressive FRCNN (OGAN-PFRCNN) model is proposed, which enhances the FRCNN by integrating the Feature Pyramid Network (FPN) and RoI-grid attention strategy for TMJ-OA identification. First, the training images are fed to the ResNet101 for feature mining, which provides Multi-Scale Feature Map (MSFM) from the dental panoramic scans. Those features are then passed to the FPN with the RoI-grid attention strategy, which encodes richer characteristics by considering standard attention and graph-based point functions into a combined formulation. Then, those characteristics are fused at various levels to get a useful MSFM, which increases the network efficiency significantly. Moreover, such a Feature Map (FMap) is used to train the PFRCNN model, which is later applied to recognize the test scans into either healthy or TMJ-OA. At last, the testing outcomes show that the OGAN-PFRCNN attains 96.2% accuracy on the panoramic dental X-ray database compared to the FRCNN model.