Localization Improvements in Faster Residual Convolutional Neural Network Model for Temporomandibular Joint – Osteoarthritis Detection

To recognize the Osteoarthritis of the Temporomandibular Joint (TMJ-OA) from panoramic dental X-ray images, deep learning algorithms are widely used these days. Among others, an Optimized Generative Adversarial Network with Faster Residual Convolutional Neural Network (OGAN-FRCNN) was recently achieved better FRCNN training by providing more synthetic images for TMJ-OA recognition. However, the localization of a small condyle OA region was ineffective because of the complex background, occlusion and low-resolution images. Hence, this article proposes the OGAN with Progressive Localization-improved FRCNN (PLFRCNN) model for TMJ-OA recognition. First, the OGAN can augment the number of panoramic X-ray scans. Then, ResNet101 can extract the Feature maps (F-maps) at multiple levels, followed by the Feature Pyramid Network (FPN) with a Region-of-Interest (RoI)-grid attention for multiscale F-map extraction. Those F-maps are given to the Modified Region Proposal Network (MRPN), which applies a multiscale convolution feature fusion and an Improved Non-Maximum Suppression (INMS) scheme for creating the Region Proposals (RPs) with more information. To resolve the localizing variance and obtain the proposal F-maps, the improved RoI pooling based on bilinear interpolation merges both F-maps and RPs. Moreover, the fully connected layer is used to classify those F-maps into corresponding classes and localize the target Bounding Box (BB). Additionally, the BB regression in the TMJ-OA localization stage is enhanced by the new Intersection-Over-Union (IOU) loss function. Finally, the test outcomes reveal that the OGAN-PLFRCNN model attains an accuracy of 98.18% on the panoramic dental X-ray corpus, in contrast to the classical CNN models.