Optimized Adversarial Network with Faster Residual Deep Learning for Osteoarthritis Classification in Panoramic Radiography

Abstract

Temporomandibular joint osteoarthritis (TMJ-OA) is a degenerative disorder affecting the TMJ and is distinguished by the gradual deterioration of the joint's interior surfaces. To identify and classify the TMJ-OA from panoramic dental X-ray images, many deep learning models were developed. Amongst, a faster region-based convolutional neural network (FRCNN) can find the condylar area and recognize its abnormalities by learning adequate features with a limited number of images. Nonetheless, the accuracy was not effective for larger databases. Hence in this article, an optimized generative adversarial network (OGAN) model is proposed to create larger panoramic dental X-ray images for TMJ-OA recognition. This GAN model utilizes the generator to produce synthetic panoramic dental images and trains the discriminator to decide whether the created images are real or counterfeit. Besides, an Elephant Herding Optimization (EHO) algorithm is adopted to select the most optimal hyperparameters of the GAN according to the clan and separating factors. Then, the created synthetic panoramic images are added to the actual database and partitioned into 2 distinct collections: training and test collections. The training collection is utilized to train the FRCNN, which extracts the condylar area from every image and identifies the abnormalities. Further, the trained model is tested using the test set to analyze the efficiency of TMJ-OA recognition. Finally, the experimental results exhibit that the OGAN-FRCNN model achieves an accuracy of 94.59% on the panoramic dental X-ray database, whereas the classical models such as VGG16,

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

“…Amongst, 65% of scans are applied for learning and the residual 35% are applied for testing. Also, the obtained efficiency is compared with the existing models: FRCNN [19], OGAN-FRCNN [20], DetectNet [21], ANN [25], HNN [26] and DenseNet121 [28] regarding the below metrics:…”

“…With the help of the DetectNet structure and the DIGITS, a deep learning framework [21] was created to recognize and classify radiolucent cancers in the mandible on panoramic radiography images. The first step was to gather panoramic images of people with mandibular radiolucent tumors.…”