A Novel Deep Learning Approach for Classification of Abnormal Teeth in Panoramic X-rays

Abstract

Dental professionals can identify common dental conditions including caries with the use of panoramic and periapical radiograph techniques. In most cases, panoramic and periapical pictures are used by dentists to physically diagnose problematic teeth. Unnoticeable aberrant teeth can result via manual diagnosis for a number of reasons, including inexperience and carelessness brought on by a tremendous workload. Therefore, in order to avoid these drawbacks, advanced computer vision technologies along with Data-driven image enhancement techniques are required. Convolutional Neural Networks (CNNs) are recognized for their ability to utilize multiple convolutional layers, resulting in superior classification accuracy. Consequently, they represent one of the most effective methodologies for diagnosing dental abnormalities. However, the classification process using CNNs necessitates a substantial dataset of images to ensure adequate training and to achieve satisfactory performance outcomes. This study proposes a modified output layer for the ResNet-50, MobileNet V2, and Lightweight Convolutional Neural Networks (LWCNN) models to classify panoramic and periapical images into four distinct categories: quadrant, quadrant-enumeration, quadrant-enumeration-disease, and unlabeled. The unsharp filter, histogram equal, and complement image were among the methods applied to the chosen image samples in order to create a different perspective on the dataset. Training and testing accuracy for ResNet-50, MobileNet V2, and LWCNN were 38.56, 62.13, 63.91, 34.06, 22.83, and 26.38, respectively. On the other hand, the CNN models in the second scenario obtained training and testing accuracy of 99.80, 99.60, 99.90, 93.20, 96.50, and 96.60 respectively. This outcome demonstrates that the suggested LWCNN model can successfully aid in the classification of abnormal teeth.