A Study on Deep Learning Algorithm Conditions for Improving the Accuracy of Carious Tooth Detection in Panoramic Radiographic Images

Abstract

This study conducted research on an object detection module to distinguish images of carious teeth, a common dental condition, and further sought conditions to support AI in identifying these images more effectively. During the training process using U-Net, 10 process modules were created with the DEEP:PHI program, and training and evaluation were conducted using data augmentation. The Dice and miou values were derived to compare and analyze the accuracy of carious tooth detection based on the degree of area overlap. As a result, the Pre-processing Patch 768×768 module achieved the highest values, with a Dice value of 0.6725 and an miou value of 0.5066. This demonstrates that the use of patch extraction techniques with pre-processing in data augmentation and larger image sizes can improve the accuracy of area segmentation. This study represents an initial attempt to integrate dental panoramic radiographic images with AI, with the expectation that it will serve as a reference for future research. Furthermore, it is anticipated to serve as a foundation for radiologists to apply AI technology to the imaging field, aiming for academic and technological advancements.