Automated Dental Image Analysis by Deep Learning on Small Dataset

Abstract

Dental radiography provides important evidence for clinical diagnosis, treatment and quality evaluation. Much effort has been spent on developing digitalized dental X-ray image analysis systems for clinical quality improvement. In this paper, we present the datasets, procedures, and results conducted to evaluate dental treatment qualities using periapical dental X-ray images taken before and after the operations. In order to support dentists to make clinical decisions, we propose a tool pipeline for automated clinical quality evaluation. We build a dataset with 196 patients' periapical dental radiography images before and after the treatments. Radiography images are labelled as cases that are `getting better', `getting worse' and `have no explicit change' by designated dental experts. Our proposal includes an automatic method with the medical knowledge to crop the ROIs for clinical evaluation - the apical adjacent regions, and then pairs of ROIs are fed into a CNN to train the model for automated clinical quality evaluation. Our approach achieves the F1 score of 0.749, which is comparable to the performance of expert dentists and radiologists.