Deep learning–based ROI hybrid classification of periodontal disease in dental radiographs

Abstract

The objective was to develop a decision-support application that assists dentists in identifying cavities’ bone loss resulting from Periodontal Disease for suitable dental procedures. The dataset consists of 290 periapical X-ray images, including Maxillary anterior-posterior and Mandibular anterior-posterior views, gathered from real clinical examinations. The first objective of the thesis was achieved through binary classification models, which categorize images into two groups: normal teeth and problem teeth.

The methodology included image preprocessing using CLAHE, Region-of-interest(ROI) single tooth root annotation, and neural network modeling, with attention mechanisms guided by the ROI boundaries. Five experimental models were developed from simple neural networks with full images, a pretrained model with full images, ROI guided attention, ROI guided attention with fine-tuning, and the pretrained model with ROI cropped images.

The results indicate highly unstable accuracies for the simple neural networks at an accuracy of 60%. In contrast, DenseNet121 produced a stable accuracy of around 70%. The optimal performance appeared to be the guided cropped attention regions of interest model, which reached an accuracy of 86%, followed by the more reliable full image guided attention 82%.