Tooth segmentation in panoramic dental radiographs using deep convolution neural network -Insights from subjective analysis

In the last few years, dentistry has witnessed a phenomenal advancement in artificial intelligence. The importance of teeth segmentation in dental radiographs has increased since it enables medical practitioners to conduct examinations more precisely and accurately in dentistry and helps them develop the most effective treatment strategy for their patients. In this research work, TUFT and UFBA dental data sets have been combined and used to train UNet, UNet ++ with ResNet 50 pre-trained model, UNet with mobileNet as encoder, and Deeplabv2 models for teeth segmentation. Evaluation of their performance in teeth segmentation using panoramic dental radiographs is carried out. Also, a subjective analysis of the model's predicted mask output from the practitioners is carried out. UNet++ with the combined data set and after fine-tunning hyperparameter gives the best results (IOU 0.8619 and Dice coefficient 0.9258). Also, it is observed that the use of the post-processing technique, 'Residual dense spatial-asymmetric attention' for deblurring the output images improved the result. According to the findings of the subjective study, the practitioner's satisfaction index is 4.2 on a scale of 5, which emphasizes the need for practitioners feedback in model building to ensure the clinical usability of the proposed system.