Masked Self-Supervised Pretraining for Semantic Segmentation of Dental Radiographs

In dental diagnosis and treatment planning, teeth segmentation is essential. Semantic segmentation of 2D dental radiographs helps to analyze dental structures precisely, detect anomalies in teeth, and evaluate oral health issues. However, creating segmentation masks is a time-consuming process and is prone to inaccuracies due to complex tooth structures. In this research, we propose a self-supervised learning approach for teeth segmentation using Modified ResUNet and Random Block Masking as the pretext task, where random blocks in dental radiographs are masked, and the model is trained to reconstruct the entire radiograph. Additionally, we utilize only 20% of the samples in the datatset for training. Our proposed approach outperforms state-of-the-art models such as U-Net, and PSPNet and performs comparably to LinkNet, trained on 80% of the samples in the dataset. Modified ResUNet trained on our approach is able to produce an accurate segmentation mask even when the ground truth mask contains errors.