Automatic Detection and Segmentation of Tooth Cracks Based on Improved Mask R-CNN

Early diagnosis and intervention of cracked teeth are crucial for preventing further dental damage. However, the detection of cracked teeth remains challenging for dental clinicians due to the subtle, complex, and irregular feature of these cracks. To address this issue, we propose an improved Mask R-CNN instance segmentation network for the automatic detection and segmentation of cracked teeth. Firstly, the backbone network was replaced with ResNeXt-50(32×4d) to enhance the extraction of local features specific to cracks. Secondly, we introduced a Crack Feature Enhancement Module (CFEM), utilizing Bayesian optimization to fine-tune its hyperparameters, which leverages the pixel intensity differences between cracked and non-cracked regions to increase the sensitivity of the Feature Pyramid Network (FPN) to the complex features of cracks while suppressing irrelevant background information. Additionally, the mask head was redesigned into an encoder-decoder structure incorporating dynamic snake convolutions, which enables better capture of crack edge details and the integration of both deep and shallow feature information, with deep supervision applied to adjust the loss function weights. Extensive experiments and comprehensive evaluations demonstrate that our method outperforms the current state-of-the-art techniques. Furthermore, experiments on real intraoral images validate the effectiveness of our approach in detecting tooth cracks. Our model enables more accurate and earlier detection of cracked teeth, improving patient outcomes by allowing for timely interventions, reducing the need for invasive treatments, and preserving dental structure. Our code and datasets are available at https://github.com/YCHuang18/ToothCrack.