DFTNet: Differential Feature Fusion Model for Tooth Segmentation in CBCT Images

In dental medicine, precise tooth segmentation is crucial for many clinical applications. The three-dimensional anatomical structure of teeth is clearly visible in CBCT images, where the teeth, roots, and surrounding hard tissues form a complex structure. Manually annotating the dental regions is time-consuming. Therefore, this study proposes a tooth segmentation method based on CBCT images. The model achieves the instance segmentation task by detecting and segmenting the teeth. Information loss is unavoidable during the upsampling and downsampling processes, impacting the final segmentation results. To mitigate this problem, this paper presents a differential feature fusion module. This module utilizes differential feature maps to allow the network to learn specific regions, increasing the significance of the areas where information is lost during sampling. This enables the model to learn more targeted features when fusing different scale features, ultimately improving the quality of segmentation. The experimental results show that the proposed approach effectively resolves the tooth segmentation issue in CBCT images, yielding outstanding segmentation outcomes.