CISA-UNet: Dual auxiliary information for tooth segmentation from CBCT images

With the rise of artificial intelligence approaches, fully automatic tooth segmentation models from Cone-beam Computed Tomography (CBCT) images become more popular for dental clinical diagnosis. Recently, many deep learning-based tooth segmentation techniques from CBCT images have been proposed. However, fully segmenting teeth regions remains a challenging task due to unclear boundaries in the images, high similarity between adjacent teeth, and image noise caused by intensity inhomogeneity. In this paper, we propose a novel tooth segmentation algorithm called CISA-UNet, based on auxiliary information of edges and contrast enhancement maps. This auxiliary information highlights key details such as the surface and roots of the teeth, that can provide clear tooth contours and shape information for the network. Furthermore, a sliding window strategy is employed to crop input images into smaller blocks, to effectively reduce computational resources. Moreover, we design a dual attention mechanism based on ConvNext Block and improved spatial attention module, that can improve tooth segmentation by learning spatial and channel information. To evaluate the performance of the proposed method, we utilize a publicly available CBCT tooth segmentation dataset. Comprehensive evaluations, comparisons, and ablation experiments indicate that the proposed CISA-UNet can fully segment teeth regions from CBCT images and improve segmentation accuracy for digital dentistry compared with the existing state-of-the-art models. The proposed model achieves a better tooth segmentation accuracy with Dice Similarity Coefficient of 94.10% and Pixel Accuracy of 99.94%.