Advanced Teeth Instance Segmentation Using Contextual Spatial Attention and Deeper Mask Head

Dental image analysis has advanced significantly with deep learning, yet achieving accurate teeth segmentation remains challenging, particularly with complex structures like adjacent or small teeth. This paper introduces Segmentation based on Contextual Spatial Attention Block (Seg-CSAB), a novel framework that integrates the CSAB with dilated convolutions to enhance feature extraction while maintaining high spatial resolution and an increased receptive field. Unlike traditional segmentation models, which often lose spatial information and compromise the representation of small structures due to downsampling, Seg-CSAB effectively captures fine-grained spatial details and contextual relationships between teeth, improving the understanding of smaller tooth structures and adjacent teeth. Additionally, Seg-CSAB features a deeper mask head with three additional convolutional layers (3 ×3) before the transposed convolution and final convolution layer, refining segmentation output and enhancing boundary delineation. It outperforms existing state-of-the-art (SOTA) methods, including Yolov8 and Mask R-CNN, in comprehensive segmentation, both qualitatively and quantitatively, offering substantial improvements in dental image analysis and treatment planning.