SRA-UNet: Selective Scan Residual Attention Network for Tooth Segmentation in Panoramic X-ray Images
Year: 2026
Type: conference-paper
Abstract: Tooth segmentation is a fundamental task in oral medical image analysis. It has significant applications in assisted diagnosis, orthodontic planning, and preoperative assessment. However, tooth images often exhibit blurred boundaries, adhesion between adjacent targets, uneven grayscale distributions, and local occlusions. These issues make it difficult for convolutionbased segmentation models to fully capture cross-region spatial dependencies and affect recognition accuracy in structurally complex regions. To address these problems, this study proposes SRA-UNet, a tooth image segmentation network that integrates two-dimensional selective scanning, residual connections, and an attention mechanism. Based on the U-Net backbone, the method introduces an SS2D-ResBlock, which combines two-dimensional selective scanning with residual connections in both the encoding and decoding stages. This enhances the network’s global contextual modeling and the stability of feature propagation. A CBAM attention module is embedded into the skip connections to adaptively recalibrate low-level detail features, improving the recognition of tooth contours and adjacent adhesive regions. Experimental results on two panoramic dental X-ray datasets show that SRA-UNet achieves 87.04% Dice, 77.55% IoU, and 90.71% Recall on the UFBA_UESB_DENTAL_IMAGES dataset, and 86.42% Dice, 75.79% IoU, and 89.21% Recall on Tufts Dental Database. In addition, the proposed model contains 9.36M parameters and requires 20.3G FLOPs, showing a favorable trade-off between segmentation accuracy and computational cost. These results demonstrate that SRA-UNet effectively improves tooth segmentation and handles complex boundaries and adhesive regions well.
Institution: Hangzhou Dianzi University
Language: English
Subfield: Oral Surgery
Field: Dentistry
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7201874997