A Hybrid Attention-Enhanced Network for Accurate Dental Pulp Segmentation from CBCT

Accurate segmentation of dental pulp from cone-beam computed tomography (CBCT) is essential for digital endodontic planning, virtual surgery, and anatomical visualization. However, the intrinsic challenges of low contrast between pulp and surrounding dentin, irregular anatomical morphology, and vulnerability to imaging artifacts greatly impede robust segmentation. To address these issues, we propose PulpSegNet (PSN), a hybrid attention–enhanced segmentation framework built upon the nnUNet architecture and specifically designed for fine-grained delineation of dental pulp structures. PSN incorporates a self-calibrated convolution module SCConv and a contrast-driven feature aggregation module CDFA, strategically integrated into both deep and shallow network stages to enhance structural coherence and boundary precision simultaneously. Through hierarchical feature calibration and foreground-background contrast modeling, the network effectively mitigates the difficulties associated with small-volume targets and ambiguous tissue boundaries. Comprehensive experiments conducted on a multi-class dental pulp segmentation benchmark demonstrate that PSN consistently outperforms existing approaches and establishes a new benchmark for high-precision segmentation in digital dentistry.