An Explainable Deep Learning Framework for Mandibular Canal Segmentation from Cone Beam Computed Tomography volumes

Cone Beam Computed Tomography (CBCT) is an indispensable imaging modality in oral radiology, offering comprehensive dental anatomical information. Accurate detection of the mandibular canal (MC), a crucial anatomical structure in the lower jaw, within CBCT volumes is essential to support clinical dentistry workflows, including diagnosis, preoperative treatment planning, and postoperative evaluation. In this study, we present a deep learning-based (DL) approach for MC segmentation using 3D U-Net and 3D Attention U-Net networks. We collected a unique dataset of CBCT scans from 20 anonymous hemisected mandibular bones, which were further processed for analysis. The samples were scanned using a CBCT scanner after inserting a wire through the whole length of the MC to identify its location in space (as a gold standard). Our experimental results demonstrate that the 3D Attention U-Net outperforms the standard 3D U-Net in detecting the MC’s location, with Dice similarity score, Precision, and Recall values of 0.65, 0.75, and 0.60, respectively. Unlike current DL-enabled methods for MC segmentation, which face deployment and trust challenges due to their ”black-box” nature, our approach incorporates a post-hoc visual explainability feature through the Grad-CAM++ (Gradient-weighted Class Activation Mapping) algorithm. This tool highlights important regions within the CBCT volumes that influence the model’s predictions, providing valuable insights into the segmentation process, and bridging the gap between cutting-edge DL technology and clinical practice.