[A study on the application of an edge-enhanced triple-branch neural network in three-dimensional segmentation of the...

Objective: To construct an edge-enhanced triple-branch neural network (T-Net) for automated three-dimensional(3D) segmentation of the mandibular condyle from cone-beam computed tomography (CBCT) images, aiming to improve segmentation accuracy. Methods: CBCT images of 354 patients with malocclusion (708 condyles) who attended the Department of Orthodontics, Xiangya Stomatological Hospital, Central South University, from January 2022 to January 2024 were retrospectively collected. The dataset was divided into training (424 condyles), validation (142 condyles), and test (142 condyles) sets in a ratio of 6∶2∶2. The T-Net model was built on a single-encoder and dual-decoder architecture. It introduced an edge decoder supervised by edge information as labels. Through skip connections and feature interaction between the segmentation decoder and the edge decoder, the model's capability to extract contour feature information was effectively enhanced. Manual annotation of all images using 3D Slicer software served as the gold standard. The T-Net model parameters were optimized using the validation set, and its performance was evaluated on the test set and compared with four mainstream models (3D U-Net, Attention U-Net, Swin UNETR, and V-Net). Qualitative and quantitative evaluations were performed using visual segmentation results and metrics such as Dice coefficient, intersection over union (IoU), accuracy, precision, F1-score, sensitivity, and mean absolute error (MAE). Results: The T-Net model achieved precise and complete segmentation of the condyle. Its Dice coefficient (0.976±0.011), IoU (0.953±0.021), accuracy (0.999±0.001), precision (0.979±0.010), sensitivity (0.974±0.019) and F1-score (0.976±0.011) were superior to those of the other four models. Compared with the gold standard, the T-Net model yielded the smallest MAE values for condylar morphological parameters (volume: 550 mm³, surface area: 81 mm², length: 0.473 mm, width: 0.781 mm, height: 1.876 mm). Conclusions: The T-Net model demonstrates excellent performance in the condyle segmentation task, with its metrics significantly outperforming the other four models. The model can accurately extract condylar morphological features from CBCT images to achieve precise segmentation and three-dimensional reconstruction.