Mandibular Canal Segmentation From CBCT Image Using 3D Convolutional Neural Network With scSE Attention

Abstract

In dental implant planning, the mandibular canal is an important reference for determining the safe position of the implant. Accurate and automatic segmentation of the mandibular canal from CBCT image is of great significance. However, the variable curvature of the mandibular canal and the blurred borders make the process challenging. At present, the segmentation of mandibular canal is usually carried out by experienced, doctors using manual or semi-automatic methods, which are time-consuming and have poor segmentation consistency. For these issues, this paper proposes a mandibular canal segmentation method based on 3D convolutional neural network. Firstly, a mandibular canal segmentation method based on center line combined with region growth and a mandibular canal clipping method based on fixed points are proposed to improve the efficiency of mandibular canal marking and reduce the imbalance between sample categories, respectively. Thereafter, the scSE attention module was added to the network for reducing irrelevant background information. Finally, an weighted BCE loss function was used to prevent the mental foramen and mandibular foramen areas from participating in the back-propagation calculation, and the network is more focused on the feature learning of the mandibular canal. The experimental results show that the proposed segmentation method achieves good segmentation results, with a Dice score of 85.9% and a 95% Hausdorff distance of 0.5371mm. Compared with other methods, the segmentation method proposed in this paper had a higher accuracy and a faster efficiency, which is expected to play a practical role in dental implants.INDEX TERMS Deep learning, image processing, intelligent segmentation, oral implantation.

Deep Learning for Mandibular Canal Segmentation in Digital Dental Radiographs: A Systematic Literature Review

“…This volume is generated using the α-shape algorithm. On the other hand, Du et al [51] adopted a different approach, employing regional growth and center point techniques in their annotation process.…”

“…The scSE attention module enhances the network's capacity to extract edge characteristics by reducing the presence of irrelevant background information, resulting in more precise segmentation outcomes. The scSE attention module is included in the Attention 3D U-Net network to augment its capability to segment the MC [51]. It enhances the effectiveness of feature learning for MC segmentation by reducing irrelevant background information.…”

“…Du et al [51] 3D CBCT This research presents a new segmentation methodology that combines centerline and regional growth algorithms for MC annotation. The article introduces the Attention-3D U-Net network.…”

Automatic segmentation of mandibular canal using transformer based neural networks

“…To counteract the issue of blurred boundaries, Faradhilla et al (2021) introduced a Double Auxiliary Loss (DAL) in the loss function to make the network more attentive to the target area and its boundaries, achieving a high Dice accuracy of 0.914 on their private dataset. To combat class imbalance, Du et al (2022) innovatively introduced a pre-processing step involving centerline extraction and region growing to identify the mandibular canal’s location. They used a fixed point as a reference to crop a localized region around the mandibular canal, thereby minimizing the impact of background samples.…”

Dual-Stage Deeply Supervised Attention-Based Convolutional Neural Networks for Mandibular Canal Segmentation in CBCT Scans

“…Alternatively, Du et al [29] proposed another framework based on 3D Convolutional Neural Networks (CNNs) trained using the dataset developed by Capriano et al [27]. In contrast to Capriano et al [27], they first generated the annotations by employing the centerline combined with regional growth method.…”