Robust and Accurate Mandible Segmentation on Dental CBCT Scans Affected by Metal Artifacts Using a Prior Shape Model

Abstract

Accurate mandible segmentation is significant in the field of maxillofacial surgery to guide clinical diagnosis and treatment and develop appropriate surgical plans. In particular, cone-beam computed tomography (CBCT) images with metal parts, such as those used in oral and maxillofacial surgery (OMFS), often have susceptibilities when metal artifacts are present such as weak and blurred boundaries caused by a high-attenuation material and a low radiation dose in image acquisition. To overcome this problem, this paper proposes a novel deep learning-based approach (SASeg) for automated mandible segmentation that perceives overall mandible anatomical knowledge. SASeg utilizes a prior shape feature extractor (PSFE) module based on a mean mandible shape, and recurrent connections maintain the continuity structure of the mandible. The effectiveness of the proposed network is substantiated on a dental CBCT dataset from orthodontic treatment containing 59 patients. The experiments show that the proposed SASeg can be easily used to improve the prediction accuracy in a dental CBCT dataset corrupted by metal artifacts. In addition, the experimental results on the PDDCA dataset demonstrate that, compared with the state-of-the-art mandible segmentation models, our proposed SASeg can achieve better segmentation performance.

Keywords: 3D virtual surgical planning (3D VSP); accurate mandible segmentation; convolutional neural network; oral and maxillofacial surgery.

Figure 1

Example illustrations that challenge mandible segmentation in CBCT images. (a) Original CBCT image. The mandible and teeth appear with almost invisible boundaries. (b) Example of manual annotation. Low contrast often appears in condyles. The purple region indicates the manual annotation of the mandible.

Figure 2

An example illustration that shows the comparison with SegUnet and RSegUnet. The existing state-of-the-art convolutional neural network (CNN) approaches for segmentation tasks perform poorly when the input data are strongly degenerated by noise. (a) the ground truth segmentation; (b–d) the automatic segmentation results obtained from SegUnet [27], RSegUnet [26] and the proposed SASeg. The purple region indicates the manual annotation, while the green regions indicate automatic segmentations.

Figure 3

Overview of the proposed SASeg and its corresponding unfolded computational graph. The PSFE module is leveraged to extract general shape features from a mean mandible shape, and recurrent SegUnet connections are used to conduct the slice-by-slice segmentation.

Figure 4

The procedure of generation of the mean mandible shape model. The generation consists of preparing the mandible model, generating the mesh surfaces, creating the mean mandible shape model, and calculating the corresponding mask stack.

Figure 5

Details of each unit of SASeg. Each unit of SASeg consists of the PSFE, an encoder and a decoder. The number of channels is indicated in the left of each convolution. The PSFE module is inserted at the last layer of the encoder to capture the mandible prior information.

Figure 6

Results with the fashionable segmentation approaches. The ground truth is shown in the column (a). Columns (b–f) show the mandible predictions generated by the U-Net, SegNet, SegUnet, AttUnet, and the proposed SASeg, respectively. The red rectangle indicates the zoom-in views of bad predictions.

Figure 7

Results with different ablations of our method. The ground truth is shown in column (a). Columns (b–e) illustrate the segmentation results derived from the model without RNN and PSFE modules, without the PSFE module, and without the RNN module and the proposed SASeg, respectively. The red rectangle indicates the zoom-in views of bad predictions. (w/o RNN and PSFE: without RNN and PSFE modules, w/o PSFE: without the PSFE module, and w/o RNN: without the RNN module.)

Figure 8

Results with different ablations of loss functions. The ground truth is shown in column (a). Columns (b–d) illustrate the segmentation results derived from the model without Dice loss, BCE loss and the proposed SASeg, respectively. The red rectangle indicates the zoom-in views of bad predictions. (w/o Dice: without Dice loss, w/o BCE: without BCE loss.)