RBMDC-Net: Effective Jaw Cyst Segmentation Network Using Residual Bottleneck and Multiscale Dilated Convolution

Automatic segmentation of jaw cysts in cone-beam computed tomography (CBCT) scans plays a crucial role in clinical and treatment planning, and it provides an efficient alternative to labor-intensive manual procedures. However, cysts on pathological sections often present with irregular and complex morphology, as well as extremely similar to surrounding tissue, which poses a great challenge in achieving their accurate segmentation. To overcome these limitations, this paper proposes a U-shaped image segmentation network based on residual bottleneck and multiscale dilated convolution, called RBMDC-Net. In our approach, the standard convolutional module in the U-Net encoding path is replaced with residual bottleneck module (RBM) that enables the network to extract features at various scales using multiple kernel sizes and residual connection. Next, we process each feature extracted from the encoder by passing it through a series of multiscale dilated convolution modules (MDCM), which capture more diverse and reliable features by increasing the acceptance field. In addition, we introduce the multi-level feature guidance module (MFGM) to address the issue of semantic dilution that often occurs during the progressive up-sampling. Finally, to evaluate the effectiveness of RBMDC-Net, we developed two image datasets: the first is the original images of jaw cyst collected by Quzhou People’s Hospital, and the second is its augmented dataset. These datasets encompass a wide range of jaw cyst characteristics, including size, shape, and location, as well as varying environmental conditions such as lighting and imaging quality. The experimental results demonstrate that our proposed RBMDC-Net significantly outperforms other comparable methods in the task of image segmentation. Specifically, when applied to the original jaw cyst dataset, RBMDC-Net achieved F1 of 93.13%, Mcc of 93.08% and Jaccard of 87.17%. These results represent improvements of 5.75%, 5.36%, and 9.31% over the performance of the widely used U-Net architecture. Furthermore, when tested on the augmented jaw cyst dataset, the RBMDC-Net continued to show superior performance with F1 of 91.92%, Mcc of 91.85% and Jaccard of 85.14%. These metrics were 3.03%, 2.84%, and 4.88% higher than the corresponding values obtained using U-Net. Besides, ablation experiments also indicate the effectiveness of residual bottleneck module, multiscale dilated convolution module and multi-level feature guidance module.