More Efficient Auto-Segmentation Framework Using Patient-Specific Information for CBCT-Guided Adaptive Radiotherapy

Abstract

Accurate and fast delineation of regions of interest (ROIs) is critical for online adaptive radiotherapy (ART). Due to the noise, artifacts and low contrast of soft tissue on cone-beam computed tomography (CBCT), the CBCT-guided ART (CBCTgART) needs an effective tool to assist the segmentation of ROIs. This study aims to develop an efficient auto-segmentation workflow with personalized modeling for CBCTgART based on deep learning (DL) model. Five hundred fifty-two patients with nasopharyngeal carcinoma were enrolled in this study. At the beginning, a cycle consistent adversarial network (CycleGAN) model was trained on 172 patients' CBCT/CT paired images to improve the image quality of CBCT to CT level. A generalized segmentation model was trained using the planning CT (pCT) and contour data from 530 patients. For personalized modeling, the generalized segmentation model was fine-tuned on the specific patient's pCT and contour to get the personalized model. When CBCT images were available, the trained CycleGAN model transformed the CBCT to synthetic pCT. Then the personalized auto-segmentation model generated the contour of ROIs on the synthetic pCT. We randomly selected 22 patients for model test. The proposed method (DL-personalized) was compared with other DL methods based on the same architecture of network: regular deep learning method (DL-regular), which was trained on the CBCT and corresponding contours, and generalized model in our framework (DL-generalized). So, 22 personalized, 1 generalized and 1 regular DL models were tested. The paired t-test was performed to test the significance for mean dice similarity coefficient (DSC), mean distance to agreement (MDA), and Hausdorff distance (HD) between the alternative and proposed methods. Two ROIs were included: the clinical target volume (CTV) and nasopharynx gross tumor volume (GTVnx). The proposed DL-personalized model achieved better results compared with others as shown in the table. The accuracy and robustness of our proposed framework was reliable. All of p values were under 0.01, which indicated the statistically significant difference. The proposed framework utilizing patient-specific information can improve the segmentation accuracy of ART. This method has potential to be integrated into the ART workflow to improve efficiency.