Author: Ming Chao, Thomas Chum, Tenzin Kunkyab, Yang Lei, Tian Liu, Richard G Stock, Hasan Wazir, Junyi Xia, Jiahan Zhang 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai 🌍
Purpose:
This study aims to develop effective strategies for multi-organ segmentation of pelvic cone-beam computed tomography (CBCT) images based on transformer models to facilitate adaptive radiation therapy for prostate cancer.
Methods:
We investigated two advanced transformer architectures, UNETR and Swin-UNET, within a U-Net-style segmentation framework. Each was evaluated under three distinct data scenarios: (1) CBCT images only, (2) CBCT plus planning CT (pCT), and (3) a cross-modality fusion model using CBCT and pCT as multi-channel inputs. A total of 53 CBCT sets from 15 prostate cancer patients, along with their corresponding pCT images, were acquired. Manual contours for four organs at risk (OARs) - bladder, prostate, rectum, and femur bones - were delineated on each CBCT. The dataset was split in an 80:20 ratio for training and testing, ensuring that scans from any single patient were exclusive to either set. Model performance was assessed using the Dice Similarity Coefficient (DSC) and the 95% Hausdorff Distance (HD95).
Results:
The Swin-UNETR fusion model achieved the highest overall DSC of 0.86, followed by the Swin-UNETR CBCT-only model with a DSC of 0.85. The best performing model, Swin-UNETR fusion model, achieved an organ-specific DSC scores: 0.94 for bladder, 0.80 for rectum, 0.75 for prostate and 0.96 for femur bone. Furthermore, the fusion model (Swin-UNETR) achieved the following HD95 metrics for organ-specific segmentation: bladder 4.71, rectum 62.6, prostate 7.25, and femur heads 1.47.
Conclusion:
Preliminary findings indicates that the Swin-UNETR model is capable of generating contours with clinically acceptable accuracy. When integrated with planning CT, this model outperformed other model-CT-data scenarios. Further validation on a larger cohort is warranted, as robust multi-organ CBCT segmentation could substantially streamline adaptive radiotherapy workflows for prostate cancer.