Physics-based data augmentation for improved training of cone-beam computed tomography auto-segmentation of the female pelvis

Background and Purpose: Labeling cone-beam computed tomography (CBCT) images is challenging due to poor image quality. Training auto-segmentation models without labelled data often involves deep-learning to generate synthetic CBCTs (sCBCT) from planning CTs (pCT), which can result in anatomical mismatches and inaccurate labels. To prevent this issue, this study assesses an auto-segmentation model for female pelvic CBCT scans exclusively trained on delineated pCTs, which were transformed into sCBCT using a physics-driven approach. Materials and Methods: To replicate CBCT noise and artefacts, a physics-driven sCBCT (Ph-sCBCT) was synthesized from pCT images using water-phantom CBCT scans. A 3D nn-UNet model was trained for auto-segmentation of cervical cancer CBCTs using Ph-sCBCT images with pCT contours. This study included female pelvic patients: 63 for training, 16 for validation and 20 each for testing on Ph-sCBCTs and clinical CBCTs. Auto-segmentations of bladder, rectum and clinical target volume (CTV) were evaluated using Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95). Initial evaluation occurred on Ph-sCBCTs before testing generalizability on clinical CBCTs. Results: The model auto-segmentation performed well on Ph-sCBCT images and generalized well to clinical CBCTs, yielding median DSC's of 0.96 and 0.94 for the bladder, 0.88 and 0.81 for the rectum, and 0.89 and 0.82 for the CTV on Ph-sCBCT and clinical CBCT, respectively. Median HD95′s for the CTV were 5 mm on Ph-sCBCT and 7 mm on clinical CBCT. Conclusions: This study demonstrates the successful training of auto-segmentation model for female pelvic CBCT images, without necessarily delineating CBCTs manually.