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

• DOI: 10.1016/j.phro.2025.100744
• Publication Date: 2025-03-07T17:16:19Z
• Authors: Yvonne J.M. de Hond; Paul M.A. van Haaren; An-Sofie E. Verri...; Rob H.N. Tijssen; Coen W. Hurkmans
• Keywords: Cone (formal languages); Cone-Beam CT

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 generate synthetic CBCTs (sCBCT) from planning CTs (pCT), which can result in anatomical mismatches and inaccurate labels. To prevent this issue, study assesses an model for female pelvic CBCT scans exclusively trained on delineated pCTs, were transformed into sCBCT using a physics-driven approach. replicate noise artefacts, (Ph-sCBCT) was synthesized pCT water-phantom scans. A 3D nn-UNet of cervical cancer Ph-sCBCT with contours. This included patients: 63 training, 16 validation 20 each testing Ph-sCBCTs clinical CBCTs. Auto-segmentations bladder, rectum target volume (CTV) evaluated Dice Similarity Coefficient (DSC) 95th percentile Hausdorff Distance (HD95). Initial evaluation occurred before generalizability The performed well generalized CBCTs, yielding median DSC's 0.96 0.94 the 0.88 0.81 rectum, 0.89 0.82 CTV CBCT, respectively. Median HD95's 5 mm 7 CBCT. demonstrates successful training images, necessarily delineating manually.