Auto-segmentation and dose accuracy evaluation of synthetic computed tomography generated from pelvis cone-beam computed tomography images using a cycle-generative adversarial network with a novel loss function.

BACKGROUND AND PURPOSE: Synthetic CT (sCT) images generated from cone-beam CT (CBCT) images have shown improved image quality over CBCT images. This work evaluated a deep learning (DL)-based sCT model with the potential to facilitate accurate adaptive radiotherapy for pelvis patients.

MATERIALS AND

METHODS

A DL-based cycle-Generative Adversarial Network (cycleGAN) model, incorporating a novel loss function to enhance structural image similarity, was used to generate sCT images from CBCT images acquired on a conventional c-arm Linac during intensity-modulated radiotherapy. Auto-segmentation accuracy for 10 pelvic structures was assessed compared to physician-drawn contours using Dice similarity coefficients (DSC), Hausdorff distance (HD), and mean distance-to-agreement (MDA). CT numbers for pelvic organs-of-interest were compared between sCT and planning CT (refCT). Dose, calculated using a Monte Carlo algorithm, was compared using relative dose differences to the gross tumor volume (GTV) for D98%, D50%, and D2% between sCT and recalculated refCT, alongside 3D-global gamma analysis with 3%/3mm, 2%/2mm, and 1%/2mm criteria.

RESULTS

The sCTs demonstrated high auto-segmentation accuracy with DSC > 0.8 and MDA < 2 mm, except for the small bowel and colon. CT number differences were within published tolerances to achieve less than 1% dose difference excepting structures with substantial gas variations compared to the refCT. The mean dose difference to the GTV was 0.3%±0.6 across D98%, D50%, and D2%. 3D-gamma passing rates were 97 ± 2% for the most restrictive criterion used, 1%/2mm.

CONCLUSIONS

The improved cycleGAN-based CBCT-to-sCT model generated high-quality sCT images that may be directly utilized for image segmentation and dose calculation to advance adaptive radiotherapy on conventional Linacs.