Evaluation of Deep-Learning Based Sct Derived from CBCT Images with Several Network Dimensionalities for Clinical Use In Online Adaptive Radiotherapy.
Abstract
Purpose
Synthetic CT (sCT) derived from CBCT images show potential to facilitate online adaptive radiotherapy (oART) by enhancing CBCT image quality. Compared to sCT from a 2D deep-learning (DL) model that can result in through-plane streak artifacts, we evaluate a 2.5D model that incorporates neighboring slice information into sCT generation to enhance anatomical accuracy for oART workflow suitability.
Methods
Two cycle-generative adversarial networks (cycleGANs) were trained with modified loss functions on 300 sets of pelvis images to generate sCT from CBCT with 2D and 2.5D dimensionalities. CBCT from 20 independent pelvis patients were used to compare the models. Each sCT and corresponding planning CT (refCT) was contoured with commercial auto-segmentation software, and auto-contours were compared to physician-edited contours, including clinical contours for refCT, using dice similarity coefficient (DSC) and mean distance-to-agreement (MDA). Mean HU values for contoured structures were assessed for accuracy to refCT mean HU values, and sCT dose calculation accuracy was assessed using 3D gamma analysis at 3%/2mm and 1%/2mm criteria. sCT performances were compared using t-tests.
Results
Both models showed excellent overall auto-segmentation accuracy, with DSC>0.8 and MDA<2 mm for most structures. sCT images from the 2.5D model showed fewer through-plane artifacts and less blurring than the 2D model, suggesting improved anatomical accuracy resulting in fewer large segmentation failures. Mean absolute HU deviations between refCT and 2D/2.5D sCT were similar (p=0.25): 32 and 28 HU for bony anatomy and 9 and 11 HU for soft tissue, respectively. Gamma pass rates were statistically identical, with mean pass rates of 99.5±0.5% for 3%/2mm criteria (p=0.74) and 97±2% for 1%/2mm criteria (p=0.77).
Conclusion
The 2.5D DL model produced sCT images with fewer streak artifacts than the 2D model. Both sCTs showed similar quantitative performance relative to refCTs. These findings support the use of 2.5D sCT images in clinical oART workflows.