3034 - Evaluation of a Novel 2.5D Deep Learning Synthetic CT model Derived from CBCT Images for Head and Neck Adaptive Radiotherapy

Purpose/Objective(s): Adaptive radiotherapy (ART) workflows on conventional linacs are challenged by poor onboard CBCT image quality, often resulting in re-simulation of patients or requiring purchase of modern, dedicated adaptive platforms. We evaluate the performance of a novel 2.5 dimensionality (2.5D) deep learning (DL) synthetic CT (sCT) solution to enable ART in head and neck (HN) patients using onboard CBCT images on existing linacs.

Materials/Methods: A novel cycle generative adversarial network sCT model was developed (185 HN training sets). The model employs a 2.5D to achieve smoothness and anatomical continuity by incorporating neighboring slice information during inference. Seven independent CBCT datasets were used for evaluation by comparing the sCT to the planning CT (pCT) images. Commercial DL auto-segmentation (DLAS) software was used to delineate 14 organs-at-risk (OARs)and compared to physician contours. Mean distance to agreement (MDA) and dice similarity coefficient (DSC) assessed contour acceptability and adjustment needs, with variation evaluated using Welch’s test. The CT number (HU) mean absolute error (MAE) between the pCT and sCT OARs were computed for quantitative accuracy. The treatment dose was recalculated on the sCT and compared to the pCT dose using 3D gamma analysis at 3%/2mm and 1%/2mm criteria.

Results: sCT DLAS results demonstrated similar accuracy to reference contours for 13 of 14 structures (DSC and MDA, p>0.05), suggesting that sCT DLAS contours required similar adjustment to pCT DLAS contours to attain clinical suitability. MAE for soft tissue OARs were within 21 HU of the corresponding pCT structures. Bony structures showed larger HU deviations but were within acceptable tolerances (<100). sCT calculated doses showed excellent agreement with the pCT, with gamma pass rates of 99.6±0.4% and 97.4±1.4% for 3%/2mm and 1%/2mm criteria, respectively.

Conclusion: Generation of 2.5D sCT images can overcome current CBCT image quality challenges, facilitating online adaptive ART in HN patients on existing linac systems.