ESTRO 2024 - Abstract Book

S3899

Physics - Image acquisition and processing

ESTRO 2024

be evaluated throughout the course of treatment. However, this cannot consistently be done directly on in-room cone-beam CT (CBCT), as the CT number accuracy often is insufficient for proton dose calculations. Deep learning has been shown to be superior to the traditional CBCT correction methods. One of the most explored deep learning networks is the Cycle-consistent Generative Adversarial Network (CycleGAN), which enforces anatomical preservation through a pixel-wise similarity function. Recently, the Contrastive Unpaired Translation (CUT) network has been proposed. This network instead learns a flexible and effective matching of the CBCT and CT image distributions through contrastive learning. A joint network (Cycle-consistent CUT; CycleCUT) could potentially combine the benefits of the CycleGAN and CUT networks, capturing both the local features of the specific CBCT image (through the pixel wise similarity function) and the global features of the image distributions (through contrastive learning). The aim of this study was to evaluate the quality of synthetic CTs (sCTs) generated from CBCT using a 3D CycleCUT network, with the purpose of enabling CBCT-based daily proton dose calculations. The evaluations in this study serve as the first step towards a prospective clinical implementation of the sCT generation method. A total of 94 head-and-neck cancer patients treated with proton therapy were used to train the CycleCUT network, validate the training process, and ultimately evaluate its performance on unseen data in a 77/5/12 train/validation/test split. All patients were enrolled in a randomized trial, and each patient had daily CBCTs (33-50 scans per patient) and weekly repeat CTs (reCTs; 5-6 scans per patient), resulting in 467/32 CT scans and 2781/176 CBCT scans for training/validation, respectively. For patients in the test set, a single CBCT was chosen with a criterion of being anatomically similar to a same-day reCT (verified by visual inspection). The reCT was deformably registered to the corresponding CBCT through deformable image registration in Velocity® (Varian Medical Systems, Palo Alto, CA) to create a ground-truth reCT (gt-reCT) to compare to the sCTs. The sCTs were rigidly registered and stitched onto the planning CT (pCT) to obtain a full field-of-view (FOV) for dose calculation (no beams had entrance through the stitched areas). The structure set was deformably propagated from the pCT to the sCT and gt-reCT, and the clinical proton plan was then re-calculated on the sCT and gt-reCT in Eclipse (Varian Medical Systems, Palo Alto, CA, USA). For the evaluation, the proton dose distributions on the sCT and gt-reCT were compared in terms of target coverage and organ-at-risk (OAR) doses through dose-volume-histogram (DVH) parameters. The evaluated DVH parameters included the OAR mean doses used in clinical normal tissue complication probability models for dysphagia and xerostomia. Wilcoxon signed rank test was used to determine if the differences between DVH parameters were statistically significant (p ≤ 0.05). Additionally, the average time required for creating full-FOV sCT (including CBCT to sCT conversion, stitching, and creating dicom image) was calculated to evaluate the feasibility of the sCT generation method in a clinical offline adaptive workflow. Material/Methods:

Results:

Visually the sCTs generated using the CycleCUT network had a greatly reduced image noise compared to the CBCTs, with the image quality of the sCTs being comparable to the gt-reCTs (Figure 1). No statistically significant dose differences were found between the sCT and gt-reCT for any of the DVH parameters according to the Wilcoxon signed rank test. The largest DVH difference for the targets was found for the D 99% of the intermediate-risk target (CTV2), where the median of differences was 0.5 percentage point (pp; interquartile range: 1.7 pp). The largest difference in the DVH parameters for the OAR was found for D mean of the combined submandibular glands, (submandibular tot in Figure 2) where the median of differences was 0.6 Gy (interquartile range: 1.4 Gy). The average time needed to create a full-FOV sCT image in dicom format was 1.7 minutes.