Document Type
Conference Proceeding
Publication Date
6-2023
Publication Title
Medical Physics
Abstract
Purpose: To evaluate the feasibility of deep-learning and CBCT-based online adaptive radiotherapy(ART) in thorax, Head&Neck, and pelvic regions. Methods: The study utilized ADMIRE(Advanced Medical Image Registration Engine), a research software designed to enable fast online ART through: (1)synthetic CT(sCT) generation from CBCT, (2)automatic segmentation of critical organs(OARs), and(3)deformable image registration(DIR) to transfer the target volume. ADMIRE trained a modified Cycle-Generative-Adversarial-Networks with Structural Similarity Index Map(SSIM) loss term to generate sCT, while the auto-segmentation model was based on a 3D-ResUnet with modification. Site-specific models were trained for three anatomic sites: thorax, pelvic, and head&neck, and for both networks. The study included six randomly selected patients, two for each site. The ART workflow began with sCT generation, followed by auto-contouring of relevant OARs on the sCT. The tumor volume was deformably transferred to the sCT via DIR. The RT- plan was calculated on the sCT to evaluate treatment dose. The time costs of sCT-generation, auto-contouring, and dose calculation were recorded. The quality of sCT was assessed by comparing CT#difference maps and CT#line profiles to the real-CT. VMAT plans were calculated on both real-CT and sCT(GPU collapsed cone), and dosimetric accuracy was evaluated by DVH and dose distribution difference. Results: Each of the three sub-functions took less than 15seconds to complete. In under one minute, the sCT with contours of treatment day was ready. The sCT showed good agreement with real-CT by CT number. The time of dose calculation was under one minute, and the dose difference between sCT and real-CT is minimal. The auto-segmented contours of the targets and OARs were clinically acceptable with minor editing. Conclusion: The preliminary results indicate the quality of synthetic CT and auto contours generated by ADMIRE is promising for online ART applications. The whole ART workflow can be completed in two minutes for online treatment dose evaluation.
Volume
50
Issue
6
First Page
e210
Recommended Citation
Qin A, Xu J, O'Connell N, Ding X, Chen S, Thill D. The clinical usability, efficiency, and accuracy of combing deep-learning-based synthetic CT and auto-contouring for daily CBCT-based online adaptive radiation therapy. Med Phys. 2023 Jun;50(6):e210. doi:10.1002/mp.16525.
Comments
American Association of Physicists in Medicine 65th Annual Meeting & Exhibition, July 23-27, 2023, Houston, TX