ROI-Guided Swin-Transformer Deep Learning Network for Improved Proton Dose Calculation In CBCT-Based Adaptive Proton Therapy
Abstract
Purpose
Daily cone-beam CT (CBCT) is widely used in adaptive proton therapy; however, scatter artifacts can degrade image quality and introduce proton dose calculation inaccuracies. We developed a region-of-interest (ROI)–guided Swin-Transformer deep learning (DL) network to improve CBCT image quality and, consequently, proton dose calculation accuracy.
Methods
Twenty-two head-and-neck cancer patients were included in this IRB-approved study. Data from 12 patients (7,093 CBCT–QACT image pairs) were used for training, and data from 10 patients (2,189 image pairs) were reserved for testing. The clinical target volume (CTV) mask was defined as the ROI to guide network attention. A Cycle Swin-Transformer GAN (CSTGAN) was designed for AI-based CBCT generation, incorporating ROI-based tumor-region attention. Planning-quality CT (QACT) images were used as ground truth. The CSTGAN employed adversarial and cycle-consistency objectives, with prioritized learning within the target region. An original CycleGAN model (M1) served as the baseline, while four additional models (M2–M5) incorporated incremental improvements using ROI-guided attention and Swin-Transformer architectures. Image quality was evaluated using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Proton dose calculation accuracy was assessed by comparing dose–volume histogram (DVH) metrics against QACT-based plans.
Results
The ROI-guided CSTGAN achieved a global SSIM of 0.877 ± 0.048 and a PSNR of 28.281 ± 7.614. Compared with uncorrected CBCT, target coverage error (ΔD95, referenced to QACT) was reduced in magnitude from −124.67 ± 114.84 cGy to −56.00 ± 42.24 cGy. Hotspot error (ΔD1) decreased from 45.89 ± 53.38 cGy to 19.71 ± 31.16 cGy based on DVH analysis. The CSTGAN-enhanced AI-CT demonstrated reduced interpatient variability and improved proton dose calculation accuracy relative to uncorrected CBCT, with closer agreement to QACT than CycleGAN.
Conclusion
ROI-focused CSTGAN significantly improves CBCT image quality and yields proton dose calculations closer to QACT, supporting improved accuracy in adaptive proton therapy, comparing to the traditional CybleGAN.