2102 - Improving Proton Dose Calculation Accuracy Using ROI-Focused AI Deep Learning Network for CBCT Based Adaptive Proton Therapy
Presenter(s)
Z. Qiao1, W. Yao2, Z. Zhang3, B. Zhang4, C. Beltran5, B. Y. Yi2, E. S. Paulson6, L. Ren7, R. Fang1, and M. Huang5,6; 1University of Florida, Gainesville, FL, 2Maryland Proton Treatment Center, Baltimore, MD, 3Department of Radiation Oncology, Duke University Medical Center, Durham, NC, 4Department of Radiation Oncology, University of Maryland School of Medicine, Baltimore, MD, 5Department of Radiation Oncology, Mayo Clinic, Jacksonville, FL, 6Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI, 7University of Maryland School of Medicine, Baltimore, MD
Purpose/Objective(s): Direct application of CBCT, with suboptimal image quality and range shifts, for adaptive proton therapy dose calculation continues to be challenging. Our study aims to develop a ROI based self-attention CycleGAN deep learning (DL) framework to improve CBCT image quality for proton dose calculation using AI. Our hypothesis is that the ROI-based self-attention DL network will improve CBCT image quality, maintaining daily anatomy, while improving dose calculation accuracy.
Materials/Methods: The DL framework included 23 head and neck cancer patients (7093 CBCT-QACT images) for training, and 5 patients (808 images) for testing. A tumor region of interest (ROI) based self-attention CycleGAN network, was designed specifically for AI-CBCT generation. Using the scanned CT quality as ground truth, three DL models were trained: (M1) CycleGAN (non-registration), (M2) Registration CycleGAN (registration between QACT and CBCT), (M3) ROI-based self-attention registration CycleGAN. The three levels of tumor CTV targets were selected as masks with self-attention rather than the entire images. After the DL network, image quality indexes including PSNR, SSIM and histogram were evaluated for the AI CBCTs. Proton dose calculation accuracy in D95 and D1 was also compared to the ground truth CT.
Results: All the DL models significantly improved CBCT image quality by the evaluation metrics, and resulting in images comparable to scanned CTs. M3: ROI-based and self-attention CycleGAN achieved the best image quality performance in SSIM, PSNR and Histogram ratio to CT. M2 performed better than CycleGAN M1. Proton therapy dose calculation accuracy improved with all three AI models compared to original daily CBCTs. Results were summarized in the table.
Conclusion: The ROI based self-attention CycleGAN achieved the best image enhancement and improved dose calculation accuracy in proton therapy, demonstrating the feasibility of AI-enhanced CBCT to replace the daily QACT scans or CT-on-rail for direct proton dose calculation.
Abstract 2102 - Table 1: Five testing patients' SSIM, PSNR, Histogram ratio (HU 0-800) and Proton dose calculation results based on 23 training patients through the DL networks are listed below. Scanned CT as ground truth. M1: Non-registration CycleGAN. M2: Registration CycleGAN. M3: ROI-based self-attention DL network. The best performers are in bold.| Evaluation Metrics (Mean± Std) | Original CBCT(CT as Ground Truth) | M1 AI CBCT | M2 AI CBCT | M3 AI CBCT |
| SSIM | (0.8± 0.1) | (0.8± 0.1) | (0.8± 0.0) | (0.9± 0.0) |
| PSNR | (27.3±2.0) | (27.7±2.5) | (26.9± 0.9) | (27.8± 0.8) |
| ROI SSIM | (0.7± 0.2) | (0.7± 0.2) | (0.7± 0.2) | (0.7± 0.2) |
| ROI PSNR | (26.9± 6.8) | (26.9± 6.9) | (27.7± 8.0) | (27.8± 6.2) |
| Histogram ratio (HU 0-800)(%): CBCT/CT | (19.1± 2.7)/(20.7± 3.8) | (22.2± 5.3) | (19.5± 4.8) | (20.6± 4.2) |
| Diff of D95 of CTV_high (cGy) | (-107.4± 96.4) | (-97.8± 105.4) | (-72.6± 44.0) | (-70.2± 51.8) |
| Diff of D1 of CTV_high (cGy) | (33.2± 40.1) | (24.3± 35.9) | (17.8± 20.8) | (20.4± 22.2) |