Author: Daniel H. Bushe, Arthur Lalonde, Hoyeon Lee, Harald Paganetti, Brian Winey 👨🔬
Affiliation: Universite de Montreal, Massachusetts General Hospital, Massachusetts General Hospital and Harvard Medical School, University of Hong Kong 🌍
Purpose: Improving the precision and fidelity of daily volumetric imaging is essential for enabling adaptive proton therapy (APT). While cone-beam CT (CBCT) provides daily volumetric imaging, their utility for APT is hindered by inherent image inaccuracies, including scatter artifacts. This study aims to evaluate the effectiveness of a dual projection- and image-domain scatter correction technique utilizing sequential neural networks to enable accurate dose calculations in APT.
Methods: Monte Carlo simulations were utilized to generate a dataset comprising of paired scatter-free and scatter-contaminated, or raw, CBCT projections obtained from 28 patients. A U-Net network was trained in the projection-domain to estimate the normalized scatter signal from raw CBCT projections. To enhance the model’s accuracy and address image-domain features, the corrected projections from 20 patients were reconstructed, and two additional neural networks, including a U-Net and Flip Swin Transformer U-shaped (FSTU) network, were trained in the image-domain. The quantitative accuracy of the dual-domain corrected images was benchmarked against stand-alone projection- and image-domain networks. The clinical applicability of the developed model was evaluated by comparing proton dose distributions.
Results: For all evaluated metrics, the dual-domain FSTU corrected CBCT images demonstrated superior quantitative and structural accuracy, reporting mean error, mean absolute error, root mean squared error, and structural similarity index metric values of 9.10±1.1, 27.4±1.2, 44.9±3.8 HU and 0.996±0.003, respectively. The dual-domain U-Net correction outperformed the single, projection-domain U-Net corrections across all metrics. The dose distributions computed on the dual-domain FSTU images had the highest agreement with the baseline scatter-free image dose distributions with gamma pass rates of 99.76%, 100% and 100% for the 1%/1 mm, 2%/2 mm and 3%/3 mm criteria, respectively.
Conclusion: This work introduces an innovative approach to generate robust daily volumetric imaging through the development of a novel dual-domain network and evaluates the networks utility for online APT dose calculations.