Fast Generation of Deliverable Proton Plans Via Deep Unrolling Networks
Abstract
Purpose
Proton PBS treatment planning relies on iterative numerical optimization and is time-consuming. Deep learning enables rapid inference and therefore has the potential for fast prediction of beamlet intensities from anatomical imaging inputs (e.g., CT voxels and structure masks). However, the lack of explicit physics information creates a large gap between voxel-space representations and beamlet-space treatment plans, which in practice has limited most deep learning approaches to dose prediction. To bridge this gap to inverse planning while leveraging fast deep learning inference, this study develops a learning-based unrolling strategy that reformulates the optimization process as learned iterations, preserves the original physics-based formulation, and substantially reduces computations.
Methods
A deep unrolling framework was developed by embedding trainable components into conventional optimization and reformulating it as a fixed number of learned neural network layers, replacing hundreds of iterations. This approach preserves the physics-based formulation and enables fast, direct prediction of deliverable beamlet intensities. The deep unrolling network was trained and evaluated on a prostate cancer proton PBS cohort with two beam angles (90° and 270°), including 159 patients (127 training, 16 validation, and 16 testing). Conventional optimization served as the reference, and a UNet-based dose prediction model was used as a deep learning baseline for dose quality comparison.
Results
The proposed method reduced the average planning time from 33.18 s to 3.39 s while achieving plan quality comparable to conventional numerical optimization. Compared with the UNet-based dose prediction approach, the proposed framework yielded lower voxel-wise mean absolute error (MAE), higher conformity index (CI) for target coverage, and improved performance across most clinically relevant dose-volume metrics.
Conclusion
This deep unrolling–based inverse planning strategy enables rapid generation of physically deliverable proton treatment plans while maintaining clinically acceptable plan quality, indicating its potential to enhance the efficiency and robustness of proton therapy planning workflows.