Deep Inverse Optimization with Supervised Lookahead Learning for Real-Time Proton Treatment Plan Optimization
Abstract
Purpose
Intensity-modulated proton therapy (IMPT) plan optimization is time-intensive due to its high dimensionality and the inherent non-convexity of clinical dosimetric constraints. Conventional algorithms like projected gradient descent (PGD) often require extensive iterative tuning of beam parameters to fulfill clinical constraints. To meet the critical clinical need of instantaneous treatment planning, this study introduces a deep inverse optimization (DIO) framework that surrogates traditional gradient-descent optimizer, achieving optimal plans within substantially fewer iterations.
Methods
A DIO optimizer was developed with supervised lookahead training strategy. Instead of imitating single-step updates, the network predicts the optimization state with N steps beyond based on current beamlet intensities and gradient information. For generalizability across patients, a shared-weight 1D ResNet is introduced to efficiently capture global optimization trends. A cohort of 25 consecutive liver tumor patients previously treated with IMPT was replanned using this method.
Results
The DIO method demonstrates substantial acceleration, requiring 72.65% fewer iterations to meet identical convergence criterion. While PGD takes 217 ± 25.5 s to converge, the DIO framework generates high-quality plans in 1.1 ± 0.3 s. This corresponds to an ~197× speedup, reducing planning time to sub-second level. Across the testing cohort (n=7), mean liver–GTV dose of 10.5 Gy and 11.0 Gy, chest wall D2cc of 38.5 Gy and 40.0 Gy, CTV Dmax of 106% (of prescription) and 108%, were reported for DIO and PGD, respectively. All OAR and target comparisons showed no significant differences, indicating that DIO maintained plan quality comparable to PGD with a dramatically reduced computational cost.
Conclusion
We introduced a DIO framework that markedly accelerates IMPT planning by combining physical feedback with look-ahead neural acceleration. This enables near-instantaneous generation of high-quality plans, supporting on-couch re-optimization to accommodate daily anatomy changes and setup errors. Overall, the approach moves online adaptive proton therapy closer to clinically practical real-time workflows.