Speaker
Description
Objective: Robust optimization in particle therapy conventionally uses a voxel wise-worst case approach, where the plan is optimized to meet the clinical criteria even under the worst of a set of uncertainty scenarios. This entails significant computational overhead, most of which is only needed to compute the worst-case scenario. We hypothesize that optimizing only on the nominal case and a predicted last-iteraion worst-case scenario per voxel will provide equivalent results, at significantly reduced computational load.
Methods: 47 head&neck cancer patients were selected from the patient cohort treated in the GSI carbon ion therapy pilot project, and randomly split in training (35) and validation set (12). For each patient, full robust (FRO, 9 scenarios, 3% range and 3% setup uncertainties) and non-robust (NRO) carbon ion therapy treatment plans were optimized with the experimental treatment planning system TRiP98. For every optimization iteration, the voxel-wise worst case scenario was stored. As a proof-of-concept that single-scenario worst-case (SSRO) optimization can produce equivalent results, the scenario output of the last iteration of the FRO optimization was used as input for a robust optimization on the nominal and predetermined worst-case scenario. Robust evaluation was performed on 9 scenarios and considered typical DVH metrices. A Unet model and bespoke preprocessing workflow was developed with MONAI. The loss was a composite loss of class-weighted cross entropy and dice loss.
Results: The robust optimization based on the last-iteration worst case scenario performed clinically equivalent to the conventional robust optimization, at mean minimum D95% near 98%, and all cases above 96%. SSRO saved about 70% of the runtime per iteration, while it reduced the memory for the dose-influence matrix by a factor 4. Conversely, non-robust optimization performed considerably worse, at mean minimum D95% of 91.7%. The deep-learning model achieved promising performance on the validation data, with strong overall spatial agreement, particularly in OARs, between the true and predicted worst case scenario. Still, predictions over-homogenize regions of considerable scenario variation in the label, leaving room for further improvement.
Outlook & Discussion: Robust optimization based on a single scenario is a more efficient approach to robust optimization. With further fine-tuning of the deep-learning prediction, streamlined, memory efficient robust optimization becomes possible for carbon ion therapy. Future work will expand the number of robustness scenarios and test architectures tailored to locally highly variable class labels.