Author: Christopher Ackerman, Chang Chang, Yan-Cheng Huang, Robert Kaderka, Che Lin, Hsin-Chih Lo, Iain MacEwan, Yi-Chin Tu, James Urbanic 👨🔬
Affiliation: University of California San DIego, Taiwan AI Labs, National Taiwan University, California Protons Cancer Therapy Center, University of Miami, Sylvester Comprehensive Cancer Center 🌍
Purpose: To investigate the performance of an existing AI beam angle prediction model on external patient datasets for liver proton treatments. The AI model was trained on datasets exclusively from one proton center, and tested on datasets from two other proton centers.
Methods: The AI prediction model was trained on liver planning data. The test dataset contains 28 liver proton PBS cases, 19 from center A and 9 from center B. Two plans are generated for each case: one using original human chosen beam angles and the other AI predicted beam angles. Both plans are optimized using an identical RapidPlanPT model to ensure the only variability is the human versus AI beam angles. AI beam angles were selected using a probability distribution produced by the model, which was trained to choose angles with a balance of short path length and OAR sparing. Target coverage and critical OAR constraints were compared between plans.
Results: Dosimetric statistics for Human and AI plans were similar. In most cases both plans were able to achieve good CTV coverage while also maintaining comparable OAR sparing. Some cases showed different OAR preferences between Human and AI plans and different preferences for cases from Center A versus Center B but with similar target coverages.
Conclusion: The AI selected beam angles achieved comparable dosimetric performance to the human selected angles. In some cases the AI selected angles were able to achieve superior OAR sparing compared to human selected angles, while the opposite was true in other cases, potentially reflecting institutional-level preferences for these OARs. It’s likely that the AI model is influenced by the center that provided training data, and these results may carry preferences of that center versus the outside centers’ cases evaluated here. A larger population of outside data is needed to scrutinize the model further.