2280 - Advancing Personalized Radiotherapy Planning with AI-Empowered Real-Time Automatic Treatment Planning System
Presenter(s)
S. Yan1, J. Xie1, A. Maniscalco1, Y. Zhang1, D. Nguyen2, S. B. Jiang1, and C. Shen3; 1MAIA Laboratory, Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX, 2Medical Artificial Intelligence & Automation (MAIA) lab, University of Texas Southwestern Medical Center, Dallas, TX, 3Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX
Purpose/Objective(s): Treatment planning remains a time-consuming iterative process that often forces clinicians to accept suboptimal plans due to heavy time constraints in clinic, potentially compromising the quality and outcome of radiotherapy (RT). To address this long-standing bottleneck, we propose RT-AutoTPS, an AI-driven framework that enables real-time, high-quality treatment planning. This approach allows physicians to explore multiple plan options instantly, optimizing treatment personalization.
Materials/Methods: RT-AutoTPS integrates two deep learning-based modules: TransFM, which estimates fluence maps (FMs) based on CT and dose distribution in real-time; and GeoDose, which instantaneously calculates the dose distribution via geometry-preserving embedding FMs and CT images. Specifically, TransFM is designed to approximate the inverse transformation from dose to FMs, which is a complex global operation requiring extensive computational resources. Directly implementing this transformation with fully connected layers is impractical due to excessive memory demands. Instead, we developed an innovative shuffle scheme that sequentially applies convolution operators across different input dimensions, effectively approximating the global transformation with significantly lower memory requirements. GeoDose encodes FMs into the CT domain while preserving the exact treatment delivery geometry, enabling accurate 3D dose calculation. To enhance the training efficacy, we developed an end-to-end learning scheme to augment the performance of TransFM using GeoDose. Specifically, once GeoDose was fully trained, it generated FMs-dose pairs from randomly perturbed FMs, which were used to fine-tune TransFM and improve its performance. As a proof-of-principle study, we considered VMAT for prostate cancer as the testbed. For training, 280 patients (252 for training and 28 for validation) were used and another cohort of 20 patients was independently saved for testing.
Results: RT-AutoTPS can complete planning and dose calculation within 100ms for each testing case. The dose distributions calculated by GeoDose achieved a ?-passing rate (3%/2mm) of 99.1%±0.84%. For FMs estimation, the TransFM achieved a relative error of 2.82%±1.17% in predicted FMs. The mean and maximal dose error for treatment targets in the dose calculated using estimated FMs were 0.97%±1.12% and 2.32%±2.25%, respectively.
Conclusion: RT-AutoTPS eliminates the time-consuming iterative planning process. It holds great potential of enabling personalized RT treatment with limited resources.