Author: Steve B. Jiang, Austen Matthew Maniscalco, Dan Nguyen, Chenyang Shen, Jiacheng Xie, Shunyu Yan, Ying Zhang, You Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab & Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center, The University of Texas at Dallas 🌍
Purpose: Although treatment planning systems (TPSs) can handle dose calculation and plan optimization automatically, planning for radiotherapy still requires extensive efforts and expertise from a multi-disciplinary team to obtain clinically acceptable plans. Suboptimal plans are sometimes unwillingly accepted given intense time constraint in clinical. To address this long-lasting bottleneck, we proposed RT-AutoTPS, a novel AI-driven TPS, for high-quality treatment planning in real time.
Methods: We considered VMAT for prostate cancer as the testbed of this study. RT-AutoTPS consists of two modules: real-time fluence map (FM) estimation module via domain transformation (TransFM), and real-time dose calculation module via geometry-encoding (GeoDose). Taking the targeted 3D dose and CT image as input, TransFM mimics the inverse transformation from dose to FM. As a global operator, this transformation is implicit and requires a huge amount of memory and therefore is deemed infeasible to directly model using deep neural networks. To mitigate this issue, we developed an innovative shuffling scheme using the local convolution operators along different dimensions of data to mathematically approximate the global transformation. For GeoDose, FMs were embedded into CT while preserving the treatment delivery geometry for 3D dose estimation. End-to-end training of both modules was conducted simultaneously in a unified framework, in which infinite training samples can be synthesized by randomly perturbing the FMs to improve the training performance. RT-AutoTPS was trained using 280 patients (252 for training and 28 for validation) and tested on an independent set of 20 patients.
Results: RT-AutoTPS can complete planning and dose calculation within 100ms. The final dose achieved a γ-passing rate (3%/2mm) of 97.9%±1.12%, with 0.97%±1.12% error in Dmean and 2.32%±2.25% in Dmax for treatment targets.
Conclusion: This study introduces RT-AutoTPS, which successfully tackled the challenging domain transformation problems in plan optimization and dose calculation, demonstrating great potential in real-time treatment planning.