Author: Mingli Chen, Xuejun Gu, Hao Jiang, Mahdieh Kazemimoghadam, Weiguo Lu, Qingying Wang, Zi Yang, Kangning Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Department of Radiation Oncology, Stanford University School of Medicine 🌍
Purpose: Dose prediction (DP) is essential in guiding radiotherapy planning. However, current DP models for intensity-modulated radiation therapy (IMRT) primarily rely on fixed-beam orientations and are evaluated on specific treatment sites, limiting broader clinical applicability. We proposed a deep learning-based universal DP model (UniDose) designed to accommodate various treatment sites and diverse treatment modalities, particularly IMRT with variable beam configurations.
Methods: The UniDose model, built on nnU-Net framework with Huber loss, functions as an image-to-image network for 3D DP. It contains 3 input channels: a normalized prescribed dose image, a weighted avoidance image, and a beam trace image, represented as a 3D-matrix of non-modulated beam’s eye view ray-tracing dose distribution normalized by beam count. We used a large dataset of 871 patients, collected from our institutional radiotherapy patient database, across 25 treatment sites, with prostate, liver, and brain comprising over 50%. Beam numbers ranged from 7 to 25, and the dataset was divided into 586 training, 147 validation, and 138 testing cases. To investigate feasibility, the predicted doses were used as references to generate corresponding physically feasible plans (opt-plans) with an in-house optimization engine and machine parameters. The 3%/3mm gamma passing rate (GPR) was calculated to evaluate agreements between predictions, opt-plans, and clinical plans.
Results: An average 92.36 % GPR and strong DVH consistency between predictions and opt-plans confirm the reliability and approachability of UniDose predictions. Although average GPR between predictions and clinical plans is 86.13%, most predictions and opt-plans demonstrated improved OAR sparing with comparable target coverage, highlighting UniDose's potential for higher-quality plan generation. Additionally, by adjusting weight assignments in the avoidance input channel, UniDose can effectively generate a patient-specific trade-offs between OARs and targets.
Conclusion: We developed a general and approachable DP model to guide treatment planning for various sites with arbitrary beam configurations, while effectively managing trade-offs.