325 - Universal Deep Learning Dose Prediction for IMRT Planning

Purpose/Objective(s): Dose prediction (DP) is crucial for AI-driven radiotherapy planning. However, existing DP models are primarily limited to simple scenarios, such as consistent beam configurations and specific treatment sites, restricting clinical applicability. We propose UniDose, a deep learning-based universal DP model for diverse treatment sites and modalities, with a focus on IMRT with arbitrary beam configurations.

Materials/Methods: UniDose model, built on the nnU-Net framework with Huber loss, serves as an image-to-image mapping network tailored for 3D DP. The input includes 3 channels: a normalized target prescription dose image, a weighted avoidance mask image, and a beam trace image, represented by non-modulated beam’s eye view traces with its intensity normalized by beam count. We conducted experiments on a large dataset of 871 patients, collected from our institutional radiotherapy patient database, covering 25 treatment sites, with prostate, liver and brain comprising over 50%. The number of beams varied from 7 to 25 with arbitrary orientations. The dataset was divided into 586 training, 147 validation, and 138 testing cases. To assess the realism of UniDose predictions, we converted them into physically feasible plans (Opt Plans) constrained by machine limitations, using an in-house optimization engine. The 3%/3mm gamma passing rate (GPR) was calculated to evaluate agreements between predictions, Opt Plans, and clinical plans. We also investigated the impact of different weight assignments in the avoidance input channel on dose prediction.

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, as shown in Table 1, highlighting UniDose's potential for higher quality plan generation. Additionally, by adjusting weight assignments in the avoidance input channel, UniDose can effectively tailor patient-specific trade-offs between OAR sparing and target coverage.

Conclusion: We developed a general DP model for practical heterogeneous clinical treatment scenarios with arbitrary beam configurations, while enabling quick and easy user interaction through adjustable input conditions.