Feasibility of a General-Purpose Deep Learning Dose Engine: A Multi-Site Validation Study
Abstract
Purpose
The primary objective of this study was to develop a generalized deep learning-based dose calculation engine capable of accurate, site-independent dose prediction. By utilizing a beamlet-based input strategy, we aimed to establish a computationally consistent and differentiable dose module that enables end-to-end training for autoplanning, maintaining dosimetric accuracy across diverse anatomical geometries.
Methods
A dataset of 3,600 step-and-shoot IMRT and 3D-CRT plans (6 MV) was generated from 120 patients evenly distributed across six anatomical sites. We investigated two 3D convolutional neural network architectures—a standard U-Net and a coarse-to-fine Cascade UNet—to predict 3D dose distributions directly from patient CT images and divergent MLC and jaw segment projections ("beamlets"). Models were trained using both Mean Squared Error (MSE) and Mean Absolute Error (MAE) loss functions. Performance was validated using 3D gamma analysis on an independent external cohort of 60 VMAT plans.
Results
The deep learning framework demonstrated high dosimetric accuracy across all tested configurations. The optimal model (U-Net trained with MAE loss) achieved a mean gamma passing rate of 98.9 ± 1.6% (3%/2mm, 10% threshold) on the independent test set, with the Cascade U-Net achieving a similarly high passing rate of 98.8 ± 1.6%. The model maintained robust performance across all six anatomical sites, with passing rates consistently exceeding 98%, demonstrating that the beamlet-based input strategy effectively generalizes to complex geometries without site-specific training.
Conclusion
We demonstrated that a single, site-independent deep learning model can calculate radiotherapy dose distributions with clinical accuracy. By effectively learning the relationship between patient anatomy, beam geometry, and dose distribution, this approach provides a computationally consistent and differentiable engine. This makes it highly suitable for integration into end-to-end automatic planning, as well as online ART and secondary dose verification workflows.