Toward Universal Dose Prediction: A Site-Agnostic Deep Learning Framework with Strategic Goal Array Integration for IMRT and VMAT Plans
Abstract
Purpose
Deep learning (DL)-based dose prediction has become an important component of modern radiotherapy treatment planning. However, most existing approaches depend on site-specific models, necessitating separate training for each anatomical site, which limits scalability. The purpose of this study is to develop and evaluate a universal, site-agnostic DL model capable of accurately predicting 3D dose distributions across multiple treatment sites within a single unified framework.
Methods
A retrospective dataset of 1,559 patients treated with IMRT or VMAT was assembled across eleven anatomical sites, including brain, head and neck, prostate, uterus, pelvis, liver, kidney, bladder, pancreas, lung, and breast. Patients were randomly assigned to training (80%), validation (10%), and testing (10%) cohorts. Separate site-specific dose prediction models were trained independently for each anatomical site. In parallel, a universal model was trained using pooled multi-site data to predict dose distributions across all treatment sites. Both the site-specific and universal models employed a U-Net–based deep convolutional neural network architecture. Model inputs consisted of planning CT images, the planning target volume (PTV), organ-at-risk (OAR) information, a dosimetric goal array, and a PTV distance array, with the predicted dose distribution as the model output.
Results
The universal model achieved consistent dose prediction performance across all disease sites, with dose distributions comparable to site-specific models. Dose–volume histogram (DVH) analysis demonstrated agreement in PTV coverage and OAR sparing. Quantitatively, the universal model achieved a mean absolute percentage error (MAPE) of 3.29% for PTV dose prediction across the test cohort, whereas the site-specific models achieved 3.71%.
Conclusion
This work demonstrates that incorporating the goal array into a DL–based dose prediction framework enables accurate, site-agnostic dose prediction using a single model. This design eliminates the need to retrain multiple site-dependent models, thereby improving scalability and offering a more efficient pathway toward streamlined radiotherapy planning workflows.