Abstract
BACKGROUND: Dose prediction has great potential in improving plan quality and efficiency by estimating optimal dose distribution. However, most existing deep learning (DL) based dose prediction models for intensity-modulated radiation therapy (IMRT) have been primarily developed under simplified conditions, such as fixed beam configuration and/or disease site. These constraints limit the generalizability and clinical usability of such models across the diverse scenarios encountered in real-world practice. PURPOSE: We proposed a DL-based universal dose prediction model, named UniDose, designed to accommodate a wide range of disease sites and support diverse clinical scenarios, especially for IMRT treatment plans with arbitrary beam configurations. METHODS: UniDose is built on a customized nnU-Net framework, adapted into an image-to-image mapping network tailored for 3D dose prediction and trained using the Huber loss. The network takes three generalized input channels: a normalized prescription dose map that encodes planning goals for the target, a weighted avoidance mask that consolidates multiple organs at risk (OARs) and body structures into a single channel with clinical relevance-based voxel weights, and a beam trace image that captures beam configuration using a non-modulated, cumulative dose approximation generated via a ray-tracing based algorithm. The model was trained, validated and tested on a heterogeneous dataset of 871 patients encompassing 25 disease sites and a wide spectrum of prescription doses and beam configurations. To assess the deliverability of the predicted dose, we incorporated a reference-based in-house optimization engine into the UniDose framework to generate feasible plans constrained by machine limitations. Model performance was evaluated by comparing predicted doses, optimized doses, and clinical plans using gamma passing rate (GPR) with a 3%/2 mm criteria and 10% lower dose threshold and dose-volume histogram (DVH) metrics. RESULTS: The UniDose predictions achieved an average GPR of 92.36% compared to the optimized doses and demonstrated strong DVH consistency. The average GPR between predicted and clinical doses was 86.13%. DVH comparisons showed that the predictions and the optimized dose achieved improved OAR sparing while maintaining comparable target coverage relative to clinical dose, particularly in prostate, liver, and brain cases. Case studies across six disease sites with variable beam configurations further confirmed that the predicted and optimized doses exhibited similar dose deposition patterns along beam paths, suggesting that the predicted dose is physically feasible and approachable following dose optimization. Additionally, adjusting voxel weights in the avoidance input channel enabled flexible trade-offs between OAR sparing and target coverage, supporting patient-specific treatment planning. CONCLUSIONS: UniDose demonstrates strong potential as a universal DL-based dose prediction framework capable of generalizing across diverse disease sites and beam configurations. By combining a generalized input design, robust network customization, and integration with a reference-guided optimization engine, UniDose generates physically feasible dose predictions and allows efficient user interaction through adjustable input conditions.