A universal framework for IMRT dose prediction.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Deep learning (DL) models for intensity-modulated radiation therapy (IMRT) dose prediction are often limited by fixed beam configurations and disease sites, hindering clinical generalizability.
• Existing models struggle to adapt to the diverse scenarios encountered in real-world radiation oncology practice.

Purpose of the Study:

• To develop UniDose, a universal DL-based dose prediction model for IMRT.
• To enable dose prediction across a wide range of disease sites and arbitrary beam configurations, enhancing clinical applicability.

Main Methods:

• UniDose utilizes a customized nnU-Net framework for 3D dose prediction, trained with Huber loss.
• Generalized inputs include normalized prescription dose, a weighted avoidance mask for organs at risk (OARs), and a beam trace image.
• The model was trained on 871 patients across 25 disease sites and validated using gamma passing rate (GPR) and dose-volume histogram (DVH) metrics against optimized and clinical plans.

Main Results:

• UniDose predictions achieved an average GPR of 92.36% against optimized doses and 86.13% against clinical plans.
• Predicted doses demonstrated improved OAR sparing and comparable target coverage, particularly for prostate, liver, and brain cases.
• Case studies confirmed the physical feasibility of predicted doses, and adjustable input weights allowed for flexible trade-offs in treatment planning.

Conclusions:

• UniDose shows significant potential as a universal DL framework for IMRT dose prediction, generalizing across diverse sites and beam configurations.
• The model's generalized input design and integration with an optimization engine produce physically feasible predictions.
• UniDose facilitates efficient, patient-specific treatment planning through adjustable input conditions and robust network customization.