nnDoseNet: Intuitive and Flexible Deep Learning Framework to Train and Evaluate Radiotherapy Dose Prediction Models

Radiotherapy (RT) dose optimization is often labor-intensive, requiring repeated manual adjustments to achieve clinically acceptable plans. In this work, we introduce nnDoseNet, a deep learning framework designed to automate and streamline RT dose prediction. Building on the nnU-Net segmentation engine, nnDoseNet adapts this architecture for dose regression by incorporating specialized loss functions (including dose–volume histogram terms) and multi-channel input (CT, targets, organs-at-risk, and body mask). It also supports clinically relevant evaluation metrics (e.g., gamma analysis and D95). We evaluated nnDoseNet on the OpenKBP challenge dataset comprising 340 head-and- neck cancer cases (240 for training and 100 for testing). Multiple hyperparameters (U-Net depth, patch size, batch size, and loss function) were tested. The best-performing configuration achieved a dose score of 2.579 and a DVH score of 1.540 on the test set—competitive with top submissions in the original challenge. Additional validation on an institutional cohort of 80 prostate cancer patients (45 training, 35 testing) demonstrated good agreement with clinical dose distributions (mean-squared error 0.817) and improved target coverage compared to clinical plans. By offering automated data preprocessing, systematic model training, and robust dose evaluation—all within a single framework—nnDoseNet reduces the complexity of building and testing dose prediction models. It accommodates diverse prescription doses, organ-at-risk definitions, and hardware configurations, making it a suitable benchmark for multi-institutional research. With its balance of simplicity, flexibility, and performance, nnDoseNet aims to accelerate the development, comparison, and clinical integration of advanced AI-driven dose prediction methods in radiotherapy.