Radiant: A Fully Configurable Radiotherapy Dose Prediction Framework
Abstract
Purpose
This work presents the Radiotherapy Dose Inference and Analysis Toolkit (RADIANT), an open-source, fully configurable framework for 3D radiotherapy dose prediction. Built upon the Medical Imaging Segmentation Toolkit, RADIANT supports a wide range of network architectures, loss functions, and training strategies.
Methods
We demonstrate its capabilities on cervical and prostate treatment plans generated with the Radiation Planning Assistant, as well as on head and neck cancer treatment plans from the AAPM OpenKBP challenge data. Comprehensive dose prediction modeling for cervical cancer was performed using nnUNet, FMG-Net, W-Net, ddUNet, and Swin UNETR and evaluated for clinical metrics such as dose score, mean absolute difference (MAD), and percent errors in D95, D98, and D99. Benchmark testing was also performed on a public dataset for head and neck cancer and dvh score metric was also provided for comparison with the top contenders of the OpenKBP challenge.
Results
Overall, nnU-Net gave most accurate results compared to other approaches. For the cervix, the nnU-Net model trained with MAE loss and cosine learning rate scheduler had the lowest errors in predictions with MAD =1.6 and D95 error = 0.66% on the test set. The mean dose error was within 1% for PTVs and 2% or OARs with respect to the maximum prescription dose 57.50 Gy. For the prostate, the same configuration gave MAD = 1.96 and D95 error = 0.36%. When predicting on the head and neck OpenKBP benchmark test set, dose score was 2.70 and DVH score was 1.496 (which would have been 2nd place on leaderboard).
Conclusion
This work provides an open-source, fully configurable framework for deep learning–based dose prediction. RADIANT is designed to facilitate reproducible experimentation by providing tools for data preprocessing, model configuration, training, and evaluation across multiple cancer sites and anatomical structures.