RADIANT: A fully configurable radiotherapy dose prediction framework

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Radiotherapy treatment planning is complex, requiring specialized expertise and software.
• Deep learning offers a powerful approach for predicting patient-specific radiation dose distributions.

Purpose of the Study:

• To present the Radiotherapy Dose Inference and Analysis Toolkit (RADIANT), an open-source framework for 3D radiotherapy dose prediction.
• To demonstrate RADIANT's capability in supporting diverse deep learning models and training strategies for dose prediction.

Main Methods:

• Developed RADIANT, a configurable framework built on the Medical Imaging Segmentation Toolkit.
• Applied RADIANT to cervical, prostate, and head and neck cancer treatment plans, including data from the AAPM OpenKBP challenge.
• Compared various deep learning architectures (nnU-Net, FMG-Net, W-Net, ddU-Net, Swin UNETR) using clinical metrics like dose score and DVH score.

Main Results:

• RADIANT demonstrated scalability for developing and benchmarking dose prediction models across multiple cancer sites.
• The best configuration achieved a dose score of 1.20 for cervical cancer and 1.96 for prostate cancer on test sets.
• On head and neck data, RADIANT achieved competitive results with a dose score of 2.702 and DVH score of 1.495.

Conclusions:

• RADIANT provides a fully configurable, open-source solution for deep learning-based radiotherapy dose prediction.
• The framework supports reproducible research from data preprocessing to model evaluation for diverse cancer types.
• RADIANT facilitates advancements in radiotherapy treatment planning through efficient model development and benchmarking.