Technical Note: Dose prediction for radiation therapy using feature-based losses and One Cycle Learning.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Accurate dose prediction is crucial for effective radiotherapy planning.
• The Open Knowledge-Based Planning (OpenKBP) challenge aims to advance automated treatment planning.
• Generative Adversarial Network (GAN) techniques, while powerful, can be complex and resource-intensive for model training.

Purpose of the Study:

• To detail the runner-up model from the OpenKBP challenge's dose-volume histogram (DVH) stream.
• To demonstrate a simple, reproducible training approach for dose prediction models.
• To avoid the need for advanced and costly generative adversarial network (GAN) techniques.

Main Methods:

• A U-Net architecture incorporating ResNet blocks was developed using the OpenKBP dataset (200 training, 40 validation head-and-neck patients).
• The model was trained using AdamW optimizer with a One Cycle scheduler and a combined L1 and feature loss function utilizing a pre-trained video classifier.
• Performance was evaluated on 100 test patients, assessing DVH metrics for organs at risk and target structures.

Main Results:

• The model secured 2nd place in the DVH stream and 4th in the dose stream of the OpenKBP challenge.
• Achieved a DVH score of 1.52 ± 1.06 and a dose score of 2.62 ± 1.10.
• Demonstrated mean dose differences within ±1% for predicted versus true DVH metrics across structures.

Conclusions:

• A simple U-Net based approach, incorporating ResNet blocks and feature-based losses, yielded excellent results in the OpenKBP challenge.
• The study highlights the effectiveness of common computer vision techniques like One Cycle Learning and feature losses for radiotherapy dose prediction.
• The model offers a reproducible and efficient alternative to more complex GAN-based methods.