Cross-technique transfer learning to predict the dose distribution for radiotherapy planning based on a limited...

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Accurate radiotherapy dose prediction is crucial but challenging due to limited training data and evolving techniques.
• Intensity-modulated radiotherapy (IMRT) and volumetric modulated arc therapy (VMAT) are advanced radiotherapy techniques.

Purpose of the Study:

• To develop a cross-technique transfer learning strategy for predicting radiotherapy dose distribution with minimal training samples.
• To evaluate the efficacy of this strategy compared to independent models and standard models.

Main Methods:

• A deep learning network (Res-U Net) was utilized for dose prediction.
• Models were pretrained on IMRT data and fine-tuned on VMAT data using limited samples (5 and 7 cases).
• Performance was assessed using dose-volume histograms (DVH), mean absolute error (MAE), and Dice similarity coefficient (DSC).

Main Results:

• Cross-technique models with five training samples showed clinically acceptable performance, comparable to standard models (MAE deviation: 0.15%, DSC deviation: 0.11%-0.72%).
• Performance improved with seven training samples (MAE deviation: 0.05%, DSC deviation: 0.02%-0.40%).
• Independent models trained on limited samples performed significantly worse.

Conclusions:

• Cross-technique transfer learning enables accurate and reliable dose distribution prediction for new radiotherapy techniques, even with limited sample sizes.
• This approach addresses the challenge of data scarcity in radiotherapy planning.