Site-agnostic 3D dose distribution prediction with deep learning neural networks

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Current radiation dose prediction models often require site-specific retraining and struggle with limited data, leading to suboptimal performance.
• A site-agnostic approach can leverage larger, diverse datasets for more robust model training.

Purpose of the Study:

• To develop and evaluate a site-agnostic, three-dimensional (3D) deep learning model for accurate dose distribution prediction in radiotherapy.
• To demonstrate the model's adaptability to new treatment sites using transfer learning with minimal fine-tuning.

Main Methods:

• A 3D UNet architecture was employed for dose prediction.
• The model was initially trained on prostate cancer intensity-modulated radiation therapy data (source) and then adapted to head-and-neck cancer volumetric-modulated arc therapy data (target) using transfer learning.
• Performance was evaluated using dose-volume histogram metrics for planning target volumes and organs at risk.

Main Results:

• The site-agnostic model achieved accurate dose predictions on the source site with low absolute dose errors for PTV and OARs.
• When adapted to the target site, the model demonstrated effective transfer learning, maintaining accurate predictions for PTV and OARs with limited target data.
• The adapted model showed improved performance compared to a model trained solely on the target data.

Conclusions:

• A site-agnostic deep learning model for 3D dose prediction is feasible and effective.
• Transfer learning allows for efficient adaptation of the model to new treatment sites, even with small datasets.
• This approach enhances the generalizability and applicability of dose prediction models in radiotherapy.