CBCT-guided adaptive radiotherapy using self-supervised sequential domain adaptation with uncertainty estimation.

Area of Science:

• Medical Physics
• Radiotherapy Oncology
• Artificial Intelligence in Medicine

Background:

• Adaptive radiotherapy (ART) requires accurate tumor segmentation on low-quality on-board images for treatment adaptation.
• Current manual and deep learning segmentation methods face challenges due to poor image quality and limited labeled data.

Purpose of the Study:

• To develop a novel deep neural network with attention for accurate tumor segmentation and shrinkage prediction in adaptive radiotherapy.
• To address image quality and data scarcity issues using self-supervised domain adaptation (SDA).
• To provide uncertainty estimation for sequential segmentation to aid treatment planning and model reliability.

Main Methods:

• Proposed a sequence transduction deep neural network with an attention mechanism for tumor segmentation.
• Implemented a self-supervised domain adaptation (SDA) method to transfer features from CT to CBCT.
• Incorporated uncertainty estimation for sequential segmentation.

Main Results:

• The model achieved an average Dice score of 0.92 for predicting immediate weekly tumor deformation.
• It accurately predicted tumor shrinkage up to 5 weeks in advance with a minimal Dice score reduction of 0.05.
• Incorporating predictions into re-planning reduced radiation pneumonitis risk by 35% while maintaining tumor control.

Conclusions:

• The developed deep learning model effectively tracks tumor shrinkage in adaptive radiotherapy using weekly CBCT data.
• SDA and uncertainty estimation enhance model performance and reliability in low-quality imaging scenarios.
• This approach shows significant potential for improving treatment outcomes and reducing toxicity in lung cancer patients undergoing ART.