A Transformer-Embedded Multi-Task Model for Dose Distribution Prediction.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Radiotherapy planning is currently subjective and time-consuming, relying heavily on radiologist experience.
• Achieving clinically acceptable radiotherapy plans requires iterative adjustments, impacting efficiency and consistency.

Purpose of the Study:

• To develop an automated system for accurate radiation dose distribution prediction in radiotherapy.
• To introduce a novel transformer-embedded multi-task dose prediction (TransMTDP) network.

Main Methods:

• The TransMTDP network integrates three tasks: main dose prediction, auxiliary isodose lines prediction, and auxiliary gradient prediction.
• A multi-task learning strategy with a shared encoder is employed, enhanced by isodose and gradient consistency losses.
• A transformer is embedded to capture long-range dependencies in dose maps, leveraging anatomical symmetries.

Main Results:

• The TransMTDP network demonstrated superior performance compared to state-of-the-art methods on rectum and head and neck cancer datasets.
• The multi-task approach with consistency losses improved the stability and accuracy of dose predictions.
• The transformer component effectively captured global features and long-range dependencies in dose maps.

Conclusions:

• The proposed TransMTDP network offers an automated, accurate, and efficient solution for radiotherapy dose prediction.
• This AI-driven approach has the potential to optimize radiotherapy planning, reducing subjectivity and time.
• The integration of multi-task learning and transformer architecture represents a significant advancement in medical image analysis for radiation oncology.