Predicting 3D dose distribution with scale attention network for prostate cancer radiotherapy

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate radiation dose prediction is crucial for optimizing cancer treatment planning.
• Improving the efficiency of radiation therapy treatment planning is a growing need.

Purpose of the Study:

• To propose a novel deep learning framework, Scale Attention Networks (SA-Net), for voxel-wise radiation dose prediction.
• To evaluate the effectiveness of SA-Net using distance data and CT images for prostate cancer treatment planning.

Main Methods:

• Developed a Scale Attention Network (SA-Net) incorporating a dynamic scale attention model.
• Utilized signed distance maps of organs at risk and CT images as network inputs.
• Trained and tested the model on prostate cancer cases treated with Volumetric Modulated Arc Therapy (VMAT).

Main Results:

• Achieved an average dose difference of 0.94 Gy (2.1%) compared to clinical plans.
• Demonstrated that signed distance maps provide superior input performance over binary masks.
• Validated the deep learning approach for automated treatment planning in prostate cancer radiotherapy.

Conclusions:

• The SA-Net framework offers a feasible and effective deep learning strategy for automating radiation dose prediction.
• The use of signed distance maps enhances the accuracy of deep learning-based dose prediction models.
• This approach has the potential to significantly improve radiation therapy treatment planning efficiency and accuracy.