A comparative study of deep learning-based knowledge-based planning methods for 3D dose distribution prediction of...

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Medicine

Background:

• Accurate prediction of 3D radiation dose distributions is crucial for effective cancer treatment planning.
• Knowledge-based planning (KBP) leverages past treatment data to optimize future plans.
• Deep learning offers advanced capabilities for complex medical image analysis and prediction tasks.

Purpose of the Study:

• To compare the performance of four novel deep learning-based knowledge-based planning (KBP) algorithms.
• To evaluate the accuracy of these algorithms in predicting three-dimensional (3D) dose distributions for head and neck cancer patients.
• To assess the potential clinical utility of these KBP models for improving radiotherapy workflows.

Main Methods:

• Utilized a dataset of 340 oropharyngeal cancer patients from the AAPM OpenKBP - 2020 Grand Challenge.
• Developed and trained four 3D convolutional neural network architectures (U-Net, attention U-Net, Res U-Net, attention Res U-Net) for voxel-wise dose prediction.
• Evaluated model performance on a test set using dose statistics and dose-volume indices, comparing predicted doses against ground truth.

Main Results:

• All four KBP dose prediction models demonstrated promising performance, with average mean absolute dose errors within the body contour below 3 Gy.
• Attention Res U-Net and Res U-Net showed the smallest average differences in predicting D99 for targets (0.92 Gy and 0.94 Gy, respectively).
• U-Net achieved the lowest average difference for OAR indices (0.84 Gy), while attention Res U-Net and Res U-Net also showed significant accuracy.

Conclusions:

• The developed KBP models, particularly those based on 3D U-Net architectures, exhibit comparable performance for voxel-wise dose prediction.
• These models hold potential for clinical deployment to enhance the quality and consistency of cancer patient treatment plans.
• Implementing these advanced KBP models can lead to more efficient radiotherapy workflows.