Deep learning-based dose prediction for magnetic resonance-guided prostate radiotherapy.

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Daily adaptive radiotherapy (ART) on MR-Linac requires selecting between adaptation methods like Adapt to Position (ATP) and Adapt to Shape (ATS).
• ATS necessitates daily re-contouring, which is time-consuming and resource-intensive.
• Rapid prediction of dose distribution and evaluation criteria could streamline adaptation method selection and reduce treatment times.

Purpose of the Study:

• To develop and validate a deep-learning-based dose-prediction pipeline for prostate cancer treatments using an MR-Linac.

Main Methods:

• Trained a deep learning segmentation network on 212 MR images from 35 prostate cancer patients to delineate CTV, bladder, and rectum.
• Developed a second deep learning network to predict dose distributions using segmented structures.
• Inference involved using predicted segmentations as input for dose prediction and comparing predicted doses to true doses.

Main Results:

• Segmentation network achieved median Dice Similarity Coefficient (DSC) values of 0.90 for CTV, 0.94 for bladder, and 0.87 for rectum.
• Predicted segmentations as input to the dose prediction network resulted in mean dose differences <2% for key dosimetric parameters (D98%, D95%, D2%, Dmean, V33Gy, V38Gy, V41Gy) compared to true segmentations.
• Differences were statistically insignificant for most parameters, indicating comparable performance between predicted and manual structures.

Conclusions:

• The developed deep learning dose-prediction pipeline shows minimal differences (<2%) in achieving clinical dose-volume constraints compared to using manually delineated structures.
• The pipeline is a valuable decision support tool for MR-Linac treatments, particularly when differences exceed 2%.