Deep learning-based dose map prediction for high-dose-rate brachytherapy.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Brachytherapy treatment planning is time-intensive and challenging within clinical workflows.
• Existing deep learning dose prediction models are primarily for external beam radiation therapy (EBRT) and not optimized for brachytherapy.
• A well-established deep learning model specifically for brachytherapy dose prediction is lacking.

Purpose of the Study:

• To develop and evaluate a novel Squeeze and Excitation Attention Net (SE_AN) for accurate 3D brachytherapy dose distribution prediction.
• To emphasize the importance of applicator geometry in dose prediction for brachytherapy.

Main Methods:

• A novel SE module was integrated into a Cascaded UNet architecture to enhance feature recalibration, creating the SE_AN model.
• The SE_AN model was trained and validated on 250 clinical brachytherapy cases encompassing various applicator types (vaginal, tandem and ovoid, multi-channel, free needle).
• Model performance was assessed by comparing predicted dose distributions against clinically approved plans using Mean Absolute Error (MAE) of dose-volume histogram (DVH) metrics (D2cc, D90%).

Main Results:

• The SE_AN model achieved low MAEs for critical structures: 0.37 ± 0.25 Gy for HRCTV D90%, 0.23 ± 0.14 Gy for bladder D2cc, and 0.28 ± 0.20 Gy for rectum D2cc.
• SE_AN demonstrated comparable or superior accuracy to classic UNet and Cascaded UNet models in predicting DVH metrics.
• The model's predictions closely matched dose distributions from plans created by experienced dosimetrists.

Conclusions:

• A specialized deep learning method (SE_AN) for 3D brachytherapy dose prediction has been successfully developed.
• The SE_AN model provides accurate dose predictions comparable to clinical standards.
• This technique has the potential to enhance the standardization and quality control of brachytherapy treatment planning.