FastD_M,Mcalculation in LDR brachytherapy using deep learning methods.

Area of Science:

• Medical Physics
• Radiotherapy
• Computational Biology

Background:

• Monte Carlo (MC) simulations are accurate for brachytherapy dose calculation but are computationally intensive.
• Clinical implementation of MC methods is limited by long computation times.
• Accurate dose distribution is crucial for effective low-dose rate (LDR) brachytherapy.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for predicting dose to medium (D_M,M) distributions in LDR prostate brachytherapy.
• To accelerate dose calculation by leveraging DL trained on MC simulations.
• To integrate prior physics knowledge into the DL model for improved accuracy.

Main Methods:

• A 3D U-Net convolutional neural network was trained using 2369 single-seed configurations from 44 prostate patient plans.
• The model incorporated prior knowledge of brachytherapy dose dependency using an r² kernel.
• DL and MC dose distributions were compared using dose maps, isodose lines, and dose-volume histograms.

Main Results:

• The DL model accurately predicted dose distributions, with small differences below the 20% isodose line compared to MC.
• Predicted CTV D₉₀ showed an average difference of -0.1% compared to MC.
• The DL model predicted D_M,M volumes in 1.8 ms, significantly faster than MC.

Conclusions:

• The proposed DL model offers a fast and accurate method for dose calculation in LDR prostate brachytherapy.
• The model effectively incorporates physical principles, including source anisotropy and tissue heterogeneity.
• This DL approach has the potential to significantly improve clinical brachytherapy treatment planning.