FastD_M,Mcalculation in LDR brachytherapy using deep learning methods

Objective.The Monte Carlo (MC) method provides a complete solution to the tissue heterogeneity effects in low-energy low-dose rate (LDR) brachytherapy. However, long computation times limit the clinical implementation of MC-based treatment planning solutions. This work aims to apply deep learning (DL) methods, specifically a model trained with MC simulations, to predict accurate dose to medium in medium (D_M,M) distributions in LDR prostate brachytherapy.Approach.To train the DL model, 2369 single-seed configurations, corresponding to 44 prostate patient plans, were used. These patients underwent LDR brachytherapy treatments in which¹²⁵I SelectSeed sources were implanted. For each seed configuration, the patient geometry, the MC dose volume and the single-seed plan volume were used to train a 3D Unet convolutional neural network. Previous knowledge was included in the network as anr²kernel related to the first-order dose dependency in brachytherapy. MC and DL dose distributions were compared through the dose maps, isodose lines, and dose-volume histograms. Features enclosed in the model were visualized.Main results.Model features started from the symmetrical kernel and finalized with an anisotropic representation that considered the patient organs and their interfaces, the source position, and the low- and high-dose regions. For a full prostate patient, small differences were seen below the 20% isodose line. When comparing DL-based and MC-based calculations, the predicted CTVD₉₀metric had an average difference of -0.1%. Average differences for OARs were -1.3%, 0.07%, and 4.9% for the rectumD_2cc, the bladderD_2cc, and the urethraD_0.1cc. The model took 1.8 ms to predict a complete 3DD_M,Mvolume (1.18 M voxels).Significance.The proposed DL model stands for a simple and fast engine which includes prior physics knowledge of the problem. Such an engine considers the anisotropy of a brachytherapy source and the patient tissue composition.