Click here to

Session: AI in Treatment Planning [Return to Session]

Sub-Second D(M,M) Calculation for LDR Prostate Brachytherapy Using Deep Learning Methods

Francisco Berumen-Murillo1,2*, Shirin Abbasinejad Enger3,4, Luc Beaulieu1,2, (1) Departement de physique, genie physique et optique et Centre de recherche sur le cancer, Universite Laval, Quebec, QC, CA, (2) CHU De Quebec - Universite Laval, Quebec, QC, CA, (3) Medical Physics Unit, Department of Oncology, McGill University, Montreal, QC, CA, (4) Lady Davis Institute for Medical Research, Jewish General Hospital, Montreal, QC, CA


SU-H300-IePD-F4-1 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 4

Purpose: Monte Carlo (MC) method is the gold standard for low-dose rate (LDR) brachytherapy TG-186 calculations, and it provides a full solution to tissue heterogeneity effects. 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 D(M,M) (dose to medium in medium) distributions in LDR prostate brachytherapy.

Methods: Data of 44 prostate patients was used. These patients underwent LDR brachytherapy treatment in which I-125 SelectSeed sources were implanted. The dataset included on average 54 ± 10 seeds per patient for a total of 2369 seeds. For each seed, the patient geometry, the MC dose volume, and the plan volume were used for training a 3D UNet convolutional neural network. The model included an inverse-square law kernel. The dataset was randomly split into training, validation, and test sets. The Adam optimizer was used with a learning rate of 0.001. The mean absolute error was used as the loss function in 370 epochs. MC and DL dose distributions were compared with dose maps, isodoses, and dose-volume histograms.

Results: The DL model accurately predicted dose distributions as seen in the isodoses. In single seed predictions, the anisotropy of the source was correctly depicted. For a full prostate patient, small but noticeable differences were seen in the low-dose region below the 20% isodose line. The predicted CTV D(90) metric had a relative difference of 1.0% with respect to the MC reference. The rectum D(2cc) metric had a relative difference of 3.2%. The model took 0.8 s to predict a complete D(M,M) 3D dose volume (1.18 M voxels).

Conclusion: DL-based dose prediction provides fast and accurate patient dosimetry taking into account the anisotropy of the source and the patient tissue composition.

Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) grant RGPIN201905038. FBM acknowledges support by the Fonds de Recherche du Quebec, Nature et Technologies (FRQNT). This research was enabled in part by the support provided by Calcul Quebec and Compute Canada.


Brachytherapy, Dosimetry


TH- Brachytherapy: Treatment planning using machine learning/automation

Contact Email