Session Item

August 28
10:30 - 12:30
Special Multidisciplinary
Contouring workshop
Deep learning for automated applicator reconstruction in high-dose-rate prostate brachytherapy


Deep learning for automated applicator reconstruction in high-dose-rate prostate brachytherapy

Christopher Deufel1, Luca Weishaupt1, Hisham Kamal Sayed1, Chunhee Choo1, Bradley Stish1

1Mayo Clinic, Radiation Oncology, Rochester, USA

Show Affiliations
Purpose or Objective

To develop a deep-learning-based algorithm that automates the segmentation of treatment applicators on CT images for high-dose-rate (HDR) brachytherapy prostate patients implanted with titanium needles. The automation of applicator reconstruction can improve quality and consistency, while reducing the time that the patient is under anesthesia.

Material and Methods

A U-Net deep learning algorithm was developed to perform applicator digitization on CT images for HDR prostate brachytherapy treatment planning. The algorithm was trained using approximately 50,000 patient images generated from 26 distinct patient data sets using data augmentation with random rotations and noise. The patient DICOM CT image files were obtained using a standard reconstruction filter, a 50 cm nominal field of view, 0.977 mm to 1.25 mm slice thickness, and 512 x 512-pixel resolution. CT images were cropped to a 30 cm field of view and resampled to a uniform 0.5 mm x 0.5 mm x 0.5 mm voxel resolution for this study. The data sets were divided into 16 patients for training and 10 patients for validation. Each patient was implanted with between 15 and 20 titanium needle applicators.  DICE score was used as the U-Net's loss function for training and testing the deep learning algorithm. The model was trained on a GPU cluster to produce a normalized probability map of the likelihood that a pixel lies inside of an applicator. A global threshold of 0.8 was used to discriminate the applicator points from non-applicator points. HDBScan, a density-based clustering algorithm, was used to assign the applicator points to distinct applicators.


The deep learning algorithm converged after 36 epochs and a runtime of 6 hours using a single CPU. Figure 1 illustrates the algorithm performance on a representative CT image. The distribution of the DICE scores for all applicators from 25 patients is provided in Figure 2A. The median DICE score was 0.90, which is consistent with previously published results for metal applicator digitization on CT images. Density-based clustering provided an ‘off-the-shelf’ solution for the assignment of applicator points to distinct needles (Figure 2B) 


Deep learning algorithms are an effective strategy for automating the digitization of brachytherapy applicators, and a 2-dimensional U-Net approach provided an excellent correspondence between the automated and human segmentations for prostate treatments using titanium needles. The automation of brachytherapy applicator digitization is expected to improve the consistency, efficiency, and quality of brachytherapy treatments.