Automated Computer Vision and Machine Learning Workflows in Radiation Treatment Planning

Overview

Mayo Clinic Brachytherapy, the use of radioactive implants placed near tumors, is a vital part of cervical cancer treatment. However, the lack of automated workflows makes brachytherapy planning susceptible to time constraints, human error, and inconsistencies in radiation dose quantities. This project proposes a fully automated software that does not require manual interaction with patient images, with the goal of reducing planning duration and variability. The software is equipped with a three-dimensional interface for clinical usage and features a novel assemblage of 3D computer vision tools. Tandem and ovoid applicators, the titanium catheters that ferry radiation into the body, are segmented by thresholding CT images. The isolated high-density voxels are assigned to catheter and non- catheter structures through HDBSCAN, an unsupervised machine learning algorithm that uses density-based linkage clustering. The algorithm recognized seven structures, including three catheters and four gold seeds. After clustering, the applicator contours were traced through finding the centroid point of the applicator voxels on each image slice. The contours, along with treatment reference points and ovoid surface lines, were written to a treatment plan file. Through comparing manual and automated planning, retrospectively, for 10 cervical cancer patients, the software demonstrated clinically viable results in geometric and dosimetric accuracy. The average execution time for the algorithm was less than 30 seconds. The implementation of this automated workflow, the first of its kind for tandem and ovoids applicators, can lead to safer radiation delivery and reduce the duration and variability of high-dose rate treatment planning.