Automated Needle Digitization in Ultrasound-based Prostate High Dose-rate Brachytherapy Using a Deep Learning Algorithm

Abstract

Prostate high dose-rate brachytherapy (HDR) delivers the radiation using a radioactive Ir-192 seed through interstitial needles and the needle implants are guided by a real-time ultrasound (US) imaging system. The success of this therapy relies highly on accurate digitization of the needles in the treatment planning US images. In current clinical procedures, the needles' trajectories are manually labeled by physicists, which is time-consuming and prone to errors. In recent years, deep learning techniques have brought many breakthroughs, especially in medical imaging processing. We thus propose a deep learning-based method to segment the needles accurately in the US images. The training data including the US images, needle positions and treatment plans were collected from the clinical database with IRB approval. Before training, images in the Cartesian coordinate were transferred into the polar plane and then split into smaller batches so that each contains one needle. Data augmentation techniques were used to avoid overfitting. The needle segmented neural network was based on U-Net with an encoder-decoder structure and shortcuts connecting directly the encoder layers and the corresponding decoder ones. The network structure and parameters were optimized for the needle segmentation problem. To estimate the positions of the needle tips, a separate convolutional neural network with the architecture of Vgg16 was developed and the 3D position accuracy was obtained. Data from 441 patients and 13200 needles was utilized. The needle trajectories were successfully segmented with mean errors of 0.67 mm and 0.32 mm in left-right (X) and anterior-posterior (Y) directions of the original US images and the estimation errors within 2 mm in X and Y directions are 95.4% and 99.2%, respectively. As for the needle tip localization, the positioning accuracies in X, Y, and Z (superior-inferior) are 0.72 mm, 0.37 mm and 1.88 mm, with 94.2%, 98.3% and 67.5% of the data have error of less than 2 mm, respectively. A novel method based on deep learning and neural networks was successfully developed to digitize HDR brachytherapy needles automatically. This method can be served as a valuable tool during brachytherapy procedures and has the potential to improve the efficiency and quality of treatment delivery, contributing overall to the patient care.