Premium
Deep learning applications in automatic needle segmentation in ultrasound‐guided prostate brachytherapy
Author(s) -
Wang Fuyue,
Xing Lei,
Bagshaw Hilary,
Buyyounouski Mark,
Han Bin
Publication year - 2020
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1002/mp.14328
Subject(s) - brachytherapy , deep learning , prostate brachytherapy , segmentation , artificial intelligence , prostate cancer , computer science , pixel , convolutional neural network , prostate , artificial neural network , ultrasound , medicine , nuclear medicine , computer vision , medical physics , radiology , cancer , radiation therapy
Purpose High‐Dose‐Rate (HDR) brachytherapy is one of the most effective ways to treat the prostate cancer, which is the second most common cancer in men worldwide. This treatment delivers highly conformal dose through the transperineal needle implants and is guided by a real time ultrasound (US) imaging system. Currently, the brachytherapy needles in the US images are manually segmented by physicists during the treatment, which is time consuming and error prone. In this study, we propose a set of deep learning–based algorithms to accurately segment the brachytherapy needles and locate the needle tips from the US images. Methods Two deep neural networks are developed to address this problem. First, a modified deep U‐Net is used to segment the pixels belonging to the brachytherapy needles from the US images. Second, an additional VGG‐16–based deep convolutional network is combined with the segmentation network to predict the locations of the needle tips. The networks are trained and evaluated on a clinical US images dataset with labeled needle trajectories collected in our hospital (Institutional Review Board approval (IRB 41755)). Results The evaluation results show that our method can accurately extract the trajectories of the needles with a resolution of 0.668 mm and 0.319 mm in x and y direction, respectively. 95.4% of the x direction and 99.2% of the y direction have error ≤ 2 mm. Moreover, the position resolutions of the tips are 0.721, 0.369, and 1.877 mm in x , y , and z directions, respectively, while 94.2%, 98.3%, and 67.5% of the data have error ≤ 2 mm. Conclusions This paper proposed a neural network‐based algorithm to segment the brachytherapy needles from the US images and locate the needle tip. It can be used in the HDR brachytherapy to help improve the efficiency and quality of the treatments.