Deep learning applications in automatic needle segmentation in ultrasound-guided prostate brachytherapy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• High-Dose-Rate (HDR) brachytherapy is a key prostate cancer treatment using transperineal needle implants guided by ultrasound (US).
• Manual segmentation of brachytherapy needles in US images is currently performed by physicists, a process that is both time-consuming and prone to errors.

Purpose of the Study:

• To develop and evaluate deep learning algorithms for accurate segmentation of brachytherapy needles in US images.
• To enable precise localization of brachytherapy needle tips from US images, aiming to enhance treatment accuracy and efficiency.

Main Methods:

• A modified deep U-Net architecture was employed for pixel-level segmentation of brachytherapy needles.
• A VGG-16-based deep convolutional network was integrated with the segmentation network for predicting needle tip locations.
• The algorithms were trained and validated on a clinical dataset of US images with labeled needle trajectories.

Main Results:

• The proposed method achieved accurate needle trajectory extraction with resolutions of 0.668 mm (x-direction) and 0.319 mm (y-direction).
• High accuracy was demonstrated, with errors ≤ 2 mm in 95.4% (x-direction) and 99.2% (y-direction) of cases.
• Needle tip localization showed position resolutions of 0.721 mm (x), 0.369 mm (y), and 1.877 mm (z), with 94.2%, 98.3%, and 67.5% of data having errors ≤ 2 mm, respectively.

Conclusions:

• A novel neural network-based algorithm effectively segments brachytherapy needles and localizes needle tips in US images.
• This automated approach has the potential to significantly improve the efficiency and quality of High-Dose-Rate brachytherapy for prostate cancer treatment.