Automatic needle localization in 3D ultrasound images for brachytherapy

Abstract

Needle segmentation in ultrasound images may be the indispensable step to solve other problems, such as the detection of radioactive seeds in ultrasound images for brachytherapy treatment. In this paper we propose a novel method to localize accurately curved paths of flexible needles in threedimensional (3D) ultrasound images. Our method is based on an automatic thresholding step where the bayesian classifier theory is applied to select a needle's voxels from the background. The next step consists in detecting the needle and representing its curvature. Here we propose to fit a parametric polynomial of low degree to representing the parabolic curvature. We describe a first evaluation on a data set of 9 patients.

Towards 3D Ultrasound Guided Needle Steering Robust to Uncertainties, Noise, and Tissue Heterogeneity

“…To solve this issue, we propose to use the naive Bayesian classification approach detailed in [26]. The intensity histogram is approximated by a sum of two normal distributions: a needle normal distribution N n and a background normal distribution N b .…”

Needle Segmentation in 3D Ultrasound Volumes Based on Machine Learning for Needle Steering

“…As explained in [3], the VOI histogram can be modeled using an additive Gaussian Mixture Model (GMM):…”

“…A meaningful initialization of θ is necessary for suitable convergence. More details can be found in [3].…”

“…In [2], we proposed an observer of the needle tip pose. In [3], we detailed a machine learning approach for needle localization in 3D US volumes. In this paper, we describe the interconnection of both methods for precise curved needle localization in the context of needle steering.…”

Tip Estimation Method in Phantoms for Curved Needle Using 2D Transverse Ultrasound Images

“…Chatelain et al used the particle filtering to track a robot-guided flexible needle by using 3D US [21]. In addition, a convolutional neural network with conventional image processing techniques has also been used to track and detect the needle [22] and a naive Bayesian classifier was used to localize the needle among 3D US volume voxels [23]. However, the large 3D US volumetric dataset would make it difficult to obtain and process the online data.…”