A deep learning approach for real time prostate segmentation in freehand ultrasound guided biopsy
A deep learning approach for real time prostate segmentation in freehand ultrasound guided biopsy
复制标题
DOI：
10.1016/j.media.2018.05.010
复制
发表时间：
2018-08-01
影响因子：
10.9
通讯作者：
Abolmaesumi, Purang
中科院分区：
文献类型：
--
作者：
Abu Anas, Emran Mohammad;Mousavi, Parvin;Abolmaesumi, Purang
靶向前列腺活检，结合多参数磁共振成像（mp-MRI）及其与超声的登记，是目前前列腺癌诊断的最先进技术。当今大多数靶向活检系统的配准过程在很大程度上依赖于超声图像的准确分割。自动或半自动分割通常在活检程序开始之前离线进行。在本文中，我们提出了一种在活检过程中基于深度神经网络的实时前列腺分割技术，从而为mp-MRI和超声数据的动态配准铺平了道路。除了使用卷积网络提取空间特征外，该方法还使用循环网络来利用一系列超声图像之间的时间信息。该体系结构的关键贡献之一是在循环网络中使用残差卷积来改进优化。我们还利用深层网络不同层内部和之间的循环连接，以最大限度地利用时间信息。此外，我们对输入超声序列进行密集和稀疏采样，使网络对超声伪像具有鲁棒性。我们的架构在2238张标记的经直肠超声图像上进行了训练，另外还有637张和1017张未见过的图像分别用于验证和测试。我们得到Dice相似系数均值为93%，平均表面距离误差为1.10 mm，平均Hausdorff距离误差为3.0 mm。将报告的结果与最先进技术的结果进行比较表明，拟议的方法在统计上取得了显著的改进。(C) 2018 Elsevier B.V.版权所有
Targeted prostate biopsy, incorporating multi-parametric magnetic resonance imaging (mp-MRI) and its registration with ultrasound, is currently the state-of-the-art in prostate cancer diagnosis. The registration process in most targeted biopsy systems today relies heavily on accurate segmentation of ultrasound images. Automatic or semi-automatic segmentation is typically performed offline prior to the start of the biopsy procedure. In this paper, we present a deep neural network based real-time prostate segmentation technique during the biopsy procedure, hence paving the way for dynamic registration of mp-MRI and ultrasound data. In addition to using convolutional networks for extracting spatial features, the proposed approach employs recurrent networks to exploit the temporal information among a series of ultrasound images. One of the key contributions in the architecture is to use residual convolution in the recurrent networks to improve optimization. We also exploit recurrent connections within and across different layers of the deep networks to maximize the utilization of the temporal information. Furthermore, we perform dense and sparse sampling of the input ultrasound sequence to make the network robust to ultrasound artifacts. Our architecture is trained on 2,238 labeled transrectal ultrasound images, with an additional 637 and 1,017 unseen images used for validation and testing, respectively. We obtain a mean Dice similarity coefficient of 93%, a mean surface distance error of 1.10 mm and a mean Hausdorff distance error of 3.0 mm. A comparison of the reported results with those of a state-of-the-art technique indicates statistically significant improvement achieved by the proposed approach. (C) 2018 Elsevier B.V. All rights reserved.