Patch-based corner detection for cervical vertebrae in X-ray images
Patch-based corner detection for cervical vertebrae in X-ray images
复制标题
DOI：
10.1016/j.image.2017.04.002
复制
发表时间：
2017-11-01
影响因子：
3.5
通讯作者：
Slabaugh, Greg
中科院分区：
文献类型：
--
作者：
Al Arif, S. M. Masudur Rahman;Asad, Muhammad;Slabaugh, Greg
角点包含有关 X 射线图像中椎骨大小、形状和形态的重要信息，最近的文献（Al-Arif 等人，2015）[1,2] 显示了使用基于霍夫森林的架构检测椎骨角点的良好性能。为了提供空间背景，该方法在椎骨周围生成一组 12 个补丁，并使用机器学习方法通过投票过程来预测椎体的角点。在本文中，我们在补丁生成和预测方法方面扩展了该框架。在补丁生成过程中，方形感兴趣区域已被数据驱动的矩形和梯形感兴趣区域替换，这可以更好地将补丁与椎体几何形状对齐，从而产生更具辨别力的特征向量。通过利用单核密度估计的更有效的投票过程，角点估计或预测阶段得到了改进。此外，还引入了高级且更复杂的特征向量。我们还使用不同的补丁生成方法、森林训练机制和预测方法对框架进行了全面的评估。为了将该框架的性能与更通用的方法进行比较，引入了一种新颖的基于多尺度 Harris 角点检测器的方法，该方法通过朴素贝叶斯方法结合了空间先验。所有这些方法都在 90 张 X 射线图像的数据集上进行了测试，平均角点定位误差为 2.01 毫米，与之前最先进的方法相比，定位精度提高了 33% (Al-Arif et al., 2015) [2].(1) (C) 2017 Elsevier B.V. 保留所有权利。
Corners hold vital information about size, shape and morphology of a vertebra in an x-ray image, and recent literature (Al-Arif et al., 2015) [1,2] has shown promising performance for detecting vertebral corners using a Hough forest-based architecture. To provide spatial context, this method generates a set of 12 patches around a vertebra and uses a machine learning approach to predict corners of a vertebral body through a voting process. In this paper, we extend this framework in terms of patch generation and prediction methods. During patch generation, the square region of interest has been replaced with data-driven rectangular and trapezoidal region of interest which better aligns the patches to the vertebral body geometry, resulting in more discriminative feature vectors. The corner estimation or the prediction stage has been improved by utilising more efficient voting process using a single kernel density estimation. In addition, advanced and more complex feature vectors are introduced. We also present a thorough evaluation of the framework with different patch generation methods, forest training mechanisms and prediction methods. In order to compare the performance of this framework with a more general method, a novel multi-scale Harris corner detector-based approach is introduced that incorporates a spatial prior through a naive Bayes method. All these methods have been tested on a dataset of 90 X-ray images and achieved an average corner localization error of 2.01 mm, representing a 33% improvement in localization accuracy compared to the previous state-of-the-art method (Al-Arif et al., 2015) [2].(1) (C) 2017 Elsevier B.V. All rights reserved.