Abstract

This paper describes a new method for medical data segmentation based on superpixel propagation. The proposed method is a modification of the classical region growing algorithm and partly inherits the concept of octrees. The key difference of the proposed approach is the transition to the superpixel domain, as well as more flexible conditions for adding neighbor superpixels to the region. The region formation algorithm checks superpixels for compliance with some homogeneity criteria. First, the average intensity of superpixels is compared with the intensity of a resulting region. Second, each pixel on the edges and diagonals of a superpixel is compared with a threshold value. An important feature of the proposed method is the dynamically changing (floating) size of superpixels. The resulting region is formed by constructing a spline based on the points of intersection among the superpixels external to the region. To test the accuracy of the method, we use the MRI images of the left ventricle obtained at the University of York and MRI images of brain tumors obtained at the Southern Medical University. To demonstrate the performance of our method, a set of high-resolution synthetic images was additionally created. As an accuracy estimation metric, we use the Dice similarity coefficient (DSC). For the proposed method, it corresponds to 0.93 ± 0.03 and 0.89 ± 0.07 for the left ventricle and tumor segmentation, respectively. It is demonstrated that a step-by-step reduction in the size of a superpixel can significantly speed up the method without loss of accuracy.

中文翻译：

---

基于动态变化的超像素传播的分割

摘要

本文介绍了一种基于超像素传播的医学数据分割新方法。该方法是对经典区域增长算法的一种改进，部分继承了八叉树的概念。所提出的方法的主要区别是向超像素域的过渡，以及向该区域添加相邻超像素的更灵活的条件。区域形成算法检查超像素是否符合某些均匀性标准。首先，将超像素的平均强度与所得区域的强度进行比较。其次，将超像素的边缘和对角线上的每个像素与阈值进行比较。所提出的方法的重要特征是动态改变（浮动）超像素的大小。通过基于区域外部的超像素之间的相交点构造样条来形成结果区域。为了测试该方法的准确性，我们使用了约克大学获得的左心室MRI图像和南方医科大学获得的脑肿瘤的MRI图像。为了演示我们方法的性能，还额外创建了一组高分辨率合成图像。作为准确性估算指标，我们使用骰子相似系数（DSC）。对于所提出的方法，左心室和肿瘤分割分别对应于0.93±0.03和0.89±0.07。已经证明，逐步减小超像素的尺寸可以显着加快该方法，而不会降低精度。