基于超像素与图卷积神经网络的白细胞分割

White blood cell segmentation based on superpixel and graph convolution neural n etwork

摘要:

白细胞分割是医学图像处理领域的一项富有挑战 性的任务,针对目前白细胞分割存在 的准确度不高、粘连情况不易分割等问题,将图像的分割转化为区域节点的分类问题,提出 基于图卷积神经网络的白细胞分割算法。首先将训练图像经超像素分割得到若干超像素区域 ,把每个超像素区域作为图的一个节点,并充分利用超像素区域的彩色特征以及空间邻域关 系构造稀疏加权图来训练图卷积网络,然后利用训练好的网络对测试图像进行白细胞核、质 、背景的三域一次性分类。实验数据表明,本文算法对不同类白细胞均具有较好的分割效果 。

Abstract:

White blood cell (WBC) segmentation is a challenging task in the field of medical image processing.Aiming at problems such as the low accuracy of cur rent white cell segmentation and the difficulty of segmentation of adhesion,the image segmentation is transformed into a classification problem of graph nodes and a white cell segmentation algorithm based on graph convolutional neural netw ork (GCN) is presented.First,the training samples are oversegmented by superpi xel algorithm to obtain a number of superpixel regions,each superpixel region i s used as a node of the graph,the RGB features of the superpixel region and the spatial neighborhood relationship are used to construct a sparse weighted graph to train the graph convolution network,then the trained network is used to cla ssify the experimental data into the three domains of white blood cell nucleus, qualitative and background.Experimental data show that the algorithm in this pa per has a good segmentation effect on different types of white blood cells.