CN110110807B - Leukocyte extraction and classification method based on improved K-means and convolutional neural network - Google Patents

本发明涉及一种基于改进K‑means及卷积神经网络的白细胞提取和分类方法。首先，根据细胞图像灰度分布来选定初始聚类中心，对图像所有像素按就近原则进行初始聚类；接着，对FWSA‑KM算法的欧式距离进行改进；白细胞提取之前，先进行颜色空间分解，采用有利于白细胞分割的颜色分量和改进的K‑means算法进行细胞核和细胞质的提取；接着采用分水岭算法来分离复杂粘连的部分；最后，基于卷积神经网络进行分类。本发明方法使得白细胞细胞核和细胞质分割精度分别为95.81％和91.28％，较传统分割方法有较大提高；分类准确度最大能达到98.96%，分类平均时间为0.39s，相对于现有的白细胞分类算法，CNN分类方法不仅有明显优势，同时还有很大的提升空间。

The invention relates to a white blood cell extraction and classification method based on improved K-means and convolutional neural network. First, select the initial cluster center according to the gray distribution of the cell image, and perform initial clustering on all pixels of the image according to the principle of proximity; then, improve the Euclidean distance of the FWSA-KM algorithm; before white blood cell extraction, first perform color space decomposition , the color components that are beneficial to white blood cell segmentation and the improved K-means algorithm are used to extract the nucleus and cytoplasm; then the watershed algorithm is used to separate the parts of complex adhesions; finally, the classification is based on the convolutional neural network. The method of the present invention makes the segmentation accuracy of leukocyte nucleus and cytoplasm respectively 95.81% and 91.28%, which is greatly improved compared with the traditional segmentation method; the maximum classification accuracy can reach 98.96%, and the average classification time is 0.39s, compared with the existing white blood cell classification Algorithm, CNN classification method not only has obvious advantages, but also has a lot of room for improvement.

一种基于改进K-means及卷积神经网络的白细胞提取和分类 方法A white blood cell extraction and classification based on improved K-means and convolutional neural network method

技术领域technical field

本发明涉及医学图像分割提取技术领域，特别是涉及一种基于改进K-means及卷积神经网络的白细胞提取和分类方法。The invention relates to the technical field of medical image segmentation and extraction, in particular to a white blood cell extraction and classification method based on improved K-means and convolutional neural network.

背景技术Background technique

在医学上，白细胞是人体免疫系统重要的组成部分，负责识别并吞噬非正常细胞。血常规检查中白血细胞的传统分类计数和形态分析依赖于人工计数和血液检查的专家分析，效率低且具有较强主观性。目前常用的是流式细胞仪，也不能实现白细胞自动分类，并且在临床应用中具有局限性。In medicine, white blood cells are an important part of the human immune system, responsible for identifying and engulfing abnormal cells. The traditional differential counting and morphological analysis of white blood cells in blood routine examination rely on manual counting and expert analysis of blood tests, which is inefficient and highly subjective. Currently, flow cytometry is commonly used, which cannot automatically classify white blood cells, and has limitations in clinical application.

近年来，为了更好地分割图像和识别出白细胞，研究者相继提出了一些效果较好的算法来实现白细胞的精确分割和分类算法，但是白细胞分割仍然还存在着问题，这些问题主要来源于图像色彩亮度不一，图像中存在杂质，白细胞形状多种多样，染色后细胞质与红细胞的颜色相近。现有的方法所能实现的分割精度还不能达到临床实际需要，因此在白细胞分割领域还有许多工作需要进行。In recent years, in order to better segment images and identify white blood cells, researchers have successively proposed some algorithms with better effects to realize accurate segmentation and classification algorithms of white blood cells, but there are still problems in white blood cell segmentation, which mainly come from image The color brightness is different, there are impurities in the image, the shape of white blood cells is various, and the color of cytoplasm and red blood cells after staining is similar. The segmentation accuracy achieved by the existing methods can not meet the actual clinical needs, so there is still a lot of work to be done in the field of white blood cell segmentation.

发明内容Contents of the invention

本发明的目的在于提供一种基于改进K-means及卷积神经网络的白细胞提取和分类方法，该方法可以有效地提取白细胞并且分割精度较高，最后利用卷积神经网络(CNN)进行白细胞分类和识别。The purpose of the present invention is to provide a leukocyte extraction and classification method based on improved K-means and convolutional neural network, which can effectively extract leukocytes and have high segmentation accuracy, and finally utilize convolutional neural network (CNN) to carry out leukocyte classification and identification.

为实现上述目的，本发明的技术方案是：一种基于改进K-means及卷积神经网络的白细胞提取和分类方法，包括如下步骤：To achieve the above object, the technical solution of the present invention is: a leukocyte extraction and classification method based on improved K-means and convolutional neural network, comprising the following steps:

步骤S1、在提取白细胞之前，先进行颜色空间的分解，采用有利于白细胞分割的颜色分量；Step S1, before extracting white blood cells, decompose the color space first, and use color components that are beneficial to white blood cell segmentation;

步骤S2、根据细胞图像灰度分布对K-means聚类算法进行改进，选定初始聚类中心，使图像中所有像素按就近原则进行初始聚类，并对FWSA-KM算法的欧式距离进行改进，使聚类算法的鲁棒性提高；Step S2, improve the K-means clustering algorithm according to the gray distribution of the cell image, select the initial clustering center, make all pixels in the image perform initial clustering according to the nearest principle, and improve the Euclidean distance of the FWSA-KM algorithm , to improve the robustness of the clustering algorithm;

步骤S3、采用改进的K-means算法进行细胞核和细胞质的提取；Step S3, using the improved K-means algorithm to extract the nucleus and cytoplasm;

步骤S4、采用分水岭算法来分离复杂粘连的白细胞部分；Step S4, using the watershed algorithm to separate the white blood cells with complex adhesions;

步骤S5、采用卷积神经网络对提取分离的白细胞进行实验，实现粘连白细胞的识别。Step S5, using the convolutional neural network to conduct experiments on the extracted and separated leukocytes to realize the identification of the adhered leukocytes.

在本发明一实施例中，所述步骤S1具体实现如下：In an embodiment of the present invention, the step S1 is specifically implemented as follows:

步骤S11、建立彩色模型：对白细胞染色处理，使其在色调分量(H)空间和饱和分量(S)空间对应的细胞质区域以及白细胞区域都与背景图像存在着较强的对比度；Step S11, establishing a color model: staining the white blood cells so that the cytoplasmic regions and white blood cell regions corresponding to the hue component (H) space and the saturation component (S) space have a strong contrast with the background image;

步骤S12、在后续分割中，设置饱和分量空间和色调分量空间中阈值，将白细胞的细胞核区域以及白细胞区域粗略地从细胞图像中提取出来。Step S12 , in the subsequent segmentation, setting thresholds in the saturation component space and the hue component space, and roughly extracting the nucleus area and the white blood cell area of the white blood cell from the cell image.

在本发明一实施例中，所述步骤S2具体实现如下：In an embodiment of the present invention, the step S2 is specifically implemented as follows:

步骤S21、对细胞图像灰度进行直方图分布统计来选定初始聚类中心，使得图像所有像素按就近原则进行初始聚类；Step S21, performing histogram distribution statistics on the gray scale of the cell image to select the initial clustering center, so that all pixels of the image are initially clustered according to the principle of proximity;

步骤S22、基于非欧氏距离，改进特征权重，对K-means进行改进，得到Improved-KM聚类算法，即将该算法目标函数中的χ_ik和v_jk的欧氏距离|χ_ik-v_jk|修改为非欧氏距离

由此得到改进的Improved-KM聚类算法的目标函数为：

Step S22, based on the non-Euclidean distance, improve the feature weight, improve K-means, and obtain the Improved-KM clustering algorithm, that is, the Euclidean distance between χ _ik and v _jk in the algorithm's objective function | χ _ik -v _jk |Modified to non-Euclidean distance

The objective function of the improved Improved-KM clustering algorithm is:

在本发明一实施例中，所述步骤S22中，Improved-KM聚类算法步骤如下：In an embodiment of the present invention, in the step S22, the steps of the Improved-KM clustering algorithm are as follows:

步骤S221、目标函数：

其中U＝(u_ij)_n×c是隶属度矩阵；如果第i个数据点x_i属于第j个类，则u_ij＝1，否则u_ij＝0，并且

而V＝[v₁,v₂,…,v_c]是c个聚类中心构成的矩阵；同时式子满足：Step S221, objective function:

Where U=(u _ij ) _n×c is the membership degree matrix; if the i-th data point x _i belongs to the j-th class, then u _ij =1, otherwise u _ij =0, and

And V=[v ₁ ,v ₂ ,…,v _c ] is a matrix composed of c cluster centers; at the same time, the formula satisfies:

步骤S222、最优隶属度矩阵

和聚类中心矩阵

中的元素为：

和

Step S222, optimal membership degree matrix

and cluster center matrix

The elements in are:

and

步骤S223、通过迭代求解三个最小化问题；Step S223, solving three minimization problems through iteration;

步骤S224：令

和

则a_k度量了聚类在第k维特征上总的类内紧致性，b_k度量了聚类在第k维特征上总的类间分离性度量；Step S224: make

and

Then a _k measures the total intra-class compactness of the clustering on the k-th dimension feature, and b _k measures the total inter-class separation measure of the clustering on the k-th dimension feature;

步骤S225、用新的目标函数

求解如下特征权重矩阵：Step S225, using a new objective function

Solve the following feature weight matrix:

其中满足：

Which satisfies:

步骤S226、设

为第t步迭代的特征权重，则下式可表示为第t+1步的特征权重：

其中特征权重调节差量如下：Step S226, set

is the feature weight of the t-th step iteration, then the following formula can be expressed as the feature weight of the t+1-th step:

The feature weight adjustment difference is as follows:

为了使

满足约束条件

对特征权重公式进行规范化处理，得到特征权重

because

meet the constraints

Normalize the feature weight formula to get the feature weight

在本发明一实施例中，所述步骤S3具体实现如下：In an embodiment of the present invention, the step S3 is specifically implemented as follows:

步骤S31、白细胞的提取：通过观察不同细胞图像的不同色彩模型的不同分量来选择，再进行聚类分割；Step S31, extraction of white blood cells: select by observing different components of different color models of different cell images, and then perform clustering and segmentation;

步骤S32、细胞核的提取：利用聚类后的二值图像含有大量的噪声且噪声普遍面积小的特点计算出每个连通区域的面积，剔除小于阈值的连通区域面积，填充在细胞核区域之中小的孔洞；Step S32, cell nucleus extraction: calculate the area of each connected region by using the characteristic that the clustered binary image contains a lot of noise and the noise generally has a small area, remove the connected region area smaller than the threshold, and fill in the small ones in the nucleus region holes;

步骤S33、细胞质的提取：将提取出的白细胞减去白细胞核的方法来得到细胞质部分，再进行重新恢复彩色像。Step S33, cytoplasmic extraction: subtracting the white blood cell nucleus from the extracted white blood cells to obtain the cytoplasmic part, and then restore the color image.

在本发明一实施例中，所述步骤S4具体实现如下：In an embodiment of the present invention, the step S4 is specifically implemented as follows:

步骤S41、对于有粘连在一起的白细胞，进行去噪和孔洞填充处理；Step S41, performing denoising and hole filling processing for the white blood cells that have adhered together;

步骤S42、用分水岭分割算法将粘连在一起的白细胞进行分割处理。Step S42, using the watershed segmentation algorithm to segment the cohesive white blood cells.

在本发明一实施例中，所述步骤S5中，所述卷积神经网络由输入层、卷积层、采样层、连接层和输出层构成；输入层输入需要分类的图像，由卷积层提取出相应的特征，为加速学习速度通过采样层下采样减少需要神经元个数同时保留有用信息，连接层通过激活函数将分类结果输入至输出层，输出层的维度为所需分类的类别数。In one embodiment of the present invention, in the step S5, the convolutional neural network is composed of an input layer, a convolutional layer, a sampling layer, a connection layer, and an output layer; the input layer inputs images to be classified, and the convolutional layer The corresponding features are extracted. In order to speed up the learning speed, the number of neurons required is reduced by downsampling in the sampling layer while retaining useful information. The connection layer inputs the classification results to the output layer through the activation function, and the dimension of the output layer is the number of categories to be classified. .

相较于现有技术，本发明具有以下有益效果：本发明方法使得白细胞细胞核和细胞质分割精度分别为95.81％和91.28％，较传统分割方法有较大提高；分类准确度最大能达到98.96％，分类平均时间为0.39s，相对于现有的白细胞分类算法，CNN分类方法不仅有明显优势，同时还有很大的提升空间Compared with the prior art, the present invention has the following beneficial effects: the method of the present invention makes the segmentation accuracy of leukocyte nucleus and cytoplasm respectively 95.81% and 91.28%, which is greatly improved compared with the traditional segmentation method; the maximum classification accuracy can reach 98.96%, The average classification time is 0.39s. Compared with the existing white blood cell classification algorithm, the CNN classification method not only has obvious advantages, but also has a lot of room for improvement

附图说明Description of drawings

图1为本发明方法流程图。Fig. 1 is a flow chart of the method of the present invention.

图2为本发明卷积神经网络的构建图。Fig. 2 is a construction diagram of the convolutional neural network of the present invention.

具体实施方式detailed description

下面结合附图，对本发明的技术方案进行具体说明。The technical solution of the present invention will be specifically described below in conjunction with the accompanying drawings.

如图1所示，本发明提供了一种基于改进K-means及卷积神经网络的白细胞提取和分类方法，包括如下步骤：As shown in Figure 1, the present invention provides a kind of leukocyte extraction and classification method based on improved K-means and convolutional neural network, comprises the following steps:

步骤S1、在提取白细胞之前，先进行颜色空间的分解，采用有利于白细胞分割的颜色分量；具体如下：Step S1, before extracting white blood cells, decompose the color space first, and use color components that are beneficial to white blood cell segmentation; details are as follows:

步骤S11、建立彩色模型：对白细胞染色处理，使其在色调分量(H)空间和饱和分量(S)空间对应的细胞质区域以及白细胞区域都与背景图像存在着较强的对比度；Step S11, establishing a color model: staining the white blood cells so that the cytoplasmic regions and white blood cell regions corresponding to the hue component (H) space and the saturation component (S) space have a strong contrast with the background image;

步骤S12、在后续分割中，设置饱和分量空间和色调分量空间中阈值，将白细胞的细胞核区域以及白细胞区域粗略地从细胞图像中提取出来；Step S12, in the subsequent segmentation, setting thresholds in the saturation component space and the hue component space, and roughly extracting the nucleus area and the leukocyte area of the leukocyte from the cell image;

步骤S2、根据细胞图像灰度分布对K-means聚类算法进行改进，选定初始聚类中心，使图像中所有像素按就近原则进行初始聚类，并对FWSA-KM算法的欧式距离进行改进，使聚类算法的鲁棒性提高；具体如下：Step S2, improve the K-means clustering algorithm according to the gray distribution of the cell image, select the initial clustering center, make all pixels in the image perform initial clustering according to the nearest principle, and improve the Euclidean distance of the FWSA-KM algorithm , so that the robustness of the clustering algorithm is improved; the details are as follows:

步骤S21、对细胞图像灰度进行直方图分布统计来选定初始聚类中心，使得图像所有像素按就近原则进行初始聚类；Step S21, performing histogram distribution statistics on the gray scale of the cell image to select the initial clustering center, so that all pixels of the image are initially clustered according to the principle of proximity;

步骤S22、基于非欧氏距离，改进特征权重，对K-means进行改进，得到Improved-KM聚类算法，即将该算法目标函数中的χ_ik和v_jk的欧氏距离|χ_ik-v_jk|修改为非欧氏距离

由此得到改进的Improved-KM聚类算法的目标函数为：

Step S22, based on the non-Euclidean distance, improve the feature weight, improve K-means, and obtain the Improved-KM clustering algorithm, that is, the Euclidean distance between χ _ik and v _jk in the algorithm's objective function | χ _ik -v _jk |Modified to non-Euclidean distance

The objective function of the improved Improved-KM clustering algorithm is:

所述步骤S22中，Improved-KM聚类算法步骤如下：In the step S22, the steps of the Improved-KM clustering algorithm are as follows:

步骤S221、目标函数：

其中U＝(u_ij)_n×c是隶属度矩阵；如果第i个数据点x_i属于第j个类，则u_ij＝1，否则u_ij＝0，并且

而V＝[v₁,v₂,…,v_c]是c个聚类中心构成的矩阵；同时式子满足：Step S221, objective function:

Where U=(u _ij ) _n×c is the membership degree matrix; if the i-th data point x _i belongs to the j-th class, then u _ij =1, otherwise u _ij =0, and

And V=[v ₁ ,v ₂ ,…,v _c ] is a matrix composed of c cluster centers; at the same time, the formula satisfies:

步骤S222、最优隶属度矩阵

和聚类中心矩阵

中的元素为：

和

Step S222, optimal membership degree matrix

and cluster center matrix

The elements in are:

and

步骤S223、通过迭代求解三个最小化问题；Step S223, solving three minimization problems through iteration;

步骤S224：令

和

则a_k度量了聚类在第k维特征上总的类内紧致性，b_k度量了聚类在第k维特征上总的类间分离性度量；Step S224: make

and

Then a _k measures the total intra-class compactness of the clustering on the k-th dimension feature, and b _k measures the total inter-class separation measure of the clustering on the k-th dimension feature;

步骤S225、用新的目标函数

求解如下特征权重矩阵：Step S225, using a new objective function

Solve the following feature weight matrix:

其中满足：

Which satisfies:

步骤S226、设

为第t步迭代的特征权重，则下式可表示为第t+1步的特征权重：

其中特征权重调节差量如下：Step S226, set

is the feature weight of the t-th step iteration, then the following formula can be expressed as the feature weight of the t+1-th step:

The feature weight adjustment difference is as follows:

为了使

满足约束条件

对特征权重公式进行规范化处理，得到特征权重

because

meet the constraints

Normalize the feature weight formula to get the feature weight

步骤S3、采用改进的K-means算法进行细胞核和细胞质的提取；具体如下：Step S3, using the improved K-means algorithm to extract the nucleus and cytoplasm; details are as follows:

步骤S31、白细胞的提取：通过观察不同细胞图像的不同色彩模型的不同分量来选择，再进行聚类分割；Step S31, extraction of white blood cells: select by observing different components of different color models of different cell images, and then perform clustering and segmentation;

步骤S32、细胞核的提取：利用聚类后的二值图像含有大量的噪声且噪声普遍面积小的特点计算出每个连通区域的面积，剔除小于阈值的连通区域面积，填充在细胞核区域之中小的孔洞；Step S32, cell nucleus extraction: calculate the area of each connected region by using the characteristic that the clustered binary image contains a lot of noise and the noise generally has a small area, remove the connected region area smaller than the threshold, and fill in the small ones in the nucleus region holes;

步骤S33、细胞质的提取：因为细胞质的颜色往往和白细胞或者背景图像的颜色相似，所以极难分割；因此采用将提取出的白细胞减去白细胞核的方法来得到细胞质部分，再进行重新恢复彩色像Step S33, cytoplasm extraction: because the color of cytoplasm is often similar to the color of white blood cells or the background image, it is extremely difficult to segment; therefore, the cytoplasm part is obtained by subtracting the white blood cell nucleus from the extracted white blood cells, and then the color image is restored

步骤S4、采用分水岭算法来分离复杂粘连的白细胞部分；具体如下：Step S4, using the watershed algorithm to separate the white blood cells with complex adhesions; the details are as follows:

步骤S41、对于有粘连在一起的白细胞，进行去噪和孔洞填充处理；Step S41, performing denoising and hole filling processing for the white blood cells that have adhered together;

步骤S42、用分水岭分割算法将粘连在一起的白细胞进行分割处理。Step S42, using the watershed segmentation algorithm to segment the cohesive white blood cells.

步骤S5、采用卷积神经网络对提取分离的白细胞进行实验，实现粘连白细胞的识别；如图2所示，所述卷积神经网络由输入层、卷积层、采样层、连接层和输出层构成；输入层输入需要分类的图像，由卷积层提取出相应的特征，为加速学习速度通过采样层下采样减少需要神经元个数同时保留有用信息，连接层通过激活函数(sigmoid函数等)将分类结果输入至输出层，输出层的维度为所需分类的类别数。Step S5, using the convolutional neural network to conduct experiments on the extracted and separated white blood cells to realize the identification of adherent white blood cells; as shown in Figure 2, the convolutional neural network consists of an input layer, a convolutional layer, a sampling layer, a connection layer and an output layer Composition; the input layer inputs the image that needs to be classified, and the corresponding features are extracted by the convolutional layer. In order to accelerate the learning speed, the number of neurons is reduced by downsampling the sampling layer while retaining useful information. The connection layer passes the activation function (sigmoid function, etc.) The classification result is input to the output layer, and the dimension of the output layer is the number of categories to be classified.

以上是本发明的较佳实施例，凡依本发明技术方案所作的改变，所产生的功能作用未超出本发明技术方案的范围时，均属于本发明的保护范围。The above are the preferred embodiments of the present invention, and all changes made according to the technical solution of the present invention, when the functional effect produced does not exceed the scope of the technical solution of the present invention, all belong to the protection scope of the present invention.