CN103353938B - A kind of cell membrane dividing method based on hierarchical level feature - Google Patents

Abstract

The invention relates to a cell membrane segmentation method based on hierarchical features. The method fully combines the advantages of pixel-level features and super-pixel-level features to propose hierarchical features, which can adaptively describe local context information and effectively identify cell membranes. And the complex microstructure around it, which in turn improves the recognition performance of cell membrane segmentation based on supervised learning methods. The specific process is as follows: I. In the preprocessing stage, image enhancement is performed first, and then over-segmentation is performed to obtain superpixels; II. Selection of typical sample points based on superpixels III. Feature extraction a. Pixel-level feature extraction b. Superpixel-level feature extraction c. Hierarchical feature extraction IV. Classifier classification V. post-processing stage.

Description

A Cell Membrane Segmentation Method Based on Hierarchical Features

technical field

The invention relates to the field of cell membrane segmentation under electron microscope images, in particular to a cell membrane segmentation method based on hierarchical features.

Background technique

In order to better study the memory and recognition mechanism of the brain, brain scientists need to reconstruct the central nervous system in three-dimensional space, and the first step is the segmentation of nerve cells in two-dimensional space. Obviously, the accuracy of 2D space segmentation will directly affect the reconstruction effect of 3D space. In addition, the segmentation of nerve cells also has important clinical medical value. For example, studies have found that the decline of the retina is caused by the reduction of sensory cells in nerve cells. In recent years, more and more machine learning methods based on supervised learning have been used in the segmentation of cell membranes, and have achieved good recognition results. This method is mainly divided into two steps: feature extraction and pattern recognition. The quality of feature extraction will ultimately affect the recognition performance of the pattern recognition system.

Nerve cells have intricate topological structures, and the shapes and sizes of internal organelles are different. In addition, the grayscale of microscopic images, blurred boundaries, and the influence of noise, etc., make full use of the context information of each pixel when identifying cell membranes appears to be particularly important.

When the existing technology recognizes the cell membrane based on the supervised learning method, it often only considers the information of a single pixel in the feature extraction stage, or when considering the context information, it simply uses a square window with a fixed size and shape, which cannot fully describe the local complex and changeable The microstructure greatly reduces the recognition rate.

Contents of the invention

In order to overcome the deficiencies of the above-mentioned prior art, the present invention proposes a cell membrane segmentation method based on hierarchical features, which fully combines the advantages of pixel-level features and super-pixel-level features, and proposes hierarchical features, which can be described adaptively Local context information can effectively identify the cell membrane and its surrounding complex microstructures, thereby improving the recognition performance of cell membrane segmentation based on supervised learning methods.

To achieve the above object, the present invention adopts the following technical solutions:

A cell membrane segmentation method based on hierarchical features, the specific process is as follows:

I． preprocessing stage

First perform image enhancement, and then perform image over-segmentation to obtain superpixels;

II. Selection of Typical Sample Points Based on Superpixels

III. feature extraction

a. Pixel-level feature extraction

b. Superpixel-level feature extraction

c. Hierarchical feature extraction

IV. Classifier Classification

V． post-processing stage.

In the above-mentioned I, image enhancement is carried out by using histogram equalization and Gaussian filtering technology; an improved method based on k-means proposed by RadhakrishnaAchanta et al., that is, Simple Linear Iterative Clustering (SLIC) to obtain super pixel, this method divides the entire image into superpixels with relatively uniform shape and size by synthesizing the color information and position information of local adjacent pixels; the simple linear iterative clustering method only needs to input the number of expected superpixels to divide The entire image is over-segmented into superpixels with a specified segmentation granularity.

In the above II, typical sample point selection is to randomly select three sample points in each superpixel to represent all sample points in the superpixel.

In said III, for each sample point, a total of 56-dimensional features are extracted, including 28-dimensional pixel-level features and 28-dimensional superpixel-level features; where

a Pixel-level feature extraction is:

Divide the extracted 28-dimensional pixel-level features into three categories: (1) 5-dimensional neighborhood information is 3×3 filter features: mean filter, median filter, minimum value filter, maximum value filter, variance filter; ( 2) Gaussian filter with Sigma values of 2.0, 3.5, and 4.0, Sobel filter with kernel function of 3×3, convolution filter with kernel function of 5×5, maximum and minimum eigenvalues of Hessian matrix, Laplacian operator , second-order derivative in the horizontal direction, second-order derivative in the vertical direction, the above features have a total of 10 dimensions; (3) the 12-dimensional shape descriptor Rays feature, and the composite feature Radon-like feature that combines texture and geometry in one dimension;

bSuperpixel-level feature extraction

First define and extract various pixel-level features of all pixels in a superpixel, and then use statistical methods to use the statistical features of all pixels in the superpixel as the characteristics of the superpixel for subsequent processing; by calculating The average value of the pixel-level features of all pixels in the superpixel is used as the superpixel-level feature, so there are 28-dimensional superpixel-level features;

c level feature extraction

Combine the pixel-level features obtained in step a and the superpixel-level features obtained in step b, and the two levels of feature vectors are spliced into one feature vector to form a 28+28=56-dimensional feature vector, and then combine the two levels of features , as a hierarchical level feature.

In the IV, the classifier adopts a random forest classifier.

6. the cell membrane segmentation method based on hierarchical features as claimed in claim 1, is characterized in that, in said V, on the basis of random forest return probability value, implement Ridler, the IsoData threshold segmentation method that TW and Calvard propose; For some isolated regions that are difficult to identify, the removal operation of the isolated region is performed, wherein the removal criterion of the region is a series of threshold operations based on the region attribute. The region attributes used include the region area, the Euler function, the eccentricity (Eccentricity) of the ellipse with the same standard second-order central moment as the region, and the pixel ratio (Solidity) in both the region and its minimum convex polygon.

The beneficial effects of the present invention are: a method for cell membrane segmentation based on hierarchical features is proposed, and hierarchical features are obtained by combining pixel-level features with rich feature expressions and superpixel-level features with certain semantic features. Then, the hierarchical features are used to train a random forest for cell membrane segmentation. Compared with the contextual information in the neighborhood of the original square window with fixed size and shape, the hierarchical features can adaptively adjust the neighborhood information around a single pixel, which better describes the local intricate microstructure and improves the efficiency of the application. Recognition Performance of a Pattern Recognition System for Cell Membrane Segmentation. At the same time, a typical sample point selection method based on superpixels is proposed, which simplifies the sample space and reduces the redundancy between samples.

Description of drawings

Figure 1a is a neighborhood graph based on a fixed square window.

Figure 1b is a superpixel-based neighborhood map.

Fig. 2 is a flowchart of the present invention.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

In Fig. 1a, Fig. 1b and Fig. 2, the present invention is mainly divided into five stages. For the sake of simplicity, when selecting sample points based on superpixels in the figure, only one sample point is selected in a superpixel to represent the entire superpixel. all sample points.

1. Preprocessing stage

In order to make the gray level of the image uniform, and at the same time strengthen the continuity of the cell membrane. The present invention uses histogram equalization and Gaussian filter technology to enhance the image, and then performs over-segmentation to obtain superpixels on the basis of the enhanced image. The present invention adopts a simple linear iterative clustering method (Simple Linear Iterative Clustering, SLIC) based on an improved k-means method to obtain superpixels. This method divides the image into small clusters with relatively uniform shape and size by integrating local color information and position information. Moreover, the method is simple and easy to use, and has high execution efficiency. When this method is used specifically, only the number of expected superpixels can be input to generate superpixels with a specified segmentation granularity.

2. Typical sample point selection method based on superpixels

Facing a large database of microscopic images, training a pixel-level classifier places high demands on both time and space. A superpixel is a collection of locally homogenous pixels, essentially a cluster of locally adjacent similar pixels. Utilizing this characteristic of superpixels, in order to simplify the sample space, the present invention selects sample points based on superpixels. Specifically, some sample points are randomly selected in each superpixel to represent all sample points in the entire superpixel. In the present invention, three sample points are randomly selected in each superpixel to represent all sample points in the superpixel.

3. Feature extraction: For each sample point, a total of 56-dimensional features are extracted, including 28-dimensional pixel-level features and 28-dimensional super-pixel-level features.

a. Pixel-level feature extraction

For pixel-level feature descriptors, the present invention extracts a total of 28-dimensional pixel-level features, which can be roughly divided into three categories: (1) 5-dimensional neighborhood information is a 3×3 filter feature: mean filter, medium Value filtering, min filtering, max filtering, variance filtering. (2) Gaussian filter with Sigma values of 2.0, 3.5, and 4.0, Sobel filter with kernel function of 3×3, convolution filter with kernel function of 5×5, maximum and minimum eigenvalues of Hessian matrix, Laplace calculation Son, the second derivative in the horizontal direction, the second derivative in the vertical direction, the above features have a total of 10 dimensions. (3) The 12-dimensional shape descriptor Rays feature, and the composite feature Radon-likefeature, which combines texture and geometry, is one-dimensional.

b. Superpixel-level feature extraction

While strengthening the local homogeneity of the image, superpixels can better preserve the original boundary of the image. Superpixel is a kind of local feature descriptor, which itself contains certain semantic features. In the current state of the art, there are very few related studies on superpixel-level features. From the perspective of collection, the present invention regards superpixels as a collection of local homogeneous pixel points, and uses relevant statistical means to obtain superpixel features. The present invention firstly defines and extracts various pixel-level features of all pixels in a superpixel, and then uses statistical means to use the statistical features of all pixels in the superpixel as the features of the superpixel for subsequent processing. In this way, all feature expressions that can be used at the pixel level can be used to express superpixels, thereby greatly enriching the feature expression of superpixels and improving the ability to distinguish. Considering both simplicity and efficiency, the present invention only obtains the mean value of the pixel-level features of all pixels in the superpixel as the superpixel-level features, so there are 28 superpixel-level features in total.

c. Hierarchical feature extraction

The pixel-level feature is a low-level feature, and the super-pixel-level feature is a middle-level feature. In the present invention, the two-level feature vectors of pixel-level features and super-pixel-level features are spliced into one feature vector to form a 28+28=56-dimensional feature vector, and then the two levels of features are combined as a Hierarchy-level features. Under the hierarchical feature, for each sample point, the surrounding context information is adaptive based on superpixels, which can better describe the intricate microstructure around the cell membrane pixel sample point.

4. Classifier classification

For a pattern recognition system, after defining and extracting a series of more targeted features, it is also very important to choose a good classifier. Random forest is an integrated learning method, which is easy to use, and this method can effectively deal with high-dimensional features without explicit feature selection, can handle large sample data, fast classification speed, strong anti-noise ability and non-destructive prone to overfitting. In view of the above points, the present invention selects random forest as the classifier.

5. Post-processing

Based on the probability values returned by the random forest, automatic threshold segmentation is performed. For some isolated areas that are difficult to identify, the operation of removing isolated areas is simply performed.

Claims (6)

1. based on a cell membrane dividing method for hierarchical level feature, it is characterized in that, described hierarchical level is characterized as the proper vector of the proper vector splicing of these two ranks of Pixel-level characteristic sum super-pixel level feature; Detailed process is as follows:

I. pretreatment stage;

First carry out image enhaucament, then perform over-segmentation and obtain super-pixel;

II. based on the typical sample point selection of super-pixel;

III. feature extraction;

A. Pixel-level feature extraction;

B. super-pixel level feature extraction;

C. hierarchical level feature extraction;

IV. sorter classification;

V. post-processing stages.

2. as claimed in claim 1 based on the cell membrane dividing method of hierarchical level feature, it is characterized in that, in described I, use histogram equalization and gaussian filtering technology to carry out image enhaucament; Adopt simple linear iterative clustering methods to obtain super-pixel, entire image is divided into shape, the relative homogeneous super-pixel of size one by one by the colouring information of comprehensive local vicinity points with positional information by the method; Entire image over-segmentation just can be the super-pixel of specifying segmentation granularity by the number that simple linear iterative clustering methods only need input expection super-pixel.

3., as claimed in claim 1 based on the cell membrane dividing method of hierarchical level feature, it is characterized in that, in described II, typical sample point selection be in each super-pixel Stochastic choice three sample points to represent the whole sample points in this super-pixel.

4., as claimed in claim 1 based on the cell membrane dividing method of hierarchical level feature, it is characterized in that, in described III, for each sample point, be extracted the feature of 56 dimensions altogether, comprising the Pixel-level feature of 28 dimensions, the super-pixel level feature of 28 dimensions; Wherein

The feature extraction of a Pixel-level is:

The Pixel-level feature of extract 28 dimensions is divided three classes: the neighborhood information of (1) 5 dimension is all 3 × 3 filtering characteristics: mean filter, medium filtering, mini-value filtering, maximal value filtering, variance filter; (2) Sigma value is respectively 2.0,3.5, the Gaussian filtering of 4.0, kernel function is the Sobelfilter of 3 × 3, and kernel function is the convolutional filtering of 5 × 5, and Hassian matrix is maximum, minimal eigenvalue, Laplace operator, horizontal direction second derivative, vertical direction second derivative, above feature is totally 10 dimensions; The shape descriptor Rays feature of (3) 12 dimensions, and combine the compound characteristics Radon-likefeature one dimension of texture and geometric configuration;

The feature extraction of b super-pixel level:

First define and extract the various Pixel-level features of all pixels in a super-pixel, then by statistical method, by the statistical nature of pixels all in this super-pixel, as the feature of this super-pixel, for subsequent processes; By asking for the feature of average as super-pixel level of all pixel Pixel-level features in super-pixel, thus totally 28 dimension super-pixel level features;

The feature extraction of c hierarchical level:

By a step gained Pixel-level characteristic sum b step gained super-pixel level feature, the proper vector of these two ranks is spliced into a proper vector, form the proper vector of a 28+28=56 dimension, and then the integrate features of two kinds of ranks is got up, as a kind of feature of stratum level.

5. as claimed in claim 1 based on the cell membrane dividing method of hierarchical level feature, it is characterized in that, in described IV, sorter adopts random forest sorter.

6. as claimed in claim 5 based on the cell membrane dividing method of hierarchical level feature, it is characterized in that, in described V, return on the basis of probable value at random forest, perform IsoData threshold segmentation method; For the impalpable isolated area of part, what perform isolated area removes operation, and wherein the criterion that removes in region is a series of threshold operation based on area attribute; The area attribute used has region area, Euler's function, has the eccentricity of the ellipse of identical standard second-order moment around mean with region, and the pixel ratio simultaneously in region and its minimal convex polygon.