Feature extraction algorithm for medical CT images--An algorithm for CT image feature extraction of pulmonary nodules

　Excerpt from my graduation thesis "Study on CT image feature extraction algorithm of pulmonary nodules"

The feature extraction of medical image can be considered as the necessary feature based on image content extraction, and what characteristics need to be extracted in medical image based on research needs. In contrast, medical image feature extraction is more demanding because it plays an important role in the diagnosis of doctors, so it needs rigorous and reliable characteristics. CT image feature extraction of pulmonary nodules is also a part of medical image feature extraction, which has the basic requirement of medical image feature extraction. There are other methods to extract the features of medical images, as well as specific feature extraction methods for pulmonary nodules. This section mainly introduces some commonly used CT image feature extraction methods of pulmonary nodules, which can be divided into gray characteristics, texture features and shape features .

3.1 Gray feature extraction algorithm

Medical images are basically gray-scale images, and the focus of this study is to extract CT images of pulmonary nodules, and also to grayscale images. Gray feature is a kind of image feature commonly used in medical image processing, which is simple and effective. The gray characteristic of medical image is that it has certain stability by using gray scale, is insensitive to size and direction, and can show quite strong robustness [17]. Medical images are basically gray-scale images, gray-scale features are also a kind of important features, through the gray analysis of the image, the gray histogram of the image, which can extract a lot of effective gray features. The gray histogram of medical image shows a variety of statistical characteristics of the gray distribution of medical image.

Grayscale histogram algorithm [18]:

A grayscale image is represented by F (x, Y), and the value at (x, y) represents the location pixel. A graph has m pixels, distributed in the 0-l-1 level, the gray histogram is the statistics of each level of gray-scale pixel bibliographic statistics obtained by gray-scale histogram to obtain a series of gray characteristics. The grayscale histogram is defined as

Wherein, I represents the gray level, L represents the number of gray level, the number of pixels of gray level I, M represents the total image pixel, the formula (3-1) represents each gray level of the total pixel percentage, that is, the gray level of the pixel frequency of I.

3.2 Texture Feature Extraction method

Texture feature is a kind of human vision can obviously feel the characteristics, but also is an important feature of the image, the main expression of pixels in the space distribution model description, can reflect the image representation of the surface roughness, smoothness, particle size, randomness and other properties. Texture features have regular texture and quasi-rule texture, regular texture refers to the texture of the basic elements are arranged according to certain rules, become artificial texture; quasi-rule textures are random textures that do not have a particular shape or rule, but are distributed according to Gray or color features, which are difficult to identify, [19] It can be seen only by observing the entire object, called the natural texture.

There are three types of texture feature extraction methods, statistical analysis method, structural analysis method and spectrum analysis method. [20] This section mainly describes the three types of feature extraction methods of the more typical algorithm, the content is as follows:

3.2.1 Statistical analysis method

Statistical class method is based on the statistical characteristics of local area of texture image to measure the spatial distribution of pixels, and in 1973, the Gray-scale co-occurrence matrix (GLCM) proposed by Haralick et [21] has become a very classical and effective algorithm for texture feature extraction, which has been widely used in many fields. Based on the gray-level symbiosis matrix, many statistics can be extracted as texture features. In this paper, gray-scale co-occurrence matrix is introduced as a typical application of statistical analysis method, which mainly extracts energy, contrast, correlation coefficient, entropy and local stability (deficit moment).

3.1, wherein the range has a pixel pitch d and the direction θ two parameters, set F (x, Y) represents a grayscale image, to any area of the image of R, defines s as a set of pixel pairs with a specific spatial linkage in the region, then the elements of the grayscale symbiosis matrix are expressed as [22]:

　　in order to reduce the amount of computation, the image is generally first gray-level transformation, reduce the gray scale, while reducing the number of directions, usually take 0o, 45o, 90o, 135o four direction. The Symbiosis Matrix embodies the difference between different textures, and its symbiosis matrix is obviously different for image textures with different characteristics. For example, the horizontal direction of a pixel of the gray-level joint distribution, if the main diagonal element of the symbiosis matrix is 0, it indicates that the horizontal direction of the adjacent two elements do not have the same value, that is, the texture is more delicate; weak main diagonal elements are large, it indicates that the horizontal direction texture is coarser .

Several major image texture features extracted from the gray-scale symbiosis matrix are defined as follows:

3.2.2 Structural Analysis method

The structure method is the first method to analyze the image texture feature, but it is seldom researched and used because the method is not effective. The structure method is to analyze the grain and permutation rules from the structural angle of the texture image, the first step, the extraction of the grain element, the second step, the arrangement rules of the grain elements are inferred.

3.2.3 Spectrum Analysis Method

Fourier transform is a bridge between time domain and frequency domain, which provides the possibility for frequency domain analysis. Texture features are not only represented in the time domain, but also have many texture features in the frequency domain, for example, the energy spectrum of the transformed image is an important and simple frequency domain feature [22].

3.3 Morphological Feature Extraction method

The morphological features are relative to the texture features, which is a kind of kind of characteristic which can be described by geometric characters. The morphological features of images can be divided into two categories, one is based on the morphological features of contour, and the other is based on the morphological characteristics of regions. Morphological features in Medical image processing field is also a kind of more commonly used feature extraction method, medical diagnosis and so on have great help.

3.3.1 Contour-based morphological features

Contour-based Morphological feature description method is only the contour information of extracting shape, there are two kinds of methods: continuous (global type) and discrete (structure type). In this paper, we mainly study two kinds of commonly used morphological feature extraction methods: Simple geometric descriptor and Fourier shape descriptor [23].

(1) Simple geometric descriptor algorithm:

A. Boundary Length: 4-directional connected boundary or 8-directional connected boundary is generally used to get an approximate length

B. Boundary diameter: the distance between two points farthest from the boundary

C. Curvature: Curvature is the rate of change of the slope, describing the situation of the change of the boundary direction

(2) Fourier morphological descriptor algorithm:

The Fourier shape Description method (Fourier shape descriptions) is a widely used shape descriptor, assuming that an object contour consists of a series of pixels of coordinates, where 0 s N-1, and N is the total number of pixels on the contour.

You can export 3 shape representations from these boundary points: the curvature function (curvature functions), the centroid distance (centroid distance), and the complex coordinate functions (complex coordinates function).

3.3.2 based on the morphological characteristics of region

Based on the method of morphological feature description, the whole area is taken as a whole, using all the pixels in the region to extract a series of morphological features, this section mainly introduces several simple region-based morphological feature extraction algorithms: simple geometric parameter descriptor and geometric invariant moment.

(1) Simple geometric parameter descriptor algorithm:

1. Area Area: Area area describes the basic characteristics of the size of the area, assuming that the area is the number of pixels in the region.
2. Regional center of Gravity: A global descriptor, the coordinate formula is as follows

where B is the area perimeter, it can be seen that the circular parameter is 1, to some extent this parameter can describe the area compactness, is insensitive to the scale transformation and the rotation transformation.

where B is the area perimeter, it can be seen that the circular parameter is 1, to some extent this parameter can describe the area compactness, is insensitive to the scale transformation and the rotation transformation.

(2) Geometric invariant moment algorithm:

1962, hu[24] The invariant moment theory of image recognition is presented, and the invariant moment of algebraic invariants is presented for the first time, and a set of moments which are invariant to the shift, rotation and scale of the image are derived through the nonlinear combination of geometric moments. Invariant moment is a statistic feature, and the gray distribution of image is described by using the order moment of gray scale distribution.

The p+q order of the discrete Digital image f (x, y) is defined as

　　The region-based morphological feature extraction method mainly describes the two simple extraction methods, and many other similar methods to extract the morphological features of the image. In the field of medical image feature extraction, morphological features also play an important role in the image feature extraction and prognosis has a great help.

3.4 Summary of this chapter

This chapter mainly introduces the basic knowledge of pulmonary nodules, describes the characteristics of pulmonary nodules growth, such as size and growth patterns, and then describes the clinical manifestations and pathological features of pulmonary nodules, and provides a basic supplement to understanding the medical knowledge needed for pulmonary nodules. This chapter focuses on the methods of feature extraction commonly used in medical image processing, in this paper, the CT Image feature extraction method of pulmonary nodules, which is divided into three categories: Gray feature, texture feature and morphological feature. Each kind of characteristic takes the commonly used method to extract the characteristic, through each kind of characteristic extraction method different, extracts the different lung nodule ct image characteristic.

　The next article describes the code that is implemented in the main section.

Feature extraction algorithm for medical CT images--An algorithm for CT image feature extraction of pulmonary nodules