CN117876833A - Lung CT image feature extraction method for machine learning - Google Patents

Abstract

The present application discloses a lung CT image feature extraction method for machine learning, which relates to the field of image processing technology, and includes: acquiring a lung CT image of a patient, preprocessing the lung CT image, and determining texture features according to the preprocessing result; segmenting the lung CT image according to the texture features, and determining a region of interest in the lung CT image; comparing changes in the region of interest according to a plurality of lung CT images, further determining a tumor region, and outputting metabolic features corresponding to the tumor region; determining morphological features of the tumor region according to the determination of the tumor region, and determining changes in texture features of the tumor region based on morphological changes in the tumor region; determining feature fusion of the tumor region according to the extraction result of the tumor region, and completing the extraction of tumor features; the accuracy of image recognition can be improved, and the efficiency of image recognition can be improved.

Description

A lung CT image feature extraction method for machine learning

Technical Field

The present invention relates to the technical field of image processing, and in particular to a lung CT image feature extraction method for machine learning.

Background technique

There are tumor areas in the CT images of patients with lung cancer. Tumors with cell necrosis appear as a cavity inside the tumor area with a low grayscale value. The necrotic area in the tumor area has the risk of metastasis to other tissues or organs. The more necrotic areas there are inside the tumor area, the greater the risk of cancer spread. Currently, tumor analysis of lung CT images is completed by doctors based on their own experience and judgment, which often results in a large workload and a long time. Therefore, by extracting tumor feature data from CT images to assist doctors in making judgments, the workload of doctors can be further reduced.

However, when determining the tumor area, the corresponding disease cannot be identified according to the specific situation in the tumor area, and the multiple extracted features cannot be fused to make the extracted tumor features more detailed.

Summary of the invention

The embodiment of the present application solves the problem of poor image recognition effect in the prior art by providing a lung CT image feature extraction method for machine learning, thereby improving the accuracy of image recognition.

The embodiment of the present application provides a lung CT image feature extraction method for machine learning, comprising: acquiring a lung CT image of a patient, preprocessing the lung CT image, and determining texture features according to a result of the preprocessing;

Segmenting the lung CT image according to the texture features to determine the region of interest in the lung CT image;

Based on multiple lung CT images, the changes in the region of interest are compared to further determine the tumor area and output the metabolic characteristics corresponding to the tumor area;

Determine the morphological characteristics of the tumor region according to the determination of the tumor region, and determine the change of the texture characteristics of the tumor region based on the change of the morphology of the tumor region;

According to the extraction result of the tumor area, the feature fusion status of the tumor area is determined to complete the extraction of tumor features.

Determining the region of interest in the lung CT image further includes: acquiring a first region of interest in the lung CT image;

Acquire a lung PET image according to the lung CT image at a corresponding position, and acquire a second region of interest of the lung PET image;

Determining a first interest feature of a first interest region based on correlations between the plurality of first interest regions;

Determining a second interest feature of the second interest region based on the correlation between the plurality of second interest regions;

Performing feature aggregation on the first interest feature and the second interest feature to obtain an aggregated feature between each first interest region and the second interest region;

Based on the correlation between each aggregated feature, a third feature of interest is obtained, and the third feature of interest is aggregated into a single feature to obtain an image feature corresponding to the lung CT image.

When processing the lung CT image, the lung PET image and the lung CT image are also registered. The specific implementation methods include:

Determine a control point grid corresponding to the lung CT image, where the control point grid is used to move and change the image shape so that the lung CT image can be aligned with the lung PET image;

Based on the control point grid, the B-spline basis functions of the lung PET image and the lung CT image are determined, and the basis function value of each pixel position is calculated according to the position of the control point and the definition of the B-spline basis function;

By moving the control points and calculating the basis functions, the new position of each pixel in the lung PET image is determined;

Interpolation operation is performed on the deformed pixel positions to obtain the aligned lung PET image.

Metabolic feature identification methods also include:

The ratio of the radiation concentration in the region of interest to the injected dose and the patient's weight is obtained. Based on the distribution of the radioactive drug, the average radioactive drug concentration in the region of interest is determined, and the SUV image corresponding to the region of interest is obtained; the obtained SUV image is compared with the preset threshold to further determine the tumor area.

One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

By combining CT images with PET images, the corresponding tumor areas in different situations can be determined according to the metabolic changes at the corresponding positions during image processing, so as to identify the specific characteristics of the current tumor area; by combining the texture features, morphological features, and metabolic features corresponding to the tumor area, the lung area can be effectively segmented. Baoning identifies the morphological information and metabolic activities such as tumor volume, surface area, sphericity, flatness, etc. according to the feature vectors of the corresponding region of interest, which improves the representativeness and computational efficiency of the features, and improves the efficiency and accuracy of feature extraction through the particle swarm optimization algorithm, generating clear images of tumor location and contour, and realizing accurate and efficient extraction of lung tumor features.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a lung CT image feature extraction method for machine learning;

FIG2 is a flow chart of another embodiment of a method for extracting lung CT image features for machine learning.

Detailed ways

To facilitate the understanding of the present invention, the present application will be described more comprehensively below with reference to the relevant drawings; the drawings show preferred embodiments of the present invention, but the present invention can be implemented in many different forms and is not limited to the embodiments described herein; on the contrary, the purpose of providing these embodiments is to enable a more thorough and comprehensive understanding of the disclosed content of the present invention.

It should be noted that the terms “vertical”, “horizontal”, “up”, “down”, “left”, “right” and similar expressions used in this document are only for illustrative purposes and do not represent the only implementation method.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs; the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention; the term "and/or" used herein includes any and all combinations of one or more related listed items.

As shown in FIG1 , the present application provides a lung CT image feature extraction method for machine learning, including:

S101, obtaining a grayscale image of the lungs, obtaining a grayscale histogram corresponding to the grayscale image, and determining a range of a tumor region in the grayscale histogram;

S102, obtaining edge pixel points in the tumor area, determining the distance between edge pixel points of adjacent grayscale images based on the change of grayscale images between different grayscale images, and determining the difference amount of the tumor area;

S103, detecting the edge of the tumor region according to the difference amount of the tumor region to determine the maximum diffusion area of the tumor region; and determining the distribution of the boundary between the tumor region and the surrounding tissue according to the maximum area of the tumor region;

S104, determining the extraction status of tumor features according to the distribution of tumor region boundaries.

Specifically, when identifying a tumor, the tumor is determined to be benign or malignant by identifying its corresponding position, shape, size, and boundary. By identifying the spread of the tumor, the maximum infiltration area of the current tumor can be predicted to determine the points that should be paid attention to when treating the tumor, so as to improve the treatment effect of the tumor. The tumor features extracted at this time are used to assist in the treatment of the tumor in tumor management. According to the shape of the tumor boundary, shape features such as roundness, ellipticity, and irregularity can be extracted. These features can reflect the geometric morphology of the tumor.

At this time, when processing the boundary, features such as clarity, smoothness or roughness of the boundary can be extracted; for example, by calculating the curvature, convexity and concavity of the boundary, the characteristics of the tumor boundary can be quantitatively described; the texture information inside the tumor is also one of the important features; this includes texture patterns extracted by gray-level co-occurrence matrix, Fourier transform and other methods, such as uniformity, contrast, directionality, etc.; at the same time, the size of the tumor is one of the most intuitive features, which can be quantified by calculating parameters such as the area, perimeter, and diameter of the tumor area; and the location information of the tumor in the image, such as the coordinates of the center point, the relative position to the surrounding structures, etc., are also important features; the extraction of these features will help with subsequent tasks such as tumor classification, grading, and prognosis assessment.

By analyzing the grayscale histogram in the above technical solution, the tumor area in the lung image can be more accurately identified and located. By calculating the difference value, the boundary and shape of the tumor area can be more accurately described. By using the difference value calculated previously, the edge detection of the tumor area can be performed to determine the maximum diffusion area of the tumor area, which helps doctors evaluate the severity of the tumor and formulate treatment plans.

In one embodiment of the present invention, in order to determine the accuracy of the acquired tumor area, as shown in FIG2 , the boundary of the tumor area is determined based on multiple dimensions by performing multi-dimensional fusion processing on the acquired image, so that the shape and size of the tumor area can be accurately identified.

S201, obtaining a lung CT image of a patient, preprocessing the lung CT image, and determining texture features according to the preprocessing result.

The acquired lung CT images are denoised and smoothed, and the noise and misidentified points in the CT images are marked; the lung CT images are preprocessed through image processing techniques such as Gaussian filtering.

S202, segmenting the lung CT image according to texture features to determine a region of interest in the lung CT image.

Wherein, determining the region of interest in the lung CT image further includes: acquiring a first region of interest in the lung CT image;

Acquire a lung PET image according to the lung CT image at a corresponding position, and acquire a second region of interest of the lung PET image;

Determining a first interest feature of a first interest region based on correlations between the plurality of first interest regions;

Determining a second interest feature of the second interest region based on the correlation between the plurality of second interest regions;

Performing feature aggregation on the first interest feature and the second interest feature to obtain an aggregated feature between each first interest region and the second interest region;

Based on the correlation between each aggregated feature, a third feature of interest is obtained, and the third feature of interest is aggregated into a single feature to obtain an image feature corresponding to the lung CT image.

After the doctor determines the scanning range, a CT scan is performed. If necessary, a local diagnostic scan is also performed. After the CT data is collected, the patient will be automatically sent to the PET scanning area at the rear section of the gantry for a PET emission scan from the legs to the head. The image acquisition of PET imaging mainly includes emission scanning and transmission scanning. The emission scanning methods include 2D acquisition, 3D acquisition, static acquisition, dynamic acquisition, gated acquisition, local acquisition and whole-body acquisition. After the acquisition is completed, the corresponding lung PET image is obtained to assist in lung CT image recognition and check the location and size of the tumor.

Among them, the first region of interest is used to determine the key areas in the lung CT image, such as areas of inflammation, necrosis, etc.; the first interest feature is determined based on the correlation of the first region of interest, and is used to represent the characteristics between different conditions; the second region of interest is used to determine the metabolic conditions in the tumor area in the lung PET image to identify changes in the corresponding area; the second interest feature is used to show the association between different metabolic conditions; the third interest feature is used to determine the corresponding feature conditions when each area is aggregated, and to determine some pathological conditions corresponding to different metabolisms. According to the corresponding pathological conditions, the aggregated third interest feature is the image feature that you want to identify in the end, and the extracted image feature is used to select the area for subsequent segmentation.

After preprocessing the lung CT image, the lung PET image at the corresponding position of the lung CT image is obtained, the lung PET image is divided into multiple regions, and each region is standardized; the standardized lung PET image is aligned with the lung CT image to obtain metabolic information in the tumor area, so as to determine the changes in the current tumor area and assist in identifying the spatial size of the tumor.

When processing the lung CT image, the lung PET image and the lung CT image are also registered. The specific implementation methods include:

A control point grid corresponding to the lung CT image is determined, and the control point grid is used to move and change the image shape so that the lung CT image can be aligned with the lung PET image.

Based on the control point grid, the B-spline basis functions of the lung PET image and the lung CT image are determined, and the basis function value of each pixel position is calculated according to the position of the control point and the definition of the B-spline basis function.

The new position of each pixel in the lung PET image is determined by moving the control points and calculating the basis functions.

Interpolation operation is performed on the deformed pixel positions to obtain the aligned lung PET image.

S203, comparing changes in the region of interest based on the multiple lung CT images, further determining the tumor region, and outputting metabolic features corresponding to the tumor region.

By utilizing the high spatial resolution of CT images, using methods such as threshold segmentation, region growing or deep learning, the lung tumor areas corresponding to different image features are initially located, and combined with the metabolic information in the lung PET images, the tumor areas are further confirmed by setting the SUV (standardized uptake value) threshold and other methods to identify the distribution of different feature areas to assist doctors in diagnosis. At this time, a single CT image may be affected by multiple factors such as noise, artifacts, and scanning conditions, resulting in inaccurate or ambiguous diagnostic information. By comparing multiple images, these random errors can be averaged, thereby improving the accuracy of tumor area identification. Through multiple CT images acquired at different time points, the dynamic changes of the tumor can be observed, including its size, shape, density, and interaction with surrounding tissues.

Preferably, the multiple lung CT images acquired are for the same patient, acquired at different time points, and the number of the acquired lung CT images is at least two.

Specifically, the metabolic feature recognition method also includes: fusing the acquired lung CT image with the corresponding lung PET image, integrating the information of the two images into the same spatial coordinate system, so that each tumor area has both structural information and metabolic information.

Obtaining the ratio of the concentration of the radiation in the region of interest to the injected dose and the patient's weight, determining the average concentration of the radioactive drug in the region of interest based on the distribution of the radioactive drug, and obtaining the SUV image corresponding to the region of interest; comparing the obtained SUV image with a preset threshold to further determine the tumor area;

SUV images are often referred to as Standardized Uptake Value Images in the medical field; SUV is a measurement used to quantify the extent of radioactive tracer uptake in PET (positron emission tomography) images.

Specifically, in this embodiment, the value corresponding to SUV represents the ratio of the radioactive drug concentration in the tissue to the injected dose and the patient's weight.

At this time, the patient's PET image is first obtained, which reflects the distribution of the radioactive drug in the patient's body; the area of interest, usually the suspected tumor area, is manually or automatically outlined on the PET image; the radioactive drug concentration in the area of interest is measured: the average radioactive drug concentration in the area of interest is calculated; the injection dose and patient weight are obtained: this information is usually obtained before the PET scan.

Calculate SUV: Use the above formula to calculate the SUV value of the region of interest;

SUV = (radiopharmaceutical concentration in tissue)/(injected dose/patient weight).

One or more SUV thresholds are set, for example, an area with SUV>2.5 may be considered a tumor area; the calculated SUV image is compared with the set threshold to determine and extract the tumor area; the calculation of SUV is mainly based on the uptake and distribution principles of radioactive drugs and the physical principles of PET imaging.

The pixel value in the PET image reflects the concentration of the radioactive drug, and the SUV standardizes this concentration to eliminate differences in injection dose and patient weight. The accuracy of the injection dose and scanning time is also very important for the calculation of SUV. The setting of the SUV threshold needs to be adjusted according to the specific application scenario and disease type. Different diseases and scanning conditions may require different SUV thresholds, and the obtained SUV value of the region of interest is used as the output metabolic feature.

S204, determining the morphological features of the tumor region according to the determination of the tumor region, and determining the change of the texture features of the tumor region based on the change of the morphology of the tumor region.

Based on the further determination of the tumor area, morphological features corresponding to different three-dimensional shapes, sizes, and volumes of the tumor are selected from multiple CT images; the morphological features include the volume, surface area, sphericity (similarity to a sphere), and flatness (similarity to a flat shape) corresponding to the tumor area.

According to the morphological features corresponding to the tumor area, the texture features corresponding to each morphological feature are determined to determine the change of the texture features.

Compare lung CT images at multiple time points to determine the morphological features corresponding to each time point. Based on the changes in the parameter values of the morphological features, determine the changing trend of the tumor area; for example, calculate the growth rate of the area and the rate of change of the perimeter. The growth and change of tumors is a dynamic process. By collecting images at different time points, it is possible to capture subtle changes in the tumor during this process, including changes in volume, shape, and relationship with surrounding tissues. Some tumors grow slowly in the early stages, and images at a single time point may not accurately reflect their existence or changes. By comparing multiple time points, signs of tumor growth can be discovered earlier, which helps with early diagnosis and treatment.

According to the change trend of the tumor area, the change of the texture eigenvalue is determined; for the change of the texture eigenvalue, the contrast and correlation of each texture feature are determined according to the grayscale co-occurrence matrix, Gabor filter response, and wavelet transform coefficient of the texture feature, and the texture eigenvalues at different time points or under different conditions are compared to determine the tumor growth pattern, changes in internal structure, and its association with treatment effect.

S205, determining the feature fusion status of the tumor region according to the extraction result of the tumor region, and completing the extraction of the tumor feature.

Extracting metabolic characteristics of tumors such as maximum SUV, average SUV, and SUV coefficient of variation from PET images;

The extracted morphological features, metabolic features and texture features are fused to form a multi-dimensional feature vector.

For example, feature vector = [volume, circularity, maximum SUV, mean SUV, SUV coefficient of variation, GLCM contrast, ...];

The acquired feature vector is projected into the space composed of principal components, and the first few principal components of the feature vector are selected as the feature set of information. These principal components can retain the variance information of the original feature vector to the greatest extent; the feature vector after dimensionality reduction is output; the feature selection method is used to reduce the feature dimension, remove redundant information, and improve the classification performance.

A multi-layer neural network structure is designed, in which the input layer receives the feature vector after dimensionality reduction. The network weights are trained through the back-propagation algorithm to minimize the classification error. The hyperparameters such as the network structure and learning rate are adjusted to optimize the model performance, thereby training the feature vector to improve the classification performance of tumor characteristics.

After determining the classification performance extraction of the tumor area, the trained feature vector is used as input. A confusion matrix is constructed based on the true labels of the samples in the test data set and the labels predicted by the model. Based on the confusion matrix, the accuracy, sensitivity, and specificity of the feature vector are determined. The confusion matrix is used to display the number of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN).

Calculate the ratio of all correctly predicted samples to the total number of samples in the test data set, that is, accuracy = {TP+TN}{TP+FP+TN+FN}.

The proportion of tumors that are correctly predicted to be malignant among all tumors that are actually malignant is calculated, that is, sensitivity = {TP}{TP+FN}.

The proportion of all tumors that are actually benign that are correctly predicted to be benign is calculated, that is, specificity = {TN}{TN+FP}.

The evaluation effect of the trained feature vector is determined based on accuracy, sensitivity, and specificity.

Use hyperparameter optimization techniques such as grid search to tune the parameters of the model; for example, adjust the C value and kernel function in a support vector machine, or adjust the number of layers, number of neurons, and learning rate in a neural network.

When adjusting parameters, combine cross-validation to avoid overfitting and find the parameter combination that performs best on the validation set.

In the above steps, through preprocessing and texture feature extraction, the image quality can be effectively improved, and important details in the lung structure can be highlighted. The extracted texture features are used to segment the lung CT image, and the current CT image is assisted in processing according to the acquired PET image, which can help to quickly locate and extract the area of suspected tumor. By comparing images at multiple time points, tumors can be identified more accurately and their growth and metabolic characteristics can be understood, thereby improving the performance and reliability of the entire tumor detection and analysis system.

In one embodiment of the present invention, in order to determine the obtained feature vector, the extracted tumor features can be made more detailed; the tumor area corresponding to the feature vector is obtained, different tumor areas are connected, and the corresponding connected areas are filled and processed so that the filled image can show the position and contour image of the tumor in the tumor area.

Get the smallest filling particle and generate the corresponding particle swarm; the filling particle represents a possible solution and has a speed, position, and fitness value.

The fitness of the particle swarm is calculated, and based on the fitness of each particle swarm, the current fitness value and the optimal fitness value of each filling particle are determined. If the current value is better, it is updated to the optimal fitness value of the filling particle; the global optimal position is found from the optimal fitness values of all filling particles; the particle's current speed, optimal fitness value and global optimal position are used to update the particle's speed and position until a preset number of iterations is reached or the fitness value of the global optimal position no longer changes significantly, and the final global optimal position is used for post-processing, such as connecting different tumor areas, filling and processing connected areas, etc., to generate an image showing the tumor location and contour.

By using the particle swarm algorithm to fill and process the connected areas, an intuitive and clear image of the tumor location and contour can be obtained, providing doctors with valuable diagnostic information.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims (10)

1. A method for extracting lung CT image features for machine learning, comprising: acquiring a lung CT image of a patient, preprocessing the lung CT image, and determining texture characteristics according to a preprocessing result;

dividing the lung CT image according to the texture characteristics to determine an interested region in the lung CT image;

comparing the change conditions of the interested areas according to the lung CT images, further determining the tumor areas and outputting metabolic characteristics corresponding to the tumor areas;

determining morphological characteristics of the tumor area according to the determination condition of the tumor area, and determining the change condition of texture characteristics of the tumor area based on the morphological change of the tumor area;

and determining the feature fusion condition of the tumor region according to the extraction result of the tumor region, and completing the extraction of the tumor features.

2. The method for extracting lung CT image features for machine learning according to claim 1, wherein a gray level image of the lung is obtained, a gray level histogram corresponding to the gray level image is obtained, and a range of a tumor region in the gray level histogram is determined;

acquiring edge pixel points in a tumor area, determining the distance between the edge pixel points of adjacent gray images based on the change condition of gray images among different gray images, and determining the difference amount of the tumor area; detecting edges of the tumor region according to the difference amount of the tumor region to determine the maximum diffusion area of the tumor region; determining the distribution condition of the boundary between the tumor area and surrounding tissues according to the maximum area of the tumor area; and determining the extraction condition of the tumor features according to the distribution condition of the boundary of the tumor region.

3. The method of machine learning lung CT image feature extraction of claim 1, wherein determining a region of interest in the lung CT image further comprises: acquiring a first region of interest in a lung CT image;

acquiring a lung PET image according to the lung CT image of the corresponding position, and acquiring a second interest area of the lung PET image;

determining a first interest feature of the first region of interest based on correlations between the plurality of first regions of interest;

determining a second feature of interest of the second region of interest based on correlations between the plurality of second regions of interest;

feature aggregation is carried out on the first interest features and the second interest features, and aggregation features between each first interest region and each second interest region are obtained;

and acquiring a third interest feature based on the correlation between each aggregation feature, and aggregating the third interest feature into a single feature to obtain the image feature corresponding to the lung CT image.

4. The method for extracting features from a CT image of a lung for machine learning as recited in claim 1, further comprising registering the PET image of the lung with the CT image of the lung when processing the CT image of the lung, the method comprising:

determining a control point grid corresponding to the lung CT image, wherein the control point grid is used for moving and changing the shape of the image so as to align the lung CT image with the lung PET image;

b spline basis functions of the lung PET image and the lung CT image are determined based on the control point grids, and the basis function value of each pixel position is calculated according to the positions of the control points and the definition of the B spline basis functions;

determining a new position of each pixel in the lung PET image through movement of the control point and calculation of a basis function;

and carrying out interpolation operation on the deformed pixel positions to obtain an aligned lung PET image.

5. The method for extracting features from a CT image of the lung for machine learning according to claim 1, wherein the method for identifying metabolic features further comprises:

acquiring the ratio of the concentration of the radioactive substance in the region of interest to the injection dose and the weight of the patient, determining the average concentration of the radioactive substance in the region of interest based on the distribution condition of the radioactive substance, and acquiring an SUV image corresponding to the region of interest; and comparing the acquired SUV image with a preset threshold value, so as to further determine the tumor area.

6. The method of claim 1, wherein the morphological features include volume, surface area, sphericity, flatness corresponding to the tumor region.

7. The method for extracting features from a lung CT image for machine learning according to claim 1, wherein the lung CT images at a plurality of time points are compared to determine a morphological feature corresponding to each time point, and a trend of the change in the tumor region is determined based on a change in a parameter value of the morphological feature; and determining the change condition of the texture characteristic value according to the change trend of the tumor area.

8. The method for extracting features of a lung CT image for machine learning according to claim 1, wherein the extracted morphological features, metabolic features and texture features are fused to form a multidimensional feature vector; and projecting the obtained feature vectors into a space formed by the principal components, selecting the first several principal components of the feature vectors as feature sets of information, and outputting the feature vectors after dimension reduction.

9. A method for extracting features of a lung CT image for machine learning as claimed in claim 1, wherein minimum filling particles are obtained and corresponding particle groups are generated; calculating the fitness of the particle swarm, and determining the current fitness value and the optimal fitness value of each filling particle based on the fitness of each particle swarm; and (3) finding out the global optimal position from the optimal fitness values of all the filling particles, updating the speed and the position of the particles by using the current speed, the optimal fitness values and the global optimal position of the particles, and performing post-processing by using the final global optimal position to generate an image for displaying the tumor position and the outline.

10. The method for extracting features of lung CT images for machine learning according to claim 8, wherein the trained feature vectors are used as inputs, a confusion matrix is constructed based on the real labels of the samples in the test dataset and the labels predicted by the model, and the accuracy, sensitivity and specificity of the feature vectors are determined based on the confusion matrix; and determining the evaluation effect of the trained feature vector according to the accuracy, sensitivity and specificity.