[0035] Reference figure 1 The medical image diagnosis system based on migration core matching tracking of the present invention includes: an image preprocessing module, an image feature extraction module, an image migration core matching tracking diagnosis module, and an image unmarked sample test module. These modules are connected in turn. The image migration core matching tracking diagnosis module includes an image target domain identification sample training sub-module and an image source domain identification sample training sub-module. The image target domain identification sample training submodule completes the training of the image target domain with identification samples, generates an image training classifier with target domain identification samples, and transmits the classifier to the image source domain's identified sample training submodule; The identification sample training sub-module of the image source domain uses the training classification of the target domain identification samples for the migration core matching tracking learning of the identification samples in the image source domain, and composes a new sample that meets the similar distribution of the training sample set in the target domain. A sample set of, and classify and diagnose the sample set to obtain a diagnostic classifier.
[0036] The image preprocessing module performs de-redundancy and histogram equalization enhancement processing on the input original image, and transmits the processed result to the image feature extraction module; the image feature extraction module extracts the gray level co-occurrence matrix from the processed image , Hu moments, Brushlet features, and Contourlet features are used as diagnostic samples, and then the diagnostic samples are passed to the image migration nuclear matching and tracking diagnosis module; the image migration nuclear matching and tracking diagnosis module is used to adjust the nuclear parameters of the input diagnostic samples according to the migration nuclear matching tracking learning method P, obtain the diagnostic recognition rate that meets the conditions and the updated source domain with logo training sample set; reorganize the source domain with logo training sample set and the target domain with logo training sample set into an updated target domain with logo training sample set; For the updated target domain identified training sample set, the nuclear matching tracking method is used to obtain the diagnostic classifier, and the diagnostic classifier is output to the image unmarked sample test module; the image unmarked sample test module uses the output diagnostic classifier to detect the unmarked The test samples are classified and diagnosed, and the final diagnosis result is output.
[0037] Refer to figure 2 , The breast X imaging diagnosis method based on migration nucleus matching tracking of the present invention includes the following steps:
[0038] Step 1: Using histogram equalization and mean square error standardization methods, the mammary gland X images in the original medical image set are cut and enhanced to obtain a mammary gland X image set with better visual effects.
[0039] 1a) Input the original breast X image, its size is M×N, this example chooses as image 3 An image in the original mammography X image set shown, its size is 1024×1024;
[0040] 1b) The input original breast X image, namely the breast X image, adopts the horizontal and vertical computer automatic cutting method to remove the background of the image and the artificial imprints in the image to obtain the cut breast X image, such as Figure 4 Shown
[0041] 1c) Use histogram equalization and mean square error standardization methods to remove noise on the cut mammogram X images, and obtain mammogram X images with visual effects, such as Figure 5 Shown.
[0042] Step 2: Perform gray-level co-occurrence matrix, Hu moment, Brushlet feature and Contourlet feature extraction on the obtained breast X image with visual effects.
[0043] 2A. The process of extracting the four features of the gray-level co-occurrence matrix is as follows:
[0044] 2A.1) Generate a gray-level co-occurrence matrix p for the obtained mammary gland X images with visual effects ij (s, θ), where the value of θ is 4 discrete directions: 0°, 45°, 90°, 135°, s∈[1, size], size represents the length of medical images with better visual effects or width;
[0045] 2A.2) According to the obtained gray-level co-occurrence matrix, four features are extracted from the mammary gland X image with visual effects, namely:
[0046] Angle second moment: f 1 = Σ i = 0 N - 1 Σ j = 0 N - 1 p 2 ( i , j )
[0047] entropy: f 2 = - Σ i = 0 N - 1 Σ j = 0 N - 1 p ( i , j ) log p ( i , j )
[0048] Homogeneous area: f 3 = Σ i = 0 N - 1 Σ j = 0 N - 1 p ( i , j ) / [ 1 + ( i - j ) 2 ] 2
[0049] Non-similarity: f 4 = Σ i = 0 N - 1 Σ j = 0 N - 1 | i - j | p ( i , j ) .
[0050] 2B. The specific process of extracting the seven features of Hu moments is as follows:
[0051] 2B.1) Calculate the (p+q) moment m at the point (x, y) on the obtained mammary X image with visual effects pq And (p+q) order central moment μ pq :
[0052] m pq = Σ x = 0 M - 1 Σ y = 0 N - 1 x p y q f ( x , y )
[0053] μ pq = Σ x = 0 M - 1 Σ y = 0 N - 1 ( x - x c ) p ( y - y c ) q f ( x , y )
[0054] In the formula, f(x, y) represents the pixel value at point (x, y), (x c , Y c ) Represents the center of gravity coordinates of the mammogram with better visual effects;
[0055] 2B.2) According to the obtained point (x, y) at (p+q) moment m pq And (p+q) order central moment μ pq , Calculate the normalized central moment at the point (x, y) according to the following formula:
[0056] η pq = μ pq / μ oo r
[0057] In the formula, μ oo r Represents the zero-order γ-th moment of the point (x, y), γ=(p+q)/2+1;
[0058] 2B.3) Using the normalized central moment at the point (x, y), extract the seven Hu moment features of the breast X image with better visual effects, which are defined as φ 1 , Φ 2 ,..., φ 7 ,which is:
[0059] φ 1 =η 20 +η 02
[0060] φ 2 = ( η 20 - η 02 ) 2 + 4 η 11 2
[0061] φ 3 =(η 30 -3η 12 ) 2 +(3η 21 -η 03 ) 2
[0062] φ 4 =(η 30 +η 12 ) 2 +(η 21 +η 03 ) 2
[0063] φ 5 =(η 30 -3η 12 )(η 30 +η 12 )φ x +(η 03 -3η 21 )(η 21 +η 03 )φ y
[0064] φ 6 =(η 20 -η 02 )[(η 30 +η 12 ) 2 -(η 03 +η 21 ) 2 ]+4η 11 (η 30 +η 12 )(η 03 +η 21 )
[0065] φ 7 =(3η 21 -η 03 )(η 30 +η 12 )φ x +(η 30 -3η 12 )(η 03 +η 21 )φ y ,
[0066] φ x =(η 30 +η 12 ) 2 -3(η 03 -3η 21 ) 2
[0067] among them,
[0068] φ y =(η 03 +η 21 ) 2 -3(η 30 -η 12 ) 2.
[0069] 2C. Extract Brushlet features:
[0070] On the obtained mammary X image with visual effects, a sliding window is used to perform two-layer Brushlet transformation on the elements in the mammary X image. The mammary X image is divided into 16 blocks with 12 directions, and the Brushlet is decomposed into two layers. The sub-blocks in each direction extract the mean and variance features respectively. Calculate the features of the eight subbands in the upper half of the real part to obtain a 16-dimensional feature vector. The method of calculating the mean and variance features is carried out using the following formula:
[0071] μ i = 1 RC Σ j R Σ k C I i ( j , k )
[0072] σ i = 1 RC Σ j R Σ k C ( I i ( j , k ) - μ i ) 2
[0073] In the above formula, I i (j, k) represents the decomposition coefficient of each brushlet, i = 1, 2, ..., 16; j = 1, 2, ...; R, k = 1, 2, ..., C, R, C respectively represent the coefficient of each block Number of rows and columns.
[0074] 2D. Extract Contourlet features:
[0075] On the obtained mammary gland X image with visual effects, perform 3-layer Contourlet transformation on the mammary gland X image to obtain 17 sub-bands, that is, 17-dimensional feature vectors. The number of feature dimensions depends on the number of decomposition layers and the decomposition direction of each layer. Calculate the L1 norm energy measure of each Contourlet coefficient matrix obtained by decomposition:
[0076] E = 1 MN Σ i = 1 M Σ j = 1 N | coef ( i , j ) |
[0077] Among them, M×N is the sub-band size, i, j represent the index of the coefficient in the sub-band, and coef(i, j) is the coefficient value of the i-th row and j-th column in the sub-band.
[0078] Step 3: Extract the gray-level co-occurrence matrix, Hu moment, Brushlet feature, and Contourlet feature based on the features extracted from the mammary gland X image with visual effects as the auxiliary diagnosis sample of the breast X image, where: L s Is the training sample domain of the marked breast X image of the source domain, L T Is the sample domain of the marked target mammography X image, U T Is the unidentified target breast X image sample domain, x i Indicates the feature sample of breast X image, y={1, -1} is set as the sample ID;
[0079] The above-mentioned breast X image feature samples are defined as follows:
[0080] The training samples of marked mammography X images in the source domain are: X s = { ( x i s , y ( x i s ) ) } , among them, x i s ∈ X S ( i = 1 , . . . k 1 ) , k 1 Means L s The size of the source domain;
[0081] The marked mammography X image training sample of the target domain is: X T , L = { ( x i L , y ( x i L ) ) } , among them, x i L ∈ X L ( i = 1 , . . . k 2 ) , k 2 Means L T The training sample size of the target domain;
[0082] The unidentified mammography test samples of the target domain are: X T , U = { ( x i U ) } , among them, x i U ∈ X U ( i = 1 , . . . k 3 ) , k 3 Means U T The target domain test sample size.
[0083] Step 4: Training sample X of the marked breast X image in the source domain s And the marked breast X image training sample X in the target domain T，L Perform image classification diagnosis based on migration core matching tracking, and get the updated source domain with marked breast X image sample set X s，New , The specific implementation steps are as follows:
[0084]4A) A training sample of breast X image with a mark on the target domain X T，L Perform kernel matching tracking classification first to obtain a training classifier;
[0085] 4B) Use this training classifier to train the marked breast X image training sample X in the source domain. s And the marked breast X image training sample X in the target domain T，L Perform classification diagnosis based on migration image nuclear matching tracking, and obtain marked breast X image training sample X s The diagnosis result;
[0086] The specific process is as follows:
[0087] 4B.1) Mammary X image training sample X with a mark on the target domain T，L Perform kernel matching tracking classification to obtain training classifier C 1 , Set the kernel parameter of the classifier to P;
[0088] 4B.2) Use training classifier C 1 Training sample X of the marked breast X image in the source domain s Carry out classification diagnosis, get mammography X image training sample X s The diagnosis result;
[0089] 4B.3) Set the marked breast X image training sample X s The diagnosis result and the mammography X image training sample X s The sample ID y(x i s ) Perform comparison to obtain the diagnostic recognition rate R;
[0090] 4B.4) Set the threshold d, compare the diagnosis recognition rate R and the threshold d;
[0091] If R≤d, adjust the nuclear parameter P, increase or decrease the nuclear parameter P according to the given step size s; transfer the updated nuclear parameter P to step 4B.1), and perform the nuclear matching tracking classification diagnosis again;
[0092] If R>d, stop adjusting the nuclear parameter P, and get the marked breast X image training sample X s The result of the diagnosis.
[0093] 4C) Take the marked breast X image training sample X s The diagnosis result and the training sample identification y(x i ) In contrast, reorganize the marked breast X image training samples with consistent results into the updated source domain marked breast X image sample set X s，New.
[0094] Step 5: The updated source domain has the marked breast X image sample collection X s，New Marked breast X image sample X with target domain T，L Together to form a new target domain with marked breast X image sample X T，L ′, the sample domain sample X of the breast X image with a mark for each target domain T，L ′Perform the nuclear matching tracking classification, and get the nuclear matching tracking classifier C 2.
[0095] Step 6: Use nuclear matching to track classifier C 2 For unidentified test mammogram samples X T，U Carry out classification diagnosis, get unlabeled test breast X image sample X T，U The final diagnosis result, where, when the nuclear match tracking classifier C 2 When the category identification obtained by the classification is -1, the diagnosis result of the breast X image sample is cancerous; otherwise, when the nuclear matching tracking classifier C 2 When the category identification obtained by the classification is 1, the diagnosis result of the mammography sample is normal.
[0096] Step 7: Use the diagnostic results of the breast X-imaging sample as an unlabeled breast X-imaging test sample X T，U The final output result is output.
[0097] The effect of the present invention can be further illustrated by the following simulation data of mammography:
[0098] 1. Simulation conditions
[0099] The simulation of the present invention runs on windows XP, SPI, CPU Pentium(R)4, basic frequency 2.4GHZ, and software platform Matlab7.0.1. The original mammogram X images selected for the simulation came from the public data set MIAS, and a total of 150 original mammogram X images were obtained.
[0100] 2. Simulation results
[0101] The data obtained from the public data set MIAS is preprocessed and rotated, and the image is rotated counterclockwise by 5°, 10°, 15°, and clockwise by 5°, 10°, 15° to obtain: 1050 breast X images, Among them, there are 350 cancerous images and 700 normal images. The ratio of cancerous images to normal images is 1:2. While acquiring mammograms, the image information also includes: the type of breast tissue, the location of pathological abnormalities, the size of the abnormality, the location of breast cancer, that is, the left and right breasts, the type of tumor, that is, the structural nature of benign or malignant .
[0102] Based on preprocessing and rotation, the experiment extracts Hu moments, gray level co-occurrence matrix (GLCM), Brushlet features, and Contourlet features from the image after multi-angle rotation. The image data is divided into training set and test set. The training set is composed of 300 case samples and 600 normal sample data, and the test set is composed of 50 case samples and 100 normal sample data. Three methods are used for identification. For diagnosis, these three methods are the traditional nuclear matching tracking method KMP, the boosting-based nuclear matching tracking method Boosting KMP, and the migration-based nuclear matching tracking method TLKMP. Especially for the TLKMP method, set 150 samples in the test set as the unidentified samples in the target domain, the first 250 positive samples and the first 500 negative samples in the training set as the source domain identified samples, and the last 50 positive samples and the last 100 samples A negative sample is used as a target domain with a sample, and the above three methods are used to simulate the breast data. The diagnosis results are shown in Table 1, Table 2, Table 3, and Table 4:
[0103] Table 1 Comparison of recognition rate under Hu moment feature
[0104]
[0105] Table 2 Comparison of recognition rate under GLCM features
[0106]
[0107] Table 3 Comparison of recognition rate under Brushlet features
[0108]
[0109] Table 4 Comparison of recognition rate under Contourlet feature
[0110]
[0111] From the simulation results of the above four tables, it can be seen that under the premise of the same number of test samples, after the training samples of the LTKMP method are subjected to the migration kernel matching tracking method, the total number of training samples is reduced, and the ratio of training and testing samples is reduced, but the breast The recognition rate of X images has been significantly improved, which shows that compared with the simulation results of KMP and Boosting KMP methods, the recognition rate of breast X images using the TLKMP method has been improved, and at the same time, the shortcomings of the small number of reference samples of breast X images have been alleviated. The generalization ability of image processing methods.
[0112] The above-mentioned medical breast X image processing methods all realize their functions through computer programs, and at the same time complete the auxiliary diagnosis of the migration nuclear matching tracking of the medical breast X images.
[0113] This example is implemented on the premise of the technical solution of the present invention, and a detailed implementation mode and specific operation process are given, but the protection scope of the present invention is not limited to the above-mentioned embodiment.
