A method for extracting features of CT lesions in brain gliomas
By optimizing the Gabor filter using the energy of the ring rotation-invariant gray-level co-occurrence matrix, the problems of inaccurate boundary and insufficient texture feature recognition in brain glioma detection are solved, and efficient brain glioma imaging is achieved.
Patent Information
- Application Number
- CN202410094293.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-23
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2044-01-23
AI Technical Summary
Existing technologies for detecting gliomas in the brain suffer from insufficient accuracy in boundary detection, inadequate texture feature recognition, high computational load, and low efficiency, making it difficult to achieve high-precision imaging of gliomas.
A Gabor filter with energy optimization using a ring rotation-invariant gray-level co-occurrence matrix is employed. By extracting features, the texture features of spatial pixels are mapped to the feature space. Feature vectors are extracted using the rotation-invariant gray-level co-occurrence matrix to perform texture region segmentation, thereby enabling the detection and imaging of gliomas in the brain.
It improves the accuracy and functionality of detection, reduces the amount of computation, and increases detection efficiency.
Smart Images

Figure CN118097165B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of CT detection of brain gliomas, specifically involving a method for extracting features of brain glioma CT lesion regions. This method uses a ring rotation-invariant gray-level co-occurrence matrix energy-optimized Gabor filter to complete the feature extraction of brain glioma CT lesion regions. Background Technology
[0002] Gliomas are among the most common malignant brain tumors in adults, characterized by rapid onset, high recurrence rates, and very high mortality rates. Brain gliomas primarily result from the cancerous transformation of glial cells in the brain and spinal cord, making them the most common primary malignant brain tumor, accounting for approximately 35.2% to 61.0% of all intracranial tumors. Malignant gliomas are highly invasive during their development and have a clear boundary with normal brain tissue; therefore, accurate boundary assessment is crucial in glioma detection. Medical professionals typically use CT scans or MRI to examine the boundaries of gliomas in patients.
[0003] Existing technologies for detecting brain gliomas, particularly those targeting the boundaries of malignant gliomas, can generally detect the boundaries of these tumors. However, these methods are not precise enough for lesion de-imaging and have limited functionality. They also fail to adequately represent the texture features of the lesion area, hindering medical personnel from accurately assessing the disease's progression. Furthermore, these traditional methods suffer from large measurement errors and low efficiency. Additionally, traditional methods for extracting texture features from brain gliomas involve computationally intensive processes.
[0004] To address this, this application proposes a method for detecting gliomas in the brain based on a Gabor filter optimized by an energy-optimized ring rotation-invariant gray-level co-occurrence matrix. This method maps the texture features of spatial pixels to a feature space through feature extraction and extracts feature vectors of texture information using a rotation-invariant gray-level co-occurrence matrix. Through feature classification, the pixels to which the features belong are mapped to the corresponding texture categories, thereby achieving segmentation of the texture region and realizing the goal of glioma detection and imaging. This method improves detection accuracy and functionality while reducing computational load, exhibiting high efficiency. Summary of the Invention
[0005] This invention provides a method for feature extraction of CT lesion regions of brain gliomas. It employs a ring-shaped rotation-invariant gray-level co-occurrence matrix (GLCM) energy-optimized Gabor filter for glioma detection. Feature extraction maps the texture characteristics of spatial pixels to a feature space, and a rotation-invariant GLCM extracts feature vectors to describe texture information. Feature classification maps the pixels to which the features belong to the corresponding texture categories. This achieves texture region segmentation and enables glioma detection imaging.
[0006] Specifically, a method for extracting features of CT lesions in brain gliomas includes:
[0007] S1, segment the brain CT image I1(x,y) into A×A size images, and convert the segmented images into grayscale images I2(x,y);
[0008] S2, divide I2(x,y) into N×N image blocks, and perform discrete cosine transform on each image block to obtain I3(x,y);
[0009] S3, construct Gabor filter banks of different scales in four directions, and use the Gabor filter banks to filter I3(x,y) to obtain the amplitude feature set I4(x,y); the four directions refer to the four directions at 0°, 45°, 90°, and 135°;
[0010] S4, input I4(x,y) into the gray-level co-occurrence matrix, and obtain the output in the four directions respectively to obtain the preliminary texture features of the brain CT image under multi-scale conditions;
[0011] S5, by optimizing the rotation invariance of the gray-level co-occurrence matrix using an improved LBP algorithm, accurate texture features are obtained. The improved LBP algorithm refers to... Calculate the precise texture features, wherein, Perform an operation to shift the q-bit number X p times in a clockwise or counterclockwise direction. The following calculation process is described: Taking the currently calculated pixel as the center pixel, within a neighborhood with a radius of r, q sampling points are taken and compared with the gray value of the center pixel. If the gray value of the surrounding points is greater than or equal to the gray value of the center pixel, the gray value of the surrounding points is set to 1; otherwise, it is set to 0. Binary encoding is performed in clockwise or counterclockwise order, and the binary encoding is converted into a decimal value to obtain the feature value.
[0012] As a preferred option, A=128 and N=8.
[0013] Furthermore, in step S2, the discrete cosine transform is a two-dimensional discrete cosine transform, which is calculated by decomposing it into two one-dimensional discrete cosine transform matrices in the row and column directions.
[0014] Furthermore, step S3 includes: constructing a two-dimensional Gabor filter bank with different scales in four directions; and decomposing the real part function of the two-dimensional Gabor filter into a product of multiple one-dimensional filters for smoothing filtering.
[0015] Furthermore, step S3 includes performing horizontal convolution filtering on each pixel within the image block using a one-dimensional bandpass filter function, and then performing vertical convolution filtering on the pixels within the image block using a one-dimensional lowpass filter function.
[0016] Furthermore, in step S5, LBP values are output using an equivalent pattern; texture features are represented using a joint difference distribution.
[0017] This invention patent employs a ring-shaped rotation-invariant gray-level co-occurrence matrix (GLCM) energy-optimized Gabor filter for glioma detection. It maps the texture characteristics of spatial pixels to a feature space through feature extraction, and uses a rotation-invariant GLCM to extract feature vectors describing texture information. Feature classification maps the pixels to which the features belong to the corresponding texture categories. This achieves texture region segmentation and enables glioma detection imaging. Attached Figure Description
[0018] Figure 1 This is the overall framework diagram of the present invention.
[0019] Figure 2 It is an image obtained after grayscale processing.
[0020] Figure 3 It is an image that has been segmented and subjected to discrete cosine transform on each small block.
[0021] Figure 4 This is a comparison between the original image and the filtered image using a Gabor filter.
[0022] Figure 5 This is a detailed process for feature extraction in this invention.
[0023] Figure 6 It is a feature map processed by the improved LBP algorithm.
[0024] Figure 7 It is an LBP histogram. Detailed Implementation
[0025] The present invention will now be described in detail with reference to the accompanying drawings and examples, but the specific embodiments of the present invention are not limited thereto.
[0026] This embodiment illustrates the application of the present invention in the field of brain CT lesion region feature extraction technology, specifically involving a method for brain CT lesion region feature extraction.
[0027] like Figure 1 As shown, a method for extracting features of brain glioma CT lesions includes the following steps:
[0028] S1: Segment the brain CT scan image I1(x,y) into A×A size images and convert them to grayscale images I2(x,y). An example of the grayscale processed image is shown below. Figure 2 As shown.
[0029] S2: Divide I2(x,y) into N×N image blocks, and perform a discrete cosine transform on each image block to obtain I3(x,y).
[0030] Divide the image I2(x,y) into N×N image blocks, each of which can be represented as I 2ij (x,y), i,j=1,2,...,A / N. Step S1 and the image segmentation here both belong to the image preprocessing process. Any signal can be regarded as the superposition of multiple sine or cosine signals with different amplitudes and frequencies. Then, a discrete cosine transform is performed on each image block. In terms of image processing, conventional processing methods are not good at processing detailed information, but filtering can be performed on the high-frequency part with relatively less information content so that the subsequent Gabor filter processing can obtain better feature signals.
[0031] For time series ,in, The definition of its one-dimensional discrete cosine transform is as follows:
[0032] (1)
[0033] in, , , .so It is the sum of the cosine terms corresponding to each point in the sequence.
[0034] For the image signal I2(x,y), the two-dimensional discrete cosine transform is derived from the one-dimensional discrete cosine transform, which transforms the sub-blocks I in the image. 2ij The formula for the two-dimensional discrete transformation of (x,y) is as follows:
[0035] (2)
[0036] in, , ,u,v=1,2,...,A / N.
[0037] In processing CT images, to avoid excessively large discrete cosine transform matrices that would significantly increase overall complexity, the acquired CT images are first segmented. Furthermore, the formula reveals that the calculation of each coefficient of the discrete cosine transform matrix is relatively independent. By dividing the entire matrix into blocks and further subdividing each block into units based on its coefficients, the computational load is kept relatively low.
[0038] Simultaneously, the two-dimensional discrete cosine transform matrix is decomposed into two one-dimensional discrete cosine transform matrices in the row and column directions for calculation:
[0039] (3)
[0040] (4)
[0041] in, , , , A / N.
[0042] like Figure 3 The image shown is divided into 8×8 blocks. 2ij (x, y) is then used to transform the image by performing a discrete cosine transform on each small block. The formula for the two-dimensional discrete transform is as follows:
[0043] (5)
[0044] in, , ,u,v=1,2,...,A / N.
[0045] Although the original CT image has been filtered by two-dimensional discrete cosine transform, some light and shadow noise still exists. Therefore, Gabor filter is used to optimize the image based on two-dimensional discrete cosine transform.
[0046] S3: Construct Gabor filter banks of different scales in four directions, and use the Gabor filter banks to filter I3(x,y) to obtain the amplitude feature set I4(x,y).
[0047] 3.1 Constructing a two-dimensional Gabor filter bank
[0048] The underlying information of an image can be viewed as an irregular discrete sequence in two-dimensional space. Two-dimensional Gabor filters are the most commonly used processing function in image texture processing. Gabor filters possess excellent spatial and directional characteristics. They are used to capture image texture edge features and local features at different frequencies and multiple scales and directions, including all amplitude, phase, and frequency values. Essentially, a Gabor filter is a Fourier transform with a Gaussian window applied. When processing images, the window function determines its spatial domain characteristics and locality. Moving the window function allows us to obtain spatial domain characteristic information at different locations in a grayscale image.
[0049] To accommodate the characteristics of the co-occurrence matrix algorithm, four fixed directions are used for Gabor filter direction selection, and Gabor filter banks of different scales are constructed in the four directions of 0°, 45°, 90° and 135° respectively.
[0050] The main impulse response of a Gabor filter constructed using Gaussian functions is expressed as follows:
[0051] (6)
[0052] In the formula It is the standard deviation of the Gaussian function. and These represent the relative frequency characteristics of the signal on the x-axis and y-axis coordinates, respectively. The relative frequency characteristic is radial; therefore, in the direction... choose To represent by function
[0053] (7)
[0054] The polar coordinate expression for the Gabor filter is:
[0055] (8)
[0056] (9)
[0057] in For Gaussian kernel function, It is a complex plane wave function. This represents the standard deviation of the Gaussian function in the Gabor kernel function. For the selected angle and direction, where the parameters are... This determines the size of the acceptable region for the Gabor filter kernel. The bandwidth (b) cannot be directly specified and varies with the bandwidth. The bandwidth value is a positive real number, typically 1. In this case, the relationship between the standard deviation and wavelength is: σ = 0.56λ. The smaller the bandwidth, the larger the standard deviation, the larger the Gabor shape, and the more visible parallel excitation and inhibition stripes. The wavelength (λ) value is specified in pixels, typically greater than or equal to 2, and less than one-fifth of the input image size. If the side length of the convolution kernel is smaller than the wavelength, the entire waveform cannot be included in the kernel, rendering the filtering effect at the waveform edges ineffective. Conversely, if the side length of the convolution kernel is greater than the wavelength, it will not affect the filter. In actual Gabor filter kernel selection, the half-response spatial frequency bandwidth or half-peak order in the frequency domain is chosen, i.e., the spatial location of b / 2 in the frequency domain. and The relationship is as follows:
[0058] (10)
[0059] (11)
[0060] (12)
[0061] (13)
[0062] in These are adaptive parameters.
[0063] Regarding the calculation of spatial aspect ratio:
[0064] (14)
[0065] (15)
[0066] (16)
[0067] in This represents the aspect ratio. The spatial aspect ratio determines the ellipticity of the Gabor function shape. When γ = 1, the shape is circular. When γ < 1, the shape elongates along the direction of the parallel stripes; typically, this value is 0.5.
[0068] The difference in the center frequency of the Gabor filter reflects the periodicity of image texture information and corresponds to the dominant frequency component. Given the relative frequency feature parameter z and the Gaussian function squared difference parameter... If K Gabor filter banks are needed to obtain the texture feature information of an image, then a K-dimensional vector is obtained. Feature vector relationship function, pixel The corresponding vector feature function expression can be represented as:
[0069] (17)
[0070] (18)
[0071] (19)
[0072] Where s represents the scale. Indicating direction, the discrete cosine transform of the I3(x,y) image patch is transformed into the texture feature function at (x,y) using a Gaussian window function. Therefore, I3(x) and I3(y) can be regarded as image functions in the X and Y directions, respectively.
[0073] The two-dimensional Gabor filter function is a new function obtained by multiplying a Gaussian function and a complex plane wave. The characteristic expression of the two-dimensional Gabor filter in the spatial domain can be transformed into:
[0074] (20)
[0075] in It is a Gaussian function, and by applying a window through a functional approach, the filtering range of the Gabor filter function is limited, ensuring that its function is effective within the specified range. Meanwhile... The function can be viewed as a trigonometric oscillating function. The sine function, being an even function, serves as the real part feature, used to smooth the filtered image; the cosine function, being an odd function, serves as the imaginary part feature, used for edge detection. The phase difference between the real and imaginary parts is 90°. t refers to the time series.
[0076] Performing a Fourier transform on the complex function expression of the two-dimensional Gabor filter yields the frequency domain expression of the two-dimensional Gabor filter function as follows:
[0077] (twenty one)
[0078] This function possesses excellent time-frequency localization capabilities, effectively expressing the spatial location, frequency, orientation, phase, and bandwidth characteristics of image information. Furthermore, the designed two-dimensional Gabor filter function exhibits strong bandpass properties, demonstrating good resolution in both the spatial and frequency domains. During use, different parameters are set to obtain Gabor filter banks with varying bandwidths, center frequencies, and orientations. By fusing Gabor filters from different orientations at the same scale, the optimal scale parameters are obtained at the corresponding frequencies and directions.
[0079] 3.2 Decompose the real part of the two-dimensional Gabor filter into a product of multiple one-dimensional filters for smoothing filtering.
[0080] The real part function expression is:
[0081] (twenty two)
[0082] (twenty three)
[0083] (twenty four)
[0084] but, and The corresponding relationship can be viewed as:
[0085] (25)
[0086] in These are the directional characteristic parameters of the Gabor filter. The plane wave frequency represents the waveform. In the image coordinate system The coordinates rotated clockwise The new coordinate system was then obtained. and These represent the relative frequency characteristic parameters on the x-axis and y-axis, respectively, and their values directly affect the filtering effect. The larger the values of these two parameters, the stronger the noise reduction capability of the Gabor filter.
[0087] The Gabor filter has been determined. , Given the characteristic parameter values, the expression after smoothing using a Gabor filter is:
[0088] (26)
[0089] Where N(i,j) represents the two-dimensional signal in the i-th row and j-th column, This represents the Gabor filter function, where q represents the filter width. This indicates the texture direction of the corresponding local image. This represents the frequency of local image texture features. This is the original signal. After being filtered by the Gabor filter, it becomes E(i,j). The filtering uses information from the pixels surrounding N(i,j), specifically the signal within a range of length 2c and width 2v around the i-th row and j-th column. When using the Gabor filter for smoothness filtering, it is essentially performing a two-dimensional convolution, which involves a large amount of computation, and the computational cost of the two-dimensional convolution for smoothness calculation is also enormous.
[0090] exist Under these relatively special conditions, the real part function expression of the Gabor filter can be:
[0091] (27)
[0092] (28)
[0093] Based on the process of solving the two-dimensional discrete cosine transform matrix, it can be found that the real part of the Gabor filter can be viewed as the product of two one-dimensional functions:
[0094] (29)
[0095] (30)
[0096] in This represents a one-dimensional bandpass function along the X-axis. This represents a one-dimensional low-pass function along the Y-axis. Assuming the frequency characteristic parameters of the image texture features to be processed are fixed in a certain direction in the local spatial domain, then the bandpass function along the corresponding X-axis has only one uncertain variable parameter x, and the low-pass function also has only one uncertain variable parameter y. Therefore, when the direction characteristic parameter is 90°, the process of the Gabor filter processing texture smoothness can be decomposed as follows: first, a one-dimensional bandpass filter is used to perform horizontal convolution filtering on each pixel within the local image block; then, a one-dimensional low-pass filter is used to perform horizontal convolution on the pixels within the local image block, followed by vertical convolution filtering, thus transforming the two-dimensional convolution problem into two one-dimensional convolution problems. By rotating the Gabor filter, this solution for this special case is extended to arbitrary angles, reducing the computational burden on the entire filter when processing smooth images.
[0097] When using a Gabor filter to smooth brain CT images, four characteristic parameters are determined: , , , The brain CT images are transformed to obtain the amplitude feature set of the Gabor transform of the brain CT images. A Gabor filter is then used for smoothing, with fast filtering at corresponding frequencies and directions, and the results are combined to obtain amplitude feature sets at different scales for corresponding frequencies and directions.
[0098] Filter setting direction The portion of the brain CT image that aligns with the direction of local texture features in the image to be processed, at the plane wave frequency of the sine wave. When the texture frequency of this portion of the image is the same, the real part features of the Gabor filter are used for image smoothing and restoration. A one-dimensional bandpass filter function is used to perform horizontal convolution filtering on each pixel within the local CT image block, and then a one-dimensional lowpass filter function is used to perform horizontal convolution filtering on the pixels within the local CT image block, followed by vertical convolution filtering.
[0099] The texture features of brain CT images, whether in densely textured or sparsely textured areas, do not show significant frequency differences. Therefore, the average frequency of the entire CT image is used to represent the frequency of each segment. This method reduces computational time.
[0100] S04: Input I4(x,y) into the gray-level co-occurrence matrix to obtain different features in four directions, and then extract features from brain CT images in a multi-scale context.
[0101] The gray-level co-occurrence matrix is used to find the spatial correlation characteristics of two pixels with a certain distance between them in the image space, thereby describing texture features.
[0102] Let I 2ij If (i,j) has size M×N and gray levels Ng, then the gray-level co-occurrence matrix satisfying certain spatial relationships is:
[0103] (31)
[0104] in Let P represent the number of elements in set x, and let P be an Ng×Ng matrix. If two pixels... grayscale is and grayscale The distance between them is d, and the angle between them and the horizontal axis of the coordinate system is... Then, gray-level co-occurrence matrices p with various spacings and angles can be obtained. , where p The value represents the number of times a pixel pair appears in the angular directions of 0°, 45°, 90°, and 135°, as shown in the following formula:
[0105] (32)
[0106] (33)
[0107] (34)
[0108] (35)
[0109] The values of x and y are chosen based on the characteristics of the texture periodic distribution. If Smaller values are used for images with slowly changing textures, and the values on the diagonal of the gray-level co-occurrence matrix are larger. The faster the texture changes, the smaller the values on the diagonal.
[0110] Meanwhile, the gray-level co-occurrence matrix is also derived from I4(x,y). Starting from the point of the pixel, the statistics are related to the pixel. Distance of points
[0111] (36)
[0112] If the probability P of pixels with grayscale value j appearing simultaneously is given by the probability function P:
[0113] (37)
[0114] like
[0115] like
[0116] like
[0117] like
[0118] but The relationship between the value of and the image is as follows:
[0119] (38)
[0120] Because the computation of the gray-level co-occurrence matrix is computationally intensive, in practical applications, the second-order angular moment is usually used. Contrast Correlation entropy , reverse moment difference Five of the most commonly used methods for extracting image texture features:
[0121] (39)
[0122] (40)
[0123] (41)
[0124] in
[0125] (42)
[0126] (43)
[0127] (44)
[0128] (45)
[0129] The grayscale distribution is represented by its moments, and the grayscale histogram is calculated using the following formula.
[0130] (46)
[0131] Where N1 and N2 represent the width and height of the image.
[0132] like Figure 4 This is a comparison between the original image and the filtered image obtained from the Gabor filter.
[0133] In step S04, the image texture features are described using the gray-level co-occurrence matrix (GLCM). During the analysis, it is found that the GLCM has poor rotation resistance. Furthermore, calculations show that the second moment of the angle is 0.8234, the correlation is 3.2222e+03, and the entropy is 13.6718. Therefore, the image texture features are relatively complex, and the GLCM does not achieve good rotation invariance.
[0134] S05: Optimize the rotation invariance of the gray-level co-occurrence matrix using an improved LBP algorithm.
[0135] Although the co-occurrence matrix (COM) has continuously improved its accuracy in image grayscale extraction, its resistance to rotation is relatively poor. Therefore, an improved Local Binary Pattern (LBP) algorithm was introduced based on the original COM. The grayscale value parameter of the center pixel of the cropped grayscale image is used as a reference threshold, and then compared with the neighborhood of the center pixel, thereby representing the local grayscale features through the obtained binary code.
[0136] The grayscale values around the center point of the captured grayscale image are compared with the center point's grayscale value. If the grayscale values of the surrounding points are greater than or equal to the center point's grayscale value, the grayscale values of the surrounding points are set to 1; otherwise, they are set to 0. Binary encoding is performed in clockwise / counterclockwise order, and the binary encoding is converted into decimal values to obtain feature values. The LBP algorithm is used to convert the two-dimensional matrix into one-dimensional values, achieving dimensionality reduction. According to the principle of the LBP algorithm, the joint distribution of image pixel grayscale values in the local neighborhood of the texture image is as follows:
[0137] (47)
[0138] in It is the gray value of the center pixel of the local neighborhood of the captured grayscale image. Let be a rotation counting function of order q. The corresponding pixel grayscale values are divided into q equal parts within a circular symmetrical neighborhood of radius R. A circular segmentation method is used.
[0139] The center pixel in the neighborhood The grayscale value and other points within the circular symmetrical area Subtracting the grayscale values, the expression is:
[0140] (48)
[0141] Meanwhile, assuming difference and They are independent of each other, therefore:
[0142] (49)
[0143] Texture features are represented by the following joint difference distribution:
[0144] (50)
[0145] difference It is unaffected by changes in the average light source, and the joint difference distribution is invariant to changes in grayscale.
[0146] (51)
[0147] (52)
[0148] By analyzing each symbol Assigning factors The above formula is transformed into a unique LBP value, used to describe the spatial structure of local image texture:
[0149] (53)
[0150] (54)
[0151] The LBP algorithm described above will generate Each of the q different output values corresponds to a value formed by the q pixels in the neighborhood. Several different binary modes. Grayscale values change when the image is rotated. and The relative positions will change, however, because The number of rotations changes synchronously with the value of q; that is, after the initial position is determined, the number of rotations will change from... By counting, we can determine the value corresponding to each rotation. Value, and Usually, the center is taken. The coordinates (0, r) to the right of the circle will result in different LBP values. However, rotation at any angle does not affect the relative positional relationship of 0 and 1 within the circular symmetric neighborhood. To remove rotation and obtain a unique LBP value, define:
[0152] (55)
[0153] (56)
[0154] in, The operation involves shifting the q-bit value X along the clock direction p times. For an image pixel, this means rotating the neighborhood set multiple times along the clock direction until the LBP value formed under the current rotation is minimized. As the number of rotations increases... Count each time The numerical values were confirmed, and the rotation direction was finally determined when the LBP value was the smallest.
[0155] Then, the LBP value is output through the equivalent pattern, and grayscale rotation invariance is achieved through the equivalent pattern.
[0156] (57)
[0157] (58)
[0158] (59)
[0159] This improvement reduces the number of binary modes without losing any information; the number of modes is reduced from the original... Reduce to +1 remaining patterns, which makes the feature vector less dimensional and reduces the impact of high-frequency noise.
[0160] The detailed process for extracting the features of the CT lesion region of the brain glioma described above is as follows: Figure 5 As shown.
[0161] Figure 6 It is a feature map processed by the improved LBP algorithm. Figure 7 This is an LBP histogram. It shows the center pixel of a local neighborhood of the extracted grayscale image. grayscale value, Let be a rotation counting function of order q. The corresponding pixel grayscale values are divided into q equal parts within a circular symmetrical neighborhood of radius R. A circular segmentation method is used to calculate the center pixel within the neighborhood. The grayscale value and other points within the circular symmetrical area The gray values are subtracted to calculate the sum of the differences, and the texture features are represented using a joint difference distribution.
[0162] difference The properties are unaffected by changes in the average light source. Therefore, the joint difference distribution is invariant to changes in grayscale. Rotation at any angle does not affect the relative positions of 0 and 1 within the circularly symmetric neighborhood. The process involves shifting the q bits x i times along the clock direction. For image pixels, this means rotating the neighborhood set many times along the clock direction until the LBP value formed under the current rotation is minimized. Of course, as the number of rotations changes, Count each time Once the numerical values are confirmed, the direction of rotation is determined by comparing and finding the value with the minimum LBP value. Then, the LBP value is output using an equivalent mode type to achieve grayscale rotation invariance.
Claims
1. A method for extracting features of CT lesions in brain gliomas, comprising: S1, segment the brain CT image I1(x,y) into A×A size images, and convert the segmented images into grayscale images I2(x,y); S2, divide I2(x,y) into N×N image blocks, and perform discrete cosine transform on each image block to obtain I3(x,y); S3, construct Gabor filter banks of different scales in four directions, and use the Gabor filter banks to filter I3(x,y) to obtain the amplitude feature set I4(x,y); the four directions refer to the four directions at 0°, 45°, 90°, and 135°; S4, input I4(x,y) into the gray-level co-occurrence matrix, and obtain the output in the four directions respectively to obtain the preliminary texture features of the brain CT image under multi-scale conditions; S5, by optimizing the rotation invariance of the gray-level co-occurrence matrix using an improved LBP algorithm, accurate texture features are obtained. The improved LBP algorithm refers to... Calculate the precise texture features, wherein, Perform an operation to shift the q-bit number X p times in a clockwise or counterclockwise direction. The following calculation process is described: Taking the currently calculated pixel as the center pixel, within a neighborhood with a radius of r, q sampling points are taken and compared with the gray value of the center pixel. If the gray value of the surrounding points is greater than or equal to the gray value of the center pixel, the gray value of the surrounding points is set to 1; otherwise, it is set to 0. Binary encoding is performed in clockwise or counterclockwise order, and the binary encoding is converted into a decimal value to obtain the feature value.
2. The method for extracting features of CT lesion regions of brain gliomas as described in claim 1, characterized in that, A=128。 3. The method for extracting features of CT lesions in brain gliomas as described in claim 1, characterized in that, N=8。 4. The method for extracting features of CT lesions in brain gliomas as described in claim 1, characterized in that, In step S2, the discrete cosine transform is a two-dimensional discrete cosine transform, which is calculated by decomposing it into two one-dimensional discrete cosine transform matrices in the row and column directions.
5. The method for extracting features of CT lesions in brain gliomas as described in claim 1, characterized in that, Step S3 includes: constructing a two-dimensional Gabor filter bank with different scales in four directions; and decomposing the real part function of the two-dimensional Gabor filter into a product of multiple one-dimensional filters for smoothing filtering.
6. The method for extracting features of CT lesion regions of brain gliomas as described in claim 5, characterized in that, Step S3 includes performing horizontal convolution filtering on each pixel within the image block using a one-dimensional bandpass filter function, and then performing vertical convolution filtering on the pixels within the image block using a one-dimensional lowpass filter function.
7. The method for extracting features of CT lesions in brain gliomas as described in claim 1, characterized in that, In step S4, the gray-level co-occurrence matrix outputs for I4(x, y) in the four directions are: ; ; ; ; in Let P represent the number of elements in set x, and let P be an Ng×Ng matrix. If two pixels... grayscale is and grayscale The distance between them is d, and the angle between them and the horizontal axis of the coordinate system is... , Indicates the spacing d and angle The gray-level co-occurrence matrix at time p The value indicates the number of times a pixel pair appears in the direction of an angle, such as 0°, 45°, 90°, and 135°.
8. The method for extracting features of CT lesion regions of brain gliomas as described in claim 1, characterized in that, In step S5, the LBP value is output using the equivalent mode.
9. The method for extracting features of CT lesion regions of brain gliomas as described in claim 1, characterized in that, In step S5, a joint difference distribution is used to represent texture features.
10. The method for extracting features of CT lesion regions of brain gliomas as described in any one of claims 1 to 9, characterized in that, when hour, ; otherwise, , , ; in, , It is the gray value of the center pixel of the local neighborhood of the captured grayscale image. Let be a rotation counting function of order q. The pixel grayscale values within a circular symmetrical neighborhood of radius R are divided into q equal parts, and a circular segmentation method is adopted.
Citation Information
Patent Citations
Facial expression identification method based on Gabor wavelet and gray-level co-occurrence matrix
CN105005765A
Multi-scale technology-based tumor texture feature extraction method
CN105869137A
