An image processing method based on artificial intelligence

Abstract

The application discloses an image processing method based on artificial intelligence and belongs to the technical field of image processing. The application sets a sliding window to an ultrasonic image, acquires row and column texture correlation of pixel points, converts and exponentially amplifies irrelevance, filters row and column lesion points, extracts the same target lesion points, constitutes first and second lesion target regions, cuts the region boundary, acquires sharpness and branching degree and constitutes a sequence, processes the sequence through a lesion region recognition neural network, and obtains lesion region classification. The application improves classification accuracy through accurate extraction of lesion texture and morphological features and intelligent analysis, and provides an efficient technical scheme for ultrasonic lesion diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an image processing method based on artificial intelligence. BACKGROUND

[0002] In the field of modern medical diagnosis, ultrasonic imaging technology has become an indispensable means in clinical detection due to its significant advantages such as non-invasiveness, real-time, low cost, and is widely used in the screening and diagnosis of lesions in multiple fields such as abdominal organs, cardiovascular system, gynecology and obstetrics, etc. However, the ultrasonic image itself has the characteristics of low resolution, easy to be disturbed by noise, complex tissue texture, etc., which brings great challenges to the accurate identification and classification of the lesion area.

[0003] Traditional ultrasonic image processing methods rely on manual experience for feature extraction and analysis, which is not only inefficient but also subjective, and is prone to cause deviations in diagnosis results due to differences in experience of operators, and is difficult to meet the high requirements of clinical diagnosis accuracy and consistency. Especially in the detection of some early lesions or small lesions, since the texture difference between the lesion area and the surrounding normal tissue is not obvious, the traditional method is often difficult to achieve accurate identification, which is easy to cause missed diagnosis or misdiagnosis, and delays the best treatment opportunity of the patient.

[0004] With the rapid development of artificial intelligence technology, image processing methods based on deep learning provide a new way to solve the above problems. However, the existing ultrasonic image processing technology based on artificial intelligence focuses on the overall gray scale features of the lesion area and the contrast between the lesion area and the normal area, ignoring the subtle features of texture changes, and some methods fail to effectively amplify the differences between the lesion and normal tissue in the feature processing process, resulting in insufficient feature discrimination and affecting the accuracy of lesion area division, leading to low accuracy of lesion area classification. SUMMARY

[0005] In view of the above deficiencies in the prior art, the present application provides an image processing method based on artificial intelligence to solve the problem of low accuracy of lesion area classification in the prior art.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: an image processing method based on artificial intelligence, comprising the following steps:

[0007] A sliding window is set for the ultrasonic image to obtain the row texture correlation and column texture correlation of each pixel point;

[0008] The row and column texture correlations are converted into non-correlations and are exponentially amplified to screen out row lesion points and column lesion points;

[0009] The row lesion points and column lesion points belonging to the same target are extracted to form a first lesion target area and a second lesion target area;

[0010] respectively, the sharpness and the branch degree of each segment are obtained to constitute a sharpness sequence and a branch degree sequence of the first lesion target region and the second lesion target region;

[0011] The sharpness sequence and the branch degree sequence of the first lesion target region and the second lesion target region are processed by using a lesion region identification neural network to obtain a classification of the lesion region.

[0012] Further, the process of obtaining the row texture correlation and the column texture correlation of each pixel point comprises:

[0013] A sliding window is used to slide on each row of the ultrasound image, and the sliding step is 1 pixel point;

[0014] On each row, the texture correlation of the i-th sliding window and the i+1-th sliding window is taken as the row texture correlation of a new pixel point in the i+1-th sliding window, and i is the number of sliding times;

[0015] A sliding window is used to slide on each column of the ultrasound image, and the sliding step is 1 pixel point;

[0016] On each column, the texture correlation of the i-th sliding window and the i+1-th sliding window is taken as the column texture correlation of a new pixel point in the i+1-th sliding window.

[0017] Further, the process of screening out the row lesion points and the column lesion points comprises:

[0018] The row texture correlation and the column texture correlation are respectively subtracted by 1 to obtain row texture disassociation and column texture disassociation;

[0019] The row texture disassociation and the column texture disassociation are respectively exponentially amplified to obtain a row texture disassociation amplification value and a column texture disassociation amplification value;

[0020] The row texture disassociation amplification values of each pixel point on the ultrasound image are averaged to obtain a row disassociation average value;

[0021] The column texture disassociation amplification values of each pixel point on the ultrasound image are averaged to obtain a column disassociation average value;

[0022] Pixel points with a row texture disassociation amplification value greater than the row disassociation average value are marked as row lesion points;

[0023] Pixel points with a column texture disassociation amplification value greater than the column disassociation average value are marked as column lesion points.

[0024] Further, the process of constituting the first lesion target region and the second lesion target region comprises:

[0025] The row disease points belonging to the same area range are classified into a row correlation group;

[0026] The column disease points belonging to the same area range are classified into a column correlation group;

[0027] The pixel point coordinates of the row correlation group and the column correlation group are taken as an intersection-over-union ratio;

[0028] When the intersection-over-union ratio is greater than an intersection-over-union ratio threshold value, the corresponding row correlation group is taken as a first disease target area, and the corresponding column correlation group is taken as a second disease target area.

[0029] Further, the process of constructing the sharpness sequence and the branch degree sequence of the first disease target area and the second disease target area includes:

[0030] In the first disease target area, a row disease point existing in a neighborhood range of a non-row disease point is marked as a boundary disease point;

[0031] In the second disease target area, a column disease point existing in a neighborhood range of a non-column disease point is marked as a boundary disease point;

[0032] In the first disease target area and the second disease target area, respectively, each boundary disease point is divided into multiple segments with any boundary disease point as a starting point, to obtain multiple boundary segments;

[0033] According to the distance of each boundary disease point on each boundary segment to a straight line, the sharpness and the branch degree are obtained, each sharpness belonging to the same disease target area is constructed into a sharpness sequence, and each branch degree belonging to the same disease target area is constructed into a branch degree sequence.

[0034] Further, the process of obtaining the sharpness and the branch degree includes:

[0035] Connecting the starting point and the ending point of each boundary segment, a straight line is obtained;

[0036] The distance of each boundary disease point in each boundary segment to the straight line is calculated;

[0037] The boundary disease point corresponding to the maximum distance is selected as a farthest point from the distances;

[0038] The ratio of the distance of the farthest point to the length of the straight line is taken as the sharpness of the boundary segment;

[0039] The local maximum value is selected from the distances, and the boundary disease point corresponding to the local maximum value is taken as a branch point when the local maximum value is greater than a distance threshold value;

[0040] According to the difference between the distance of each branch point in the same boundary segment and the distance threshold value, the branch degree is obtained.

[0041] Further, the branch degree acquisition process comprises: subtracting the distance threshold from the distance of each branch point to obtain a difference of each branch point, and adding the differences of the branch points corresponding to each boundary segment, and normalizing the addition result to obtain the branch degree.

[0042] Further, the lesion region recognition neural network comprises: a plurality of feature extraction units, a multiplier M1, a multiplier M2, a multiplier M3, a multiplier M4, a splicing layer, a convolution processing unit and a full connection layer.

[0043] The input end of each feature extraction unit is used for inputting a sequence.

[0044] The input end of the multiplier M1 is connected with the second output end of the first feature extraction unit and the second output end of the second feature extraction unit respectively; the input end of the multiplier M2 is connected with the first output end of the first feature extraction unit and the first output end of the second feature extraction unit respectively; the input end of the multiplier M3 is connected with the second output end of the third feature extraction unit and the second output end of the fourth feature extraction unit respectively; the input end of the multiplier M4 is connected with the first output end of the third feature extraction unit and the first output end of the fourth feature extraction unit respectively.

[0045] The input end of the splicing layer is connected with the output end of the multiplier M1, the output end of the multiplier M2, the output end of the multiplier M3 and the output end of the multiplier M4 respectively, and the output end thereof is connected with the input end of the convolution processing unit.

[0046] The input end of the full connection layer is connected with the output end of the convolution processing unit, and the output end thereof serves as the output end of the lesion region recognition neural network.

[0047] Further, each of the plurality of feature extraction units comprises: a first convolution layer, a maximum pooling layer and an average pooling layer.

[0048] The input end of the first convolution layer serves as the input end of the feature extraction unit, and the output end thereof is connected with the input end of the maximum pooling layer and the input end of the average pooling layer respectively.

[0049] The output end of the maximum pooling layer serves as the first output end of the feature extraction unit.

[0050] The output end of the average pooling layer serves as the second output end of the feature extraction unit.

[0051] Further, the convolution processing unit comprises, which are connected in sequence: a first stacked convolution module, a residual convolution module and a second stacked convolution module.

[0052] In summary, the beneficial effects of the present application are:

[0053] 1. This invention obtains the row and column texture correlation of each pixel by setting a sliding window, which can capture the difference in texture between the lesion area and normal tissue at a subtle level, making up for the deficiency of traditional methods in paying insufficient attention to texture features, and can extract more accurate lesion target areas.

[0054] 2. This invention converts the correlation between row and column textures into non-correlation and amplifies it exponentially, which can significantly amplify the originally subtle differences in lesion features, make lesion points more prominent, improve the accuracy of screening row and column lesion points, and improve the accuracy of lesion area division.

[0055] 3. After segmenting the boundary of the lesion area, this invention quantifies the boundary morphology by sharpness (degree of boundary prominence) and branching degree (boundary branching characteristics), and characterizes the lesion from a morphological perspective (e.g., the boundary of malignant tumors is more irregular and the sharpness / branching degree characteristics are more significant).

[0056] 4. This invention uses a lesion area recognition neural network to process the sharpness sequence and branching sequence of the first lesion target area and the second lesion target area, and combines the sharpness sequence and branching sequence of the lesion target area obtained from different orientations to improve the classification accuracy of the lesion target area. Attached Figure Description

[0057] Figure 1 This is a flowchart of an image processing method based on artificial intelligence;

[0058] Figure 2 A schematic diagram of the structure of a neural network for identifying lesion areas;

[0059] Figure 3 This is a schematic diagram of the feature extraction unit.

[0060] Figure 4 This is a schematic diagram of the structure of a convolution processing unit;

[0061] Figure 5 This is an ultrasound illustration of a benign cyst.

[0062] Figure 6 This is a schematic diagram of an ultrasound scan of a malignant tumor. Detailed Implementation

[0063] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0064] likeFigure 1 As shown in the figure, an artificial intelligence-based image processing method comprises the following steps:

[0065] A sliding window is set for the ultrasound image to obtain row texture correlation and column texture correlation of each pixel point;

[0066] The row and column texture correlations are converted into disassociation and exponentially amplified to screen out row lesion points and column lesion points;

[0067] The row lesion points and column lesion points belonging to the same target are extracted to form a first lesion target region and a second lesion target region;

[0068] The boundaries of the first lesion target region and the second lesion target region are respectively segmented to obtain the sharpness and branching degree of each segment, and a sharpness sequence and a branching degree sequence of the first lesion target region and the second lesion target region are formed;

[0069] The lesion region recognition neural network is used to process the sharpness sequence and the branching degree sequence of the first lesion target region and the second lesion target region to obtain the classification of the lesion region.

[0070] In this embodiment, the process of obtaining the row texture correlation and the column texture correlation of each pixel point comprises:

[0071] The sliding window is slid on each row of the ultrasound image with a sliding step of 1 pixel point;

[0072] On each row, the texture correlation of the i-th sliding window and the i+1-th sliding window is taken as the row texture correlation of the new pixel point in the i+1-th sliding window, and i is the number of sliding times;

[0073] The sliding window is slid on each column of the ultrasound image with a sliding step of 1 pixel point;

[0074] On each column, the texture correlation of the i-th sliding window and the i+1-th sliding window is taken as the column texture correlation of the new pixel point in the i+1-th sliding window.

[0075] In this embodiment, the size of the sliding window is set to 1x5, and the calculation of the texture correlation can use cosine similarity or Pearson correlation coefficient, specifically: the cosine similarity or Pearson correlation coefficient of the pixel value under the i-th sliding window and the pixel value under the i+1-th sliding window is calculated as the texture correlation. In this embodiment, on the row texture correlation, there are 5 columns in front, and on the column texture correlation, the pixel points of the 5 rows in front cannot obtain the texture correlation, and the texture correlation of these pixel points is assigned as 1, or the pixel points of the edges of the image are not considered.

[0076] The present application adopts a sliding window to slide on each row and each column of the ultrasonic image with 1 pixel point as a sliding step, so that each pixel point can be associated with the texture information of the adjacent region, the row texture association and the column texture association of each pixel point are calculated according to the distribution of the texture on the same column or the same row, so that the coincidence degree of the texture in the row or column direction is reflected.

[0077] In the embodiment, the process of screening the row lesion points and the column lesion points comprises:

[0078] The row texture inassociation and the column texture inassociation are obtained by subtracting 1 from the row texture association and the column texture association respectively;

[0079] The row texture inassociation amplification value and the column texture inassociation amplification value are obtained by exponentially amplifying the row texture inassociation and the column texture inassociation respectively;

[0080] The row inassociation average value is obtained by taking the average of the row texture inassociation amplification values of the pixel points on the ultrasonic image;

[0081] The column inassociation average value is obtained by taking the average of the column texture inassociation amplification values of the pixel points on the ultrasonic image;

[0082] The pixel points with the row texture inassociation amplification value greater than the row inassociation average value are marked as the row lesion points;

[0083] The pixel points with the column texture inassociation amplification value greater than the column inassociation average value are marked as the column lesion points.

[0084] In the embodiment, the formula of the row texture inassociation amplification value is X L = exp(x L ), wherein X L is the row texture inassociation amplification value, x L is the row texture inassociation, and exp( ) is a natural exponential function; the formula of the column texture inassociation amplification value is X A = exp(x A ), wherein X A is the column texture inassociation amplification value, and x A is the column texture inassociation.

[0085] The present application subtracts 1 from the row and column texture association to obtain the inassociation, converts the "texture coincidence degree" into "texture difference degree", exponentially amplifies the row and column texture inassociation, can use the nonlinear characteristics of the exponential function to significantly amplify the tiny inassociation difference, the amplification value of the pixel points with higher inassociation (may be lesion points) will increase sharply, and the amplification value of the normal pixel points with lower inassociation grows slowly, so that the problem of the original texture difference not being obvious is effectively solved.

[0086] In the embodiment, the process of constituting the first lesion target region and the second lesion target region comprises:

[0087] The row lesion points belonging to the same region range are classified into a row association group;

[0088] The column lesion points belonging to the same region range are classified into a column association group;

[0089] The intersection-over-union of the pixel point coordinates of the row association group and the column association group is calculated;

[0090] When the intersection-over-union is greater than an intersection-over-union threshold value, the corresponding row association group is taken as the first lesion target region, and the corresponding column association group is taken as the second lesion target region.

[0091] In the embodiment, the intersection-over-union threshold value is set to a value greater than 0, that is, the row association group and the column association group exist in partial lesion point association, which is determined to belong to the same lesion target.

[0092] In the embodiment, the process of classifying the row lesion points belonging to the same region range into a row association group comprises:

[0093] A1, taking any row lesion point as the center, detecting whether there are other row lesion points in the 5*5 neighborhood range, if there are, then these other row lesion points and the center row lesion point are classified into a row association group;

[0094] A2, traversing all row lesion points in the current row association group that are not the center, taking each point as the center in turn, detecting whether there are other row lesion points not included in the row association group in the 5*5 neighborhood range, if there are, then these new row lesion points are included in the row association group;

[0095] A3, repeating the process of A2 until all row lesion points in the row association group are detected as the center, and no new row lesion point can be included.

[0096] In the embodiment, the process of classifying the column lesion points belonging to the same region range into a column association group comprises:

[0097] B1, taking any column lesion point as the center, detecting whether there are other column lesion points in the 5*5 neighborhood range, if there are, then these other column lesion points and the center column lesion point are classified into a column association group;

[0098] B2, traversing all column lesion points in the current column association group that are not the center, taking each point as the center in turn, detecting whether there are other column lesion points not included in the column association group in the 5*5 neighborhood range, if there are, then these new column lesion points are included in the column association group;

[0099] B3, repeat the process of B2 until all column lesion points in the column correlation group are detected as the center and no new column lesion points can be included.

[0100] In the embodiment, the process of constructing the sharpness sequence and the branching degree sequence of the first lesion target region and the second lesion target region includes:

[0101] In the first lesion target region, mark the row lesion point existing in the neighborhood range as the boundary lesion point;

[0102] In the second lesion target region, mark the column lesion point existing in the neighborhood range as the boundary lesion point;

[0103] In the first lesion target region and the second lesion target region respectively, divide each boundary lesion point into multiple segments with any boundary lesion point as the starting point, to obtain multiple boundary segments;

[0104] According to the distance of the boundary lesion point on each boundary segment to the straight line, obtain the sharpness and the branching degree, and construct the sharpness sequence of each sharpness belonging to the same lesion target region, and construct the branching degree sequence of each branching degree belonging to the same lesion target region.

[0105] The boundary of the lesion region is the transition zone of the lesion and the normal tissue, and there is necessarily the feature that the lesion point is adjacent to the non-lesion point, therefore, in the embodiment, each row lesion point in the first lesion target region is taken as the center, and the non-row lesion point exists in the 3*3 neighborhood range, and the row lesion point at the center is the boundary point, and the second lesion target region is the same.

[0106] In the embodiment, the boundary segment length can be set to 10-15 pixels, that is, the continuous 10-15 boundary lesion points are divided into a boundary segment.

[0107] The process of obtaining the sharpness and the branching degree of the first lesion target region and the second lesion target region is the same, and both include:

[0108] Connect the starting point and the ending point of each boundary segment to obtain a straight line;

[0109] Calculate the distance of each boundary lesion point in each boundary segment to the straight line;

[0110] Select the boundary lesion point corresponding to the maximum distance from the distances as the farthest point;

[0111] Take the ratio of the distance of the farthest point to the length of the straight line as the sharpness of the boundary segment;

[0112] Select the local maximum value from the distances, and when the local maximum value is greater than the distance threshold, take the corresponding boundary lesion point as the branching point;

[0113] The branching degree is obtained by comparing the distance between each branch point in the same boundary segment with the distance threshold.

[0114] The equation of the straight line is: Ax + By + C = 0, A = y2 - y1, B = x1 - x2, C = x2y1 - x1y2, where the starting coordinates of the boundary segment are (x1, y1) and the ending coordinates are (x2, y2), x is the independent variable of the equation, and y is the dependent variable. The formula for calculating the distance from the nth boundary lesion point in the boundary segment to the straight line is:

[0115] , where d n Let x be the distance from the nth boundary lesion point in the boundary segment to the straight line. n ,y n ) represents the coordinates of the nth boundary lesion point in the boundary segment, and || represents the absolute value.

[0116] The local maxima satisfy the condition: d n-1 <d n And d n >d n+1 The boundary lesion point, where d n Let d be the distance from the nth boundary lesion point on the boundary segment to the straight line. n-1 d is the distance from the (n-1)th boundary lesion point on the boundary segment to the straight line. n+1 It is the distance from the (n+1)th boundary lesion point on the boundary segment to the straight line.

[0117] In this embodiment, the distance threshold is set based on experiments or experience.

[0118] In this embodiment, the process of obtaining the branch degree includes: subtracting the distance of each branch point from the distance threshold to obtain the difference of each branch point, adding the differences of each branch point corresponding to the same boundary segment, and normalizing the sum to obtain the branch degree.

[0119] The formula for branch degree is: Where μ is the branching degree, d r Let d be the distance to the r-th branch point. th denoted as the distance threshold, D as the normalization parameter, R as the number of branch points, and r as the branch point number.

[0120] This invention uses the line connecting the start and end points of a boundary segment as a reference, and selects the ratio of the distance of the farthest point to the length of the straight line as the sharpness, which can intuitively and effectively reflect the degree of protrusion and steepness of the boundary segment.

[0121] The formula for calculating sharpness is: θ i =d i / S i , where θi Let d be the sharpness of the i-th boundary segment. i S is the distance from the farthest point on the i-th boundary segment to the straight line. i Let be the length of the line.

[0122] This invention selects points with local maxima greater than a distance threshold as branch points, enabling precise identification of meaningful boundary bifurcations and eliminating interference from minor fluctuations. The branching degree is obtained by summing and normalizing the differences between the distance to each branch point and the distance threshold, reflecting not only the number of branch points but also the prominence of each branch point.

[0123] like Figure 2 As shown, the lesion area recognition neural network includes: multiple feature extraction units, multipliers M1, M2, M3, and M4, a splicing layer, a convolutional processing unit, and a fully connected layer.

[0124] The input terminal of each feature extraction unit is used to input a sequence;

[0125] The input terminals of multiplier M1 are connected to the second output terminals of the first feature extraction unit and the second feature extraction unit, respectively; the input terminals of multiplier M2 are connected to the first output terminals of the first feature extraction unit and the second feature extraction unit, respectively; the input terminals of multiplier M3 are connected to the second output terminals of the third feature extraction unit and the fourth feature extraction unit, respectively; and the input terminals of multiplier M4 are connected to the first output terminals of the third feature extraction unit and the fourth feature extraction unit, respectively.

[0126] The input of the splicing layer is connected to the output of multipliers M1, M2, M3 and M4 respectively, and its output is connected to the input of the convolution processing unit.

[0127] The input of the fully connected layer is connected to the output of the convolutional processing unit, and its output serves as the output of the lesion region recognition neural network.

[0128] like Figure 3 As shown, each feature extraction unit includes: a first convolutional layer, a max pooling layer, and an average pooling layer;

[0129] The input of the first convolutional layer serves as the input of the feature extraction unit, and its output is connected to the input of the max pooling layer and the input of the average pooling layer, respectively.

[0130] The output of the max pooling layer is used as the first output of the feature extraction unit.

[0131] The output of the average pooling layer serves as the second output of the feature extraction unit.

[0132] In this invention, the first convolutional layer is a one-dimensional convolution, the max pooling layer and the average pooling layer are one-dimensional pooling layers, the kernel size of the first convolutional layer is 1×3, the pooling window of the max pooling layer is 1×3, and the pooling window of the average pooling layer is 1×3.

[0133] This invention captures the "extreme features" of lesion morphology by focusing peak features in the sequence using max pooling, while average pooling preserves the overall trend of the sequence. The invention processes one sequence per feature extraction unit, extracting peak features and overall trend features. A multiplier M1 fuses the overall trend features output by the first and second feature extraction units, and a multiplier M2 fuses the peak features output by the first and second feature extraction units, thereby enhancing the overall trend and peak features of the first lesion target region.

[0134] Multiplier M3 is used to fuse the overall trend features output by the third feature extraction unit and the overall trend features output by the fourth feature extraction unit. Multiplier M4 is used to fuse the peak features output by the third feature extraction unit and the peak features output by the fourth feature extraction unit, thereby enhancing the overall trend features and peak features of the second lesion target area.

[0135] The concatenation layer is used to concatenate the outputs of multipliers M1, M2, M3, and M4 to form a 4×M feature matrix, where M is the length of the feature output by the multipliers. The convolutional processing unit is used to further extract features from the output of the concatenation layer. The fully connected layer performs classification based on the output of the convolutional processing unit.

[0136] like Figure 4 As shown, the convolution processing unit includes, in sequence: a first stacked convolution module, a residual convolution module, and a second stacked convolution module.

[0137] Each stacked convolutional module includes a second convolutional layer and a third convolutional layer, which are connected sequentially. The kernel size of both the second and third convolutional layers is 3×3.

[0138] The residual convolution module uses standard residual blocks, which include two convolutional layers and an adder.

[0139] like Figure 5 The benign cyst shown has clear borders, as... Figure 6 The malignant tumors shown have more irregular borders, higher sharpness, and more branching, while benign lesions have relatively regular borders. These sequence characteristics provide a more detailed and discriminative basis for lesion classification.

[0140] In this embodiment, the classification types include: benign lesions and malignant lesions.

[0141] This invention obtains the row and column texture correlation of each pixel by setting a sliding window, which can capture the difference in texture between the lesion area and normal tissue at a subtle level, making up for the shortcomings of traditional methods that do not pay enough attention to texture features, and can extract more accurate lesion target areas.

[0142] This invention converts the correlation between row and column textures into non-correlation and amplifies it exponentially, which can significantly amplify the originally subtle differences in lesion features, make lesion points more prominent, improve the accuracy of screening row and column lesion points, and improve the accuracy of lesion area division.

[0143] After segmenting the boundary of the lesion area, this invention quantifies the boundary morphology by sharpness (degree of boundary prominence) and branching degree (boundary branching characteristics), thus characterizing the lesion from a morphological perspective (e.g., malignant tumors have more irregular boundaries and more significant sharpness / branching characteristics).

[0144] This invention employs a lesion region recognition neural network to process the sharpness and branching sequences of the first and second lesion target regions, and combines the sharpness and branching sequences of the lesion target regions obtained from different orientations to improve the classification accuracy of the lesion target regions.

[0145] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An image processing method based on artificial intelligence, characterized in that, Includes the following steps: A sliding window is set for the ultrasound image to obtain the row texture correlation and column texture correlation of each pixel; The correlation between row and column textures is converted into non-correlation and then exponentially amplified to filter out row lesion points and column lesion points; Extract the lesion points in the row and column that belong to the same target to form the first lesion target region and the second lesion target region; The boundaries of the first and second lesion target regions are segmented respectively, and the sharpness and branching degree of each segment are obtained to form the sharpness sequence and branching degree sequence of the first and second lesion target regions. A lesion region identification neural network is used to process the sharpness and branching sequences of the first and second lesion target regions to obtain the classification of the lesion regions. The lesion region identification neural network includes: multiple feature extraction units, multipliers M1, M2, M3, and M4, a splicing layer, a convolutional processing unit, and a fully connected layer; The input terminal of each feature extraction unit is used to input a sequence; The input terminals of multiplier M1 are connected to the second output terminals of the first feature extraction unit and the second feature extraction unit, respectively; the input terminals of multiplier M2 are connected to the first output terminals of the first feature extraction unit and the second feature extraction unit, respectively; the input terminals of multiplier M3 are connected to the second output terminals of the third feature extraction unit and the fourth feature extraction unit, respectively; and the input terminals of multiplier M4 are connected to the first output terminals of the third feature extraction unit and the fourth feature extraction unit, respectively. The input of the splicing layer is connected to the output of multipliers M1, M2, M3 and M4 respectively, and its output is connected to the input of the convolution processing unit. The input of the fully connected layer is connected to the output of the convolutional processing unit, and its output serves as the output of the lesion region recognition neural network. Each feature extraction unit includes: a first convolutional layer, a max pooling layer, and an average pooling layer; The input of the first convolutional layer serves as the input of the feature extraction unit, and its output is connected to the input of the max pooling layer and the input of the average pooling layer, respectively. The output of the max pooling layer is used as the first output of the feature extraction unit. The output of the average pooling layer serves as the second output of the feature extraction unit. The convolution processing unit includes, in sequence: a first stacked convolution module, a residual convolution module, and a second stacked convolution module.

2. The image processing method based on artificial intelligence according to claim 1, characterized in that, The process of obtaining the row texture correlation and column texture correlation of each pixel includes: A sliding window is used to slide across each row of the ultrasound image, with a sliding step of 1 pixel. In each row, the texture association between the i-th sliding window and the (i+1)-th sliding window is used as the row texture association of the new pixel in the (i+1)-th sliding window, where i is the number of the sliding number. A sliding window is used to slide across each column of the ultrasound image, with a sliding step of 1 pixel. In each column, the texture correlation between the i-th sliding window and the (i+1)-th sliding window is used as the column texture correlation of the new pixel in the (i+1)-th sliding window.

3. The image processing method based on artificial intelligence according to claim 1, characterized in that, The process of screening out-of-line lesion points and column lesion points includes: Subtracting row texture correlation and column texture correlation from 1 respectively yields row texture non-correlated and column texture non-correlated results; The row texture non-association and column texture non-association are exponentially amplified to obtain the row texture non-association amplification value and the column texture non-association amplification value, respectively; The mean value of the uncorrelated magnification of the row texture of each pixel in the ultrasound image is obtained by taking the mean value of the row uncorrelated magnification. The mean value of the uncorrelated column texture magnification of each pixel in the ultrasound image is obtained by taking the mean value of the uncorrelated column texture. Pixels whose row texture non-association magnification value is greater than the row non-association mean value are marked as row lesion points; Pixels whose uncorrelated magnification value is greater than the uncorrelated mean value of the column texture are marked as column lesions.

4. The image processing method based on artificial intelligence according to claim 1, characterized in that, The process of constructing the first and second lesion target regions includes: Lesions belonging to the same region are grouped into a single line association group; Lesions belonging to the same region are grouped into a single lesion association group; Take the intersection and union of the pixel coordinates of the row association group and the column association group; When the crossover ratio is greater than the crossover ratio threshold, the corresponding row association group is taken as the first lesion target area, and the corresponding column association group is taken as the second lesion target area.

5. The image processing method based on artificial intelligence according to claim 1, characterized in that, The process of constructing the sharpness sequence and branching sequence of the first and second lesion target regions includes: In the first lesion target region, mark the lesion points that are not lesion points within the neighborhood of the target region as boundary lesion points; In the second lesion target region, mark the column lesions that have non-column lesion points within their neighborhood as boundary lesion points; In the first and second lesion target regions respectively, each boundary lesion point is divided into multiple segments, starting from any boundary lesion point, to obtain multiple boundary segments; Based on the distance from the lesion point on the upper boundary of each boundary segment to the straight line, the sharpness and branching degree are obtained. The sharpness of each lesion belonging to the same lesion target area is formed into a sharpness sequence, and the branching degree of each lesion target area is formed into a branching degree sequence.

6. The image processing method based on artificial intelligence according to claim 1 or 5, characterized in that, The process of obtaining sharpness and branching includes: Connect the start and end points of each boundary segment to obtain a straight line; Calculate the distance from each boundary lesion point in each boundary segment to the straight line; Select the boundary lesion point corresponding to the maximum distance from all distances as the farthest point; The ratio of the distance to the farthest point to the length of the line is used as the sharpness of the boundary segment; Local maxima are selected from various distances, and the corresponding boundary lesion points are taken as branch points when the local maxima are greater than the distance threshold. The branching degree is obtained by comparing the distance between each branch point in the same boundary segment with the distance threshold.

7. The image processing method based on artificial intelligence according to claim 6, characterized in that, The process of obtaining the branch degree includes: subtracting the distance of each branch point from the distance threshold to obtain the difference of each branch point, adding the differences of each branch point corresponding to the same boundary segment, and normalizing the sum to obtain the branch degree.

Images