Schistosomiasis Suspected Lesion Area Calibration System Based on Deep Learning

Abstract

The present invention discloses a schistosomiasis suspicious lesion area calibration system based on deep learning, which relates to the technical field of image fusion and improves the positioning accuracy of lesion areas in B-ultrasound image data. The present invention performs three-channel feature extraction operations on the B-ultrasound image data of each lesion, and then sequentially extracts the suspicious lesion area, grade fibrosis curve, and displacement change vector from the B-ultrasound image data of each lesion. A multi-layer image detection network is set up, and the B-ultrasound image data of the lesion at different shooting angles is input into the multi-layer image detection network. According to the suspicious lesion area, grade fibrosis curve, and displacement change vector of the B-ultrasound image data of each lesion, fusion detection is performed on the B-ultrasound image data of each lesion, and then a multi-view dynamic lesion fiber image is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of image fusion, and specifically to a schistosomiasis suspicious lesion area calibration system based on deep learning. Background Art

[0002] Schistosomiasis is a parasitic disease that seriously endangers human health. Timely and accurate detection and diagnosis are crucial for the prevention and treatment of schistosomiasis. Currently, B-ultrasound examination is one of the commonly used means in the diagnosis of schistosomiasis, which can visually present the lesion conditions in the patient's body. However, the existing B-ultrasound diagnosis methods for schistosomiasis have certain limitations.

[0003] On the one hand, traditional B-ultrasound image analysis often can only obtain images from a single or a few fixed angles, making it difficult to comprehensively reflect the true characteristics and dynamic changes of the lesions. Since the morphology, position, and characteristics of the lesions may vary at different angles, the images from a single angle may miss some important information, leading to misdiagnosis or missed diagnosis.

[0004] On the other hand, for the recognition and analysis of the lesion areas in B-ultrasound images, most rely on the subjective judgment and experience of doctors, lacking objective and accurate quantitative criteria and automated processing methods. This not only increases the workload of doctors, but also there may be certain differences in the diagnostic results among different doctors, affecting the accuracy and consistency of the diagnosis. Therefore, a schistosomiasis suspicious lesion area calibration system based on deep learning is provided. Summary of the Invention

[0005] In order to solve the above technical problems, the purpose of the present invention is to provide a schistosomiasis suspicious lesion area calibration system based on deep learning.

[0006] In order to achieve the above purpose, the present invention provides the following technical solutions:

[0007] A schistosomiasis suspicious lesion area calibration system based on deep learning includes a B-ultrasound image acquisition module for detection, a lesion area positioning module, and an image fusion detection module;

[0008] The B-ultrasound image acquisition module for detection is used to obtain a number of lesion B-ultrasound image data of a patient from multiple shooting angles;

[0009] The lesion area positioning module is used to perform three-channel feature extraction operations on each lesion B-ultrasound image data, and then sequentially extract the suspicious lesion areas, grade fibrosis curves, and displacement change vectors from each lesion B-ultrasound image data;

[0010] The image fusion detection module is used to set up a multi-layer image detection network, input the B-ultrasound image data of lesions at different shooting angles into the multi-layer image detection network, and perform fusion detection on each B-ultrasound image data of lesions according to the suspicious lesion areas, grade fibrosis curves, and displacement change vectors of each B-ultrasound image data of lesions, so as to obtain a multi-view dynamic lesion fiber image.

[0011] Further, the acquisition process of the B-ultrasound image data of lesions includes:

[0012] The detection B-ultrasound image acquisition module configures six multi-parameter B-ultrasound devices controlled by a free manipulator, and each multi-parameter B-ultrasound device is internally equipped with a convex array probe and a high-frequency linear array probe;

[0013] Before B-ultrasound image acquisition, the convex array probe in each multi-parameter B-ultrasound device first scans the whole liver area of the patient and synchronously generates corresponding three-dimensional liver volume images;

[0014] Obtain the image features of the schistosomiasis lesion area through the Internet, then traverse the suspicious lesion areas in the three-dimensional liver volume image according to the image features, and then, according to the distribution of the suspicious lesion areas in the three-dimensional liver volume image in the patient's liver area at each shooting angle, each multi-parameter B-ultrasound device performs multi-angle composite scanning on the patient through the high-frequency linear array probe;

[0015] Furthermore, when the long axis direction of the probe of each high-frequency linear array probe deflects no more than 15° to the left and right, a number of B-ultrasound image data of lesions are sequentially taken at each shooting angle in chronological order.

[0016] Further, the shooting angles of the multi-angle composite scanning include the sub-xiphoid sagittal section, the left subcostal transverse section, the right subcostal margin section, the right intercostal oblique section, and the directions parallel and obliquely cut to the liver long axis of the high-frequency linear array probe.

[0017] Further, the three-channel feature extraction operation includes structural channel feature extraction, texture channel feature extraction, and dynamic channel feature extraction.

[0018] Further, the process of the structural channel feature extraction includes:

[0019] Perform grayscale processing on each B-ultrasound image data of lesions, and then establish a corresponding curvature feature matrix according to each grayscale pixel in the grayscale B-ultrasound image data of lesions. The curvature feature matrix is expressed as , where , , , respectively represent the second-order partial derivative and the first-order partial derivative in the x direction, and the second-order partial derivative and the first-order partial derivative in the y direction of the grayscale pixel;

[0020] Furthermore, eigenvalue decomposition is performed based on the curvature feature matrix of each grayscale pixel to obtain eigenvalues α1 and α2, and then the corresponding feature response values are obtained through the eigenvalues α1 and α2 of each grayscale pixel;

[0021] Since when there are fibrous structures in the B-ultrasound image data of the lesion, the second derivative in the x-axis direction is small, while the second derivative in the y direction is large, the feature response values of the grayscale pixels located in the fibrous structure are significantly higher than those of the grayscale pixels in the non-fibrous structure;

[0022] Furthermore, a feature response value threshold is set, and the grayscale pixels with feature response values less than or equal to the feature response value threshold are removed, and the grayscale pixels with feature response values less than or equal to the feature response value threshold are retained, and then the lesion suspicious regions are segmented from each B-ultrasound image data of the lesion.

[0023] Further, the process of texture channel feature extraction includes:

[0024] Perform grayscale processing on the lesion suspicious regions in each B-ultrasound image data of the lesion, and preset the texture detection path and texture inspection angle;

[0025] Texture is formed by the repeated occurrence of grayscale in spatial positions. Therefore, when the grayscale pixels in a grayscale image correspond to the same object, there is a specific grayscale relationship between two grayscale pixels even if they are separated by a certain distance and direction;

[0026] Generate the gray-level co-occurrence matrix corresponding to the lesion suspicious region from the texture inspection angle according to the texture detection path. It should be noted that each element in the gray-level co-occurrence matrix represents the probability that any grayscale element and other grayscale elements appear simultaneously in the lesion suspicious region under the texture detection path and a specific texture inspection angle;

[0027] Obtain the energy values of each grayscale element according to the grayscale matrix of each grayscale element, and map the energy values of each grayscale element in the form of point cloud distribution to the lesion suspicious region;

[0028] When the patient's liver tissue is normal, the energy value distribution in the lesion suspicious region has strong regularity and uniformity. When there is a pile of schistosome fibers in the patient's liver tissue, the energy distribution of each grayscale element in the lesion suspicious region is disordered, and the energy value is smaller than that in the normal condition;

[0029] Furthermore, an energy threshold is set according to the energy value under the normal condition of the liver tissue. Then, for the grayscale elements in the lesion suspicious region with energy values less than the energy threshold, the corresponding grayscale elements are marked as suspicious lesion elements, otherwise the corresponding grayscale elements are ignored;

[0030] Set multiple energy value level intervals, and set the same label for the suspicious lesion elements within the same energy value level interval. Connect the adjacent suspicious lesion elements with the same label in the suspicious lesion area in sequence, and then divide several texture boundary lines in the suspicious lesion area;

[0031] Mark the grayscale pixels between two texture boundary lines at adjacent plane positions, and then draw several grade fibrosis curves in the suspicious lesion area. It should be noted that the grade of each grade fibrosis curve is equal to the average value of the grades of the energy value level intervals associated with the corresponding two texture boundary lines.

[0032] Further, the process of dynamic channel feature extraction includes:

[0033] Arrange the suspicious lesion areas corresponding to several B-ultrasound image data of the lesion at the same shooting angle in chronological order, obtain the maximum width of each grade fibrosis curve in the suspicious lesion area, and the time series displacement value between the corresponding grayscale pixels in the grade fibrosis curves in adjacent time series, and count the maximum value of the time series displacement values among them;

[0034] Then, establish a dynamic tracking window with the set maximum width as the length and the maximum value of the time series displacement value as the width. Then, capture the displacement change positions of the grayscale pixels in each grade fibrosis curve in chronological order through the dynamic tracking window, and generate corresponding displacement change vectors.

[0035] Further, the process of data image fusion for the B-ultrasound image data of the lesion includes:

[0036] The image fusion detection module is provided with multiple layers of image detection networks for data image fusion of the B-ultrasound image data of the lesion at the same shooting angle;

[0037] According to the shooting angle and chronological order of each B-ultrasound image data of the lesion, input the suspicious lesion areas in each B-ultrasound image data into the multiple layers of image detection networks from six directions and chronological order;

[0038] Set several dynamic detection neurons according to the number of grayscale pixels in the suspicious lesion area, and input the grayscale elements in the suspicious lesion area into each dynamic detection neuron in chronological order;

[0039] Set several image detection time points according to the number of B-ultrasound image data of the lesion collected from each shooting angle by the detection B-ultrasound image acquisition module. Then, whenever an image detection time point ends, the dynamic detection neuron moves to the corresponding position at the next image detection time point according to the displacement change vector included in the grayscale elements it contains;

[0040] Meanwhile, each shooting angle is successively used as the main shooting angle, and the remaining shooting angles are used as secondary shooting angles. As a result, the dynamic detection neurons of the main shooting angle do not perform any operations, and for the remaining ones, according to the included angle between the dynamic detection neurons of the secondary shooting angles and those of the main shooting angle, angle transformation is performed on the dynamic detection neurons of the secondary shooting angles, thereby generating dynamic lesion fiber images under different shooting perspectives;

[0041] Further, the process of detecting the fusion result of the B-ultrasound image data of the lesion includes:

[0042] Map the dynamic lesion fiber images at each shooting angle onto a multi-layer image detection network according to their corresponding main shooting angles. Since each dynamic lesion fiber image corresponds to the lesion area of the same patient;

[0043] Set a pixel difference threshold, compare the pixel values of the grayscale pixels corresponding to the same position and the same time sequence with each other. If the pixel differences between the pixel values of each grayscale pixel are all below the pixel difference threshold, directly accumulate and average the pixel values of the grayscale pixels corresponding to the same position and the same time sequence, and then reassign them to each grayscale pixel;

[0044] Otherwise, eliminate the grayscale pixels whose pixel differences between more than two grayscale pixels are greater than or equal to the pixel difference threshold. Then, accumulate and average the pixel values of the remaining grayscale pixels, and reassign them to each grayscale pixel;

[0045] Repeat the above pixel value averaging operation to obtain a multi-perspective dynamic lesion fiber image.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1. Through the three-channel feature extraction operation, the present invention successively extracts the lesion suspicious area, the grade fibrosis curve, and the displacement change vector from each B-ultrasound image data of the lesion, improves the efficiency of feature extraction, reduces the interference of human factors, and through the analysis of the grade fibrosis curve and the displacement change vector, the development status and dynamic changes of the lesion can be deeply understood.

[0048] 2. By performing fusion detection on the B-ultrasound image data of the lesion at different shooting angles, the detection ability of the lesion is enhanced, and a clearer and more accurate multi-perspective dynamic lesion fiber image is obtained. It helps to accurately judge the nature and development stage of the lesion, and improves the diagnostic accuracy and treatment effect of schistosomiasis. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the schematic diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0050] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation manners, structures, features and their effects of the present invention as follows.

[0051] As Figure 1 shown, a schistosomiasis suspicious lesion area calibration system based on deep learning includes a detection B-ultrasound image acquisition module, a lesion area positioning module, and an image fusion detection module;

[0052] The detection B-ultrasound image acquisition module is used to acquire a number of lesion B-ultrasound image data of a patient from multiple shooting angles;

[0053] The lesion area positioning module is used to perform three-channel feature extraction operations on each lesion B-ultrasound image data, and then sequentially extract the lesion suspicious area, the grade fibrosis curve, and the displacement change vector from each lesion B-ultrasound image data;

[0054] The image fusion detection module is used to set up a multi-layer image detection network, input the lesion B-ultrasound image data at different shooting angles into the multi-layer image detection network, and perform fusion detection on each lesion B-ultrasound image data according to the lesion suspicious area, the grade fibrosis curve, and the displacement change vector of each lesion B-ultrasound image data, so as to obtain a multi-view dynamic lesion fiber image.

[0055] Further, the working principle of the present invention is illustrated by the following embodiments:

[0056] The detection B-ultrasound image acquisition module is configured with six multi-parameter B-ultrasound devices controlled by a free robotic arm, and each multi-parameter B-ultrasound device is internally provided with a convex array probe and a high-frequency linear array probe;

[0057] Before B-ultrasound image acquisition, the convex array probe in each multi-parameter B-ultrasound device first scans the whole liver area of the patient and synchronously generates the corresponding liver three-dimensional volume image;

[0058] Obtain the image features of the schistosomiasis lesion area through the Internet, and then traverse the lesion suspicious area in the liver three-dimensional volume image according to the image features. Then, according to the distribution of the lesion suspicious area in the liver three-dimensional volume image in the patient's liver area at each shooting angle, each multi-parameter B-ultrasound device performs multi-angle composite scanning on the patient through the high-frequency linear array probe;

[0059] The shooting angles of the multi-angle composite scanning include the sub-xiphoid sagittal section, the left subcostal transverse section, the right costal margin section, the right intercostal oblique section, and the directions parallel and oblique to the long axis of the liver of the high-frequency linear array probe. Furthermore, when the long axis direction of each high-frequency linear array probe deflects no more than 15° to the left and right, a number of lesion B-ultrasound image data are sequentially taken at each shooting angle in chronological order.

[0060] Further, the B-ultrasound image acquisition module sends all the B-ultrasound image data of the lesions to the lesion area localization module;

[0061] The lesion area localization module performs a three-channel feature extraction operation on any B-ultrasound image data of the lesions. The three-channel feature extraction operation includes structural channel feature extraction, texture channel feature extraction, and dynamic channel feature extraction;

[0062] Among them, the process of structural channel feature extraction includes:

[0063] Perform grayscale processing on each B-ultrasound image data of the lesions, and then establish a corresponding curvature feature matrix according to each grayscale pixel in the grayscale B-ultrasound image data. The curvature feature matrix is expressed as where , , , The grayscale pixels respectively represent the second-order partial derivative and the first-order partial derivative in the x direction, and the second-order partial derivative and the first-order partial derivative in the y direction;

[0064] Then, perform eigenvalue decomposition according to the curvature feature matrix of each grayscale pixel to obtain eigenvalues α1 and α2, and then obtain the corresponding feature response values through the eigenvalues α1 and α2 of each grayscale pixel;

[0065] Since when there are fibrous structures in the B-ultrasound image data of the lesions, the second-order derivative along the x-axis direction is small, while the second-order derivative in the y direction is large, making the feature response values of the grayscale pixels located in the fibrous structures significantly higher than those of the grayscale pixels in non-fibrous structures;

[0066] Then, set a feature response value threshold, eliminate the grayscale pixels whose feature response values are less than or equal to the feature response value threshold, retain the grayscale pixels whose feature response values are less than or equal to the feature response value threshold, and then segment the suspicious lesion areas in each B-ultrasound image data of the lesions.

[0067] Among them, the process of texture channel feature extraction includes:

[0068] Perform grayscale processing on the suspicious lesion areas in each B-ultrasound image data of the lesions, and preset a texture detection path and a texture inspection angle. The texture inspection angles include 0°, 45°, 90°, and 135°;

[0069] Texture is formed by the repeated appearance of grayscale in spatial positions. Therefore, when the grayscale pixels in a grayscale image correspond to the same object, there is a specific grayscale relationship between two grayscale pixels even if they are separated by a certain distance and direction;

[0070] Generate a gray-level co-occurrence matrix corresponding to the suspicious lesion area from the texture inspection angle according to the texture detection path. It should be noted that each element in the gray-level co-occurrence matrix represents the probability that any gray-level element and other gray-level elements appear simultaneously in the suspicious lesion area under the texture detection path and a specific texture inspection angle;

[0071] Obtain the energy value of each gray-level element according to the gray-level matrix of each gray-level element, and map the energy values of each gray-level element to the suspicious lesion area in the form of a point cloud distribution;

[0072] When the patient's liver tissue is normal, the energy value distribution in the suspicious lesion area has strong regularity and uniformity. When there is a pile of schistosome fibrous bodies in the patient's liver tissue, the energy distribution of each gray-level element in the suspicious lesion area is disordered, and the energy value is smaller than that in the normal condition;

[0073] Furthermore, set an energy threshold according to the energy value in the normal condition of the liver tissue. Then, for the gray-level elements in the suspicious lesion area whose energy values are less than the energy threshold, mark the corresponding gray-level elements as suspicious lesion elements, otherwise ignore the corresponding gray-level elements;

[0074] Set multiple energy value level intervals, and set the same label for the suspicious lesion elements in the same energy value level interval. Connect the adjacent suspicious lesion elements with the same label in the suspicious lesion area in sequence, and then divide several texture boundary lines in the suspicious lesion area;

[0075] Mark the gray-level pixels between two texture boundary lines at adjacent plane positions, and then draw several grade fibrosis curves in the suspicious lesion area. It should be noted that the grade of each grade fibrosis curve is equal to the average value of the grades of the energy value level intervals associated with the corresponding two texture boundary lines.

[0076] Among them, the process of dynamic channel feature extraction includes:

[0077] Arrange the suspicious lesion areas corresponding to several B-ultrasound image data of the lesion at the same shooting angle in chronological order, obtain the maximum width of each grade fibrosis curve in the suspicious lesion area, and the time-series displacement value between the corresponding same gray-level pixels in the grade fibrosis curves in adjacent time sequences, and count the maximum value of the time-series displacement values among them;

[0078] Furthermore, establish a dynamic tracking window with the set maximum width as the length and the maximum value of the time-series displacement value as the width. Then, capture the displacement change positions of the gray-level pixels in each grade fibrosis curve in chronological order through the dynamic tracking window, and generate corresponding displacement change vectors;

[0079] When all the B-ultrasound image data of the lesions complete the three-channel feature extraction operation, the lesion area positioning module sends the results of all the three-channel feature extraction operations to the two-stage inspection module.

[0080] Furthermore, the image fusion detection module is provided with a multi-layer image detection network for performing data image fusion on the B-ultrasound image data of the lesions at the same shooting angle. The specific process includes:

[0081] According to the shooting angles and time sequence of each B-ultrasound image data of the lesions, the suspicious lesion areas in each B-ultrasound image data are input into the multi-layer image detection network from six directions and in time sequence;

[0082] A number of dynamic detection neurons are set according to the number of grayscale pixels in the suspicious lesion area, and the grayscale elements in the suspicious lesion area are input into each dynamic detection neuron in each time sequence;

[0083] A number of image detection time points are set according to the number of B-ultrasound image data of the lesions collected by the detection B-ultrasound image acquisition module at each shooting angle. Then, whenever an image detection time point ends, the dynamic detection neuron moves to the corresponding position at the next image detection time point according to the displacement change vector contained in the grayscale elements it contains;

[0084] At the same time, each shooting angle is taken as the main shooting angle in turn, and the remaining shooting angles are taken as the secondary shooting angles. Thus, the dynamic detection neurons of the main shooting angle do not perform any operations, and the remaining dynamic detection neurons of the secondary shooting angles are subjected to angle transformation according to the angle between them and the dynamic detection neurons of the main shooting angle, so as to generate dynamic lesion fiber images under different shooting perspectives;

[0085] The dynamic lesion fiber images at each shooting angle are mapped onto the multi-layer image detection network according to their corresponding main shooting angles. Since each dynamic lesion fiber image corresponds to the lesion area of the same patient;

[0086] A pixel difference threshold is set, and the pixel values of the grayscale pixels at the same position and in the same time sequence are compared with each other. If the pixel differences between the pixel values of each grayscale pixel are all below the pixel difference threshold, the pixel values of the grayscale pixels at the same position and in the same time sequence are directly added up and averaged, and then the averaged value is re-assigned to each grayscale pixel;

[0087] Otherwise, the grayscale pixels with pixel differences between more than two grayscale pixels greater than or equal to the pixel difference threshold are excluded, and then the pixel values of the remaining grayscale pixels are added up and averaged, and the averaged value is re-assigned to each grayscale pixel;

[0088] Repeat the above pixel value averaging operation to obtain multi-view dynamic lesion fiber images.

[0089] The above are only the preferred embodiments of the present invention and do not impose any formal restrictions on the present invention. Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments by using the disclosed technical content within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any brief modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A deep learning-based schistosomiasis suspected lesion area calibration system, characterized by: It includes a B-ultrasound image acquisition module, a lesion area positioning module and an image fusion detection module; The detection B-ultrasound image acquisition module is used to acquire a plurality of B-ultrasound image data of lesions of the patient from multiple shooting angles; The lesion area positioning module is used to perform a three-channel feature extraction operation on each lesion B-ultrasound image data, and then extract the lesion suspicious area, grade fibrosis curve and displacement change vector from each lesion B-ultrasound image data in sequence; The three-channel feature extraction operation includes structural channel feature extraction, texture channel feature extraction and dynamic channel feature extraction; The image fusion detection module is used to set up a multi-layer image detection network, input the lesion B-ultrasound image data under different shooting angles into the multi-layer image detection network, and perform fusion detection on each lesion B-ultrasound image data according to the lesion suspicious area, grade fibrosis curve and displacement change vector of each lesion B-ultrasound image data, so as to obtain a multi-view dynamic lesion fiber image; The process of extracting the structural channel features includes: Gray-scale processing is performed on each lesion B-ultrasound image data, and a corresponding curvature feature matrix is ​​established according to each gray-scale pixel in the gray-scaled lesion B-ultrasound image data; Perform eigenvalue decomposition according to the curvature feature matrix of each grayscale pixel to obtain its corresponding characteristic response value, set a characteristic response value threshold, remove the grayscale pixels whose characteristic response values ​​are less than or equal to the characteristic response value threshold, and retain the grayscale pixels whose characteristic response values ​​are greater than the characteristic response value threshold, thereby segmenting the suspicious lesion area in each lesion B-ultrasound image data; The process of extracting the texture channel features includes: Set the texture detection path and texture inspection angle, and generate the gray level co-occurrence matrix corresponding to the suspicious lesion area from the texture inspection angle according to the texture detection path; According to the grayscale matrix of each grayscale element, the energy value of each grayscale element is obtained, and the energy value of each grayscale element is mapped to the suspicious lesion area in the form of point cloud distribution, and then the energy threshold is set according to the energy value of the liver tissue under normal conditions. For the grayscale element in the suspicious lesion area whose energy value is less than the energy threshold, the corresponding grayscale element is recorded as a suspicious lesion element, otherwise the corresponding grayscale element is ignored; Setting multiple energy value level intervals, and setting the same label for suspicious lesion elements in the same energy value level interval, connecting adjacent suspicious lesion elements with the same label in the suspicious lesion area in sequence, and then dividing a number of texture boundary lines in the suspicious lesion area; Annotate the grayscale pixels between two texture boundary lines on adjacent plane positions, and then draw several graded fibrosis curves in the suspicious lesion area; The process of extracting dynamic channel features includes: The suspicious lesion areas corresponding to several lesion B-ultrasound image data taken at the same shooting angle are arranged in chronological order, the maximum width of each grade fibrosis curve in the suspicious lesion area is obtained, and the time series displacement value between the same grayscale pixel in the grade fibrosis curve in the adjacent time sequence is obtained, and the maximum value of the time series displacement value is counted; Then, a dynamic tracking window is established with the maximum width set as the length and the maximum value of the time series displacement value as the width, and then the displacement change position of the grayscale pixel in each level fiberization curve in the time sequence is captured in sequence through the dynamic tracking window, and the corresponding displacement change vector is generated.

2. The system for demarcating suspected schistosomiasis lesion areas based on deep learning according to claim 1, characterized in that: The acquisition process of the lesion B-ultrasound image data includes: The detection B-ultrasound image acquisition module is equipped with six multi-parameter B-ultrasound devices controlled by free mechanical arms, each of which has a built-in convex array probe and a high-frequency linear array probe; Before B-ultrasound image acquisition, the convex array probe in each multi-parameter B-ultrasound device first scans the patient's entire liver area and simultaneously generates a corresponding three-dimensional volume image of the liver; The image features of the schistosomiasis lesion area are obtained through the Internet, and then the suspicious lesion area is traversed in the three-dimensional volume image of the liver according to the image features. Then, based on the distribution of the suspicious lesion area in the three-dimensional volume image of the liver in the patient's liver area at various shooting angles, various multi-parameter B-ultrasound devices perform multi-angle composite scans on the patient through high-frequency linear array probes, and then take a number of lesion B-ultrasound image data at various shooting angles in chronological order.

3. The deep learning-based schistosomiasis suspected lesion area calibration system according to claim 2, characterized in that: The shooting angles of the multi-angle composite scan include the subxiphoid sagittal section, the left subcostal transverse section, the right subcostal section, the right intercostal oblique section, and the high-frequency linear array probe parallel to and oblique to the long axis of the liver.

4. The system for demarcating suspected schistosomiasis lesion areas based on deep learning according to claim 3, characterized in that: The process of data image fusion of lesion B-ultrasound image data includes: According to the shooting angle and time sequence of each lesion B-ultrasound image data, the suspicious lesion area in each lesion B-ultrasound image data is input into the multi-layer image detection network from six directions and time sequence; A number of dynamic detection neurons are set according to the number of grayscale pixels in the suspicious lesion area, and a number of image detection time points are set, and then each time an image detection time point ends, the dynamic detection neuron moves to a corresponding position at the next image detection time point according to the displacement change vector contained in the grayscale element contained in it; At the same time, each shooting angle is taken as the main shooting angle, and the remaining shooting angles are taken as secondary shooting angles, so that the dynamic detection neurons of the main shooting angle do not perform any operation, and the remaining dynamic detection neurons of the secondary shooting angle are transformed according to the angle between the secondary shooting angle and the dynamic detection neurons of the main shooting angle, thereby generating dynamic lesion fiber images under different shooting angles.

5. The deep learning-based schistosomiasis suspected lesion area calibration system according to claim 4, characterized in that: The process of detecting the fusion results of the lesion B-ultrasound image data includes: A pixel difference threshold is set, and the pixel values ​​of the grayscale pixels corresponding to the same position and the same time sequence are compared with each other. If the pixel difference between the pixel values ​​of each grayscale pixel is below the pixel difference threshold, the pixel values ​​of the grayscale pixels corresponding to the same position and the same time sequence are directly accumulated and averaged, and then reassigned to each grayscale pixel; Otherwise, the grayscale pixels whose pixel difference between two grayscale pixels is greater than or equal to the pixel difference threshold are eliminated, the pixel values ​​of the retained grayscale pixels are accumulated and averaged, and then each grayscale pixel is reassigned to obtain a multi-view dynamic lesion fiber image.

Images