Kidney ultrasonic image classification method and equipment
Through the classification method of renal ultrasound images, the similarity measurement of distance-weighted graphs and polar coordinate graphs with reference image data sets is solved in the existing technology of low sensitivity and artificial interpretation dependence, and high efficiency is achieved , accurate renal ultrasound image classification reduces the doctor's labor intensity and false negative risk.
Patent Information
- Application Number
- CN202411260807.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2024-11-19
AI Technical Summary
In the diagnosis of chronic kidney disease, the sensitivity of indicators such as serum creatinine, estimated glomerular filtration rate and urinary albumin to creatinine ratio is low, and the renal biopsy is an invasive examination, which has risks and false negative risks. , Ultrasound image diagnosis relies on doctors’ manual interpretation, is subjective and inefficient.
A kidney ultrasound image classification method is adopted to generate distance-weighted graphs and polar coordinate graphs by collecting and preprocessing ultrasound images, and a feature extractor is used to vectorize the image, and similarity measurement is performed with the reference image data set, and classification results are output. Reduce the labor intensity of doctors and improve classification accuracy.
It realizes accurate classification of renal ultrasound images without relying on a large number of doctors to interpret, improves diagnostic efficiency and accuracy, reduces the risk of false negatives, and provides a non-invasive, cost-effective imaging diagnosis method.
Smart Images

Figure CN118982690A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical imaging technology, and in particular to a classification method and device for kidney ultrasonic images. Background Art
[0002] Chronic kidney disease is a progressive disease characterized by changes in kidney structure and function caused by a variety of reasons. Pathological manifestations include renal parenchymal damage and a reduction in functional nephrons. Currently, chronic kidney disease is incurable. Early detection of chronic kidney disease and the use of the best treatment plan can slow the progression of renal dysfunction, prevent complications, and reduce malignant consequences. However, chronic kidney disease generally has no obvious symptoms in the early stages, and less than 5% of patients are aware of their disease status in the early stages. When clinical symptoms appear, renal function is already severely impaired. As the disease progresses, chronic kidney disease can lead to a series of complications such as anemia, metabolic disorders, and secondary hyperparathyroidism, and eventually develop into end-stage renal disease, which can only be treated by renal dialysis or renal transplantation, but the prognosis is usually poor and the treatment cost is expensive. Timely and accurate diagnosis of chronic kidney disease, assessment of renal function, degree of damage, and pathological type can provide important references for the formulation and implementation of timely and effective treatment plans, delay disease progression, and improve patient prognosis.
[0003] At present, the diagnosis and staging of chronic kidney disease in clinical practice are based on the patient's renal function and renal damage indicators, with the main indicators being serum creatinine, estimated glomerular filtration rate, and urine albumin to creatinine ratio. However, serum creatinine is easily affected by a variety of endogenous and exogenous factors such as race, drugs, diet, and muscle content, and will only change when renal function is damaged to a certain extent, with low sensitivity; the calculation of estimated glomerular filtration rate depends on serum creatinine, and the results have certain deviations, and cannot measure the renal function of each kidney; the urine albumin to creatinine ratio shows potential in indicating early renal damage, but is easily affected by a variety of factors such as blood pressure level, diabetes status, and drugs, with large individual differences, and the degree of popularization is significantly inconsistent in different regions. Renal biopsy is currently the gold standard for the clinical diagnosis of chronic kidney disease, but it is an invasive examination method with risks such as hematuria, perinephric hematoma, and puncture site infection. It is not easily accepted by patients and cannot be performed repeatedly in clinical diagnosis and treatment. In addition, due to the small amount of samples obtained during the puncture process, renal biopsy faces the problem of sampling bias, which cannot fully reflect the overall condition of the kidney and has a greater risk of false negatives. Renal ultrasound has been widely adopted as the preferred imaging diagnostic method for chronic kidney disease because it is non-invasive, cost-effective, and can provide results immediately. However, the diagnosis of ultrasound images relies on the manual interpretation of doctors, which is highly subjective, has high requirements for diagnostic experience, is inefficient, and difficult to quantify. Summary of the invention
[0004] In order to solve the above technical problems, the present invention provides a method and device for classifying kidney ultrasound images which can reduce the workload of doctors and have high classification accuracy.
[0005] The technical solution of the present invention is:
[0006] A classification method for kidney ultrasound images, comprising:
[0007] S1: collecting a group of ultrasound images of the kidney, manually annotating the collected group of ultrasound images, selecting the most typical ultrasound image of the kidney as a first ultrasound image of the kidney, and marking the long axis endpoint of the first ultrasound image of the kidney;
[0008] S2: preprocessing the first renal ultrasound image to obtain a second renal ultrasound image;
[0009] S3: generating a distance weighted map and a polar coordinate map of the second kidney ultrasound image, and vectorizing the distance weighted map and the polar coordinate map through a feature extractor;
[0010] S4: performing similarity measurement on the image features of the second kidney ultrasound image and the reference image data set, and outputting the classification result when the measurement value exceeds a set threshold;
[0011] The reference image data set includes kidney images collected from people of different age groups and in different health conditions, and steps S1-S3 are executed to obtain the reference image data set.
[0012] Preferably, in step S2, the first kidney ultrasound image obtained in step S1 is denoised by median filtering or Gaussian filtering, the ultrasound image is truncated at the 95% site of the intensity histogram of the ultrasound image, the midpoint between the end points of the long axis of the kidney is used as the midpoint of the image, the image is cropped to a uniform size, and the pixel intensity of the ultrasound image is normalized.
[0013] In any of the above solutions, preferably, the second kidney ultrasound image acquired in step S2 is rotated, mirror-flipped, and noise is added.
[0014] In any of the above schemes, preferably, the long axis endpoints of the second renal ultrasound image obtained in S2 are randomly sampled within the neighborhood.
[0015] In any of the above schemes, preferably, the midpoint between the two manually marked kidney long axis endpoints is taken as the kidney center point C, the distance between the kidney long axis endpoints is taken as the kidney long diameter L, and the angle between the kidney long axis endpoints and the horizontal axis is taken as the initial angle φ;
[0016] The distance weighted graph is obtained by the following formula:
[0017]
[0018] Where I(x, y) and I′(x, y) are the pixel intensities of the renal ultrasound image and the distance weighted graph at point (x, y), d is the distance between point (x, y) and the center point C of the kidney, and λ is the attention coefficient;
[0019] The initial polar coordinate plot is obtained by the following formula:
[0020] P(θ′,d)=I(x,y)if d≤λL
[0021] θ′=θ-φ
[0022] Among them, P(θ′, d) is the pixel intensity of the initial polar coordinate map at point (θ′, d), θ is the angle between the line connecting point (x, y) and the kidney center point C and the horizontal axis, d is the distance between point (x, y) and the kidney center point C, and λ is the attention coefficient.
[0023] Preferably, in any of the above solutions, measuring the similarity with the reference image set includes:
[0024] At least one similarity measurement algorithm among Euclidean distance, cosine similarity, Pearson correlation coefficient, and edge detection is used to obtain a similarity set, and the maximum similarities obtained by various algorithms are accumulated. When the accumulated value exceeds a set threshold, the classification result is output.
[0025] In any of the above schemes, it is preferred that the feature extractor is a combination of a convolutional neural network and a multi-layer perceptron, and the distance-weighted map and the polar coordinate map of the second kidney ultrasound image are used as different views, which are simultaneously input into different branches of the convolutional neural network through a multi-view learning method, and view-level feature fusion is performed at the high level of the convolutional neural network.
[0026] In any of the above solutions, preferably, the infoNCE function is used as the loss function to evaluate the current performance of the feature extractor, and the calculation formula is:
[0027]
[0028] Among them, z i and z j Represents the image features of training samples from the same patient; z k Represents z i All negative samples and z j ; τ is a positive temperature parameter; N is z i The number of negative samples plus 1.
[0029] In any of the above solutions, it is preferred that the method comprises: a memory and a processor;
[0030] The memory is used to store program instructions;
[0031] The processor is used to call program instructions, and when the program instructions are called, any one of the above-mentioned kidney ultrasound image classification methods is executed.
[0032] Compared with the prior art, the present invention collects a group of ultrasound images of the kidneys through a device, and the collected group of ultrasound images of the kidneys are processed to form a first kidney ultrasound image. The first kidney ultrasound image is processed to form a second kidney ultrasound image. The image features of the second kidney ultrasound image and the reference image data set are measured for similarity, and when the measurement value exceeds a set threshold, the classification result is output. Among them, the reference image data set includes kidney images collected from people of different age groups and different health conditions, and the reference image data set is obtained by executing the above steps. During the classification process, doctors do not need to spend too much time to identify the image images, which reduces the labor intensity of doctors and has high classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is a schematic diagram of the flow of the classification method of kidney ultrasound images of the present invention;
[0034] Figure 2 It is a flow chart of a feature extractor training method of the kidney ultrasound image classification method of the present invention. DETAILED DESCRIPTION
[0035] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0036] Embodiment 1:
[0037] When diagnosing chronic kidney disease with ultrasound, cross-sectional images are important observations. By marking the long axis of the kidney on ultrasound and evaluating the size of the kidney through the cross-sectional image where the long axis of the kidney is located, certain diseases can be indicated.
[0038] like Figure 1 In the illustrated embodiment, the ultrasound device in step S1 can be a general-purpose Doppler ultrasound diagnostic instrument. A set of ultrasound images of the kidneys are acquired by medical staff, wherein the number of acquired images is 2-8. After acquiring a set of ultrasound images of the kidneys, an experienced ultrasound physician selects the most representative kidney ultrasound image from each kidney ultrasound image as the first kidney ultrasound image, and marks the long axis endpoint of the first kidney ultrasound image. At this time, the first kidney ultrasound image shows the shapes of the kidney section, renal columns, renal pyramids, renal cortex, and renal sinus.
[0039] In step S2, the first renal ultrasound image obtained in step S1 is preprocessed by truncation, size cropping, normalization, etc. to form a second renal ultrasound image. The second renal ultrasound image formed at this time has a unified measurable form. Among them, when truncating the first renal ultrasound image, a simple threshold truncation algorithm, a piecewise linear truncation algorithm, or an adaptive truncation algorithm can be used. For size cropping, a direct cropping algorithm based on pixel coordinates can be used. Preferably, the image size is 16cm×16cm. When normalizing, preferably, a minimum-maximum normalization algorithm, a global threshold algorithm, or a local threshold algorithm can be used.
[0040] Step S3 generates a distance weighted graph and a polar coordinate graph of the second kidney ultrasound image, and vectorizes the distance weighted graph and the polar coordinate graph through a feature extractor to prepare for similarity measurement between the second kidney ultrasound image of the patient to be diagnosed and the reference image dataset. In the construction of the distance weighted graph and the polar coordinate graph, the midpoint between the two manually marked kidney long axis endpoints is used as the kidney center point C, the distance between the kidney long axis endpoints is used as the kidney long diameter L, and the angle between the kidney long axis endpoints and the horizontal axis is used as the initial angle φ;
[0041] The distance weighted graph is obtained by the following formula:
[0042]
[0043] Where I(x, y) and I′(x, y) are the pixel intensities of the renal ultrasound image and the distance weighted graph at point (x, y), d is the distance between point (x, y) and the center point C of the kidney, and λ is the attention coefficient;
[0044] The initial polar coordinate plot is obtained by the following formula:
[0045] P(θ′,d)=I(x,y)if d≤λL
[0046] θ′=θ-φ
[0047] Where P(θ′, d) is the pixel intensity of the initial polar coordinate graph at point (θ′, d), θ is the angle between the line connecting point (x, y) and the kidney center point C and the horizontal axis, d is the distance between point (x, y) and the kidney center point C, and λ is the attention coefficient. Preferably, the value is 0.6. The size of the initial polar coordinate graph is adjusted to the distance weighted value by the interpolation algorithm. Figure 1 A polar coordinate diagram is obtained, preferably, bicubic interpolation is used.
[0048] In step S4, the image features of the second kidney ultrasound image of the patient are measured for similarity with the reference image dataset to find out which sample in the reference dataset matches the shape features of the patient's renal cross-section, renal columns, renal pyramids, renal cortex, and renal sinus. When the measurement value exceeds the set threshold, the classification result is output.
[0049] Among them, the reference image data set in this embodiment includes kidney images of people of different age groups and different health conditions. When the reference image data set is established, the vectorized data of the distance weighted map and polar coordinate map of the second kidney ultrasound image of people of different age groups and different health conditions are obtained by executing steps S1-S3.
[0050] Embodiment 2:
[0051] On the basis of Example 1, in order to ensure the accuracy of feature extraction in steps S2 and S3, the first kidney ultrasound image obtained in step S1 is denoised by median filtering or Gaussian filtering, the ultrasound image is truncated at the 95% site of the intensity histogram of the ultrasound image, the midpoint between the end points of the long axis of the kidney is used as the midpoint of the image, the image is cropped to a uniform size, and the pixel intensity of the ultrasound image is normalized.
[0052] In order to further improve the accuracy of feature extraction in step S3, the second kidney ultrasound image obtained in step S2 is rotated, mirror-flipped, and noise is added (to enhance the display of dark areas).
[0053] In order to further improve the accuracy of feature extraction in step S3, the long axis endpoint of the second kidney ultrasound image obtained in step S2 is randomly sampled in the neighborhood. Specifically, a point is randomly selected in the 9×9 neighborhood of the kidney long axis endpoint as the new endpoint to reduce the doctor's error in marking the long axis endpoint.
[0054] Embodiment 3:
[0055] In order to improve the accuracy of similarity measurement between the image features of the second kidney ultrasound image and the reference image dataset, at least one similarity measurement algorithm among Euclidean distance, cosine similarity, Pearson correlation coefficient, and edge detection is used to obtain a similarity set, and the maximum similarity values obtained by various algorithms are accumulated. When the accumulated value exceeds the set threshold, the classification result is output, so as to more comprehensively evaluate the similarity between the second kidney ultrasound image of the patient and the second kidney ultrasound image in the reference image dataset.
[0056] Embodiment 4:
[0057] On the basis of any one of embodiments 1-3, Figure 2In the illustrated embodiment, in order to improve the performance of the feature extractor, the feature extractor is a combination of a convolutional neural network and a multi-layer perceptron. In a specific application, the distance weighted map and the polar coordinate map of the second kidney ultrasound image are used as different views, and are simultaneously input into different branches of the convolutional neural network through a multi-view learning method. Feature fusion at the view level is performed at the high level of the convolutional neural network. The end of the convolutional neural network is a global average pooling layer, which outputs an initial feature vector. Preferably, the backbone of the convolutional neural network uses Densenet, and the dimension of the output initial feature vector is 1024, and the weight is initialized by the Xavier method. The multi-layer perceptron maps the initial feature vector to a high dimension first, preferably setting the dimension to 2048, and then reduces the dimension to a lower dimension, preferably setting the dimension to 512, to obtain image features. Among them, the feature extractor is built using Python 3.6 language based on the Pytorch deep learning framework.
[0058] In this embodiment, the infoNCE function is used as the loss function to evaluate the current performance of the feature extractor, and the calculation formula is:
[0059]
[0060] Among them, z i and z j Image features representing training samples from the same patient (positive sample pairs); z k Represents z i All negative samples and z j ; A is a positive temperature parameter, preferably set to 0.5; N is z i The number of negative samples plus 1. sim(A,B) is the cosine similarity function, and the calculation formula is:
[0061]
[0062] Wherein, A and B are two non-zero vectors; A·B represents the dot product of A and B; ‖A‖ and ‖B‖ are the Euclidean lengths of A and B respectively. If the loss function value does not meet the training end condition, the parameters of the convolutional neural network and the multilayer perceptron of the feature extractor are adjusted and the iterative training is continued until the value of the loss function meets the training end condition, thereby obtaining a trained feature extractor.
[0063] Embodiment 5:
[0064] A device for kidney ultrasound images comprises: a memory and a processor, wherein the memory is used to store program instructions, and the processor is used to call the program instructions. When the program instructions are called, the kidney ultrasound image classification method described in the above embodiment is executed.
[0065] The embodiments described above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention.
Claims
1. A method for classifying renal ultrasound images, characterized in that: include: S1: collecting a group of ultrasound images of the kidney, manually annotating the collected group of ultrasound images, selecting the most typical ultrasound image of the kidney as a first ultrasound image of the kidney, and marking the long axis endpoint of the first ultrasound image of the kidney; S2: preprocessing the first renal ultrasound image to obtain a second renal ultrasound image; S3: generating a distance weighted map and a polar coordinate map of the second kidney ultrasound image, and vectorizing the distance weighted map and the polar coordinate map through a feature extractor; S4: performing similarity measurement on the image features of the second kidney ultrasound image and the reference image data set, and outputting the classification result when the measurement value exceeds a set threshold; The reference image data set includes kidney images collected from people of different age groups and in different health conditions, and steps S1-S3 are executed to obtain the reference image data set.
2. The method for classifying renal ultrasound images according to claim 1, characterized in that: In step S2, the first kidney ultrasound image obtained in step S1 is denoised by using a median filter or a Gaussian filter, the ultrasound image is truncated at the 95% site of the intensity histogram of the ultrasound image, the midpoint between the end points of the kidney's long axis is used as the image midpoint, the image is cropped to a uniform size, and the pixel intensity of the ultrasound image is normalized.
3. The method for classifying renal ultrasound images according to claim 1 or 2, characterized in that: The second kidney ultrasound image obtained in step S2 is rotated, mirror-flipped, and noise is added.
4. The method for classifying renal ultrasound images according to claim 3, characterized in that: The long-axis endpoints of the second kidney ultrasound images obtained in S2 were randomly sampled within the neighborhood.
5. The method for classifying renal ultrasound images according to claim 1, characterized in that: The midpoint between the two manually marked endpoints of the kidney's long axis was taken as the kidney's center point C, the distance between the endpoints of the kidney's long axis was taken as the kidney's long diameter L, and the angle between the line connecting the endpoints of the kidney's long axis and the horizontal axis was taken as the initial angle φ; The distance weighted graph is obtained by the following formula: Where I(x, y) and I′(x, y) are the pixel intensities of the renal ultrasound image and the distance weighted graph at point (x, y), d is the distance between point (x, y) and the center point C of the kidney, and λ is the attention coefficient; The initial polar coordinate plot is obtained by the following formula: P(θ′,d)=I(x,y)if d≤λL θ′=θ-φ Among them, P(θ′, d) is the pixel intensity of the initial polar coordinate map at point (θ′, d), θ is the angle between the line connecting point (x, y) and the kidney center point C and the horizontal axis, d is the distance between point (x, y) and the kidney center point C, and λ is the attention coefficient.
6. The method for classifying kidney ultrasound images according to claim 5, characterized in that: Similarity measures with a reference image set include: At least one similarity measurement algorithm among Euclidean distance, cosine similarity, Pearson correlation coefficient, and edge detection is used to obtain a similarity set, and the maximum similarities obtained by various algorithms are accumulated. When the accumulated value exceeds a set threshold, the classification result is output.
7. The method for classifying renal ultrasound images according to claim 1, characterized in that: The feature extractor is a combination of a convolutional neural network and a multi-layer perceptron. The distance weighted map and the polar coordinate map of the second kidney ultrasound image are used as different views, which are simultaneously input into different branches of the convolutional neural network through a multi-view learning method, and view-level feature fusion is performed at the high level of the convolutional neural network.
8. The method for classifying renal ultrasound images according to claim 7, characterized in that: The infoNCE function is used as the loss function to evaluate the current performance of the feature extractor. The calculation formula is: Among them, z i and z j Represents the image features of training samples from the same patient; z k Represents z i All negative samples and z j ; τ is a positive temperature parameter; N is z i The number of negative samples plus 1.
9. A device for renal ultrasound imaging, characterized in that: include: Memory and processor; The memory is used to store program instructions; The processor is used to call program instructions, and when the program instructions are called, the method for classifying kidney ultrasound images described in any one of claims 1-8 is executed.
