Image diagnosis method and system based on feature information retrieval

Abstract

The present invention provides an image diagnosis method and system based on feature information retrieval, which construct a neural network model, and use a number of historical Doppler images with diagnosis results to form a training sample set to train the neural network model; use the trained neural network model to extract feature data and classification data from the training samples and encode and store them in a feature database; obtain the Doppler image to be diagnosed; extract feature data and classification data from the Doppler image to be diagnosed based on a pre-trained neural network model; obtain training samples with feature data similar to that of the Doppler image to be diagnosed from the database; the algorithm is simple, the diagnosis efficiency is improved, and the false detection rate and missed detection rate are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to an image diagnosis method and system based on feature information retrieval. Background Art

[0002] Image feature information retrieval belongs to the field of computer vision. It is currently widely used in face recognition and gallery search, but rarely in the medical field. As machine vision has only recently been used in the medical field, there are no large-scale mature application algorithms. Most of the current algorithms have many disadvantages, such as complexity and slowness. Summary of the invention

[0003] Based on the technical problems existing in the prior art, the present invention provides an image diagnosis method and system based on feature information retrieval, aiming to solve the technical problems of low calculation efficiency and low precision in the prior art.

[0004] An image diagnosis method based on feature information retrieval, establishing a feature database, and then performing image diagnosis based on the feature database;

[0005] The steps to establish the feature database include:

[0006] Step A1, constructing a neural network model, and using a number of historical Doppler images with diagnosis results to form a training sample set, and using the training sample set to train the neural network model;

[0007] Step A2, using the trained neural network model to extract feature data and classification data from the training sample set, and encode and store them in the feature database;

[0008] The imaging diagnostic steps include:

[0009] Step B1, obtaining a Doppler image to be diagnosed;

[0010] Step B2, extracting feature data and classification data from the Doppler image to be diagnosed based on a pre-trained neural network model;

[0011] Step B3, obtaining training samples similar to the feature data of the Doppler image to be diagnosed from the database;

[0012] Among them, the classification data includes property classification labels and feature quantification classification labels. The property classification labels include abnormal labels for representing abnormal training samples and normal labels for representing normal training samples; the feature quantification classification includes feature category labels of training samples and corresponding levels.

[0013] Furthermore, step A1 includes:

[0014] Step A11, acquiring a number of historical Doppler images with diagnosis results;

[0015] Step A12, labeling each historical Doppler image with a property classification label and a feature quantification classification label;

[0016] Step A13, dividing the historical Doppler images into a training sample set and a test sample set;

[0017] Step A14, using a neural network model to extract feature data and classification data of each training sample in the training sample set;

[0018] Step A15, optimizing and training the neural network model based on the property classification labels and the feature quantization classification labels;

[0019] Step A16, testing the accuracy of the neural network model after each training by using a test sample set to train and form a neural network model.

[0020] Further, step A15 includes:

[0021] Step A151, according to the nearest neighbor distance algorithm, using the extracted feature data, respectively select a first preset number of training samples of the nearest neighbor distance of the same property classification from the training sample set;

[0022] Step A152, performing gradient optimization on the neural network model using a first loss function to minimize the nearest neighbor distance of a first preset number of selected training samples;

[0023] Step A153, calculating a first feature center composed of all normal training samples and a second feature center composed of all abnormal training samples;

[0024] Step A154, using the second loss function to further gradient optimize the neural network model to maximize the distance between the first feature center and the second feature center.

[0025] Furthermore, step A15 also includes:

[0026] Step A155, clustering the feature data of the same feature category label of the training sample set respectively, further performing gradient optimization on the neural network using the third loss function, and obtaining the level corresponding to each feature category of each training sample;

[0027] Step A16, repeat steps A151-A155 until the accuracy of the trained neural network model tested by the test sample set reaches a preset value.

[0028] Furthermore, the feature categories include inner diameter size, inner diameter smoothness, intima-media thickness, intima-media echogenicity, luminal plaque, plaque size, and plaque echogenicity.

[0029] Furthermore, the formula of the first loss function is as follows:

[0030]

[0031] in,

[0032] The feature center of all training samples classified with the same property;

[0033] f(x) is the feature data of training sample x;

[0034] The second loss function expression is:

[0035]

[0036] in,

[0037] Center(P) is the feature center of all normal training samples;

[0038] Center(N) is the feature center of all abnormal training samples.

[0039] Further, step B3 includes:

[0040] Step B31, calculating the nearest neighbor distance between the feature data of the Doppler image to be diagnosed and the feature data in the training sample in the database;

[0041] Step B32, sorting the training samples according to the nearest neighbor distance;

[0042] Step B33, obtaining a second preset number of training sample outputs that are ranked first.

[0043] An image diagnosis system based on feature information retrieval, including the aforementioned image diagnosis system based on feature information retrieval, including:

[0044] A model building module is used to build a neural network model, and use a number of historical Doppler images with diagnosis results to form a training sample set, and use the training sample set to train the neural network model;

[0045] The feature storage module is connected to the model building module and is used to extract feature data and classification data from the training sample set using the trained neural network model and encode and store them in the feature database;

[0046] An image acquisition module, used for acquiring a Doppler image to be diagnosed;

[0047] A feature extraction module, connected to the image acquisition module, is used to extract feature data and classification data from the Doppler image to be diagnosed based on a pre-trained neural network model;

[0048] An image retrieval module, connected to the feature extraction module and the database, respectively, for acquiring a sample image having feature data similar to that of the Doppler image to be diagnosed from the feature database;

[0049] Among them, the classification data includes property classification labels and feature quantification classification labels. The property classification labels include abnormal labels for representing abnormal training samples and normal labels for representing normal training samples; the feature quantification classification includes feature category labels in the training samples and their corresponding levels.

[0050] Furthermore, the model building module includes:

[0051] An acquisition unit, used for acquiring a number of historical Doppler images with diagnosis results;

[0052] A labeling unit, connected to the image acquisition unit, is used to label each historical Doppler image with a property classification label and a feature quantification classification label; a division unit, connected to the labeling unit, is used to divide the Doppler image into a training sample set and a test sample set;

[0053] The feature extraction unit and the connection division unit are used to extract feature data and classification data of each training sample in the training sample set by using a neural network model;

[0054] A training unit, connected to the feature extraction unit, is used to optimize the training of the neural network model based on the property classification label and the feature quantization classification label;

[0055] The testing unit is respectively connected to the partitioning unit and the training unit, and is used to test the accuracy of the neural network model after each training through a test sample set, so as to train and form the neural network model.

[0056] Furthermore, the training unit includes:

[0057] The sample screening subunit selects a first preset number of training samples of the nearest neighbor distance of the same property classification label respectively by using the extracted feature data according to the nearest neighbor distance algorithm;

[0058] A first training subunit, connected to the sample screening subunit, is used to perform gradient optimization on the neural network model using a first loss function to minimize the nearest neighbor distance of a first preset number of selected training samples;

[0059] A center calculation subunit, used to calculate a first feature center composed of all normal training samples and a second feature center composed of all abnormal training samples;

[0060] The second training subunit is respectively connected to the first training subunit and the center calculation subunit, and is used to further gradient optimize the neural network model using a second loss function to maximize the distance between the first feature center and the second feature center.

[0061] The beneficial technical effect of the present invention is to provide an image diagnosis method and system based on feature information retrieval, train images through hierarchical classification, use the trained model to extract feature data of training samples for storage, then extract the image to be diagnosed, obtain several stored similar sample images to assist in the qualitative diagnosis of the image to be diagnosed, the algorithm is simple, the diagnostic efficiency is improved, and the false detection rate and missed detection rate are effectively reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A flow chart of feature extraction steps of an image diagnosis method based on feature information retrieval according to the present invention;

[0063] Figure 2 A flowchart of image diagnosis steps of an image diagnosis method based on feature information retrieval according to the present invention;

[0064] Figure 3-5 A flow chart of model training steps of an image diagnosis method based on feature information retrieval according to the present invention;

[0065] Figure 6 A schematic diagram of a module of an image diagnosis system based on feature information retrieval according to the present invention;

[0066] Figure 7 A schematic diagram of a model building module of an image diagnosis system based on feature information retrieval according to the present invention;

[0067] Figure 8 A schematic diagram of a training unit module of an image diagnosis system based on feature information retrieval according to the present invention;

[0068] Fig. 9 A schematic diagram of an image retrieval module of an image diagnosis system based on feature information retrieval according to the present invention; DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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.

[0070] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0071] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0072] See also Figure 1-2 , the present invention provides an image diagnosis method based on feature information retrieval, establishes a feature database, and then performs image diagnosis based on the feature database;

[0073] The steps to establish the feature database include:

[0074] Step A1, constructing a neural network model, and using a number of historical Doppler images with diagnosis results to form a training sample set, and using the training sample set to train the neural network model;

[0075] Step A2, using the trained neural network model to extract feature data and classification data from the training sample set, and encode and store them in the feature database;

[0076] The imaging diagnostic steps include:

[0077] Step B1, obtaining a Doppler image to be diagnosed;

[0078] Step B2, extracting feature data and classification data from the Doppler image to be diagnosed based on a pre-trained neural network model;

[0079] Step B3, obtaining training samples similar to the feature data of the Doppler image to be diagnosed from the database;

[0080] Among them, the classification data includes property classification labels and feature quantification classification labels. The property classification labels include abnormal labels for representing abnormal training samples and normal labels for representing normal training samples; the feature quantification classification includes feature category labels of training samples and corresponding levels.

[0081] See also Figure 3 , further, step A1 comprises:

[0082] Step A11, acquiring a number of historical Doppler images with diagnosis results;

[0083] Step A12, labeling each historical Doppler image with a property classification label and a feature quantification classification label;

[0084] Step A13, dividing the historical Doppler images into a training sample set and a test sample set;

[0085] Step A14, using a neural network model to extract feature data and classification data of each training sample in the training sample set;

[0086] Step A15, optimizing and training the neural network model based on the property classification labels and the feature quantization classification labels;

[0087] Step A16, testing the accuracy of the neural network model after each training by using a test sample set to train and form a neural network model.

[0088] See also Figure 4 , further, step A15 includes:

[0089] Step A151, according to the nearest neighbor distance algorithm, using the extracted feature data, respectively select a first preset number of training samples of the nearest neighbor distance of the same property classification from the training sample set;

[0090] Step A152, performing gradient optimization on the neural network model using a first loss function to minimize the nearest neighbor distance of a first preset number of selected training samples;

[0091] Step A153, calculating a first feature center composed of all normal training samples and a second feature center composed of all abnormal training samples;

[0092] Step A154, using the second loss function to further gradient optimize the neural network model to maximize the distance between the first feature center and the second feature center.

[0093] Furthermore, step A15 also includes:

[0094] Step A155, clustering the feature data of the same feature class label of the training sample set respectively, further performing gradient optimization on the neural network using the third loss function, and obtaining the level corresponding to each feature class of each training sample;

[0095] Step A16, repeat steps A151-A155 until the accuracy of the trained neural network model tested by the test sample set reaches a preset value.

[0096] The present invention performs two-level classification on the Doppler image data set, the first level is for property classification labels, the second level is for quantitative classification of the features of each module in the image, and a feature extraction neural network is constructed. There are two loss functions, corresponding to the two levels respectively. For the first level, in the feature space, the nearest neighbor distance of the nearest 20 samples of the same property is minimized, and the center distance of samples of different properties is maximized. For the second level, each module feature label is clustered.

[0097] In the selection of abnormal and normal training samples, the screening principles include:

[0098] (1) Try to ensure that the number of positive and negative samples is even;

[0099] (2) Try to ensure that the second-level quantitative parameters of the sample are fully covered: for example, there are 10,000 cases of carotid artery ultrasound data, and 500 cases of tumor-negative and 500 cases of tumor-positive are selected; among the negative samples, 150 nodules are small in size, 200 nodules are medium in size, and 150 nodules are large in size; among the positive samples, 150 nodules are small in size, 200 nodules are medium in size, and 150 nodules are large in size.

[0100] Furthermore, the feature categories include inner diameter size, inner diameter smoothness, intima-media thickness, intima-media echogenicity, luminal plaque, plaque size, and plaque echogenicity.

[0101] Furthermore, the formula of the first loss function is as follows:

[0102]

[0103] in,

[0104] The feature center of all training samples classified with the same property;

[0105] f(x) is the feature data of training sample x;

[0106] The second loss function expression is:

[0107]

[0108] in,

[0109] Center(P) is the feature center of all normal training samples;

[0110] Center(N) is the feature center of all abnormal training samples.

[0111] See also Figure 5 , further, step B3 comprises:

[0112] Step B31, calculating the nearest neighbor distance between the feature data of the Doppler image to be diagnosed and the feature data in the training sample in the database;

[0113] Step B32, sorting the training samples according to the nearest neighbor distance;

[0114] Step B33, obtaining a second preset number of training sample outputs that are ranked first.

[0115] The second preset number is ten.

[0116] See also Figure 6 The present invention further provides an image diagnosis system based on feature information retrieval, including the aforementioned image diagnosis system based on feature information retrieval, including:

[0117] A model building module (1) is used to build a neural network model, and use a number of historical Doppler images with diagnosis results to form a training sample set, and use the training sample set to train the neural network model;

[0118] A feature storage module (2) is connected to the model building module (1) and is used to extract feature data and classification data from the training samples using the trained neural network model and encode and store them in a database (3);

[0119] An image acquisition module (4), used for acquiring a Doppler image to be diagnosed;

[0120] A feature extraction module (5) is connected to the image acquisition module (4) and is used to extract feature data and classification data from the Doppler image to be diagnosed based on a pre-trained neural network model;

[0121] An image retrieval module (6) is connected to the feature extraction module (5) and the database (3) respectively, and is used to obtain a sample image having feature data similar to that of the Doppler image to be diagnosed from the feature database (3);

[0122] Among them, the classification data includes property classification labels and feature quantification classification labels. The property classification labels include abnormal labels for representing abnormal training samples and normal labels for representing normal training samples; the feature quantification classification includes feature category labels in the training samples and their corresponding levels.

[0123] See also Figure 7 , further, the model building module (1) includes:

[0124] An acquisition unit (11) is used to acquire a plurality of historical Doppler images having diagnosis results;

[0125] A labeling unit (12), connected to the image acquisition unit (11), is used to label each historical Doppler image with a property classification label and a feature quantification classification label;

[0126] A division unit (13), connected to the labeling unit (12), is used to divide the Doppler image into a training sample set and a test sample set;

[0127] A feature extraction unit (14), connected to the division unit (13), is used to extract feature data and classification data of each training sample in the training sample set using a neural network model;

[0128] A training unit (15), connected to the feature extraction unit (13), is used to optimize the training of the neural network model based on the property classification label and the feature quantization classification label;

[0129] The testing unit (16) is respectively connected to the dividing unit (13) and the training unit (15) and is used to test the accuracy of the neural network model after each training through a test sample set, so as to train and form the neural network model.

[0130] See also Figure 8 , further, the training unit (15) comprises:

[0131] A sample screening subunit (150) selects a first preset number of training samples of nearest neighbor distances with the same property classification label using the extracted feature data according to a nearest neighbor distance algorithm;

[0132] A first training subunit (151), connected to the sample screening subunit (150), is used to perform gradient optimization on the neural network model using a first loss function to minimize the nearest neighbor distance of a first preset number of selected training samples;

[0133] A center calculation subunit (153), used for calculating a first feature center formed by all normal training samples, and a second feature center formed by all abnormal training samples;

[0134] The second training subunit (154) is connected to the first training subunit (151) and the center calculation subunit (153) respectively, and is used to use a second loss function to further perform gradient optimization on the neural network model to maximize the distance between the first feature center and the second feature center.

[0135] Furthermore, the training unit (15) further comprises:

[0136] The third training subunit (155) is connected to the second training subunit (154) and is used to cluster the feature data of the same feature class labels of the training samples respectively, and further perform gradient optimization on the neural network using a third loss function to obtain the level corresponding to each feature class of each training sample.

[0137] Furthermore, the feature categories include inner diameter size, inner diameter smoothness, intima-media thickness, intima-media echogenicity, luminal plaque, plaque size, and plaque echogenicity.

[0138] Furthermore, the formula of the first loss function is as follows:

[0139]

[0140] in,

[0141] The feature center of all training samples classified with the same property;

[0142] f(x) is the feature data of training sample x.

[0143] Furthermore, the second loss function expression is:

[0144]

[0145] in,

[0146] Center(P) is the feature center of all normal training samples;

[0147] Center(N) is the feature center of all abnormal training samples.

[0148] Furthermore, the formula of the third loss function is as follows:

[0149]

[0150] in,

[0151] is the feature center of all training samples;

[0152] z(x i ) is the training sample x i feature data.

[0153] Furthermore, the first preset number is 20.

[0154] See also Fig. 9 , further, an image retrieval module (6) is connected to the feature extraction module (5) and the database (3) respectively, and is used to obtain a sample image similar to the feature data of the Doppler image to be diagnosed from the database (3);

[0155] The image retrieval module (6) includes:

[0156] An image calculation unit (61) is used to calculate the nearest neighbor distance between the feature data of the Doppler image to be diagnosed and the feature data in the training sample in the database;

[0157] A sorting unit (62), connected to the image calculation unit (61), for sorting the training samples according to the nearest neighbor distance;

[0158] An image output unit (63) is connected to the sorting unit (63) and is used to obtain a second preset number of training samples that are sorted first for output.

[0159] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. An image diagnosis method based on feature information retrieval, characterized in that: Establishing a feature database, and then performing image diagnosis based on the feature database; The steps of establishing the feature database include: Step A1, constructing a neural network model, and using a number of historical Doppler images with diagnosis results to form a training sample set, and using the training sample set to train the neural network model; Step A2, extracting feature data and classification data from the training sample set using the trained neural network model, and encoding and storing the data in the feature database; The imaging diagnostic steps include: Step B1, obtaining a Doppler image to be diagnosed; Step B2, extracting feature data and classification data from the Doppler image to be diagnosed based on a pre-trained neural network model; Step B3, acquiring the training sample close to the characteristic data of the Doppler image to be diagnosed from the database; The classification data includes property classification labels and feature quantification classification labels, wherein the property classification labels include abnormal labels for representing abnormal training samples and normal labels for representing normal training samples; the feature quantification classification includes feature category labels of training samples and corresponding levels.

2. The image diagnosis method based on feature information retrieval according to claim 1, characterized in that: The step A1 comprises: Step A11, acquiring a number of historical Doppler images with diagnosis results; Step A12, marking each of the historical Doppler images with the property classification label and the feature quantification classification label; Step A13, dividing the historical Doppler images into the training sample set and the test sample set; Step A14, using a neural network model to extract feature data and classification data of each training sample in the training sample set; Step A15, optimizing and training the neural network model based on the property classification label and the feature quantization classification label; Step A16, testing the accuracy of the neural network model after each training by using the test sample set to train and form the neural network model.

3. The image diagnosis method based on feature information retrieval according to claim 2, characterized in that: The step A15 comprises: Step A151, selecting a first preset number of training samples of the same property classification in the nearest neighbor distance from the training sample set using the extracted feature data according to the nearest neighbor distance algorithm; Step A152, performing gradient optimization on the neural network model using a first loss function to minimize the nearest neighbor distance of the selected first preset number of training samples; Step A153, calculating a first feature center formed by all the normal training samples and a second feature center formed by all the abnormal training samples; Step A154, using a second loss function to further gradient optimize the neural network model to maximize the distance between the first feature center and the second feature center.

4. The image diagnosis method based on feature information retrieval according to claim 3, characterized in that: The step A15 further includes: Step A155, clustering the feature data of the same feature category label of the training sample set respectively, further performing gradient optimization on the neural network using a third loss function, and obtaining the level corresponding to each feature category of each training sample; The step A16 repeats the steps A151-A155 until the accuracy of the trained neural network model tested by the test sample set reaches a preset value.

5. The image diagnosis method based on feature information retrieval according to claim 1, characterized in that: The feature categories include inner diameter size, inner diameter smoothness, intima-media thickness, intima-media echogenicity, luminal plaque, plaque size, and plaque echogenicity.

6. The image diagnosis method based on feature information retrieval according to claim 3, characterized in that: The formula of the first loss function is as follows: in, The feature centers of all the training samples classified with the same property; f(x) is the feature data of training sample x; The second loss function expression is: in, Center (P) is the feature center of all normal training samples; Center(N) is the feature center of all abnormal training samples.

7. The image diagnosis method based on feature information retrieval according to claim 1, characterized in that: The step B3 comprises: Step B31, calculating the nearest neighbor distance between the feature data of the Doppler image to be diagnosed and the feature data in the training sample in the database; Step B32, sorting the training samples according to the nearest neighbor distance; Step B33, obtaining the output of the second preset number of training samples that are ranked first.

8. An image diagnosis system based on feature information retrieval, characterized in that: An image diagnosis method based on feature information retrieval as claimed in any one of claims 1 to 7, comprising: A model building module, used to build a neural network model, and use a number of historical Doppler images with diagnosis results to form a training sample set, and use the training sample set to train the neural network model; A feature storage module, connected to the model building module, is used to extract feature data and classification data from the training sample set using the trained neural network model and encode and store them in a feature database; An image acquisition module, used for acquiring a Doppler image to be diagnosed; A feature extraction module, connected to the image acquisition module, for extracting feature data and classification data from the Doppler image to be diagnosed based on the pre-trained neural network model; An image retrieval module, connected to the feature extraction module and the database, respectively, for acquiring a sample image similar to the feature data of the Doppler image to be diagnosed from the feature database; Among them, the classification data includes property classification labels and feature quantification classification labels, the property classification labels include abnormal labels for representing abnormal training samples and normal labels for representing normal training samples; the feature quantification classification includes feature category labels in training samples and corresponding levels.

9. An image diagnosis system based on feature information retrieval as claimed in claim 8, characterized in that: The model building module includes: An acquisition unit, used for acquiring a number of historical Doppler images with diagnosis results; a labeling unit connected to the image acquisition unit and used to label each of the historical Doppler images with the property classification label and the feature quantification classification label; a division unit connected to the labeling unit and used to divide the Doppler image into the training sample set and the test sample set; A feature extraction unit, connected to the division unit, for extracting feature data and classification data of each training sample in the training sample set by using the neural network model; A training unit, connected to the feature extraction unit, for optimizing and training the neural network model based on the property classification label and the feature quantization classification label; The testing unit is respectively connected to the dividing unit and the training unit, and is used to test the accuracy of the neural network model after each training through the test sample set, so as to train and form the neural network model.

10. An image diagnosis system based on feature information retrieval as claimed in claim 9, characterized in that: The training unit comprises: A sample screening subunit selects a first preset number of training samples of the same property classification label of the nearest neighbor distance by using the extracted feature data according to the nearest neighbor distance algorithm; A first training subunit, connected to the sample screening subunit, for performing gradient optimization on the neural network model using a first loss function to minimize the nearest neighbor distance of the first preset number of training samples selected; A center calculation subunit, used for calculating a first feature center formed by all the normal training samples and a second feature center formed by all the abnormal training samples; The second training subunit is respectively connected to the first training subunit and the center calculation subunit, and is used to use a second loss function to further gradient optimize the neural network model to maximize the distance between the first feature center and the second feature center.

Images