Training method and system of medical image recognition model
By designing the training method of feature extractor and classifier in the medical image recognition model, combined with the noise reduction processing technology, the problems of simple structure and slow convergence speed of BP neural network classifier are solved, and the recognition accuracy and training efficiency of the model are improved.
Patent Information
- Application Number
- CN202411691669.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-22
- Publication Date
- 2025-03-18
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The structure of the BP neural network classifier in the convolutional neural network model is too simple, requiring multiple iterations and a large number of training samples. The gradient-decreasing training method causes the model to converge slowly and easily fall into local optimality, affecting the recognition accuracy.
A training method for medical image recognition models is adopted, including training of feature extractors and classifiers. The feature extractor calculates residuals through forward propagation and backpropagation, updating weights and bias terms until converges. The classifier is designed as a four-layer network, and the predicted value is obtained by calculating the output quotient of the summation layer. At the same time, the training image is denoised, and wavelet transform and hard/soft thresholding are performed using the improved threshold function.
It improves the convergence speed and recognition accuracy of the model, reduces training time, enhances classification processing capabilities, and ensures the real-time and accuracy of the model. At the same time, noise reduction processing improves image clarity and detail retention, and enhances the accuracy of model training.
Smart Images
Figure CN119649101A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing technology, and more specifically, to a training method and system for a medical image recognition model. Background Art
[0002] With the rapid development of computer technology, the resolution and imaging speed of medical images have been greatly improved, and the role of medical images in medical diagnosis has become more and more significant. Medical image recognition technology has become an important way to improve the accuracy and reliability of medical diagnosis. The types of medical images mainly include X-ray imaging, magnetic resonance imaging, nuclear medicine imaging and ultrasonic imaging. In current medical imaging diagnosis, the diagnosis and identification of lesions in medical images are mainly based on the doctor's experience. Therefore, the accuracy of medical image recognition is limited by factors such as the doctor's subjectivity, cognitive differences among doctors, and their time and energy.
[0003] Medical image recognition is an interdisciplinary field that integrates medical imaging, mathematical modeling, computer technology, and other disciplines. In the era of medical image big data, massive and complex image data brings two new problems: on the one hand, the medical image data to be processed has higher dimensions and requires models with stronger learning and adaptability; on the other hand, medical image big data is more dispersed and fragmented, and the data structure is more complex, often requiring the integration of different information.
[0004] The BP neural network classifier structure in the convolutional neural network model is too simple and requires multiple iterations and a large number of training samples. At the same time, the BP neural network adopts the gradient reduction training method, which leads to slow model convergence and easy to fall into local optimality, affecting the final recognition accuracy. Therefore, a training method and system for medical image recognition model are proposed. Summary of the invention
[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a training method and system for a medical image recognition model. The technical problem to be solved by the present invention is that the BP neural network classifier structure in the convolutional neural network model is too simple and requires multiple iterations and a large number of training samples. At the same time, the BP neural network adopts a gradient reduction training method, which leads to slow model convergence speed and easy to fall into local optimality, affecting the final recognition accuracy.
[0006] To achieve the above object, the present invention provides the following technical solution: a training method for a medical image recognition model, comprising the following steps:
[0007] Feature extractor training:
[0008] Obtain training set sample images and input the training sample images into the network model;
[0009] After the network model is forward propagated, the error between the output layer output and the target label is calculated;
[0010] Determine whether the network model converges. If the network model converges, the network model after feature extractor training is obtained;
[0011] Classifier training:
[0012] Input the training image into the feature extractor after feature extraction training, and obtain the feature vector of the training image;
[0013] Input the feature vector into the classifier to be trained, and finally obtain the trained classifier;
[0014] Complete the training of the network model and obtain the final image recognition model.
[0015] As a further solution of the present invention: the classifier is a variation of an artificial neural network, specifically composed of a four-layer network, namely an input layer, a pattern layer, a summation layer and an output layer, wherein the number of neurons in the input layer is equal to the dimension of the feature vector extracted by the feature extractor from the image sample and passed to the pattern layer, the number of neurons in the pattern layer is the same as that in the input layer, different pattern layer neurons correspond to different sample data, the summation layer has only two neurons, which are fully connected to the pattern layer neuron nodes, and the output layer obtains the predicted value of the model by calculating the two output quotients of the summation layer.
[0016] As a further solution of the present invention: the output value of the model of the classifier is:
[0017]
[0018] Among them, X i and Y i is the sample observation value; σ is the smoothing factor; n is the number of samples.
[0019] As a further solution of the present invention: when the feature extractor is trained, if the network model has not converged, the residual of the output layer is calculated, the residual is transferred backward from the output layer, and the residual value is calculated layer by layer;
[0020] The updated values of the weights and biases are then calculated until they are the same as the values of the input layer.
[0021] As a further solution of the present invention: the formula for residual calculation is:
[0022]
[0023] Among them, nl refers to the output layer; y is the sample label; a is the output value; m is the number of samples in each batch; f is the activation function; z is the upper neuron of the output layer;
[0024] Calculate the updated values of weights and bias items, the formula is:
[0025]
[0026] in, is the bias term of the lth layer; α is the learning rate.
[0027] As a further solution of the present invention: the sample images of the acquired training set are subjected to denoising processing before being input into the network model. The denoising processing of the images is specifically as follows:
[0028] Assume the noisy model: f(k) = s(k) + n(k), k = 0, 1, 2, ... N-1
[0029] Among them, s(k) is the original image signal, and n(k) is Gaussian white noise;
[0030] Perform discrete wavelet transform on f(k): w i,k =W s (j,k)+W n (j,k),j=0,1,2,…,J;k=0,1,2,…,N
[0031] where w i,k , W s (j,k) and W n (j, k) are the noisy image signal, the original image signal and the wavelet coefficients of the noise at the nth layer, and the total number of pixels of the image with the maximum decomposition layer number J and N, respectively;
[0032] The hard threshold function expression is:
[0033]
[0034] The soft threshold function expression is:
[0035]
[0036] in To estimate the wavelet coefficients, wj ,k is the decomposition wavelet coefficient, λ is the threshold;
[0037] The improved threshold function,
[0038] α is the adjustment coefficient, and 0≤α≤1, when |w j,k When |→λ, That is, the improved threshold function is continuous at the threshold point; when |w j,k When |→∞, That is, with |w j,k |Gradually increasing, With w j,kThe deviation between them becomes smaller and smaller, solving the constant deviation shortcoming of the soft threshold function.
[0039] A training system for a medical image recognition model, comprising:
[0040] A data acquisition unit, used to acquire sample images used for training;
[0041] A noise reduction processing unit, used for performing noise reduction processing on training images;
[0042] A feature extractor training unit, used for training a feature extractor;
[0043] The classifier training unit is used to train the classifier.
[0044] Image recognition, also known as image pattern recognition, is the specific application of pattern recognition technology in the field of images. It is a technology that establishes an image recognition model for input image information, analyzes and extracts image features, and then establishes a classifier to perform classification and recognition based on image features.
[0045] Common methods of image recognition include: Bayesian classification, template matching, etc. Bayesian classification assumes that the features of an image are independent and calculates the posterior probability of the features to classify the image. Template matching compares the sample features with a specific template and determines whether they are of the same type based on the similarity.
[0046] Noise is inevitably introduced during the imaging process of medical images, and physical artifacts appear in the images, causing the pathological area to be blurred and interfering with disease diagnosis. The types of noise generated by different medical imaging devices are also different. In ultrasound imaging, a large number of particle beams interfere with each other, and the echo signals are scattered and superimposed, resulting in the appearance of non-real tissue signals, which appear as speckle noise of different brightness on the image. Compared with the additive noise represented by Gaussian white noise, this multiplicative noise has a greater impact on image quality. During CT imaging, the field offset effect causes image quality to degrade, generating quantum noise that approximately obeys the Poisson distribution. In low-dose CT images, due to the reduction in the number of photons received by the detector, blurred boundaries and strip artifact noise are more likely to occur during image reconstruction. The noise in MRI mainly comes from the equipment circuit and the unconscious movements of the patient during the acquisition process; this leads to a decrease in the image signal-to-noise ratio and blurred edges. The noise in MRI is mainly thermal noise and physiological noise.
[0047] The beneficial effects of the present invention are:
[0048] 1. The classifier used in the present invention starts from the probability density function and uses the training samples to directly calculate the regression value of the dependent variable to the independent variable. There is no need for iterative operation, which is time-consuming and has ideal calculation efficiency. The classification processing ability of the classification model is guaranteed, and the convergence speed of the model training process is fast, which improves the recognition accuracy of the recognition model, and the real-time performance and accuracy are better.
[0049] 2. The present invention uses a denoising algorithm to perform denoising processing on the images used for training, thereby reducing the noise in the images used for training and making the images clearer. At the same time, the detailed information of the images and the texture characteristics in the images are protected, making the training of the image recognition model more accurate. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 It is a schematic diagram of the connection of the system of the present invention;
[0051] Figure 2 This is a schematic diagram of feature extractor training of the present invention. DETAILED DESCRIPTION
[0052] 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.
[0053] As shown in the figure, Example 1:
[0054] The present invention provides a training method for a medical image recognition model, comprising the following steps:
[0055] Feature extractor training:
[0056] Obtain training set sample images and input the training sample images into the network model;
[0057] After the network model is forward propagated, the error between the output layer output and the target label is calculated;
[0058] Determine whether the network model converges. If the network model converges, the network model after feature extractor training is obtained;
[0059] Classifier training:
[0060] Input the training image into the feature extractor after feature extraction training, and obtain the feature vector of the training image;
[0061] Input the feature vector into the classifier to be trained, and finally obtain the trained classifier;
[0062] Complete the training of the network model and obtain the final image recognition model.
[0063] The classifier is a variation of an artificial neural network, consisting of four layers of networks, namely the input layer, pattern layer, summation layer, and output layer. The number of neurons in the input layer is equal to the dimension of the feature vector extracted by the feature extractor from the image sample and passed to the pattern layer. The number of neurons in the pattern layer is the same as that in the input layer. Different pattern layer neurons correspond to different sample data. The summation layer has only two neurons, which are fully connected to the neuron nodes in the pattern layer. The output layer obtains the predicted value of the model by calculating the two output quotients of the summation layer.
[0064] The obtained training set sample images are denoised before being input into the network model.
[0065] A training system for a medical image recognition model, comprising:
[0066] A data acquisition unit, used to acquire sample images used for training;
[0067] A noise reduction processing unit, used for performing noise reduction processing on training images;
[0068] A feature extractor training unit, used for training a feature extractor;
[0069] The classifier training unit is used to train the classifier.
[0070] Embodiment 2:
[0071] The present invention provides a training method for a medical image recognition model, comprising the following steps:
[0072] Feature extractor training:
[0073] Obtain training set sample images and input the training sample images into the network model;
[0074] After the network model is forward propagated, the error between the output layer output and the target label is calculated;
[0075] Determine whether the network model converges. If the network model converges, the network model after feature extractor training is obtained;
[0076] Classifier training:
[0077] Input the training image into the feature extractor after feature extraction training, and obtain the feature vector of the training image;
[0078] Input the feature vector into the classifier to be trained, and finally obtain the trained classifier;
[0079] Complete the training of the network model and obtain the final image recognition model.
[0080] The classifier is a variation of an artificial neural network, consisting of four layers of networks, namely the input layer, pattern layer, summation layer, and output layer. The number of neurons in the input layer is equal to the dimension of the feature vector extracted by the feature extractor from the image sample and passed to the pattern layer. The number of neurons in the pattern layer is the same as that in the input layer. Different pattern layer neurons correspond to different sample data. The summation layer has only two neurons, which are fully connected to the neuron nodes in the pattern layer. The output layer obtains the predicted value of the model by calculating the two output quotients of the summation layer.
[0081] The output value of the classifier model is:
[0082]
[0083] Among them, X i and Y i is the sample observation value; σ is the smoothing factor; n is the number of samples.
[0084] When the feature extractor is trained, if the network model has not converged, the residual of the output layer is calculated, and the residual is transferred backward from the output layer to calculate the residual value layer by layer;
[0085] The updated values of the weights and biases are then calculated until they are the same as the values of the input layer.
[0086] The formula for residual calculation is:
[0087]
[0088]
[0089] Among them, nl refers to the output layer; y is the sample label; a is the output value; m is the number of samples in each batch; f is the activation function; z is the upper neuron of the output layer;
[0090] Calculate the updated values of weights and bias items, the formula is:
[0091]
[0092] in, is the bias term of the lth layer; α is the learning rate.
[0093] The obtained training set sample images are denoised before being input into the network model.
[0094] A training system for a medical image recognition model, comprising:
[0095] A data acquisition unit, used to acquire sample images used for training;
[0096] A noise reduction processing unit, used for performing noise reduction processing on training images;
[0097] A feature extractor training unit, used for training a feature extractor;
[0098] The classifier training unit is used to train the classifier.
[0099] Embodiment 3:
[0100] The present invention provides a training method for a medical image recognition model, comprising the following steps:
[0101] Feature extractor training:
[0102] Obtain training set sample images and input the training sample images into the network model;
[0103] After the network model is forward propagated, the error between the output layer output and the target label is calculated;
[0104] Determine whether the network model converges. If the network model converges, the network model after feature extractor training is obtained;
[0105] Classifier training:
[0106] Input the training image into the feature extractor after feature extraction training, and obtain the feature vector of the training image;
[0107] Input the feature vector into the classifier to be trained, and finally obtain the trained classifier;
[0108] Complete the training of the network model and obtain the final image recognition model.
[0109] The classifier is a variation of an artificial neural network, consisting of four layers of networks, namely the input layer, pattern layer, summation layer, and output layer. The number of neurons in the input layer is equal to the dimension of the feature vector extracted by the feature extractor from the image sample and passed to the pattern layer. The number of neurons in the pattern layer is the same as that in the input layer. Different pattern layer neurons correspond to different sample data. The summation layer has only two neurons, which are fully connected to the neuron nodes in the pattern layer. The output layer obtains the predicted value of the model by calculating the two output quotients of the summation layer.
[0110] The output value of the classifier model is:
[0111]
[0112] Among them, X i and Y i is the sample observation value; σ is the smoothing factor; n is the number of samples.
[0113] When the feature extractor is trained, if the network model has not converged, the residual of the output layer is calculated, and the residual is transferred backward from the output layer to calculate the residual value layer by layer;
[0114] The updated values of the weights and biases are then calculated until they are the same as the values of the input layer.
[0115] The formula for residual calculation is:
[0116]
[0117] Among them, nl refers to the output layer; y is the sample label; a is the output value; m is the number of samples in each batch; f is the activation function; z is the upper neuron of the output layer;
[0118] Calculate the updated values of weights and bias items, the formula is:
[0119]
[0120]
[0121] in, is the bias term of the lth layer; α is the learning rate.
[0122] The obtained training set sample images are subjected to denoising before being input into the network model. The denoising process of the images is as follows:
[0123] Assume the noisy model: f(k) = s(k) + n(k), k = 0, 1, 2, ... N-1
[0124] Among them, s(k) is the original image signal, and n(k) is Gaussian white noise;
[0125] Perform discrete wavelet transform on f(k): w i,k =W s (j,k)+W n (j,k),j=0,1,2,…,J;k=0,1,2,…,N
[0126] where w i,k , W s (j,k) and W n (j, k) are the noisy image signal, the original image signal and the wavelet coefficients of the noise at the nth layer, and the total number of pixels of the image with the maximum decomposition layer number J and N, respectively;
[0127] The hard threshold function expression is:
[0128]
[0129] The soft threshold function expression is:
[0130]
[0131] in To estimate the wavelet coefficients, wj ,k is the decomposition wavelet coefficient, λ is the threshold;
[0132] The improved threshold function,
[0133] α is the adjustment coefficient, and 0≤α≤1, when |w j,k When |→λ, That is, the improved threshold function is continuous at the threshold point; when |w j,k When |→∞, That is, with |w j,k |Gradually increasing, With w j,k The deviation between them becomes smaller and smaller, solving the constant deviation shortcoming of the soft threshold function.
[0134] A training system for a medical image recognition model, comprising:
[0135] A data acquisition unit, used to acquire sample images used for training;
[0136] A noise reduction processing unit, used for performing noise reduction processing on training images;
[0137] A feature extractor training unit, used for training a feature extractor;
[0138] The classifier training unit is used to train the classifier.
[0139] Secondly: In the drawings of the embodiments disclosed in the present invention, only the structures related to the embodiments disclosed in the present invention are involved, and other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of the present invention can be combined with each other;
[0140] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A training method for a medical image recognition model, characterized in that: The following steps are involved: Feature extractor training: Obtain training set sample images and input the training sample images into the network model; After the network model is forward propagated, the error between the output layer output and the target label is calculated; Determine whether the network model converges. If the network model converges, the network model after feature extractor training is obtained; Classifier training: Input the training image into the feature extractor after feature extraction training, and obtain the feature vector of the training image; Input the feature vector into the classifier to be trained, and finally obtain the trained classifier; Complete the training of the network model and obtain the final image recognition model.
2. The method for training a medical image recognition model according to claim 1, characterized in that: The classifier is a variation of an artificial neural network, specifically consisting of a four-layer network, namely an input layer, a pattern layer, a summation layer and an output layer, wherein the number of neurons in the input layer is equal to the dimension of the feature vector extracted from the image sample by the feature extractor and passed to the pattern layer, the number of neurons in the pattern layer is the same as that in the input layer, different pattern layer neurons correspond to different sample data, the summation layer has only two neurons, which are fully connected to the pattern layer neuron nodes, and the output layer obtains the predicted value of the model by calculating the two output quotients of the summation layer.
3. The method for training a medical image recognition model according to claim 2, characterized in that: The output value of the classifier model is: Among them, X i and Y i is the sample observation value; σ is the smoothing factor; n is the number of samples.
4. The method for training a medical image recognition model according to claim 1, characterized in that: When the feature extractor is trained, if the network model has not converged, the residual of the output layer is calculated, and the residual is transferred backward from the output layer to calculate the residual value layer by layer; The updated values of the weights and biases are then calculated until they are the same as the values of the input layer.
5. The method for training a medical image recognition model according to claim 4, characterized in that: The formula for residual calculation is: Among them, nl refers to the output layer; y is the sample label; a is the output value; m is the number of samples in each batch; f is the activation function; z is the upper neuron of the output layer; Calculate the updated values of weights and bias items, the formula is: in, is the bias term of the lth layer; α is the learning rate.
6. The method for training a medical image recognition model according to claim 1, characterized in that: The obtained training set sample images are subjected to denoising before being input into the network model. The denoising process of the images is as follows: Assume the noisy model: f(k) = s(k) + n(k), k = 0, 1, 2, ... N-1 Among them, s(k) is the original image signal, and n(k) is Gaussian white noise; Perform discrete wavelet transform on f(k): w i,k =W s (j,k)+W n (j,k),j=0,1,2,…,J;k=0,1,2,…,N where w i,k , W s (j,k) and W n (j, k) are the noisy image signal, the original image signal and the wavelet coefficients of the noise at the nth layer, and the total number of pixels of the image with the maximum decomposition layer number J and N, respectively; The hard threshold function expression is: The soft threshold function expression is: in To estimate the wavelet coefficients, wj ,k is the decomposition wavelet coefficient, λ is the threshold; The improved threshold function, α is the adjustment coefficient, and 0≤α≤1, when |w j,k When |→λ, That is, the improved threshold function is continuous at the threshold point; when |w j,k When |→∞, That is, with |w j,k |Gradually increasing, With w j,k The deviation between them becomes smaller and smaller, solving the constant deviation shortcoming of the soft threshold function.
7. A training system for a medical image recognition model, characterized in that: include: A data acquisition unit, used to acquire sample images used for training; A noise reduction processing unit, used for performing noise reduction processing on training images; A feature extractor training unit, used for training a feature extractor; The classifier training unit is used to train the classifier.