Machine Learning-Based Automatic Medical Image Classification Method and System

Abstract

The present invention relates to the technical field of medical image classification, and specifically relates to an automatic medical image classification method and system based on machine learning. Medical image data is collected and the quantity of medical image data is enhanced; the enhanced medical image data is subjected to dimensionality reduction processing; the dimensionality-reduced features are classified to obtain the categories of medical image data. Data augmentation uses a generative adversarial network based on quantum interference to simulate the complex patterns and textures of real medical images through modulation and control in multiple dimensions, increasing the quantity and diversity of medical image data while maintaining the detail quality of the images. The dimensionality reduction intensity is adaptively adjusted to optimize the preservation of information. The classifier uses a fractional-order neural network based on dynamic penalty terms, combines dynamic penalty terms and gated vector regularization techniques to optimize the classification accuracy and the generalization ability of the model. The deficiencies of the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image classification, and in particular, to an automatic medical image classification method and system based on machine learning. Background Art

[0002] With the development of medical technology, medical images play a crucial role in clinical practice. However, traditional image classification relies on the professional judgment of radiologists, which is not only time-consuming and laborious, but also prone to human errors when faced with a large amount of image data.

[0003] The invention patent with the publication number CN113744845A proposes a medical image processing method, device, electronic device and medium based on artificial intelligence. The method includes: in response to a processing instruction for a plurality of medical images of a target patient, obtaining an identification task corresponding to each medical image; determining the image factors of each identification task; constructing an image factor hierarchical structure model according to the hierarchical relationship corresponding to the image factors; obtaining a preset judgment matrix of the image, and calculating a combined weight corresponding to the identification task based on the judgment matrix and the image factor hierarchical structure model; generating a task scheduling set for a plurality of identification tasks according to the combined weight corresponding to the identification task; sequentially obtaining the target medical image corresponding to each identification task in the task scheduling set in a preset storage space; and calling an AI edge computing device to process the sequentially obtained target medical images. The processing efficiency of medical images is improved. The invention patent with the publication number CN111210414B proposes a medical image analysis method, computer device and readable storage medium. The method includes: obtaining a medical image to be analyzed; inputting the medical image into a preset classification model to obtain a classification result of the medical image; performing a reduction operation: performing a reduction operation on the medical image to obtain a reduced medical image; inputting the reduced medical image into the classification model to obtain a classification result of the reduced medical image; if the accuracy of the classification result of the reduced medical image is greater than the accuracy of the classification result of the medical image, determining the reduced medical image as the region of interest of the medical image; if the accuracy of the classification result of the reduced medical image is not greater than the accuracy of the classification result of the medical image, changing the direction of the reduction operation and repeating the reduction operation. This method improves the accuracy of the determined region of interest.

[0004] However, although the above patents introduce machine learning methods into medical image analysis and improve the processing speed and accuracy, there are still some key problems. First, due to the high-dimensionality and complexity of medical images, the existing machine learning models still need to be improved in terms of feature processing and classification accuracy. In addition, the existing technologies also have deficiencies in data augmentation, the stability of model training, and generalization ability. Summary of the Invention

[0005] The object of the present invention is to provide a method and system for automatic classification of medical images based on machine learning, which solves the problems existing in the prior art.

[0006] To achieve the above object,

[0007] In the first aspect, the present invention provides a method for automatic classification of medical images based on machine learning, including the following steps:

[0008] Collect medical image data and enhance the quantity of medical image data, wherein the enhancement of the quantity of medical image data is realized by a generative adversarial network algorithm based on quantum interference;

[0009] Perform dimensionality reduction processing on the enhanced quantity of the medical image data extracted; wherein, an autoencoder algorithm based on a relaxation factor is used as a feature dimensionality reduction model;

[0010] Classify the dimensionality-reduced features to obtain the categories of medical image data.

[0011] Further, the training steps of the generative adversarial network algorithm based on quantum interference include:

[0012] Initialize the generator and discriminator of the generative adversarial network;

[0013] Through the preparation stage of the quantum state, the generator simulates several medical image states, wherein the medical image states include images of health conditions or pathological features; through a quantum interference simulation strategy, the medical image states will be coherently superimposed to generate first medical image data with new characteristics;

[0014] The generator outputs the first medical image data; the first medical image data is slightly perturbed for the pixels of the image through a chaotic sequence generation module to increase the diversity of the first medical image data;

[0015] The discriminator learns to identify the differences between the first medical image data and the real medical image data, and adjusts the parameters and improves the recognition accuracy according to the differences;

[0016] Repeat the iteration until the first medical image data generated by the generator makes it difficult for the discriminator to distinguish between true and false.

[0017] Further, in the process of the repeated iteration, the backpropagation algorithm is used to update the parameters of the generator and the discriminator, and the update method is expressed as:

[0018]

[0019]

[0020] In the formula, Generator parameters updated for this iteration, Discriminator parameters updated for this iteration, is the learning rate of the generative adversarial network, and represent the gradients of the loss function of the generative adversarial network with respect to the generator parameters and discriminator parameters, respectively, is the learning rate of the discriminator, and are the weight parameters of the generator and discriminator, respectively;

[0021] Furthermore, a dynamic adjustment strategy is adopted to adjust the learning rate of the discriminator. Let be the gradient of the loss function of the generative adversarial network with respect to the discriminator parameters, expressed as:

[0022]

[0023] where the calculation method of the learning rate of the discriminator is expressed as:

[0024]

[0025] In the formula, represents the change rate of the historical gradient of the discriminator, is the adjustment sensitivity parameter of the discriminator, controlling the response speed of the adjustment factor, and e is the natural number;

[0026] During the training process, the quality and diversity of the generated first medical image data are continuously evaluated until the evaluation function value is greater than the preset value, and the iteration stops.

[0027] Furthermore, the diversity of the first medical image data is increased; specifically expressed as:

[0028]

[0029]

[0030] In the formula, is the generator function; is the output image of the generator; represents the chaotic perturbation function applied to the image, is the coefficient for adjusting the chaotic perturbation intensity; is the final image generated by the generator in this iteration; is the quantum state after quantum interference simulation.

[0031] Furthermore, the calculation of the chaotic perturbation function applied to the image utilizes the combination of the logistic map and the hyperbolic tangent function in chaos theory.

[0032] Furthermore, for the autoencoder algorithm model based on the relaxation factor, the training steps include:

[0033] Initialize the autoencoder network structure. Autoencoding includes an encoder and a decoder. Among them, the encoder maps the high-dimensional medical image data feature vector to a low-dimensional representation space, and the decoder reconstructs the high-dimensional input data from the low-dimensional space;

[0034] Set the initial value of the dynamic relaxation factor and its adjustment strategy, where the relaxation factor is adjusted during the training process according to the error feedback;

[0035] Input the medical image data feature vector, which is converted into a representation in the low-dimensional space through the encoder, and then reconstructed into an output close to the high-dimensional input data through the decoder;

[0036] Calculate the loss function of the autoencoder, including the reconstruction error, the dimensionality reduction error, and the regularization term;

[0037] Perform backpropagation according to the loss function to update the weight and bias parameters in the autoencoder network;

[0038] Repeat the above steps iteratively until the preset stop iteration condition is satisfied.

[0039] Furthermore, for the autoencoder algorithm model based on the relaxation factor, the training steps also include:

[0040] The calculation method of the regularization term of the loss function is expressed as:

[0041]

[0042] In the formula, and respectively represent the Frobenius norm and the L1 norm, which are used for the regularization of the weights and biases, is the first regularization term coefficient of the autoencoder, is the weight matrix of the th layer of the encoder; is the bias of the th layer of the encoder, is the second regularization term coefficient of the autoencoder, is the total number of layers of the encoder, MRr represents the cumulative error through all layers of the encoder, and Rr is the regularization term of the autoencoder;

[0043] The calculation method of the cumulative error through all layers of the encoder is expressed as:

[0044]

[0045]

[0046] In the formula, represents the forward propagation function of the th layer, is the output of the th hidden layer of the encoder, is the output of the th hidden layer of the encoder, is the activation function of the th layer of the encoder, is the weight coefficient of the local reconstruction error, is the local mapping in the low-dimensional space, an adaptive low-dimensional representation, expressed as:

[0047]

[0048] In the formula, is the local mapping in the low-dimensional space, is the weight matrix of the encoder; is the bias of the encoder, is the sample set of the current batch input into the autoencoder, is the th neighbor point, is the activation function of the th layer of the encoder, is the th medical image data point to the th neighbor point's weighting coefficient, satisfying , and the calculation method of the weighting coefficient is expressed as:

[0049]

[0050] In the formula, is the scale parameter of the distance, controlling the image range of the weighting coefficient of the points in the neighborhood.

[0051] Furthermore, classify the features after dimensionality reduction to obtain the medical image data category. The steps further include:

[0052] Initialize the parameters of the fractional-order neural network;

[0053] Input the features of the medical image data after dimensionality reduction into the fractional-order neural network, process through the fractional-order activation function, and transfer layer by layer to the output layer;

[0054] Calculate the loss function according to the actual output and the expected output;

[0055] Calculate the gradient of the loss function with respect to each parameter through the backpropagation algorithm, and update the weights and biases of the fractional-order neural network;

[0056] Adjust the fractional-order parameters according to the loss function to optimize the learning efficiency and accuracy of the model. The adjustment method is expressed as:

[0057]

[0058] In the formula, is the fractional-order adjustment coefficient, which is used to control the update speed and amplitude of the fractional-order parameters. is the gradient of the loss function with respect to the fractional-order parameters, which is used to guide the adaptive adjustment of the fractional-order parameters. is the order of the fractional-order of the activation function of the th layer of the fractional-order neural network.

[0059] Repeat the above steps iteratively until the preset iteration stop condition is satisfied.

[0060] In a second aspect, a machine learning-based medical image automatic classification system includes:

[0061] A medical image data acquisition unit, which is used to collect medical image data and increase the quantity of medical image data.

[0062] A medical image data mining unit, which is used to perform dimensionality reduction processing on the enhanced medical image data.

[0063] A medical image data classification unit, which is used to classify the features after dimensionality reduction to obtain the medical image data categories.

[0064] In a machine learning-based medical image automatic classification method and system of the present invention, data augmentation uses a generative adversarial network based on quantum interference to simulate the complex patterns and textures of real medical images through modulation and control in multiple dimensions, increasing the quantity and diversity of medical image data while maintaining the detail quality of the images. Feature extraction uses a neural network based on dynamic adaptive oscillation to optimize the exploration of the high-dimensional parameter space by simulating the nonlinear oscillation behavior in physical phenomena, allowing the neural network to effectively utilize local extrema. Feature dimensionality reduction uses an autoencoder based on a relaxation factor to adaptively adjust the dimensionality reduction intensity by adjusting the training intensity of the autoencoder to meet the training requirements at different stages and optimize the preservation of information. The classifier uses a fractional-order neural network based on a dynamic penalty term, combining the dynamic penalty term and the gated vector regularization technique to optimize the classification accuracy and the generalization ability of the model. It solves the deficiencies of the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art.

[0066] Figure 1 It is the flowchart of the automatic medical image classification method based on machine learning in the first embodiment of the present invention;

[0067] Figure 2 It is the principle block diagram of the automatic medical image classification system based on machine learning in the second embodiment of the present invention;

[0068] Figure 3 It is the experimental effect diagram of revealing the coupling effect between parameters and verifying the robustness of the algorithm to key hyperparameters through three-dimensional parameter space scanning.

[0069] In the figure: 201 - medical image data acquisition unit, 202 - medical image data mining unit, 203 - medical image data classification unit. Detailed implementation manners

[0070] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0071] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of the present invention described here can be implemented in an order different from those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily need to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0072] The first embodiment of the present application is as follows:

[0073] Please refer to Figure 1 and Figure 2 , wherein, Figure 1 It is the training flowchart of the neural network algorithm based on dynamic adaptive oscillation in the first embodiment of the present invention. Figure 2 It is the step flowchart of the automatic medical image classification method based on machine learning in the first embodiment of the present invention.

[0074] The present invention provides a method for automatic classification of medical images based on machine learning, comprising the following steps:

[0075] S101: Collect medical imaging data and increase the amount of medical imaging data;

[0076] Specifically, the source of the medical image data collection of the present invention is a medical image database, including but not limited to CT, MRI and X-ray images, and the medical image data collection complies with the privacy protection standards of the Health Insurance Portability and Accountability Act to ensure the anonymization of patient information;

[0077] The medical image data acquisition method of the present invention includes exporting medical image data directly from the hospital's image storage and transmission system, and collecting through a cooperative medical team in specific diagnosis and treatment activities, and all medical image data are stored in DICOM (Digital Imaging and Communications Standard) format;

[0078] The medical imaging data acquisition type of the present invention is grayscale image, with a pixel depth of 16 bits and a resolution of not less than 256x256 pixels. The acquisition content focuses on common disease hallmarks, such as tumors, inflammation and other signs;

[0079] The medical image data annotation categories of the present invention cover multi-level classification from general anatomical structures to specific pathological changes, specifically organ types under different pathological conditions. The annotation process is completed by professional doctors to ensure the accuracy of the annotation;

[0080] It is understandable that in the task of the present invention, the collection, acquisition, annotation and preprocessing of medical image training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model. The present invention uses a generative adversarial network algorithm based on quantum interference to generate samples, thereby realizing medical image data expansion;

[0081] Quantum interference modulates and controls medical imaging data in multiple dimensions to mimic the complex patterns and textures in real medical images, so that the generated images can be closer to the statistical characteristics of real images in detail.

[0082] Specifically, the training process of the quantum interference-based generative adversarial network algorithm is as follows:

[0083] 1. Initialize the generator and discriminator of the generative adversarial network. The initialization method is expressed as:

[0084] ,

[0085] In the formula, To obey a specific distribution; and are the weight parameters of the generator and discriminator, respectively, represents the initial variance of the weights; represents a normal distribution with a mean of 0 and a variance of ; represents the normal distribution.

[0086] 2. First, the generator simulates various possible medical image states through the quantum state preparation stage. Each state represents an image of a specific health condition or pathological feature. Based on the quantum interference strategy, these states will be coherently superimposed to generate medical images with new characteristics. The way of simulating quantum interference is expressed as:

[0087] and In the formula, represents the quantum state. Specifically, different possible states of the medical image are expressed through qubits; is the quantum state before quantum interference simulation; is the quantum state after quantum interference simulation; represents the operation unit of the quantum interferometer; is the control parameter of the quantum interferometer; and are the amplitude modulation and phase modulation parameters, respectively, is the index of the quantum state characteristic dimension; is the imaginary unit.

[0088] Furthermore, the calculation methods of the amplitude modulation and phase modulation parameters are expressed as:

[0089] and In the formula, is the angle parameter dynamically generated by the generator network according to the input medical image data; is a value randomly sampled from the uniform distribution to simulate the randomness of the qubit phase and increase the detail richness of the generated image.

[0090] 3. According to the result of quantum interference, the generator outputs preliminary medical image data. These medical image data then pass through a chaotic sequence generation module, which slightly perturbs the pixels of the image to increase the diversity of the image, expressed as:

[0091] and In the formula, is the generator function; is the output image of the generator; represents the chaotic perturbation function applied to the image, is the coefficient that adjusts the intensity of the chaotic perturbation; The final image generated by the generator in this iteration.

[0092] Furthermore, the calculation of the chaotic perturbation function utilizes the combination of the logistic map and the hyperbolic tangent function in chaos theory to enhance the randomness and complexity of the generated image. The calculation method is expressed as: In the formula, is the perturbation amplitude control coefficient, is the parameter for adjusting the intensity of the non-linear perturbation, represents element-wise multiplication; is the hyperbolic tangent function. Preferably, is set to 0.2.

[0093] 4. The task of the discriminator is to distinguish between the generated image and the real image. It continuously adjusts its parameters by learning to recognize the subtle differences between the generated image and the real image to improve the recognition accuracy. During the adversarial training process, the calculation method of the loss function is expressed as: In the formula, is the loss function of the generative adversarial network; represents the expectation; is the distribution of real medical image data, is the distribution of generated medical image data, represents the real medical image; represents the medical image data distribution of the real medical image; represents the medical image data distribution of the final image generated by the generator in this iteration; is the discriminator function.

[0094] 5. In each iteration, the performance of the generator and the discriminator is evaluated, and the parameters are adjusted according to the feedback of the discriminator. This process is repeated continuously until the generator can generate images of sufficient quality that the discriminator can hardly distinguish between true and false. The backpropagation algorithm is used to update the parameters of the generator and the discriminator, and the update method is expressed as:

[0095] ,

[0096] In the formula, is the parameter of the generator updated in this iteration; is the parameter of the discriminator updated in this iteration; is the learning rate of the generative adversarial network, and respectively represent the gradients of the loss function of the generative adversarial network with respect to the generator parameter and the discriminator parameter; is the learning rate of the discriminator.

[0097] Further, a dynamic adjustment strategy is adopted to adjust the learning rate of the discriminator, and the speed of gradient descent is adaptively adjusted according to the characteristics of the input medical image data in each iteration, so as to improve the stability and efficiency of training. Let be the gradient of the loss function of the generative adversarial network with respect to the discriminator parameters, which is expressed as:

[0098]

[0099] Further, the calculation method of the learning rate of the discriminator is expressed as:

[0100]

[0101] In the formula, represents the change rate of the historical gradient of the discriminator; is the adjustment sensitivity parameter of the discriminator, which controls the response speed of the adjustment factor. Preferably, is set to 2.

[0102] Further, the calculation method of the change rate of the historical gradient of the discriminator is expressed as:

[0103] In the formula, is the -th iteration of the gradient of the loss function of the generative adversarial network with respect to the discriminator parameters; is the -th iteration of the gradient of the loss function of the generative adversarial network with respect to the discriminator parameters.

[0104] 6. During the training process, continuously evaluate the quality and diversity of the generated images to ensure that the newly generated medical image data matches the real medical image data in terms of visual and statistical characteristics. Finally, evaluate the quality of the generated images to ensure that the classification quality standard is met. The evaluation method is expressed as:

[0105]

[0106] In the formula, is the generated image quality evaluation function. When the evaluation function value is greater than the preset value, it means that the generated medical image data meets the classification quality standard. Then stop the iteration, otherwise continue the iteration; and are the pixel values of the generated image and the reference real image respectively, is the total number of pixels in the image; is the index of the image pixel point.

[0107] After the training of the medical image data augmentation model is completed, the trained medical image data augmentation model is used to increase the number of samples. In one embodiment, assume that the original collected samples are 800, and the medical image data augmentation model generates 200 samples through augmentation. Then, the augmented medical image data set contains 1000 samples.

[0108] S102: Perform dimensionality reduction on the enhanced medical image data;

[0109] Specifically, through feature dimensionality reduction, the medical image data is further compressed to achieve the full mining of medical image data.

[0110] The present invention uses an autoencoder based on a relaxation factor as a feature dimensionality reduction model. The autoencoder realizes the compression and reconstruction of medical image data through an encoder and a decoder, aiming to learn the optimal representation of medical image data in a low-dimensional space;

[0111] Different from traditional autoencoders, the present invention adopts a dynamic relaxation factor in the structure of the autoencoder, allowing the autoencoder model to adaptively adjust the dimensionality reduction intensity during the training process to ensure the maximum retention of information, optimize the generalization ability of the autoencoder model, and at the same time reduce the risk of overfitting.

[0112] Specifically, the training process of the autoencoder algorithm based on the relaxation factor is as follows:

[0113] 1. Initialize the network structure of the autoencoder. The autoencoder includes two parts: an encoder and a decoder. The encoder is responsible for mapping high-dimensional input data to a low-dimensional representation space. The high-dimensional input data is the feature vector of medical image data. The decoder then reconstructs the high-dimensional input data from this low-dimensional space. The parameter update method for the encoder is expressed as:

[0114] In the formula, is the weight matrix of the th layer of the encoder; is the bias of the th layer of the encoder; means subject to a specific distribution; means a normal distribution; means the initialized variance.

[0115] 2. Before model training, set the initial value of the dynamic relaxation factor and its adjustment strategy. The dynamic relaxation factor is adjusted according to the error feedback during the training process to adapt to the training requirements of different stages. The adjustment method is expressed as:

[0116] In the formula, is the The relaxation factor for one training cycle; is the initial relaxation factor; is the relaxation factor decay rate; represents the current iteration number. Preferably, is set to 1, is set to 0.5.

[0117] 3. The feature vector of the input medical image data is converted into a representation in a low-dimensional space through an encoder, and then these low-dimensional representations are reconstructed into an output close to the high-dimensional input data through a decoder. The conversion process from the input layer to the hidden layer can be expressed as:

[0118] In the formula, is the output of the th hidden layer of the encoder; is the feature vector of the medical image data input to the encoder; is the th activation function of the encoder.

[0119] 4. Calculate the loss function of the autoencoder to evaluate the performance of the current autoencoder network. The loss function of the autoencoder includes reconstruction error, dimensionality reduction error, and regularization term, and the calculation method is expressed as:

[0120] In the formula, is the loss function of the autoencoder; is the reconstructed data output by the decoder, is the feature vector of the medical image data input to the encoder, is the th relaxation factor for one training cycle, represents the deviation between the output of the encoder and its ideal low-dimensional representation; is the L2 norm; is the regularization term of the autoencoder.

[0121] Furthermore, the calculation method of the regularization term of the loss function is expressed as:

[0122]

[0123] In the formula, and respectively represent the Frobenius norm and the L1 norm, which are used for the regularization of weights and biases; is the first regularization term coefficient of the autoencoder, is the weight matrix of the th layer of the encoder; is the bias of the th layer of the encoder, is the coefficient of the second regularization term of the autoencoder; is the total number of layers of the encoder; represents the cumulative error through all layers of the encoder.

[0124] Furthermore, the calculation method of the cumulative error through all layers of the encoder is expressed as:

[0125]

[0126]

[0127] In the formula, represents the forward propagation function of the layer; is the output of the hidden layer of the layer of the encoder, is the output of the hidden layer of the layer of the encoder, is the activation function of the layer of the encoder; is the weight coefficient of the local reconstruction error; is the local mapping in the low-dimensional space. Preferably,

[0128] Furthermore, since the traditional autoencoder does not fully consider the local characteristics of different medical image data points in the low-dimensional space, the medical image data points are samples corresponding to medical image data. The present invention adopts a local adaptive reconstruction method based on a low-dimensional subspace, and adaptively adjusts the representation of each medical image data point in the low-dimensional space according to its local geometric characteristics during the dimensionality reduction process, strengthening the recognition and retention of the local structure of medical image data by the model, and improving the accuracy and flexibility of feature dimensionality reduction. Specifically, the local mapping of the medical image data point in the low-dimensional space is defined as This mapping is not only generated by a global dimensionality reduction strategy, but is obtained by weighted reconstruction of the features of all medical image data points in the sample set input to the autoencoder in the current batch, resulting in an adaptive low-dimensional representation, expressed as:

[0129]

[0130] In the formula, is the local mapping in the low-dimensional space, is the weight matrix of the encoder; is the bias of the encoder, is the sample set input to the autoencoder in the current batch; is the th neighbor point; is the activation function of the layer of the encoder; is the weighting coefficient of the th medical image data point to the th neighbor point, satisfying .

[0131] Furthermore, the weighting coefficient adopts a distance-based weighting strategy, such that the neighbor medical image data points closer to the medical image data point contribute more to its local reconstruction, and the neighbor medical image data points farther away contribute less. The calculation method is expressed as:

[0132]

[0133] In the formula, is the scale parameter of the distance, controlling the influence range of the points in the neighborhood on the weighting coefficient; exp() is the natural exponential function. Preferably, is set to 0.1.

[0134] 5. Perform backpropagation according to the loss function to update the weight and bias parameters in the autoencoder network. The gradient descent update method for the weights and biases of the encoder is expressed as:

[0135]

[0136]

[0137] In the formula, is the learning rate of the autoencoder; and are the gradients of the loss function of the autoencoder with respect to the weight and bias respectively.

[0138] 6. Repeat the above steps until the preset stop iteration condition is satisfied, which indicates that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0139] S103: Classify the dimension-reduced features to obtain the medical image data category.

[0140] Specifically, through the classifier model, classify the medical image data after feature dimension reduction, and then obtain the medical image data category;

[0141] The present invention uses a fractional-order neural network based on a dynamic penalty term to classify the medical image data after feature dimension reduction. The dynamic penalty term is to impose a penalty constraint on the weights of the fractional-order neural network during the training process of the fractional-order neural network, so as to adjust the fractional-order parameters and optimize the classification accuracy;

[0142] Specifically, the training process of the fractional-order neural network based on the dynamic penalty term is as follows:

[0143] 1. Initialize the parameters of the fractional-order neural network, and the initialization method is expressed as:

[0144]

[0145]

[0146] In the formula, represents the weight of the th layer of the fractional-order neural network; represents the bias of the th layer of the fractional-order neural network; is the variance of the weight initialization of the fractional-order neural network; represents being subject to a specific distribution; represents the normal distribution. Preferably, is set to 0.01.

[0147] 2. Input the medical image data features after dimensionality reduction processing into the fractional-order neural network, process them through the fractional-order activation function, and transmit them layer by layer to the output layer, which is expressed as:

[0148]

[0149] In the formula, is the output of the th layer of the fractional-order neural network; is the output of the th layer of the fractional-order neural network; is the fractional-order activation function; is the order of the fractional order of the activation function of the th layer of the fractional-order neural network; represents element-wise multiplication; is the gating vector of the th layer of the fractional-order neural network.

[0150] Furthermore, an adaptive feature identification gating mechanism is adopted to enhance the feature selection ability in the training process of the classifier model. Specifically, for each layer in the fractional-order neural network of the classifier, calculate the gating vector, and the element value thereof determines the transmission intensity of the corresponding feature. The calculation method is expressed as:

[0151]

[0152] In the formula, is the Sigmoid function; and are the gating weight parameter and the gating bias parameter respectively, which are training parameters used to learn which features are most important for the current task.

[0153] 3. Calculate the loss function based on the actual output and the expected output of the fractional-order neural network. The present invention adopts a loss function containing a dynamic penalty term to adjust the fractional-order parameters and optimize the classification accuracy. At the same time, the L1 regularization term of the gating vector is adopted to promote the generalization ability of the model. The calculation method is expressed as:

[0154]

[0155] In the formula, is the loss function of the fractional-order neural network; is the one-hot encoding of the true label of the th sample; is the classification output of the fractional-order neural network for the th sample; is the number of layers of the fractional-order neural network; is the regularization parameter of the fractional-order neural network, which controls the weight decay to prevent overfitting; represents the square of the Frobenius norm of the weight of the th layer; is the weight of the gating vector regularization, which balances the model complexity and performance; is the L1 norm. Preferably, is set to 0.2, is set to 0.3.

[0156] 4. Through the backpropagation algorithm, calculate the gradient of the loss function with respect to each parameter, and update the weights and biases of the fractional-order neural network, which is expressed as:

[0157]

[0158]

[0159] In the formula, is the parameter update operation; is the learning rate of the fractional-order neural network; and are the gradients of the loss function of the fractional-order neural network with respect to the weights and biases.

[0160] 5. Adjust the fractional-order parameters according to the loss function, so as to optimize the learning efficiency and accuracy of the model. The adjustment method is expressed as:

[0161]

[0162] In the formula, is the fractional-order adjustment coefficient, which is used to control the update speed and amplitude of the fractional-order parameters; is the gradient of the loss function with respect to the fractional-order parameters, which is used to guide the adaptive adjustment of the fractional-order parameters. Preferably, Set to 0.2.

[0163] 6. Repeat the above steps iteratively until the preset stop iteration condition is satisfied, which indicates that the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0164] After the training is completed, use the trained machine learning model to process new medical image samples to achieve automated classification, and the classification result is the output category of the medical image data classification unit.

[0165] Data augmentation uses a generative adversarial network based on quantum interference to simulate the complex patterns and textures of real medical images through modulation and control in multiple dimensions, increasing the quantity and diversity of medical image data while maintaining the detail quality of the images. Feature dimensionality reduction uses an autoencoder based on a relaxation factor to adapt to the training requirements at different stages by adjusting the training intensity of the autoencoder, adaptively adjusting the dimensionality reduction intensity, and optimizing the preservation of information. The classifier uses a fractional-order neural network based on dynamic penalty terms, combining dynamic penalty terms and gated vector regularization techniques to optimize the classification accuracy and the generalization ability of the model. This solves the deficiencies of the prior art.

[0166] The second embodiment of this application is as follows:

[0167] Based on the first embodiment, please refer to Figure 2 , where Figure 2 is the principle block diagram of the machine learning-based medical image automatic classification system according to the second embodiment of the present invention.

[0168] The machine learning-based medical image automatic classification system of this embodiment includes a medical image data acquisition unit 201, a medical image data mining unit 202, and a medical image data classification unit 203.

[0169] For this specific embodiment, the medical image data mining unit 202 is respectively connected to the medical image data acquisition unit 201 and the medical image data classification unit 203;

[0170] The medical image data acquisition unit 201 is used to collect medical image data and enhance the quantity of medical image data;

[0171] The medical image data mining unit 202 is used to extract features from the enhanced medical image data and perform dimensionality reduction processing on the extracted features;

[0172] The medical image data classification unit 203 is used to classify the dimensionality-reduced features to obtain disease categories.

[0173] Using a machine learning-based medical image automatic classification system according to this embodiment, the medical image data acquisition unit 201 mainly collects medical image data and enhances the medical image data. The way to enhance the medical image data is to increase the quantity of medical image data through sample augmentation; further compress the medical image data to achieve full excavation of the medical image data. The medical image data classification unit 203 classifies the medical image data after feature dimensionality reduction through a classifier model, and then obtains the category of the medical image data. The medical image data acquisition unit 201, the medical image data mining unit 202, and the medical image data classification unit 203 perform model training according to the above-mentioned machine learning-based medical image automatic classification method. After the model training is completed, the trained machine learning model is used to process new medical image samples to achieve automated classification, and the classification result is the output category of the medical image data classification unit 203. Data augmentation uses a generative adversarial network based on quantum interference to simulate the complex patterns and textures of real medical images through modulation and control in multiple dimensions, increasing the quantity and diversity of medical image data while maintaining the detail quality of the images. Feature dimensionality reduction uses an autoencoder based on a relaxation factor, and adjusts the training intensity of the autoencoder to adapt to the training needs at different stages, adaptively adjusts the dimensionality reduction intensity, and optimizes the preservation of information. The classifier uses a fractional-order neural network based on a dynamic penalty term, combines the dynamic penalty term and the gated vector regularization technique, and optimizes the classification accuracy and the generalization ability of the model. It solves the deficiencies of the prior art.

[0174] Experimental example

[0175] As Figure 3 described above, to deeply study the effectiveness of the joint optimization of hyperparameters of the autoencoder algorithm based on the relaxation factor, in the dynamic relaxation factor autoencoder, the relaxation factor decay rate and the learning rate are the core hyperparameters. The two jointly control the parameter update amplitude (the learning rate directly affects the step size of gradient descent), the change of regularization intensity (the relaxation factor decay rate determines the time-varying characteristics of the regularization term), and the exploration-exploitation balance (the dynamic relaxation factor adjusts the strategy of the model during training). Therefore, in this experiment, through three-dimensional parameter space scanning, the coupling effect between parameters is revealed, the robustness of the algorithm to key hyperparameters is verified, and the optimal parameter combination is located.

[0176] In this experiment, the range of parameters is selected as follows:

[0177] Relaxation factor decay rate ∈ [0.1, 1.0]: covering the full scenario from slow decay to fast decay;

[0178] Learning rate ∈ [0.001, 0.01]: including the typical learning rate interval for deep learning;

[0179] Sampling points 50×50: Ensure the parameter space resolution.

[0180] The loss surface modeling is expressed as:

[0181] Theoretical terms: The first two terms reflect the independent effects of parameters (as the learning rate increases, the loss decreases; as the learning rate increases, the loss decreases);

[0182] Coupling terms: Simulate the non - linear interaction between parameters;

[0183] Noise terms: Simulate the actual training fluctuations.

[0184] In this experiment, through cold - hot color mapping, the low - high loss regions are visually displayed with a blue - red gradient, and the global minimum point is highlighted.

[0185] In terms of the parameter coupling effect, in the high relaxation factor decay rate + low learning rate region (upper right), the loss surface rises steeply, indicating that when the relaxation factor decays rapidly, an appropriate learning rate compensation is required; in the low relaxation factor decay rate + high learning rate region (lower left), periodic fluctuations occur, reflecting that the parameter combination causes gradient oscillations; for the optimal channel, a low - loss corridor is formed along the relaxation factor decay rate = 0.45 - 0.55, indicating that there is an optimal ratio relationship between parameters.

[0186] To verify the algorithm robustness of the auto - encoder, for the robust band of the relaxation factor decay rate, in the interval of relaxation factor decay rate = 0.4 - 0.6, the loss value remains below 0.6 (cold - color region); for the learning rate tolerance, when the learning rate > 0.005, the loss surface shows a bifurcated mutation, indicating that the learning rate needs to be precisely controlled.

[0187] From the experiment, it can be seen that the theoretical optimal point is (relaxation factor decay rate = 0.5, learning rate = 0.005), and the gradient along the relaxation factor decay rate axis is greater than that along the learning rate axis, proving that the relaxation factor strategy is more sensitive to performance.

[0188] Furthermore, according to the surface characteristics, a parameter adjustment strategy is given, as shown in Table 1:

[0189] Table 1 Optimization suggestion table for parameter adjustment strategy based on surface characteristics

[0190]

[0191] Furthermore, to verify the necessity of the dynamic relaxation factor, the experimental data shows that the optimal relaxation factor decay rate = 0.5 corresponds to the relaxation factor decaying according to decay, satisfying the requirement of decaying to 0.67% of the initial value at t = 10, ensuring a reduction in the regularization strength in the later stage of training.

[0192] Furthermore, a learning rate sensitivity analysis was conducted. The calculation of the second derivative of the loss with respect to the learning rate showed a sudden change in curvature when the learning rate > 0.008, corresponding to the phenomenon of surface bifurcation.

[0193] Compared with the traditional method, using the traditional fixed relaxation factor (relaxation factor decay rate = 0) as a control, the results are shown in Table 2:

[0194] Table 2 Comparison results of the dynamic relaxation factor decay rate and the fixed relaxation factor decay rate = 0

[0195]

[0196] The experimental results show that there is a significant coupling effect between the relaxation factor decay rate and the learning rate, and joint optimization is required. The dynamic relaxation factor mechanism makes the loss surface smoother and the optimal region wider. In the interval of relaxation factor decay rate = 0.45 - 0.55 and learning rate = 0.004 - 0.006, the algorithm shows strong robustness.

[0197] What is disclosed above is only one or more preferred embodiments of the present application, and the scope of the rights of the present application cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present application still fall within the scope covered by the present application.

[0198] Those of ordinary skill in the art will realize that the embodiments described herein are for the purpose of assisting the reader in understanding the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations without departing from the essence of the present invention based on the technical revelations disclosed in the present invention, and these deformations and combinations are still within the scope of protection of the present invention.

Claims

1. A method for automatic classification of medical images based on machine learning, characterized in that: The following steps are involved: Collect medical imaging data and increase the amount of medical imaging data, wherein the amount of medical imaging data is increased through a generative adversarial network algorithm based on quantum interference; The quantum interference simulation method is expressed as: In the formula, ψ c Representing quantum states, specifically, expressing different possible states of medical images through quantum bits; is the quantum state before quantum interference simulation; is the quantum state after quantum interference simulation; U c () represents the operation unit of the quantum interferometer; φ c is the control parameter of the quantum interferometer; and are the amplitude modulation and phase modulation parameters respectively, j c is the index of the characteristic dimension of the quantum state; i * is an imaginary unit; The calculation method of the amplitude modulation and phase modulation parameters is expressed as: In the formula, It is the angle parameter dynamically generated by the generator network based on the input medical imaging data; is a value randomly sampled from the uniform distribution U(0,1); Performing dimensionality reduction processing on the quantity of the extracted and enhanced medical image data; wherein an autoencoder algorithm based on a relaxation factor is used as a feature dimensionality reduction model; The training process of the relaxation factor-based autoencoder algorithm is as follows: Initialize the autoencoder network structure; Before model training, the initial value of the dynamic relaxation factor and its adjustment strategy are set. The dynamic relaxation factor is adjusted according to error feedback during the training process. The specific adjustment method is expressed as: r，t =λ r，0 exp(-α r int(t)), α r >0In the formula, λ r，t is the relaxation factor of the tth training cycle; r，0 is the initial relaxation factor; α r is the relaxation factor decay rate; int() represents the current number of iterations; The input medical image data feature vector is converted into a low-dimensional space representation through the encoder; The loss function of the autoencoder is calculated to evaluate the performance of the current autoencoder network. The loss function of the autoencoder includes reconstruction error, dimensionality reduction error and regularization term. The calculation method is expressed as: Where, L r () is the loss function of the autoencoder; is the reconstructed data output by the decoder, x r is the feature vector of medical imaging data input to the encoder, λ r，t is the relaxation factor for the tth training cycle, Δz r represents the deviation between the encoder output and its ideal low-dimensional representation; || || is the L2 norm; R r is the regularization term of the autoencoder; The features after dimensionality reduction are classified to obtain the categories of medical image data; wherein, a fractional-order neural network based on a dynamic penalty term is used to classify the medical image data after feature dimensionality reduction, and the dynamic penalty term is a penalty constraint on the weight of the fractional-order neural network during the training process of the fractional-order neural network; The training process of the fractional-order neural network based on dynamic penalty term is as follows: Initialize the parameters of the fractional neural network; The medical image data features after dimensionality reduction are input into the fractional-order neural network, processed by the fractional-order activation function, and passed to the output layer layer by layer, which is expressed as: In the formula, is the output of the lth layer of the fractional-order neural network; is the output of the l-1th layer of the fractional-order neural network; is a fractional activation function; is the fractional order of the activation function of the lth layer of the fractional-order neural network; ⊙ represents element-by-element multiplication; is the gate vector of the lth layer of the fractional-order neural network; The loss function is calculated based on the actual output and expected output of the fractional-order neural network. The specific calculation method is expressed as: Where, L u ( ) is a fraction The loss function of the neural network of order y u，i is the one-hot encoding of the true label of the i-th sample; is the classification output of the fractional-order neural network for the i-th sample; L rea is the number of layers of the fractional-order neural network; u It is the regularization parameter of the fractional-order neural network, which controls the weight decay to prevent overfitting; represents the square of the Frobenius norm of the weight of the lth layer; g is the weight of the gated vector regularization, balancing model complexity and performance; || ||1 is the L1 norm; Through the back propagation algorithm, the gradient of the loss function for each parameter is calculated, and the weights and biases of the fractional-order neural network are updated; the fractional-order parameters are adjusted according to the loss function to optimize the learning efficiency and accuracy of the model. The specific adjustment method is expressed as: In the formula, γ eut is the fractional-order adjustment coefficient, which is used to control the update speed and amplitude of the fractional-order parameters; It is the gradient of the loss function with respect to the fractional-order parameter, which is used to guide the adaptive adjustment of the fractional-order parameter.

2. The method for automatic classification of medical images based on machine learning according to claim 1, characterized in that: The training steps of the quantum interference-based generative adversarial network algorithm include: Initialize the generator and discriminator of the generative adversarial network; The generator simulates a plurality of medical image states through a quantum state preparation stage, wherein the medical image states include images of health conditions or pathological characteristics; through a quantum interference simulation strategy, the medical image states are coherently superimposed to generate first medical image data with new characteristics; The generator outputs first medical image data; the first medical image data is slightly disturbed by the pixels of the image through the chaotic sequence generation module to increase the diversity of the first medical image data; The discriminator learns to identify the difference between the first medical image data and the real medical image data, adjusts parameters according to the difference and improves recognition accuracy; The iteration is repeated until the first medical image data generated by the generator makes it difficult for the discriminator to distinguish the authenticity.

3. The method for automatic classification of medical images based on machine learning according to claim 2, characterized in that: In the repeated iterative process, the back propagation algorithm is used to update the generator parameters and the discriminator parameters, and the updating method is expressed as: In the formula, is the generator parameter updated for this iteration, is the discriminator parameter updated in this iteration, η c is the learning rate of the generative adversarial network, and They represent the gradients of the loss function of the generative adversarial network with respect to the generator parameters and the discriminator parameters, respectively. is the learning rate of the discriminator, and are the weight parameters of the generator and the discriminator respectively.

4. The method for automatic classification of medical images based on machine learning according to claim 3, characterized in that: Adopt a dynamic adjustment strategy to the learning rate of the discriminator Make adjustments, set To generate the gradient of the loss function of the adversarial network with respect to the discriminator parameters, it is expressed as: Among them, the calculation method of the discriminator's learning rate is expressed as: In the formula, represents the rate of change of the discriminator's historical gradient, k c is the adjustment sensitivity parameter of the discriminator, which controls the response speed of the adjustment factor, and e is a natural number; During the training process, the quality and diversity of the generated first medical image data are continuously evaluated until the iteration is stopped when the evaluation function value is greater than a preset value.

5. The method for automatic classification of medical images based on machine learning according to claim 2, characterized in that: Increase the diversity of the first medical image data; specifically expressed as: In the formula, G c ( ) is the generator function; is the output image of the generator; Δ c ( ) represents the chaotic perturbation function applied to the image, β c is the coefficient that adjusts the intensity of chaotic disturbance; The final image generated by the generator in this iteration; It is the quantum state after quantum interference simulation.

6. The method for automatic classification of medical images based on machine learning according to claim 5, characterized in that: The calculation of the chaotic perturbation function applied to the image utilizes the combination of the logical mapping and the hyperbolic tangent function in the chaos theory.

7. The method for automatic classification of medical images based on machine learning according to claim 1, characterized in that: Based on the relaxation factor autoencoder algorithm model, the training steps include: Initializing an autoencoder network structure, the autoencoder includes an encoder and a decoder, wherein the encoder maps a high-dimensional medical image data feature vector to a low-dimensional representation space, and the decoder reconstructs high-dimensional input data from the low-dimensional space; Setting an initial value of a dynamic relaxation factor and its adjustment strategy, wherein the relaxation factor is adjusted during the training process according to error feedback; The input medical image data feature vector is converted into a representation of a low-dimensional space through an encoder, and then reconstructed into an output close to the high-dimensional input data through a decoder; Calculate the loss function of the autoencoder, including reconstruction error, dimensionality reduction error, and regularization term; Perform back propagation according to the loss function to update the weights and bias parameters in the autoencoder network; Repeat the above steps until the preset stop iteration condition is met.

8. The method for automatic classification of medical images based on machine learning according to claim 7, characterized in that: Based on the relaxation factor autoencoder algorithm model, the training steps also include: The calculation method of the regularization term of the loss function is expressed as: In the formula, || || F and || ||1 represent the Frobenius norm and L1 norm, respectively, for the regularization of weights and biases, ρ r is the first regularization coefficient of the autoencoder, is the weight matrix of the encoder layer l; is the bias of the encoder layer l, μ r is the second regularization coefficient of the autoencoder, L cf is the total number of layers of the encoder, MRr represents the cumulative error through all layers of the encoder, and Rr is the regularization term of the autoencoder; The calculation method of the accumulated error through all layers of the encoder is expressed as: In the formula, represents the forward propagation function of the lth layer, is the output of the l-1th hidden layer of the encoder, is the output of the encoder’s lth hidden layer, is the activation function of the first layer of the encoder, γ rec is the weight coefficient of the local reconstruction error, is a local mapping of the low-dimensional space, x r is the feature vector of medical imaging data input to the encoder, is the reconstructed data output by the decoder; || || is the L2 norm; The adaptive low-dimensional representation is expressed as: In the formula, is a local mapping of the low-dimensional space, W er is the weight matrix of the encoder; b er is the encoder bias, N r is the sample set input into the autoencoder for the current batch, For nth r Neighborhood points, is the activation function of the Lth layer of the encoder, is the rth medical image data point for the nth r The weighted coefficients of neighbor points satisfy The weighting coefficient calculation method is expressed as: In the formula, σ ras It is the scale parameter of the distance, which controls the image range of the weighted coefficient of the point pairs within the neighborhood.

Images