A method, system, device, and medium for classifying low-quality medical images.

Abstract

This invention discloses a method, system, device, and medium for classifying low-quality medical images, belonging to the field of artificial intelligence medical technology. Its purpose is to solve the technical problem of low accuracy in classifying benign and malignant low-quality medical images in existing technologies. The method includes: acquiring sample images, probabilistic language conversion, constructing an nmODE network model, training the nmODE network model, and real-time image classification. During probabilistic language conversion, a probabilistic language terminology set is constructed, using different probabilistic terms to characterize different brightness levels of the image. The acquired sample images are then probabilistically converted using this terminology set to obtain single-channel probabilistic language representations of the sample images. Based on the multiple single-channel probabilistic language representations of the sample images, a multi-channel feature map that can be input into the model is obtained. Probabilistic language conversion enhances the linguistic expressiveness of low-quality medical image features, improving the accuracy of low-quality medical image classification.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence medical technology, and relates to medical image classification, and more particularly to a method, system, device and medium for classifying low-quality medical images. Background Technology

[0002] Due to limitations in equipment hardware, complex acquisition environments, and improper operation, low-quality medical image data is often generated. Using these low-quality medical images to train classification models can lead to overfitting, feature extraction bias, and weakened generalization ability, increasing the complexity and difficulty of classifying low-quality medical images and further exacerbating the risk of misdiagnosis by doctors. Therefore, the processing and improvement of low-quality medical image data classification has become a crucial area requiring breakthroughs in the field of medical image data classification.

[0003] Currently, classification models for medical image data mainly rely on traditional CNN models, such as ResNet, which achieves image classification through end-to-end training.

[0004] Patent application No. 202411711397.5 discloses a medical visual question-answering method and system based on multi-task modeling. The method includes: extracting image features from the medical image to be analyzed using a visual encoder; inputting preliminary question commands and historical question-answering data into a large dialogue model to generate question commands, which are then processed by a text encoder to obtain text features; fusing the obtained text features and image features, performing self-attention image importance weighting and image-text alignment to obtain a fused image-text representation; inputting this representation into a multi-target output projection layer for multi-task prediction, outputting a question answer, image category, and important region mask; inputting the question answer and image category into the large dialogue model, combining the dialogue context and multi-turn interaction data, and incorporating the important region mask to finally generate a detailed diagnostic opinion. The visual encoder uses a ResNet architecture and is pre-trained on medical image data to enable medical image feature extraction. It utilizes convolutional neural network (CNN) layers to hierarchically encode image features, with each convolutional layer progressively extracting features at different levels, ultimately forming high-dimensional image features.

[0005] For processing low-quality medical images, image enhancement is the primary method. Image feature enhancement methods mainly include horizontal flipping, random rotation, or brightness perturbation of the data. Patent application number 202410466402.4 discloses a deep learning-based blood glucose prediction method, device, equipment, and storage medium. In the training process of the image quality enhancement model, the CycleGAN model can be selected as the generative adversarial network model in the diffusion model. The CycleGAN model first synthesizes low-quality images based on high-quality images, forming a training image pair with the high-quality images. Based on the synthesized low-quality images, the U-Net model is selected as the basic enhancement model in the diffusion model. The diffusion model uses the U-Net model to iteratively learn the inverse mapping from low-quality images to high-quality images in multiple steps, minimizing the difference between the output image and the original high-quality image. In practical applications of the trained image quality enhancement model, the model continuously adds Gaussian noise to the input eye image to destroy low-quality components. Then, it uses a base enhancement model to iteratively remove the noise to restore a higher-quality eye image, thereby achieving the goal of enhancing the image quality of low-quality eye images.

[0006] As mentioned in the aforementioned patent application, while image enhancement techniques can address the problems of low-quality medical images to some extent, low-quality medical data inherently suffers from significant noise, low contrast, loss of structural details, and severe artifact interference. These characteristics often result in a lack of substantial feature information for lesion classification. Therefore, the decision-making problem in classifying lesions (benign or malignant) using low-quality medical image data lies in the semantic level, manifested as a lack of data feature information. Thus, addressing the accuracy of lesion (benign or malignant) classification in low-quality medical images from a semantic perspective is particularly important. Summary of the Invention

[0007] The purpose of this invention is to address the technical problem of low accuracy in classifying lesions (benign or malignant) in low-quality medical images in the prior art, and to provide a method, system, device, and medium for classifying low-quality medical images. This involves calculating membership degrees based on triangular fuzzy sets and constructing an image brightness dataset, converting low-quality medical images into probabilistic linguistic representations, enhancing the linguistic expressive power of low-quality medical image features, and thereby improving the accuracy of low-quality medical image classification.

[0008] To achieve the above objectives, the present invention specifically adopts the following technical solution: A method for classifying low-quality medical images, comprising the following steps: Step S1: Obtain the sample image; Medical sample images are acquired, blurred to varying degrees, and the benign or malignant lesions in the medical sample images are labeled to obtain label data. Step S2, probabilistic language conversion; Construct a probabilistic language terminology set for characterizing the brightness and darkness of an image, and use different probabilistic terms in the terminology set to characterize different brightness and darkness of the image; The sample image after blurring in step S1 is converted into a probabilistic language using a probabilistic language terminology set to obtain a single-channel probabilistic language representation of the sample image; based on the multiple single-channel probabilistic language representations of the sample image, the final multi-channel feature map is obtained. Step S3: Construct the nmODE network model; Construct the nmODE network model; Step S4: Train the nmODE network model; Input the multi-channel feature maps and corresponding label data of each sample image obtained in step S2 into the nmODE network model constructed in step S3 to train the nmODE network model. Step S5: Real-time image classification; The medical image to be classified is acquired and probabilistic language conversion is performed to obtain a multi-channel feature map. The multi-channel feature map is then input into the nmODE network model trained in step S4. The nmODE network model outputs the classification result of benign or malignant lesions.

[0009] Furthermore, in step S2, the specific method for obtaining the probabilistic linguistic representation results of each sample image is as follows: Step S2-1: Construct a probabilistic language terminology set Probabilistic Language Terminology Set Five probabilistic languages ​​are used to describe the different levels of brightness in an image; Step S2-2: Calculate the membership degree of the triangular fuzzy set for the low-quality medical sample images and convert it into a probabilistic language representation to obtain the multi-channel feature map of each sample image.

[0010] Furthermore, in step S2-1, the probabilistic language terminology set is constructed. Represented as: ; Among them, the probabilistic language terminology set The pixel ranges corresponding to each probabilistic language are: .

[0011] Furthermore, in step S2-2, the formula for calculating membership degree is: ; in, , , These represent the starting node, peak node, and ending node of the triangular fuzzy set, respectively.

[0012] Further, in step S3, the constructed nmODE network model includes multiple independently trainable intelligent units and a voter. The output of each intelligent unit is output as a classification result after passing through the voter. Each intelligent unit includes an input layer, a linear layer W1, a neuron ODE layer, a linear layer W2, and an output layer arranged sequentially.

[0013] Furthermore, the ordinary differential equation of the neuron's ODE layer is: ; in, The output value of the memory neuron representing the presentation state. Represents the set of learning weight parameters. This represents the input neuron representing the external input variable. Indicates the deviation value. Indicates the number of input neurons. This indicates the number of memory neurons.

[0014] Furthermore, in step S4, the specific steps for training the nmODE network model are as follows: Step S4-1, Input the training set ; Step S4-2, initialize parameters; The parameters include the linear layer weights between input neurons and memory neurons. Linear layer weights between memory neurons and decision neurons Bias terms Learning rate and Network status and time nodes ; Step S4-3: Select training batches ; Step S4-4: For each training sample in the batch, solve the corresponding one-dimensional ordinary differential equation. ,get ; Step S4-5: Calculate the derivative of the loss with respect to the output of the linear layer W2. ; Step S4-6: Solve the three-dimensional inverse differential equation to calculate the gradient. ; ; ; in, This represents the Lagrange multipliers introduced during the solution process. The output value of the memory neuron representing the presentation state. This represents the set of learning weight parameters for the linear layer w1 of the neural network. This represents the input neuron representing the external input variable. Indicates the deviation value. Indicates the number of input neurons. Represents the Lagrange multiplier. Represents the gradient; Step S4-7, for and Perform gradient summation; Step S4-8, according to the update rules Update weights Weight and bias terms ; in, and All represent the learning rate. Indicates a time point. This represents the derivative of the loss with respect to the output of the linear layer W2. This indicates that the nmODE network is at time [time]. state, Indicates when The gradient at which = 0 is multiplied by the input variable of the input neuron. This represents the gradient when t=0; Step S4-9: Repeat steps S4-3 to S4-8 until the weight is reached. Weight and bias terms convergence; Step S4-10, return the weights Weight and bias terms .

[0015] A classification system for low-quality medical images, comprising: The sample image acquisition module is used to acquire medical sample images, perform different degrees of blurring on the medical sample images, and label the benign and malignant lesions in the medical sample images to obtain label data; The probabilistic language conversion module is used to construct a probabilistic language terminology set for describing the brightness and darkness of an image, and to use different probabilistic languages ​​in the probabilistic language terminology set to describe different brightness and darkness of the image. The sample images after blurring in the sample image acquisition module are converted using a probabilistic language terminology set to obtain single-channel probabilistic language representation results of the sample images; based on the multiple single-channel probabilistic language representation results of the sample images, the final multi-channel feature map is obtained. The nmODE network model building module is used to build nmODE network models. The nmODE network model training module is used to input the multi-channel feature maps and corresponding label data of each sample image obtained by the probabilistic language conversion module into the nmODE network model constructed by the nmODE network model construction module to train the nmODE network model. The real-time image classification module is used to acquire medical images to be classified and perform probabilistic language conversion to obtain multi-channel feature maps. These multi-channel feature maps are then input into the nmODE network model after training, and the nmODE network model outputs the classification results of benign and malignant lesions.

[0016] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method described above.

[0017] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described method.

[0018] The beneficial effects of this invention are as follows: 1. In this invention, a probabilistic language terminology set for characterizing the brightness of an image is constructed, which can be used to convert low-quality medical images into probabilistic language. The membership degree of the low-quality medical sample image is calculated by triangular fuzzy set, and it is converted into a probabilistic language representation form to obtain a multi-channel feature map of the image, which is used as the input of the network model. This can enhance the linguistic expression ability of low-quality medical image features, thereby improving the accuracy of low-quality medical image classification.

[0019] 2. In this invention, the network model adopts an innovative nmODE network model, which utilizes a neural memory ordinary differential equation network to learn a decoupled structure that separates input neurons and memory neurons. This enables the capture of key classification features from blurry, noisy, low-quality medical images, achieving stable and accurate classification results. This is of great significance to the development of data-driven smart healthcare. By using this classification method, the accuracy of medical record image analysis can be improved, and robust and accurate auxiliary diagnostic systems can be developed, thereby improving doctors' work efficiency and reducing the risk of misdiagnosis.

[0020] 3. In this invention, an efficient network learning algorithm (training method) is developed for training the nmODE network model. This algorithm uses three-dimensional ODE to calculate the gradient; this method demonstrates that nmODE... 2 It is decoupled, avoiding the gradient vanishing problem, thereby improving computational efficiency and accuracy in classifying low-quality medical images. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a schematic diagram of the nmODE network model in this invention; Figure 3 This is a schematic diagram of the structure of the intelligent unit in this invention; Figure 4 This is a schematic diagram of the ordinary differential equation of the neuron ODE layer in this invention; Figure 5 This is a schematic diagram of the probabilistic language representation conversion of grayscale images in this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0023] Therefore, all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] Example 1 This embodiment provides a classification method for low-quality medical images, used to classify lesions in low-quality medical images as benign or malignant. This embodiment uses breast MRI images as an example to classify the benign or malignant nature of lesion regions in breast images. In this embodiment, the benign or malignant nature of lesion regions in the image is divided into seven categories: "1", "2", "3", "4A", "4B", "4C", and "5"; where categories 1-3 represent benign findings, category 4 is further subdivided into 4A (lowly suspicious), 4B (moderately suspicious), and 4C (highly suspicious), and category 5 represents a high probability of malignancy.

[0025] When classifying lesions as benign or malignant in low-quality medical images (breast MRI images), such as Figure 1 As shown, the specific steps include: Step S1: Obtain the sample image; Medical sample images are acquired, blurred to varying degrees, and the benign or malignant lesions in the medical sample images are labeled to obtain label data.

[0026] In this embodiment, the medical sample images and their corresponding labels are all from the management systems of various hospitals. In the hospital management system, each breast cancer classification dataset has a set of images. The names of these images represent a characteristic character, and they are accompanied by an associated JSON document. This JSON document contains detailed information about three attribute values: image, category, and annotation. The image attribute consists of the image's height, width, ID, and name. The category attribute is used to determine the category to which the image belongs (i.e., label data) based on the image ID. The annotation attribute contains detailed information about each image in the dataset.

[0027] Blurring medical sample images to varying degrees can be applied. This can involve applying multiple degrees of blurring to the same image or different images, allowing those skilled in the art to choose the appropriate method. This simulates low-quality medical sample images. Therefore, using the blurred images as input to a network model enables the model to learn the ability to segment and classify lesions in low-quality medical images and output the benign or malignant classification of the lesions.

[0028] Because the sample size of low-quality breast medical images in each category is imbalanced after blurring, oversampling is necessary. Specifically: First, load the breast cancer classification dataset from the hospital system, obtain the number of images for each category, and blur them; second, take the largest number of images, ensuring that the oversampling value for each category is at least close to the maximum value. Then, divide the maximum value by the number of images in each category. This yields the oversampled values ​​for each category. Finally, the training set is dynamically expanded by a preset oversampling value, so that the sum of the original samples and the duplicate samples equals the number of images multiplied by their oversampling value plus 1.

[0029] oversampling value The calculation formula is: ; in, This represents the number of images in the i-th category.

[0030] Step S2, probabilistic language conversion; Construct a probabilistic language terminology set for characterizing the brightness and darkness of an image, and use different probabilistic terms in the terminology set to characterize different brightness and darkness of the image; The blurred sample image from step S1 is subjected to probabilistic language transformation using a probabilistic language terminology set to obtain a single-channel probabilistic language representation of the sample image. Based on the multiple single-channel probabilistic language representations of the sample image, the final multi-channel feature map is obtained. Specifically: Step S2-1: Construct a probabilistic language terminology set Probabilistic Language Terminology Set Five probabilistic terms are used to describe the different levels of brightness in an image. Specifically: First, a probabilistic linguistic terminology set needs to be defined beforehand to describe the brightness and darkness of an image. This terminology set defines five piecewise linear transformation parameters that map pixel values ​​to different intervals to extract texture or density information related to pathological features, and constructs an antisymmetric probability function for probabilistic linguistic classification. Specifically, the probabilistic linguistic terminology set... Terminology elements in ~ These describe different levels of brightness for each pixel. The pixel range for "dark" is -63.75 to 63.75, represented in probabilistic language. The pixel range corresponding to "slightly darker" is 0 to 127.5, which is represented in probabilistic language. The pixel range corresponding to "general" is 63.75-191.25, which is represented in probabilistic language. The pixel range corresponding to "slightly brighter" is 127.5 to 255, which is represented in probabilistic language. The pixel range corresponding to "bright" is 191.25 to 318.75, which can be represented in probabilistic language. Therefore, this probabilistic language terminology set It can be represented as: .

[0031] Probability Language Terminology Set The pixel ranges corresponding to each probabilistic language are shown in Table 1.

[0032] Table 1 Probabilistic Language - Pixel Correspondence Table .

[0033] Among them, the five probabilistic languages ​​are composed of five triangular fuzzy sets of the same shape. The pixel is characterized such that the sum of the membership degrees of each pixel in each fuzzy set is 1.

[0034] Step S2-2: Calculate the membership degree of the triangular fuzzy set for the low-quality medical sample images and convert it into a probabilistic language representation to obtain the multi-channel feature map of each sample image.

[0035] For each low-quality medical sample image, extract each single-channel grayscale image of each sample image (the sample image is an RGB image, so each image has three single-channel grayscale images); calculate the membership degree of a triangular fuzzy set for each single-channel grayscale image, and treat this membership degree as the probability of an element in a probabilistic linguistic terminology. That is, the probabilistic linguistic representation of a pixel in a given single-channel grayscale image is: ; in, - These correspond to the probability of that pixel in the corresponding probability language.

[0036] When calculating the membership degree of a triangular fuzzy set, the triangular fuzzy set... The formula for calculating the membership degree is: ; in, , , These represent the starting node, peak node, and ending node of the triangular fuzzy set, respectively.

[0037] According to a specific rule, the probabilistic linguistic representation of each single-channel grayscale image is transformed into 5 independent channels (that is, the 5 probabilities are merged into five independent channels, and the probability in a certain independent channel is used as the value of that independent channel), resulting in a multi-channel feature map of each single-channel grayscale image. The multi-channel feature map corresponding to each single-channel grayscale image of each sample image will be used for the training of subsequent models.

[0038] Taking pixel 48 as an example, its membership degrees below the five probabilistic terms mentioned above are 0.25, 0.75, 0, 0, and 0, respectively. If we consider each membership degree as the probability of a probabilistic term element, then pixel 48 can be converted into a probabilistic representation: ; Using the probabilistic language description method for pixels described above, the entire image is converted into a probabilistic language representation. For a single-channel grayscale image, the five probability values ​​are treated as five independent channels. For a three-channel RGB image, the converted probabilistic language representation contains 15 channels. The specific implementation is as follows... Figure 5 As shown.

[0039] Step S3: Construct the nmODE network model; An nmODE network model is constructed, using the multi-channel feature maps of a single-channel grayscale image as input. The nmODE network model can output the benign or malignant classification of lesions in low-quality medical images, such as... Figure 2 As shown.

[0040] The nmODE network model comprises multiple independently trainable intelligent units and a voter. The output of each intelligent unit is processed by the voter to output the classification result. This nmODE network model aggregates multiple intelligent units (neural memory ordinary differential equation network) using an ensemble learning approach. Each intelligent unit is trained independently, and only when outputting the final decision result is the voter used to determine the benign or malignant lesion classification of the low-quality medical image.

[0041] like Figure 3 As shown, this intelligent unit comprises a five-layer structure: an input layer, a linear layer W1, a neuron ODE layer, a linear layer W2, and an output layer. First, the input layer is formatted as [batch size, image dimension]. The linear layer W1 maps the input image data to the hidden space. The neuron ODE layer further performs nonlinear transformations on the hidden features and uses the odeint solver to solve the ordinary differential equations of neural memory. The linear layer W2 maps the hidden space to the output layer, which outputs the final classification result.

[0042] The neuron ODE layer uses a new neuron, whose one-dimensional neural network mapping equation is:

[0043] in, , Both represent the sensory input of neurons.

[0044] Based on this, such as Figure 4 As shown, the ordinary differential equation (neural memory) of the neuronal ODE layer is expressed as: ; in, The output value of the memory neuron representing the presentation state. Represents the set of learning weight parameters. This represents the input neuron representing the external input variable. Indicates the deviation value. Indicates the number of input neurons. This indicates the number of memory neurons.

[0045] After using this neuron, the neuronal ODE layer has the following characteristics: (1) Set It is the invariant set of ordinary differential equations; The definition of an invariant set is: if for any The dynamic system from The starting trajectory Will always remain in the collection Among them, the dynamic meaning of invariant sets is: when the system state enters the set... The system state will remain unchanged indefinitely. The proof of the invariant set property of ordinary differential equations is as follows: Given any state belonging to this invariant set The solution to an ordinary differential equation can be expressed in the form of: ; in, This indicates the upper limit of the training time interval. Represents a time interval between 0 and t. This represents the state of the i-th memory neuron at time t=0. This represents the state of the i-th memory neuron at time t=s. Indicates the number of memory neurons. Let w represent the set of learned weight parameters for the linear layer w1 of the neural network. This represents the i-th deviation value in n dimensions. Input neurons represent external input variables.

[0046] because Then we have: ; ; ; (2) An ordinary differential equation has one and only one equilibrium point; An equilibrium point is the point where the system state no longer changes; that is, when the system reaches an equilibrium point, its state derivative is 0, and the system is at rest. The proof of the equilibrium point characteristics of ordinary differential equations is as follows: Consider a one-dimensional ordinary differential equation. If the equation has one and only one equilibrium point, then it suffices to prove the function... Uniqueness of the solution. Clearly, The first derivative of is greater than or equal to 0, which indicates that It is a monotonically non-decreasing function. Next, consider... , and Three scenarios: 1) And only in Time to take, etc.; 2) And only in Time to take, etc.; 3) Due to It is monotonically non-decreasing, therefore there must exist a point... Make

[0047] Therefore, the function The existence of a unique solution indicates that the ordinary differential equation has one and only one equilibrium point.

[0048] (3) An ordinary differential equation has one and only one global attractor; A global attractor is defined as follows: For a dynamical system, an attractor is a set of states that gradually converge over time. This set can be a point, a trajectory, or even other more complex structures. Once the system enters the neighborhood of an attractor, it will not leave that state. If an attractor has an attraction for all initial conditions in the phase space, it is called a global attractor. A global attractor must be an equilibrium point. Therefore, mathematically, the first derivative of a global attractor at that point is zero, indicating that there is no trend of change at that point. The proof is as follows. According to the uniqueness of the equilibrium point of a one-dimensional ordinary differential equation, there must exist a point... To satisfy The following constructs a Lyapunov function to characterize the current state and equilibrium point of the system. Distance between: ;

[0049]

[0050] in, Let represent the nonlinear solution of p = sin2(p + γ) at time t; Using trigonometric identities Then we have:

[0051] when , Approximately ,therefore, ; in, Let t be a function of t. k (k=1,2,⋯) is a time series such that | |=1.

[0052] in, It is located in and A function between, therefore: ; This indicates that the derivative of the state function is less than 0, therefore when The system state will approach ,prove It is a global attractor.

[0053] Step S4: Train the nmODE network model; The multi-channel feature maps and corresponding label data of each sample image obtained in step S2 are input into the nmODE network model constructed in step S3 to train the nmODE network model.

[0054] The specific steps for training the nmODE network model are as follows: Step S4-1, Input the training set ; Step S4-2, initialize parameters; The parameters include the linear layer weights between input neurons and memory neurons. Linear layer weights between memory neurons and decision neurons Bias terms Learning rate and Network status and time nodes ; Step S4-3: Select training batches ; Step S4-4: For each training sample in the batch, solve the corresponding one-dimensional ordinary differential equation. ,get ; When using the odeint solver to solve ordinary differential equations of neural memory, the solution process is as follows: 1) Initialize the state vector All elements are initialized to 0 as initial conditions for the ODE solver; 2) Define equations using the NeuronODE module. ,in This represents the change of state over time. For neuron state variables, and For nonlinear parameters in the ODE dynamic equation, Parameters generated for linear layer W1; 3) Call the odeint solver to solve the ordinary differential equation within the time interval. The ODE numerical solution is calculated internally, and the solution at the last time point is taken as the output.

[0055] This means that the linear layer w1 maps the input (probabilistically processed) multi-channel feature image data to the hidden layer, that is, the input layer image dimension... Convert to hidden layer image dimensions ,in This represents the dimension of the image after the input channel feature map has been flattened. Representing the dimension of memory neurons. Using the ODE equation; The dimensions of the hidden layer image were nonlinearly transformed and forward propagated to the neuron ODE for solution.

[0056] Step S4-5: Calculate the derivative of the loss with respect to the output of the linear layer W2. ; This means that the output of the ODE layer is mapped to the final output classification layer through the linear layer w2, which is the hidden layer dimension that has undergone nonlinear transformation. Transform into the final output layer dimension ;

[0057] Where s is the activation function softmax, and the output a(t) is the decision neuron.

[0058] Step S4-6: Solve the three-dimensional inverse differential equation to calculate the gradient. ; ; ; in, This represents the Lagrange multipliers introduced during the solution process. The output value of the memory neuron representing the presentation state. This represents the set of learning weight parameters for the linear layer w1 of the neural network. This represents the input neuron representing the external input variable. Indicates the deviation value. Indicates the number of input neurons. Represents the Lagrange multiplier. Represents the gradient; Step S4-7, for and Perform gradient summation; Step S4-8, according to the update rules Update weights Weight and bias terms ; in, and All represent the learning rate. Indicates a time point. This represents the derivative of the loss with respect to the output of the linear layer W2. This indicates that the nmODE network is at time [time]. state, Indicates when The gradient at which = 0 is multiplied by the input variable of the input neuron. This represents the gradient when t=0;

[0059] The cost function is:

[0060] For each neuron i, the gradient is obtained by solving the three-dimensional invODE. ,use Calculate gradient .function and There is a relationship .

[0061] Step S4-9: Repeat steps S4-3 to S4-8 until the weight is reached. Weight and bias terms convergence; Step S4-10, return the weights Weight and bias terms .

[0062] Step S5: Real-time image classification; The medical image to be classified is acquired and probabilistic language conversion is performed to obtain a multi-channel feature map. The multi-channel feature map is then input into the nmODE network model trained in step S4. The nmODE network model outputs the classification result of benign or malignant lesions.

[0063] Example 2 This embodiment provides a classification system for low-quality medical images, including: The sample image acquisition module is used to acquire medical sample images, perform different degrees of blurring on the medical sample images, and label the benign and malignant lesions in the medical sample images to obtain label data; The probabilistic language conversion module is used to construct a probabilistic language terminology set for describing the brightness and darkness of an image, and to use different probabilistic languages ​​in the probabilistic language terminology set to describe different brightness and darkness of the image. The sample images after blurring in the sample image acquisition module are converted using a probabilistic language terminology set to obtain single-channel probabilistic language representation results of the sample images; based on the multiple single-channel probabilistic language representation results of the sample images, the final multi-channel feature map is obtained. The nmODE network model building module is used to build nmODE network models. The nmODE network model training module is used to input the multi-channel feature maps and corresponding label data of each sample image obtained by the probabilistic language conversion module into the nmODE network model constructed by the nmODE network model construction module to train the nmODE network model. The real-time image classification module is used to acquire medical images to be classified and perform probabilistic language conversion to obtain multi-channel feature maps. These multi-channel feature maps are then input into the nmODE network model after training, and the nmODE network model outputs the classification results of benign and malignant lesions.

[0064] Example 3 A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform steps of a classification method for low-quality medical images.

[0065] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0066] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Of course, the memory may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is commonly used to store the operating system and various application software installed on the computer device, such as the program code for the classification method of low-quality medical images. Furthermore, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0067] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is used to run program code stored in the memory or process data, such as running program code for a classification method of low-quality medical images.

[0068] Example 4 A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform steps of a classification method for low-quality medical images.

[0069] The computer-readable storage medium stores an interface display program that can be executed by at least one processor to cause the at least one processor to perform the steps of the classification method for low-quality medical images as described above.

[0070] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the low-quality medical image classification method described in the embodiments of this application.

Claims

1. A method for classifying low-quality medical images, characterized in that, Includes the following steps: Step S1: Obtain the sample image; Medical sample images are acquired, blurred to varying degrees, and the benign or malignant lesions in the medical sample images are labeled to obtain label data. Step S2, probabilistic language conversion; Construct a probabilistic language terminology set for characterizing the brightness and darkness of an image, and use different probabilistic terms in the terminology set to characterize different brightness and darkness of the image; The sample image after blurring in step S1 is converted into a probabilistic language using a probabilistic language terminology set to obtain a single-channel probabilistic language representation of the sample image; based on the multiple single-channel probabilistic language representations of the sample image, the final multi-channel feature map is obtained. Step S3: Construct the nmODE network model; Construct the nmODE network model; Step S4: Train the nmODE network model; Input the multi-channel feature maps and corresponding label data of each sample image obtained in step S2 into the nmODE network model constructed in step S3 to train the nmODE network model. Step S5: Real-time image classification; The medical image to be classified is acquired and probabilistic language conversion is performed to obtain a multi-channel feature map. The multi-channel feature map is then input into the nmODE network model trained in step S4. The nmODE network model outputs the classification result of benign and malignant lesions. In step S2, the specific method for obtaining the probabilistic linguistic representation results of each sample image is as follows: Step S2-1: Construct a probabilistic language terminology set Probabilistic Language Terminology Set Five probabilistic languages ​​are used to describe the different levels of brightness in an image; Step S2-2: Calculate the membership degree of the triangular fuzzy set for the low-quality medical sample images and convert it into a probabilistic language representation to obtain the multi-channel feature map of each sample image. In step S2-2, the formula for calculating membership degree is: ； in, , , These represent the start node, peak node, and end node of the triangular fuzzy set, respectively. Step S3: The constructed nmODE network model includes multiple independently trainable intelligent units and a voter. The output of each intelligent unit is output as a classification result after passing through the voter. Each intelligent unit includes an input layer, a linear layer W1, a neuron ODE layer, a linear layer W2, and an output layer arranged sequentially. The ordinary differential equation of the neuronal ODE layer is: ； in, The output value of the memory neuron representing the presentation state. Represents the set of learning weight parameters. This represents the input neuron representing the external input variable. Indicates the deviation value. Indicates the number of input neurons. This indicates the number of memory neurons.

2. The method for classifying low-quality medical images as described in claim 1, characterized in that: In step S2-1, the probabilistic language terminology set is constructed. Represented as: ； Among them, the probabilistic language terminology set The pixel ranges corresponding to each probabilistic language are: 。 3. The method for classifying low-quality medical images as described in claim 1, characterized in that: In step S4, the specific steps for training the nmODE network model are as follows: Step S4-1, Input the training set ; Step S4-2, initialize parameters; The parameters include the linear layer weights between input neurons and memory neurons. Linear layer weights between memory neurons and decision neurons Bias terms Learning rate and Network status and time nodes ; Step S4-3: Select training batches ; Step S4-4: For each training sample in the batch, solve the corresponding one-dimensional ordinary differential equation. ,get ; Step S4-5: Calculate the derivative of the loss with respect to the output of the linear layer W2. ; Step S4-6: Solve the three-dimensional inverse differential equation to calculate the gradient. ； ； ； in, This represents the Lagrange multipliers introduced during the solution process. The output value of the memory neuron representing the presentation state. This represents the set of learning weight parameters for the linear layer w1 of the neural network. This represents the input neuron representing the external input variable. Indicates the deviation value. Indicates the number of input neurons. Represents the Lagrange multiplier. Represents the gradient; Step S4-7, for and Perform gradient summation; Step S4-8, according to the update rules Update weights Weight and bias terms ; in, and All represent the learning rate. Indicates a time point. This represents the derivative of the loss with respect to the output of the linear layer W2. This indicates that the nmODE network is at time [time]. state, Indicates when The gradient at which = 0 is multiplied by the input variable of the input neuron. This represents the gradient when t=0; Step S4-9: Repeat steps S4-3 to S4-8 until the weight is reached. Weight and bias terms convergence; Step S4-10, return the weights Weight and bias terms .

4. A classification system for low-quality medical images, characterized in that, include: The sample image acquisition module is used to acquire medical sample images, perform different degrees of blurring on the medical sample images, and label the benign and malignant lesions in the medical sample images to obtain label data; The probabilistic language conversion module is used to construct a probabilistic language terminology set for describing the brightness and darkness of an image, and to use different probabilistic languages ​​in the probabilistic language terminology set to describe different brightness and darkness of the image. The sample images after blurring in the sample image acquisition module are converted using a probabilistic language terminology set to obtain single-channel probabilistic language representation results of the sample images; based on the multiple single-channel probabilistic language representation results of the sample images, the final multi-channel feature map is obtained. The nmODE network model building module is used to build nmODE network models. The nmODE network model training module is used to input the multi-channel feature maps and corresponding label data of each sample image obtained by the probabilistic language conversion module into the nmODE network model constructed by the nmODE network model construction module to train the nmODE network model. The real-time image classification module is used to acquire medical images to be classified and perform probabilistic language conversion to obtain multi-channel feature maps. The multi-channel feature maps are then input into the nmODE network model after training. The nmODE network model outputs the classification results of benign and malignant lesions. In the probabilistic language conversion module, the specific method for obtaining the probabilistic language representation results of each sample image is as follows: Step S2-1: Construct a probabilistic language terminology set Probabilistic Language Terminology Set Five probabilistic languages ​​are used to describe the different levels of brightness in an image; Step S2-2: Calculate the membership degree of the triangular fuzzy set for the low-quality medical sample images and convert it into a probabilistic language representation to obtain the multi-channel feature map of each sample image. In step S2-2, the formula for calculating membership degree is: ； in, , , These represent the start node, peak node, and end node of the triangular fuzzy set, respectively. In the nmODE network model construction module, the constructed nmODE network model includes multiple independently trainable intelligent units and a voter. The output of each intelligent unit is output as a classification result after passing through the voter. Each intelligent unit includes an input layer, a linear layer W1, a neuron ODE layer, a linear layer W2, and an output layer arranged sequentially. The ordinary differential equation of the neuronal ODE layer is: ； in, The output value of the memory neuron representing the presentation state. Represents the set of learning weight parameters. This represents the input neuron representing the external input variable. Indicates the deviation value. Indicates the number of input neurons. This indicates the number of memory neurons.

5. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that: The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 3.

Images