Artificial intelligence-based diabetic nephropathy patient retinopathy image classification method

Abstract

The application discloses a diabetic nephropathy patient retinopathy image classification method based on artificial intelligence, relates to the technical field of medical images, and comprises the following steps: collecting retinal images and carrying out denoising and contrast adjustment to generate standardized images; using a generative adversarial network guided by medical prior knowledge, combining microaneurysm morphology and hemorrhage area texture features to perform feature constraint, generating retinal images with enhanced pathological features, and constructing an amplified training data set; training a deep convolutional neural network based on the data set, using a multi-scale feature extraction layer to identify microaneurysm and hemorrhage lesions, and outputting preliminary classification results; combining clinical physician feedback to construct a reinforcement learning environment, dynamically adjusting classification decision thresholds and optimizing lesion classification parameters. The method disclosed by the application adopts cross-modal correlation analysis, fuses retinal lesion features and biochemical time series data, and improves the accurate classification ability of diabetic retinopathy images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical images, in particular to an image classification method for retinopathy of diabetic nephropathy patients based on artificial intelligence. BACKGROUND

[0002] With the development of deep learning technology, medical image analysis methods based on artificial intelligence have made significant progress in retinopathy detection and classification. Currently, mainstream AI methods usually rely on large-scale retinal image datasets for training to improve the generalization ability of the model. However, due to the limited access to medical data, especially in the diabetic nephropathy patient population, high-quality, fully annotated datasets of retinopathy are extremely scarce. In addition, the subtle features of the lesion area in the retinal image (microaneurysms, hemorrhage) are difficult to be accurately identified by traditional computer vision algorithms, affecting the accuracy of disease classification.

[0003] Existing retinal image classification methods mainly rely on standard deep learning models, but have the following shortcomings: first, the scarcity of training data leads to insufficient generalization ability of the model when facing different patient populations; second, existing methods lack feature constraint mechanisms in identifying microaneurysms and hemorrhagic lesions, and are easily disturbed by noise, affecting the classification accuracy; third, the classification model may lead to misdiagnosis or missed diagnosis without the optimization of doctor feedback. Therefore, there is an urgent need for an intelligent classification method that can enhance pathological features, optimize classification accuracy, and combine real-time feedback from doctors to improve the detection and grading ability of retinopathy in diabetic nephropathy patients. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an image classification method for retinopathy of diabetic nephropathy patients based on artificial intelligence to solve the problems of low classification accuracy, difficulty in extracting lesion area features, and lack of doctor feedback optimization.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an image classification method for retinopathy of diabetic nephropathy patients based on artificial intelligence, which comprises collecting retinal images and performing denoising and contrast adjustment to generate standardized images;

[0008] Input into a generative adversarial network guided by medical prior knowledge, use the microaneurysm morphological parameter library and the hemorrhagic area texture feature atlas for feature constraint, generate retinal images with enhanced pathological features and construct an expanded training dataset;

[0009] The deep convolutional neural network is trained based on an expanded training data set, a multi-scale feature extraction layer is used to identify microaneurysms and hemorrhagic lesions in a retinal image, and an initial diagnosis result containing lesion grading is output;

[0010] An enhanced learning environment is constructed by combining real-time feedback correction data of clinical doctors, a sensitivity-specificity balance factor is used to dynamically adjust the classification decision threshold, and optimized grading parameters are generated;

[0011] Load the optimized grading parameters, combine the retinal image lesion features and patient biochemical time series data, and generate a three-level retinal image classification result of diabetic retinopathy through cross-modal correlation analysis. As a preferred solution of the diabetic retinopathy image classification method based on artificial intelligence for diabetic nephropathy patients described in the present application, wherein,

[0012] The retinal image is added with Gaussian noise to generate a noisy image, and the original retinal image and the corresponding noisy retinal image are labeled to generate a retinal image data set;

[0013] A U-shaped convolutional neural network is used for denoising training to minimize the pixel difference between the denoised retinal image and the original retinal image;

[0014] Based on the CLAHE algorithm, the denoised retinal image is divided into small blocks, each small block is processed through histogram equalization, the local contrast is adjusted, the processing results of the sub-regions are merged, and the retinal image with balanced contrast is output.

[0015] As a preferred solution of the diabetic retinopathy image classification method based on artificial intelligence for diabetic nephropathy patients described in the present application, wherein:

[0016] Edge detection is performed on the standardized retinal image, the edges of the microaneurysms are extracted through pixel compression, the equidistant sampling points of the microaneurysm lesion profile are extracted, the shape features of the low frequency components are preserved through discrete Fourier transform, and the shape descriptor is calculated;

[0017] For the texture of the hemorrhagic area, a rotationally invariant local binary pattern is calculated using a ring neighborhood, and a differentiable direction encoding layer is used to capture multi-directional texture features;

[0018] The generator based on the morphological features of the microaneurysm and the texture features of the hemorrhagic area adopts a dual-path feature fusion architecture, the backbone network uses a pre-trained EfficientNet-B4 model, and the generative adversarial network.

[0019] As a preferred solution of the diabetic retinopathy image classification method based on artificial intelligence for diabetic nephropathy patients described in the present application, wherein:

[0020] Three convolution kernels arranged in parallel dynamically fuse different scale lesion features through channel attention mechanism.

[0021] As a preferred scheme of the diabetes nephropathy patient retinopathy image classification method based on artificial intelligence, the sensitivity-specificity balance factor comprises,

[0022] The threshold adjustment amount and the evaluation reinforcement learning state vector are output by the reinforcement learning Actor-Critic network to adjust the classification decision threshold through dynamic adjustment of the sliding window mean of the doctor feedback reward.

[0023] As a preferred scheme of the diabetes nephropathy patient retinopathy image classification method based on artificial intelligence, the feature splicing and fusion comprises,

[0024] The retinal image lesion features and blood glucose time series data are spliced in the channel dimension, and the time convolution kernel is used to extract the cross-modal correlation features.

[0025] As a preferred scheme of the diabetes nephropathy patient retinopathy image classification method based on artificial intelligence, the construction of the augmented training data set comprises,

[0026] The pathological feature enhanced retinal image is synthesized based on the generative adversarial network guided by medical prior knowledge;

[0027] The synthesized image is subjected to spatial domain elastic deformation processing, and the deformation parameters generate random shifts based on Gaussian distribution, simulating the position and shape changes of the lesion area.

[0028] As a preferred scheme of the diabetes nephropathy patient retinopathy image classification method based on artificial intelligence, the generation of the pathological feature enhanced retinal image comprises,

[0029] The standard lesion retinal image in the medical knowledge base is used for real-time verification; when the microaneurysm shape or hemorrhage area texture of the generated image is different from the similarity of the standard template, a dynamic correction mechanism is started to adjust the two-way feature fusion weight of the generator to generate an enhanced retinal image.

[0030] In a second aspect, the present application provides a computer device comprising a memory and a processor, the memory storing a computer program, wherein the computer program is executed by the processor to implement any step of the diabetes nephropathy patient retinopathy image classification method based on artificial intelligence according to the first aspect of the present application.

[0031] In a third aspect, the present application provides a computer readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements any step of the method for classifying diabetic retinopathy images of patients with diabetic nephropathy based on artificial intelligence according to the first aspect of the present application.

[0032] The present application has the following beneficial effects: a generative adversarial network (GAN) guided by medical prior knowledge is used to generate retinal images with enhanced pathological features in combination with a microaneurysm morphological parameter library and a hemorrhagic area texture feature atlas, and an expanded training data set is constructed to alleviate the problem of training data scarcity. Secondly, a deep convolutional neural network is trained based on the expanded data set, multi-scale feature extraction layers are used to improve the recognition ability of micro lesions and reduce misdiagnosis and missed diagnosis. In addition, a reinforcement learning mechanism is used to dynamically adjust the classification decision threshold through a sensitivity-specificity balance factor, and the grading parameters are optimized in combination with real-time feedback from clinicians to enhance the reliability and explainability of classification. Finally, cross-modal correlation analysis is used to fuse retinal lesion features and patient biochemical time series data to further improve the accurate classification ability of diabetic retinopathy, providing a scientific basis for early diagnosis and personalized treatment of patients with diabetic nephropathy. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0034] Fig. 1 Flowchart of the method for classifying diabetic retinopathy images of patients with diabetic nephropathy based on artificial intelligence in Example 1.

[0035] Fig. 2 Flowchart of the core process of medical GAN data enhancement in Example 1.

[0036] Fig. 3 Flowchart of dynamic threshold optimization in Example 1.

[0037] Fig. 4 Flowchart of multi-modal fusion process in Example 1. DETAILED DESCRIPTION

[0038] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0039] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0040] It is also noted that, as used herein, "one embodiment" or "an embodiment" means implementation that can include a particular feature, structure, or characteristic. However, such a phrase is not used to imply that all embodiments have that feature, structure, or characteristic. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the purview of one skilled in the art to effect such feature, structure, or characteristic in connection with other embodiments whether or not explicit

[0041] Embodiment 1, Reference Figs. 1-4 The embodiment provides an artificial intelligence-based diabetic nephropathy retinopathy image classification method, comprising the following steps:

[0042] S1, collecting retinal images and performing denoising and contrast adjustment to generate standardized images, comprising the following steps:

[0043] S1.1, using a high-resolution fundus camera to set the resolution to 300-600 DPI, selecting blue and red light combination illumination, reducing glare, enhancing the visible area of blood vessels and lesions in the image, adjusting the exposure time and aperture to ensure uniform image exposure and avoid overexposure or underexposure, starting the autofocus function to clearly focus the retinal area and reduce blurring caused by inaccurate focusing.

[0044] Preferably, during the collection process, at least three high-resolution (e.g., image resolution of 20 microns / pixel) retinal images at different angles are taken each time.

[0045] S1.2, adding noise to the retinal image using a Gaussian noise term to generate a noisy image, and labeling the original retinal image and the corresponding noisy retinal image pair to generate a retinal image dataset.

[0046] Where the Gaussian noise term is represented as:

[0047]

[0048] In the formula, I noisy (x,y) is the gray value of the noisy retinal image at coordinates (x,y), I(x,y) is the gray value of the original retinal image at coordinates (x,y), (x,y) is the two-dimensional coordinates in the retinal image, N(0,σ 2 ) is a normal distribution with mean 0 and variance, ∈ is a noise term and is labeled as a normal distribution, is the variance parameter of the noise intensity.

[0049] S1.3, de-noising using a U-shaped convolutional neural network, minimizing the difference between the noisy retinal image and the true noise-free retinal image by optimizing an L2 loss function.

[0050] An improved U-Net architecture is used for de-noising the noisy retinal image. During the de-noising process, a four-level encoding-decoding structure is used. The encoder extracts multi-scale features by convolution and dilated convolution at each level to capture the noise distribution pattern. The decoder uses transposed convolution to gradually restore the image details and remove noise interference.

[0051] The four-level encoding is represented as:

[0052] DilationRate = [1, 2, 4, 8];

[0053] where the optimization of the L2 loss function is represented as:

[0054]

[0055] where, is the original clear retinal image, is the de-noised retinal image, N p is the pixel value of the retinal image, and m is the summation index variable.

[0056] S1.4, based on the CLAHE algorithm (Contrast Limited Adaptive Histogram Equalization), first divide the input de-noised retinal image into 8x8 pixel blocks, process each block by histogram equalization, adjust the local contrast, limit the histogram of each block after adjustment to avoid over-enhancement, then merge the processing results of the sub-regions, and output the contrast equalized retinal image.

[0057] where the CLAHE algorithm is represented as:

[0058] I en = CLAHE(I; ClipLimit = 2.0, TileSize = 8x8);

[0059] where, en is the enhanced retinal image, CLAHE is the contrast limited adaptive histogram equalization algorithm, I is the input retinal image, ClipLimit = 2.0 is the histogram clipping value, and TileSize = 8x8 is the local processing block size.

[0060] S1.5, using RGB color space conversion to convert the contrast equalized retinal image to Lab color space, adjusting the color using the gray world assumption, and outputting the standardized retinal image.

[0061] S2, input to the medical prior knowledge guided generative adversarial network, using the microaneurysm morphological parameter library and the hemorrhagic area texture feature atlas for feature constraint, generating a retinal image with enhanced pathological features and constructing an expanded training data set, which includes the following steps:

[0062] S2.1, edge detection is performed on the standardized retinal image, the edges of the microaneurysm are extracted by pixel compression, the shape features of the microaneurysm lesions are further quantified, 200 equidistant sampling points of the microaneurysm lesion contour are extracted, which are unfolded into a complex sequence, and the low-frequency component (unit: cycle / mm) describing the essence of the lesion shape is calculated by discrete Fourier transform.

[0063] Wherein the 200 equidistant sampling points of the lesion contour are unfolded into a complex sequence, which is represented as:

[0064] z s =u s +jv s ;

[0065] In the formula, z s is the spatial coordinates of the lesion contour, u s is the u coordinate of the s-th sampling point, jv s is the jv coordinate of the s-th sampling point, and s is the sampling point index.

[0066] Wherein, the representation is calculated by discrete Fourier transform:

[0067]

[0068] In the formula, is the index of the frequency component, k=0, 1, 2, …, 20 is selected in the low frequency part to quantify the lesion shape feature, is the k fieq th Fourier coefficient low frequency component of the lesion shape feature, N s is the total sampling point, is a complex exponential function, z n is the complex definition of the contour point coordinates.

[0069] S2.2, use the rotation invariant deep LBP operator to model the texture of the hemorrhagic area, define a radius ring neighborhood (for example, r=3 pixels x 20 microns / pixel=60 microns) on a 50 micron scale, and use a differentiable direction encoding layer to capture multi-directional texture features. It is represented as:

[0070] LBP RI =min{ROR(LBP P ,θ)|θ∈θ};

[0071] In the formula, LBP RITo extract the texture features of the hemorrhagic area, P is the sampling point of the hemorrhagic area texture, min is the minimum value of all rotation results {ROR(LBP P ,θ)} is the LBP P result after i rounds of rotation, θ is the set of rotation angles, ROR is the cyclic operation, LBP P is the comparison result of the current hemorrhagic area pixel and the hemorrhagic area domain pixel.

[0072] S2.3 Based on the morphological features of microaneurysms and the texture features of the hemorrhagic area, the generator adopts a two-way feature fusion architecture, and the backbone network extracts multi-scale features through the pre-trained EfficientNet-B4.

[0073] The morphological verification layer is used to perform connected component analysis and real-time detection of generated features, which is represented as:

[0074]

[0075] In the formula, ψ circ is the circularity of the lesion area, A is the area of the lesion area, is the logical perimeter of the lesion.

[0076] Where the feature fusion is represented as:

[0077] F fusion =α·F low +(1-α)·F high ;

[0078] In the formula, F low is the low-level microaneurysm edge and hemorrhagic area texture feature, F high is the high-level lesion feature, F fusion is the fused multi-scale feature vector, and α is the circularity ψ circ of the lesion area, which is dynamically calculated.

[0079] When the pathological morphological structure is detected to be abnormal, the generator triggers feature rectification, which is adjusted and represented as:

[0080]

[0081] In the formula, is the shallow edge texture feature of the retinal image, l is the layer index in the generator, is the deep semantic lesion analogy feature of the retinal image, is the Laplacian operator, and is the element-wise multiplication, is the adjustment of the low-level feature and the high-level feature weight.

[0082] The overall image is captured using a spectral normalization convolution layer, and the similarity between the generated image and the features in the medical knowledge base is calculated within a sliding window, represented as:

[0083]

[0084] In the formula, is the feature of the retinal image at position (u, v), is the corresponding feature at position (u, v) in the medical knowledge base, (u, v) is the position index, D local (X p ) is the local feature of the retinal generated image, H·W is the height and width of the retinal image.

[0085] S2.4, based on the retinal generated image and the retinal image, a composite loss function is constructed, the difference between the feature distributions is measured by introducing the Wasserstein distance, and the texture feature distribution is constrained by combining the KL divergence.

[0086] where the composite loss function is constructed, the difference between the feature distributions is measured by the Wasserstein distance, and is expressed as:

[0087]

[0088] In the formula, L W is the difference between the retinal image distribution and the retinal generated image distribution, X r is the clinically collected retinal image, P r is the probability distribution of the retinal image, P g is the probability distribution of the retinal generated image, X g is the retinal pathological image synthesized by the adversarial network.

[0089] where the texture feature distribution is constrained by combining the KL divergence, and is expressed as:

[0090]

[0091] In the formula, L KL is the difference between the hemorrhagic area texture and the generated hemorrhagic area texture, p real (x) is the statistical distribution of the real hemorrhagic area texture, p ge n(c) is the statistical distribution of the generated hemorrhagic area texture, c is the number of texture categories.

[0092] NSGA-II algorithm (such as non-dominated sorting genetic algorithm-II) is used for multi-objective optimization, and the weights of each target loss are dynamically adjusted during the training process.

[0093]

[0094] In the formula, w(m) Weight distribution for the mth target, M is the total number of targets, m is the static target index R (m) Performance score for the mth target, η is the steepness of the control weight distribution, R (n) Performance score for the current training target, n is the dynamic target index.

[0095] S2.5, using differential equations to simulate the development of the disease, real-time blood glucose data (G(t) is included in the generation process, and a time series of retinal pathological images is generated.

[0096] Wherein, the differential equation is expressed as:

[0097]

[0098] In the formula, is the lesion diameter change rate, φ(τ) 2 is the simulation time step of lesion development, lesion diameter, ∈ is the random growth factor, τ is the lesion development time step, 0.025 is the blood glucose influence coefficient, and 0.0018 is the self-limiting coefficient.

[0099] The time series of pathological images is generated by numerical integration, and the retinal images of the lesion development process are output, and in the data enhancement process, spatial enhancement and time domain enhancement are combined,

[0100] Wherein, the spatial domain enhancement adopts random elastic deformation, which is expressed as:

[0101]

[0102] In the formula, is the generated retinal spatial deformation, which simulates the different positions and forms of the lesion area in the retinal image, σ deform is the displacement standard deviation of elastic deformation, Δx and Δy are independently sampled.

[0103] Wherein, the time domain enhancement is expressed as:

[0104] X interp = αX t +(1-α)X t+Δt ,α~U(0,1);

[0105] In the formula, X t is the state of retinal lesions at time t, X t+Δt is the state of retinal lesions at time t+Δt, α~U is a smoothing transition factor for mixing adjacent time steps, X interp is the intermediate state of the simulation of the continuous development of retinal lesions.

[0106] All the lesion images and enhanced images generated by the above method constitute the augmented training data set.

[0107] S3, training a deep convolutional neural network by expanding the training data set, using a multi-scale feature extraction layer to identify microaneurysms and hemorrhagic lesions in retinal images, and outputting an initial diagnosis result containing lesion classification,

[0108] S3.1, extracting enhanced retinal images from the expanded training data set, and dividing them into training, validation, and test sets. Each image is labeled with the location and type of microaneurysm and hemorrhagic lesion (mild, moderate, severe, etc.).

[0109] Each input image will be standardized to a fixed size and normalized to have a value range of [0, 1] for the input data. In this step, the RGB color space channels of the image will be converted to a format suitable for a multi-layer convolutional neural network.

[0110] S3.2, based on the deep convolutional neural network, introduce multi-scale feature extraction to extract features from retinal images. In the extraction process, small size convolution kernels (such as 3x3) are used for feature extraction. The multi-layer convolutional neural network extracts low-level features (such as edges, textures) from the retinal image in the early stage, and identifies microaneurysms and hemorrhagic lesions at different scales.

[0111] In order to enhance the recognition ability of the trained convolutional neural network for lesions of different sizes, cross-scale convolution layers (such as 1x1 convolution, 3x3 convolution, and 5x5 convolution) are used for feature fusion. The size of the feature map is reduced by the max-pooling layer to retain the feature information of the lesion area. The extracted lesion area features are mapped to the target classification label through multiple fully connected layers.

[0112] S3.3, use the cross-entropy loss function to optimize the output of the model, and use the Adam optimizer to optimize the network. Use the augmented retinal image as input, pass it through the deep convolutional neural network for forward propagation, get the network output, calculate the loss function, and adjust the network weights.

[0113] S3.4, the trained convolutional neural network outputs microaneurysms and hemorrhagic areas. In the output layer of the network, a multi-class softmax function is used to output the lesion type and severity level. The output of each lesion area is a three-class, corresponding to mild, moderate and severe.

[0114] S3.5, evaluate the trained convolutional neural network by cross-validation method, and adjust the structure of the network (adjust the learning rate and increase the regularization to reduce the overfitting phenomenon) according to the evaluation result.

[0115] After the convolutional neural network is trained, a new retinal image is input, and the model outputs the location and type of the lesion area and gives an initial diagnosis result (mild, moderate, severe) of the lesion grading.

[0116] S4. Combining real-time feedback correction data of clinicians to construct a reinforcement learning environment, dynamically adjusting the classification decision threshold by a sensitivity-specificity balance factor to generate optimized grading parameters;

[0117] S4.1, combined with real-time feedback of clinicians, define the doctor's correction action as a discrete reward signal, and input the initial diagnosis result of lesion grading to construct a time series clinical feedback data set.

[0118] Wherein, the definition of the doctor's correction action as a discrete reward signal is:

[0119]

[0120] In the formula, r k is the quantification of the doctor's decision-making behavior.

[0121] Wherein, and input the initial diagnosis result of lesion grading to construct a time series clinical feedback data set is:

[0122]

[0123] In the formula, is the standardized retinal image, is the initial diagnosis result of lesion grading, t clin is the time stamp.

[0124] Wherein, constructing a reinforcement learning environment is:

[0125]

[0126] In the formula, S t is the reinforcement learning state vector, P t is the output probability vector of the convolutional neural network, θ t is the current lesion classification threshold, is the reward mean value in the sliding window, is the input image lesion feature vector.

[0127] Wherein, the sensitivity-specificity balance factor is adopted, and the balance factor is defined as β t ∈[0,1] is:

[0128]

[0129] In the formula, k=2 is the sensitivity-specificity balance factor adjustment intensity, μ bmax is the long-term baseline of the sliding window mean value.

[0130] When is > μ bmax , β(t)→1, to promote sensitivity weight.

[0131] When is < μ bmax , β(t)→0, to promote specificity weight.

[0132] Wherein, the dynamic adjustment of the classification decision threshold is expressed as:

[0133] R t = β t ·Sen t +(1-β t )·Spe t -λ|θ t -θ t-1 |;

[0134] In the formula, Sen t is the sensitivity weight, Spe t is the specificity weight, and λ=0.1 is the threshold change penalty coefficient.

[0135] S4.3, dynamically adjust the classification decision threshold.

[0136] The Actor-Critic network is used to output the threshold adjustment amount and evaluate the reinforcement learning state vector, and the policy gradient is used for updating.

[0137] Wherein, the threshold adjustment amount output by the Actor network is expressed as:

[0138] a t = π(S t ; φ) = tanh(W2·ReLU(W1S t +b1)+b2)

[0139] In the formula, a t is the threshold adjustment action, φ is the Actor network parameter set, W1 is the input layer weight matrix, W2 is the hidden layer weight matrix, ReLU is the hidden layer nonlinear transformation, b2 is the hidden layer bias scalar, and tanh is the compression output to [-1, 1].

[0140] Wherein, the Critic network evaluation of the reinforcement learning state vector is expressed as:

[0141] V(S t ; ξ) = W c ·ReLU(W b S t +b b )+b c ;

[0142] where V(S t ) is the estimated reinforcement learning state value, ξ is the Critic network parameter set, W b is the input layer weight matrix, W c is the hidden layer weight matrix, b b is the input layer bias vector, and b c is the hidden layer bias scalar.

[0143] where and the policy gradient is updated as:

[0144]

[0145] where is the classification decision threshold policy gradient, J is the maximum long-term reward expectation, T is the time step window, is the reward value calculated by the dynamic balance factor, is the baseline reduction classification decision threshold policy gradient, and N is the batch size.

[0146] S4.4, generating optimized grading parameters.

[0147] Specifically, based on the results of dynamically adjusting the classification decision threshold, the NSGA-II algorithm is used to screen the threshold parameters with the best classification performance, the Pareto optimal solution generation method is used to quantify the sensitivity and specificity relationship under different thresholds on the validation set, and a trade-off curve is constructed.

[0148] Through dynamic-static parameter fusion, the dynamic policy parameters generated by reinforcement learning are integrated with the global optimal static threshold to generate deployable optimized grading parameters.

[0149] S5 loads the optimized grading parameters, combines the retinal image lesion features and patient biochemical time series data, and generates a three-level classification result of diabetic retinopathy through cross-modal correlation analysis.

[0150] S5.1, based on the generated optimized grading parameters (selected by the NSGA-II algorithm), the retinal image lesion features (microaneurysm morphology, hemorrhagic texture) and patient biochemical time series data (blood glucose concentration, glycosylated hemoglobin) are input into cross-modal fusion, and the spatiotemporal correlation features are extracted through a time convolution kernel to generate retinal lesion fusion features.

[0151] S5.2, for the retinal lesion fusion features, a deep convolutional neural network (DCNN) is used to output three-class classification results (mild, moderate, and severe).

[0152] The embodiment also provides a computer device suitable for the diabetic nephropathy patient retinopathy image classification method based on artificial intelligence, which comprises a memory and a processor.

[0153] The computer device can be a terminal, and the computer device comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved through WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse.

[0154] The embodiment also provides a storage medium having a computer program stored thereon, and the computer program is configured to be executed by a processor to implement the diabetic nephropathy patient retinopathy image classification method based on artificial intelligence. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0155] To sum up, the application generates a retinal image with enhanced pathological characteristics by using a generative adversarial network (GAN) guided by medical prior knowledge, combining a microaneurysm morphological parameter library with a hemorrhagic area texture feature atlas, and constructing an augmented training data set to alleviate the problem of scarce training data. Secondly, based on the augmented data set, a deep convolutional neural network is trained to improve the recognition ability of micro lesions by using multi-scale feature extraction layers and to reduce misdiagnosis and missed diagnosis. In addition, a reinforcement learning mechanism is used to dynamically adjust the classification decision threshold through a sensitivity-specificity balance factor and to optimize the grading parameters combined with real-time feedback from clinicians to enhance the reliability and explainability of classification. Finally, cross-modal correlation analysis is used to fuse retinal lesion features and patient biochemical time series data to further improve the accurate classification ability of diabetic retinopathy, providing a scientific basis for early diagnosis and personalized treatment of diabetic nephropathy patients.

[0156] It should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. An artificial intelligence-based image classification method for retinopathy in diabetic nephropathy patients, characterized by: include, Acquire retinal images and perform noise reduction and contrast adjustment to generate standardized images; Standardized images are input into a generative adversarial network guided by medical prior knowledge. Feature constraints are performed using a microaneurysm morphology parameter library and a texture feature map of the hemorrhage area to generate retinal images with enhanced pathological features and construct an augmented training dataset. A deep convolutional neural network is trained based on an augmented training dataset. A multi-scale feature extraction layer is used to identify microaneurysms and hemorrhagic lesions in retinal images with enhanced pathological features, and the output is an initial diagnostic result including lesion grading. Based on the initial diagnosis results, a reinforcement learning environment is constructed by combining real-time feedback and correction data from clinicians. A sensitivity-specificity balance factor is used to dynamically adjust the classification decision threshold and generate optimized grading parameters. By loading and optimizing grading parameters, and combining retinal image lesion features with patient biochemical time-series data, a three-level retinal image classification result for diabetic retinopathy is generated through cross-modal association analysis.

2. The image classification method for retinopathy in diabetic nephropathy patients based on artificial intelligence as described in claim 1, characterized in that: The aforementioned noise reduction and contrast adjustment refer to Noise is added to the retinal images using Gaussian noise to generate noisy images. The original retinal images are then paired with the corresponding noisy retinal images to generate a retinal image dataset. U-shaped convolutional neural networks were used for denoising training to optimize and minimize the pixel difference between the denoised retinal image and the original retinal image. Based on the CLAHE algorithm, the denoised retinal image is segmented into small blocks, each block is processed by histogram equalization, the local contrast is adjusted, the processing results of the sub-regions are merged, and the contrast-equalized retinal image is output.

3. The image classification method for retinopathy in diabetic nephropathy patients based on artificial intelligence as described in claim 1, characterized in that: The generative adversarial network refers to, Edge detection is performed on standardized retinal images. The edges of microaneurysms are extracted at the pixel level. Iso-sampling points of the microaneurysm lesion contour are extracted. The shape features describing the low-frequency components are preserved by discrete Fourier transform, and the shape descriptor is calculated. For the texture of the bleeding region, a rotation-invariant local binary pattern is calculated using a ring neighborhood, and a differentiable orientation coding layer is used to capture multi-directional texture features; The generator based on the morphological features of microaneurysms and the texture features of the hemorrhage area adopts a dual-path feature fusion architecture. The backbone network uses a pre-trained EfficientNet-B4 model, and the generative adversarial network is used.

4. The image classification method for retinopathy in diabetic nephropathy patients based on artificial intelligence as described in claim 1, characterized in that: The multi-scale feature extraction layer includes, Three convolutional kernels are configured in parallel, and lesion features at different scales are dynamically fused through a channel attention mechanism.

5. The image classification method for retinopathy in diabetic nephropathy patients based on artificial intelligence as described in claim 1, characterized in that: The sensitivity-specificity balance factor includes, The threshold is dynamically adjusted based on the sliding window mean of the doctor's feedback reward. The threshold adjustment amount is output by the reinforcement learning Actor-Critic network and the reinforcement learning state vector is evaluated to adjust the classification decision threshold.

6. The image classification method for retinopathy in diabetic nephropathy patients based on artificial intelligence as described in claim 1, characterized in that: The constructed augmented training dataset includes, Generative adversarial networks guided by medical prior knowledge synthesize retinal images with enhanced pathological features. Spatial domain elastic deformation processing is performed on the synthetic image, and the deformation parameters are generated as random displacements based on Gaussian distribution to simulate the position and morphological changes of the lesion area. Temporal interpolation enhancement is performed on time-series pathological images. Intermediate states are generated by randomly linearly mixing images from adjacent time steps to construct and augment the training dataset.

7. The image classification method for retinopathy in diabetic nephropathy patients based on artificial intelligence as described in claim 1, characterized in that: Generating retinal images with enhanced pathological features includes, Real-time verification is performed using standard lesion retinal images from a medical knowledge base; when the similarity between the microaneurysm morphology or hemorrhage area texture of the generated image and the standard template is different, a dynamic correction mechanism is activated to adjust the dual-path feature fusion weights of the generator and generate an enhanced retinal image.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the artificial intelligence-based image classification method for diabetic nephropathy patients as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the image classification method for retinopathy in diabetic nephropathy patients based on artificial intelligence as described in any one of claims 1 to 7.

Images