Eye fundus image classification method and device, equipment and storage medium

Abstract

The invention relates to the technical field of image recognition, and discloses a fundus image classification method, device and equipment and a storage medium, and the method comprises the steps: constructing a dual-resolution input tensor, extracting a structural feature graph representing anatomical topology through a structural perception channel, and calculating the local Shannon entropy of the structural feature graph to generate a texture complexity mapping graph; generating a basic suppression mask based on the structural feature map, and generating an exemption compensation mask based on the texture complexity mapping map; according to the method, feature orthogonal correction operation is executed in a classification feature channel, background feature response is attenuated through a basic suppression mask, meanwhile, high-entropy features are reinjected through an exemption compensation mask, and the classification probability is output based on a corrected focus residual feature map. The technical problem that weak lesion features are submerged in a strong structure background is solved, and mimicry lesion features are accurately reserved while normal anatomical structure interference is inhibited.

Description

Technical Field

[0001] This invention relates to a method, apparatus, device, and storage medium for classifying fundus images, belonging to the field of image recognition technology. Background Technology

[0002] Current methods for classifying fundus images utilize algorithms to analyze digital retinal images to identify diabetic retinopathy, glaucoma, and other fundus diseases. Mainstream deep convolutional neural networks extract semantic information representing lesion features from the image and output classification results based on feature response intensity, typically relying on high-frequency, strong features for rapid decision-making. Convolutional neural networks processing fundus images generally assume that the network automatically focuses on key classification regions. However, the statistical distribution of pixels in fundus images exhibits a significant inversion of features. Normal anatomical structures such as the optic disc and main vascular network have high-contrast edge gradients, occupying the majority of high-frequency information in the image. Early exudative lesions of microaneurysms are key lesion features represented by tiny pixel clusters with weak gradient responses. Convolutional kernel weight updates tend to respond to high-gradient, large-area normal structural features, reducing the loss function. This leads to the hierarchical transmission of features from small lesions being suppressed and averaged by strong structural features. Industry solutions introduce attention mechanisms or use ultra-high-resolution inputs to enhance weak features, increasing model computational complexity and memory consumption, making it difficult to achieve real-time inference with portable fundus camera screening equipment limited by computing power.

[0003] In complex cases of proliferative diabetic retinopathy, neovascularization grows attached to the main vessel wall of the optic disc, exhibiting a topological similarity to normal vessels. Morphological feature extraction methods struggle to distinguish between the ordered tubular structures of normal vessels and the disordered tubular structures of neovascularization. This can lead to misjudgment of lesion features that overlap with normal structures, resulting in background noise suppression and missed detection of severe lesions. Simply pursuing the detection of minute lesions sacrifices global structural information, ignoring the geometric characteristics of vascular tortuosity and diameter ratio that reflect the overall health of the fundus, thus limiting the accuracy of algorithms in multi-disease comprehensive diagnosis and lesion grading. The challenges faced by the aforementioned deep learning models mainly lie in how to accurately and efficiently separate the contradictory features of target lesions from non-target anatomical structures. For lesion types with significant morphological differences, existing technologies attempt to seek breakthroughs from the perspectives of traditional machine learning and quantitative feature analysis, but limitations remain. For example, [the following text appears to be a separate, unrelated section:] [Publication number] Chinese invention patent application CN116524257A discloses a method and apparatus for classifying fundus images and an image processing method. This method extracts suspicious lesion regions based on a preset color threshold and uses traditional machine learning classification models to determine whether the image contains lesion regions of myelinated retinal nerve fibers based on the first area, gray level, shape and location of the region. Although this method has a classification-assisted diagnostic effect for the specific lesion type of myelinated retinal nerve fibers, it is prone to failure when dealing with complex micro-mimicking lesions such as early microaneurysms or neovascularizations overlapping with blood vessels. The pattern of relying on fixed thresholds and predefined features is prone to failure. The feature engineering-driven classification method cannot adaptively capture the discriminative contextual semantics and texture disorder learned by deep networks, and the preset feature extraction process does not solve the problems of feature submersion and over-suppression in multi-scale and strongly structured backgrounds.

[0004] Therefore, how to suppress the interference of normal anatomical structures to highlight small lesions while accurately preserving the characteristics of overlapping complex lesions, and integrating global geometric information to improve comprehensive diagnostic capabilities, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A fundus image classification method, comprising the following steps:

[0006] Step 101: Obtain fundus image data to be processed, perform multi-scale sampling operation on fundus image data, and construct a first resolution input tensor and a second resolution input tensor. The spatial dimension of the first resolution is lower than that of the second resolution.

[0007] Step 102: Input the first resolution input tensor into the preset structure perception channel, and extract the structural feature map representing the anatomical topology of the fundus through convolution operation;

[0008] Step 103: Perform a sliding window scan on the structural feature map, calculate the local Shannon entropy of the feature response values ​​within each window, and generate a texture complexity map;

[0009] Step 104: Generate a basic inhibition mask for the anatomical structure based on the structural feature map, and extract regions in the texture complexity map whose values ​​are higher than a preset complexity threshold to generate an exemption compensation mask.

[0010] Step 105: Input the second resolution input tensor into the classification feature channel to extract the original feature map, and perform feature orthogonal correction operation; the feature orthogonal correction operation includes: calculating the Hadamard product of the original feature map and the basic suppression mask to generate preliminary suppression features, and calculating the weighted product of the original feature map and the exemption compensation mask, and superimposing the weighted product onto the preliminary suppression features to generate lesion residual feature map;

[0011] Step 106: Input the residual feature map of the lesion into the fully connected classifier, calculate and output the classification probability vector of the fundus image data.

[0012] Preferably, step 104, which generates the basic suppression mask for anatomical structures, includes: step 201, performing spatial upsampling on the structural feature map to make its spatial size consistent with the original feature map; step 202, using a non-linear activation function to map the upsampled structural feature map into a probability distribution map with values ​​between zero and one; step 203, calculating the difference matrix between the values ​​and the probability distribution map, defining the difference matrix as the basic suppression mask, where the numerical values ​​of the elements in the basic suppression mask represent the retention weights for corresponding pixel positions belonging to non-anatomical background structures. The calculation of the local Shannon entropy of the feature response values ​​within each window in step 103 is performed based on the following formula: Where H(x) is the local Shannon entropy, n is the total number of pixels within the sliding window, and x i For the first in the sliding window The feature response grayscale value of each pixel, p(x) i ) represents the probability of the gray value appearing within the sliding window; local Shannon entropy is used to quantify the statistical characteristics of the texture disorder of the structure feature map within a local region.

[0013] Preferably, the orthogonal correction operation performed in step 105 follows the following logical gating rules: for any target pixel position in the original feature map, if the corresponding value of the position in the basic suppression mask is lower than the preset suppression threshold and the corresponding value in the exemption compensation mask is zero, then the feature response value of the position is set to zero or attenuated; if the corresponding value of the position in the exemption compensation mask is non-zero, then the original feature response value of the position is retained or a gain operation is performed on it. The structure-aware channel contains a convolutional layer, and the parameters of the convolutional kernel in the convolutional layer are set during the initialization phase to have a gradient directional derivative response to tubular and circular topologies. The structural feature map is output by the structure-aware channel, which contains the spatial location probability distribution data of the optic disc region and the retinal main vascular network region.

[0014] Preferably, the method further includes a process for global geometric semantic-assisted classification using features extracted from the structure-aware channel. The process includes: step 601, performing a global average pooling operation on the deep feature tensor used to generate the structural feature map in the structure-aware channel to convert the three-dimensional feature tensor into a one-dimensional global structure description vector; step 602, performing a global pooling operation on the lesion residual feature map to generate a lesion feature vector; step 603, performing concatenation processing on the global structure description vector and the lesion feature vector in the channel dimension to construct a joint feature vector; and step 604, inputting the joint feature vector into a fully connected classifier.

[0015] Preferably, in step 105, the operation of performing element-wise feature response attenuation on the original feature map using the basic suppression mask is based on the negative correlation mapping relationship between the value of each element in the basic suppression mask and the feature response intensity at the corresponding position in the structural feature map. The negative correlation mapping relationship maps the high-response anatomical structural features to weight coefficients close to zero. The width and height dimensions of the first resolution input tensor are one-quarter of the width and height dimensions of the second resolution input tensor, respectively. The structure-aware channel and the classification feature channel are independent of each other in the network architecture. There is only a unidirectional data flow injection during the feature orthogonal correction operation stage. The structure-aware channel does not receive the reverse gradient propagation from the classification feature channel.

[0016] Preferably, the global structure description vector includes geometric topological feature data that characterizes the overall tortuosity and branch density of the fundus vascular network; the splicing process in step 603 uses the geometric topological feature data as a global context constraint and jointly encodes it with the local texture abnormality features in the lesion feature vector.

[0017] Preferably, the classification feature channel includes a deep convolutional neural network; the classification probability vector output in step 106 is generated by a fully connected classifier through a Softmax function mapping, and each dimension of the classification probability vector corresponds to a confidence value of a preset retinal disease grading category.

[0018] A fundus image classification device includes: a multi-scale sampling construction module for acquiring fundus image data to be processed, performing multi-scale sampling operations on the fundus image data, and constructing a first-resolution input tensor and a second-resolution input tensor respectively, wherein the spatial dimension of the first resolution is lower than that of the second resolution; a structural feature extraction module for inputting the first-resolution input tensor into a preset structural perception channel, and extracting structural feature maps representing the inherent anatomical topology of the fundus through convolution operations; a texture entropy calculation module for performing local sliding window scanning on the structural feature maps, calculating the local Shannon entropy of the feature response values ​​within each window, and generating a texture complexity mapping map; and a mask generation module for... The system generates a basic suppression mask for anatomical structures based on structural feature maps and extracts regions with values ​​higher than a preset complexity threshold from the texture complexity map to generate an exemption compensation mask. A feature orthogonal correction module is used to input a second-resolution input tensor into the classification feature channel to extract the original feature map and perform feature orthogonal correction operations. These operations include performing element-wise feature response attenuation on the original feature map using the basic suppression mask and performing weighted back-injection on the attenuated feature map using the exemption compensation mask to generate a lesion residual feature map. A classification decision output module is used to input the lesion residual feature map into a fully connected classifier, calculate and output the classification probability vector of the fundus image data.

[0019] A fundus image classification device includes a processor and a memory, which are communicatively connected to the processor for storing computer programs; the processor is used to execute the fundus image classification method when running the computer programs.

[0020] 10. A storage medium for classifying fundus images, wherein a computer program is stored thereon, characterized in that the computer program, when executed by a processor, implements the aforementioned fundus image classification method.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. In fundus image classification, the image classification feature extraction process sets up a structure perception channel and a classification feature channel in parallel. Utilizing the physical characteristics of the high-frequency gradient response of the physiological structure of the optic disc and main vascular network, the structure perception channel extracts the probability distribution map representing the inherent anatomical structure and generates an inverse inhibition mask. The mask and the original feature map of the classification channel are weighted element by element. High gradient background signals are physically filtered out in the feature space. Based on the anatomical prior subtraction logic, the feature response of low-frequency weak lesions such as microaneurysms or early exudations is relatively enhanced under the condition of reduced background noise. Without increasing the sensor hardware cost or light intensity, the algorithm improves the signal-to-noise ratio and detection sensitivity of small lesion targets.

[0023] 2. To address the challenge of misjudgment caused by the high similarity between the topological morphology of neovascularization and normal blood vessels in proliferative retinal lesions, a local sliding window scan is performed on the structural feature map during the suppression operation to calculate texture entropy. Utilizing the statistical difference between the low-entropy ordered texture of laminar growth in normal blood vessels and the high-entropy disordered texture of turbulent growth in neovascularization, high-entropy regions are automatically identified and locked to generate exemption masks. The original feature components corresponding to the exemption masks are then reinjected into the suppressed feature map. This achieves the calibration of deep learning morphological judgment logic using parametric statistical features. This mechanism ensures that the algorithm suppresses interference from normal physiological structures while retaining information on overlapping lesions with complex texture features, thus ensuring the completeness of features for complex pathological staging tasks.

[0024] 3. This invention overcomes the limitations of traditional mask-generated branches, which only assist in spatial positioning. It utilizes deep feature tensors extracted from the structure perception channel in multiple dimensions. Spatially, tensor generation suppresses masking to eliminate line-of-sight obstruction. Semantically, global pooling eliminates spatial location information, collapsing the tensor into a global description vector representing macroscopic geometric attributes such as vascular tortuosity and branch density. This global description vector is concatenated with the lesion feature vector extracted by the classification channel and the decision layer channel. This allows the classifier to output results based on the joint probability distribution of the presence of microscopic lesions and abnormal vascular morphology. The processing method does not increase the computational overhead of the additional feature extraction backbone network. By integrating local pathological features with global anatomical context, it improves the comprehensive diagnostic capability of image classification methods for systemic diseases such as hypertensive retinopathy. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the feature orthogonal correction classification process based on dual-resolution input according to the present invention;

[0026] Figure 2 A comparison chart of the diagnostic accuracy of multiple diseases using global geometric semantic features introduced in this invention;

[0027] Figure 3 This is the timing logic diagram generated for structural feature extraction and texture complexity mapping in this invention. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0029] This invention provides a method, apparatus, device, and storage medium for classifying fundus images. It classifies fundus images using parallel structure-aware channels and classification feature channels, combined with texture entropy features, to obtain fundus image data to be processed. Multi-scale sampling is performed on this fundus image data to construct a first-resolution input tensor and a second-resolution input tensor, where the spatial dimension of the first resolution is lower than that of the second resolution. In this embodiment, the original fundus image is defined as the second-resolution input tensor with a size of H×W. A bilinear interpolation downsampling algorithm is used to reduce the spatial size of the original image to H / 4×W / 4 to construct the first-resolution input tensor. It employs asymmetric resolution input, utilizing low-resolution information to preserve large-scale topological connectivity features of the vascular network and optic disc, while reducing the computational load of the structure-aware channel. The structure-aware channel and the classification feature channel operate in parallel with independent weight sets. Channel connections are only injected with unidirectional data flow during the feature orthogonality correction operation stage. To ensure the independence of the structural feature map as a non-learnable geometric prior, a gradient blocking instruction is executed on the computation graph after the mask data flows out of the structure-aware channel. The tf.stop_gradient operator or hardware no-operation instruction is used to prevent the back gradient propagation of the classification feature channel from being written into the weights of the structure-aware channel.

[0030] The first resolution input tensor is input into a preset structure-aware channel, which is composed of a convolutional neural network. Through convolution operations, it extracts structural feature maps representing the anatomical topology of the fundus. The convolution kernel parameters of the convolutional layers within this channel are set during initialization to have gradient directional derivative responses to tubular and circular topologies. Specifically, the convolution kernels are initialized using parameter distribution to ensure Gaussian second-order derivative responses to linear structures at 0°, 45°, 90°, and 135°. The structural feature map output by the structure-aware channel contains the spatial location probability distribution data of the optic disc region and the retinal main vascular network region. Local texture complexity analysis is performed on the structural feature map. A sliding window scan is performed on the structural feature map, calculating the local Shannon entropy of the feature response values ​​within each window to generate a texture complexity mapping map. For any pixel in the structural feature map... The local Shannon entropy H(x) of the centered local window is calculated based on the following formula: Where H(x) is the local Shannon entropy, n is the total number of pixels within the sliding window, and x i For the first in the sliding window The feature response grayscale value of each pixel, p(x) i The probability of the gray value appearing within the sliding window is represented by the local Shannon entropy quantification structure feature map, which is a statistical feature of the texture disorder in the local region. The difference between the laminar smooth texture of normal retinal blood vessels and the turbulent chaotic texture of neovascularization is used to distinguish them in the statistical dimension.

[0031] Based on the structural feature map and texture complexity map, a basic suppression mask and an exemption compensation mask are generated. For the basic suppression mask, spatial upsampling is performed on the structural feature map to make its spatial size consistent with the original feature map (the feature map corresponding to the second resolution input tensor). A non-linear activation function is used to map the upsampled structural feature map into a probability distribution map with values ​​between zero and one. The difference matrix between the values ​​and this probability distribution map is calculated, and this difference matrix is ​​defined as the basic suppression mask. The numerical values ​​of the elements in the basic suppression mask represent the retention weights of the corresponding pixel positions belonging to non-anatomical background structures. For the exemption compensation mask, regions with values ​​higher than a preset complexity threshold in the texture complexity map are extracted to generate the exemption compensation mask. The second resolution input tensor is input into the classification feature channel to extract the original feature map. The feature map is generated, and a feature orthogonal correction operation is performed. This operation includes calculating the Hadamard product of the original feature map and the basic suppression mask to generate preliminary suppression features, and using the weight coefficients of the basic suppression mask to attenuate the feature response of the corresponding anatomical structure region in the original feature map; calculating the weighted product of the original feature map and the exemption compensation mask, and superimposing the weighted product onto the preliminary suppression features to generate the lesion residual feature map. This correction operation follows the following logical gating rules: for any target pixel position in the original feature map, if the corresponding value of the position in the basic suppression mask is lower than the preset suppression threshold and the corresponding value in the exemption compensation mask is zero, the feature response value of the position is set to zero or attenuated; if the corresponding value of the position in the exemption compensation mask is non-zero, the original feature response value of the position is retained or a gain operation is performed on it.

[0032] Basic suppression mask M inh Strictly defined as Where σ is the nonlinear activation function and U is the upsampling operation F struct It is a structural feature map, and the orthogonal correction operation of features follows the formula. , where F orig It is the original feature map F res The residual feature map of the lesion is shown. The Hadamard product β is the compensation gain coefficient, calibrated to 1.2 to ensure the relative gain of the exempted compensation feature. The suppression threshold α in the logic gating rule is calibrated to 0.7. When M... inh Element values ​​are lower than α and M comp When the corresponding value is 0, F will be... resThe feature response at this location is directly set to zero to achieve maximum background filtering. In the classification decision stage, global geometric semantic-assisted classification is performed. Global average pooling is performed on the deep feature tensor used to generate the structural feature map in the structure perception channel to convert the three-dimensional feature tensor into a one-dimensional global structural description vector. This global structural description vector contains geometric topological feature data that characterizes the overall tortuosity and branch density of the fundus vascular network. Global pooling is performed on the lesion residual feature map to generate the lesion feature vector. The global structural description vector and the lesion feature vector are concatenated in the channel dimension to construct a joint feature vector. The joint feature vector is input into a fully connected classifier. Through the Softmax function mapping, the classification probability vector of the fundus image data is calculated and output. Each dimension of this classification probability vector corresponds to the confidence value of a preset retinal lesion grading category.

[0033] Example 1: In the application of fundus screening in grassroots communities, the portable fundus camera faces the situation where retinal images are accompanied by high-frequency vascular structure interference and low-frequency uneven illumination due to hardware limitations and light fluctuations, and the gray-scale gradient of microaneurysms is easily masked by the strong edge response of the main vascular network. The classification method provided by this invention uses a dual-resolution parallel processing mechanism to start feature extraction. The structure-aware channel processes the first resolution input tensor and uses a convolution kernel initialized with a Gabor filter to lock the anatomical structure of the optic disc and the main vascular network, outputting spatial location probability distribution data representing background interference. At the same time, the classification feature channel processes the second resolution input tensor to analyze the details of small lesions.

[0034] The orthogonal feature correction operation, as a key interface connecting the two channels, transforms the structural feature map generated from the structure-aware channel through upsampling and nonlinear mapping into a basic suppression mask. The original feature map in the classification feature channel is then weighted point-by-point using Hadamard product operations to attenuate the feature response values ​​of the optic disc and large vessel regions, thereby improving the relative significance of microaneurysm features in terms of signal-to-noise ratio. For proliferative neovascularization, the system simultaneously calculates the local Shannon entropy of the structural feature map. When a region is detected that has vascular morphology but its local entropy value exceeds a preset complexity threshold, it is identified as a high-entropy abnormal region, and an exemption compensation mask is generated to retain the original feature response of that region in the classification feature channel, preventing erroneous deletion due to morphological suppression. The lesion residual feature map, while reducing background noise, retains the feature information of micro-lesions and complex neovascularization. Through a global geometric semantic-assisted classification process, the global topological features of vascular tortuosity implicit in the structure-aware channel are encoded into a vector and concatenated with the lesion residual feature vector. The fully connected classifier outputs the lesion grading result based on this joint feature vector containing micro-lesions and macro-structures.

[0035] Example 2: In the technical verification experiment of the primary care fundus lesion auxiliary diagnosis system, to address the technical challenge of low recognition rate of small lesions in the context of strong anatomical structures, a verification platform based on public fundus image datasets such as EyePACS was constructed. This platform simulated a real primary care screening environment, introducing low-resolution input and illumination perturbations. The experimental platform was configured with an edge computing simulation environment with limited computing power to verify the algorithm's operating efficiency under low-resource conditions. The core parameter settings followed a strict engineering decision-making logic chain: For the input resolution of the structure perception channel, based on the average diameter of the optic disc and main vascular network (approximately 150 to 200 micrometers) and the physical resolution of the image sensor, a balance was struck between computational load and structural feature preservation, setting the downsampling factor to 4 times (i.e., resolution of 1 / 4 of the original image) to ensure that key topological connectivity is preserved while reducing GFLOPs; For the sliding window size for texture entropy calculation, based on the average width of normal retinal vessels and the typical size of microaneurysms (30 to 100 micrometers), the window size was set to... Pixel.

[0036] After the experiment was launched, experimental groups with a multi-dimensional control system were constructed: the experimental group of this invention adopted a complete dual-stream feature orthogonal correction model; control group 1 removed the structure-aware channel and orthogonal suppression module, retaining only the classification feature channel; control group 2 retained the structure-aware channel but removed the texture entropy compensation mechanism. For image data, Gaussian white noise (signal-to-noise ratio 20dB) and an uneven illumination mask were actively superimposed to simulate low-quality imaging conditions. During the experiment, the performance of each group in terms of microaneurysm detection rate, specificity, and overall classification accuracy was monitored. Data showed that in control group 1 without suppression, the micro lesion feature response was submerged in the high-frequency signal of the strong vascular background, resulting in a low detection rate; while the experimental group of this invention showed a significant background silencing effect in the feature space, the vascular region response was significantly attenuated, and the relative intensity of micro lesion features was increased by more than 30%. To quantify and verify the above observations, Table 1 presents the key performance indicators of different experimental groups on the test set.

[0037] Table 1: Comparison of Key Performance Indicators for Different Experimental Groups

[0038]

[0039] A detailed analysis of the data in Table 1 reveals that while control group 2 improved specificity through structural inhibition (from 88.5% to 96.2%), the detection rate improvement was limited (78.4%). Furthermore, retrospective analysis showed that missed detections were mostly concentrated in the neovascularization area around the optic disc, confirming that simple morphological inhibition carries a risk of false negatives. In contrast, the detection rate of the sample group in this invention jumped to 89.7% after introducing texture entropy compensation, while maintaining extremely high specificity (95.8%), and the computational cost increased only slightly by 0.1 GGFLOPs.

[0040] Example 3: This example combines Figures 1 to 3 This describes a method, apparatus, device, and storage medium for classifying fundus images, such as... Figure 1 As shown, the fundus image data to be processed is obtained as the raw input. A multi-scale sampling module divides the data into a high-spatial-dimensional second-resolution input tensor that retains the original size and a low-spatial-dimensional first-resolution input tensor with a size reduced to H / 4×W / 4. The data is processed through a parallel dual-channel architecture. The classification feature channel uses a deep convolutional neural network to extract the original feature map containing background interference from the second-resolution input tensor, while the structure-aware channel uses convolutional kernels initialized with directional derivative responses to extract structural feature maps representing the probability distribution of anatomical structures from the first-resolution input tensor. Figure 1 On the one hand, a basic suppression mask representing the background and retaining weights is generated through upsampling and nonlinear mapping. On the other hand, a sliding window scan is performed through the texture entropy calculation module to calculate the local Shannon entropy, thereby generating an exemption compensation mask that marks the high-entropy lesion features. Finally, in the feature orthogonal correction module, the original feature map is subjected to Hadamard product suppression using the basic suppression mask according to the logic gating rules, and weighted backinjection is performed using the exemption compensation mask. The generated lesion residual feature map is processed by a fully connected classifier and a Softmax function, and the output is a classification probability vector containing the lesion confidence value.

[0041] like Figure 2 As shown, the horizontal axis covers four disease types: diabetic retinopathy, hypertensive retinopathy, glaucoma, and age-related macular degeneration. The vertical axis represents the diagnostic accuracy as a percentage. The graph contains two trend lines: the dashed line represents the baseline model performance without incorporating global structural features, and the solid line represents the model performance after incorporating global structural features such as vascular tortuosity and branch density. The data point distribution shows that, across all four disease types, the diagnostic accuracy indicated by the solid line is higher than that of the dashed line. Figure 3 As shown, the first resolution input tensor is input into the convolutional layer. Convolution is performed using the directional derivative response parameters set to have gradients for tubular and circular topologies during the initialization phase. This generates a structural feature map containing the spatial location probabilities of the optic disc region and the main vascular network region, which is then passed to the sliding window scanner. The system traverses all window positions to locate local window regions and extract feature response values. After the data flows to the entropy calculation module, the local Shannon entropy is calculated according to the formula to quantify the texture disorder. The calculated entropy value is written to the corresponding position to complete the construction of the texture complexity mapping map.

[0042] Example 4: In an engineering scenario used to validate the calibration procedure for core parameters of a classifier, the technical team needed to address the issue of how to scientifically set the inhibition coefficient α in the structure inhibition gating unit. This coefficient directly determines the intensity of inhibition on anatomical structures. Its setting involves a crucial technical trade-off between background noise filtering and information loss due to excessive suppression. If the α value is too low, the high-frequency signal residue from background blood vessels will result in insufficient improvement in the signal-to-noise ratio of small lesions; if the α value is too high, it may incorrectly suppress lesion features due to structural flow prediction errors. To address this contradiction, a standardized parameter calibration procedure based on the statistical distribution of feature responses was developed. A calibration dataset containing normal fundus images and typical lesion images (microaneurysms, exudates) was constructed. For each image, the structural feature map F was calculated. struct Compared with the original classification feature map F cls Define an objective function J(α) to quantify the balance between suppression effect and information retention. This function consists of two parts: the first part is the background region, i.e., F. struct High response region feature map after suppression The average response intensity in the sample is intended to minimize background interference; the second part is the lesion area in the sample. The feature contrast in the model aims to maximize the significance of lesions by performing gradient scanning of α with a step size of 0.1 in the interval [0,1] and calculating the target function value corresponding to each α value.

[0043] Experimental results show that when α increases from 0.1 to 0.6, the average response intensity of the background area decreases approximately linearly, and the lesion contrast improves. However, when α exceeds 0.8, some small lesions located at the edge of blood vessels begin to show nonlinear decay, resulting in a sharp decrease in lesion contrast. This indicates that the region has entered the over-suppression range. Based on this inverted U-shaped performance curve, the optimal range for the objective function J(α) to reach its extreme value is determined to be [0.6, 0.8]. Finally, according to this procedure, the default value of the inhibition coefficient α is set to 0.7.

[0044] Example 5: In a standardized engineering calibration scenario in a newly deployed environment, to address the image baseline shift caused by batch differences in fundus camera sensors and variations in lighting conditions at the medical center, a pre-deployment calibration process needs to be performed. This involves acquiring a set of calibration image sequences containing a standard grayscale card and typical fundus structures to quantify the lighting distribution characteristics and sensor noise levels of the current imaging environment. By analyzing the calibration image sequence, the background illumination field model and noise variance parameters are extracted and used as initialization parameters input to the structure perception channel. The preset convolution kernel weights are then fine-tuned. Using a built-in self-testing program, the calibration images are automatically classified and tested to verify whether the system's feature extraction response to the standard target meets the expected baseline range under the current parameter settings. If the response deviation exceeds a preset threshold, an iterative parameter calibration mechanism is automatically triggered until the system output stabilizes within the specified tolerance range.

[0045] In addition, for the setting of the core threshold in the texture entropy calculation module, an offline calibration and data filling procedure was constructed. A large-scale fundus image dataset covering various lesion types and severity was collected, and senior ophthalmologists performed fine annotation on the lesion areas and normal structures in the dataset. A benchmark truth library was constructed, and the local Shannon entropy distribution of the structural feature map was statistically analyzed. Entropy probability density curves of normal blood vessel texture and neovascular texture were plotted. By calculating the intersection of the two curves, the optimal entropy threshold that minimizes the classification error rate was determined and used as the default threshold parameter in the texture complexity mapping process. Based on this threshold, a texture feature lookup table was constructed to map different entropy value intervals to the corresponding lesion detection probabilities. When the system was deployed, the lookup table was solidified into the inference engine to support rapid decision-making in the real-time inference process. This procedure ensures that the setting of the texture entropy threshold is based on objective big data statistical laws, rather than subjective experience.

[0046] Example 6: In the technical verification and procedural solidification scenario of a large-scale fundus screening system, it is necessary to ensure the reproducibility and stability of the fundus image classification method under cross-device and cross-environment conditions. For the initialization of convolution kernel parameters in the structure perception channel, a standardized offline calibration procedure is constructed. Using a public dataset containing typical fundus structures such as optic discs, blood vessels, microaneurysms, and exudations, the frequency, direction, and scale parameters of the Gabor filter bank are iteratively optimized using the gradient descent algorithm to maximize its edge response intensity to tubular and circular topological structures. Once the optimal parameter set is determined, it is solidified as the system's default initialization configuration to ensure the determinism of the feature extraction starting point. For the inhibition coefficient α and entropy compensation gain β involved in the feature orthogonal correction operation, an adaptive parameter calibration process based on statistical distribution is established. A set of calibration images covering different lighting conditions and lesion degrees is collected, and their structural feature maps and texture complexity maps are calculated. By analyzing the distribution differences between normal blood vessels and neovascularization in texture entropy space and feature response intensity, the optimal decision boundary for distinguishing between the two is determined, and then the baseline values ​​of α and β are quantified.

[0047] In addition, to address potential sensor noise and uneven illumination during actual deployment, a pre-calibration procedure was designed. During system startup, a set of standard grayscale images is automatically acquired, and the illumination distribution model and noise variance of the current environment are calculated. If the environmental parameters are detected to deviate from the preset benchmark by more than the tolerance threshold, the system will automatically trigger the parameter fine-tuning mechanism, preprocess and compensate the input tensor according to the illumination model, and adjust the generation logic of the suppression mask accordingly.

[0048] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fundus image classification method characterized by, The method comprises the following steps: Step 101, obtaining eye fundus image data to be processed, performing multi-scale sampling operation on the eye fundus image data, constructing a first resolution input tensor and a second resolution input tensor, and the spatial dimension of the first resolution is lower than that of the second resolution; Step 102, inputting the first resolution input tensor into a preset structure perception channel, and extracting a structure feature map representing the anatomical topology of the eye fundus by convolution operation; Step 103, performing sliding window scanning on the structure feature map, calculating the local Shannon entropy of the feature response value in each window, and generating a texture complexity map; Step 104, generating a basic suppression mask for the anatomical structure based on the structure feature map, and extracting a region with a value higher than a preset complexity threshold in the texture complexity map to generate an exemption compensation mask; Step 105, inputting the second resolution input tensor into a classification feature channel to extract an original feature map, and performing feature orthogonal correction operation; The feature orthogonal correction operation comprises: calculating the Hadamard product of the original feature map and the basic suppression mask to generate a preliminary suppression feature, and calculating the weighted product of the original feature map and the exemption compensation mask, and superimposing the weighted product on the preliminary suppression feature to generate a lesion residual feature map; Step 106, inputting the lesion residual feature map into a fully connected classifier to calculate and output a classification probability vector of the eye fundus image data.

2. The fundus image classification method of claim 1, wherein, The step of generating the basic suppression mask for the anatomical structure in step 104 includes: step 201, performing a spatial upsampling process on the structure feature map to make the spatial size of the structure feature map consistent with the spatial size of the original feature map; step 202, mapping the upsampled structure feature map to a probability distribution map with a value in the interval of zero to one by using a nonlinear activation function; step 203, calculating a difference matrix of the value and the probability distribution map, and defining the difference matrix as the basic suppression mask, the element value in the basic suppression mask representing a reserved weight of the corresponding pixel position belonging to the non-anatomical structure background; the local Shannon entropy of the feature response value in each window in step 103 is calculated based on the following formula: Wherein, H(x) is the local Shannon entropy, n is the total number of pixels in the sliding window, x i is the feature response gray value of the i th pixel in the sliding window, and p(x i ) is the occurrence probability of the gray value in the sliding window; the local Shannon entropy is used to quantify the statistical feature of the texture disorder degree of the structure feature map in the local region.

3. The fundus image classification method of claim 1, wherein, The feature orthogonal correction operation in step 105 follows the following logic gating rules: for any target pixel position in the original feature map, if the corresponding value in the basic suppression mask is lower than a preset suppression threshold, and the corresponding value in the exemption compensation mask is zero, the feature response value of the position is set to zero or attenuated; if the corresponding value in the exemption compensation mask is non-zero, the original feature response value of the position is retained or gain operation is performed thereon; the structure perception channel comprises a convolution layer, and the convolution kernel parameters in the convolution layer are set to have gradient directional derivative response to tubular and circular topological structures in the initialization stage; the structure feature map is output by the structure perception channel, which comprises spatial position probability distribution data of the optic disc region and the retinal main blood vessel network region.

4. The fundus image classification method of claim 1, wherein, The method further comprises a process of global geometric semantic auxiliary classification using the features extracted by the structure perception channel, which comprises: step 601, performing global average pooling operation on the deep feature tensor used to generate the structure feature map in the structure perception channel, converting the three-dimensional feature tensor into a one-dimensional global structure description vector; step 602, performing global pooling operation on the lesion residual feature map to generate a lesion feature vector; step 603, performing splicing processing on the global structure description vector and the lesion feature vector in the channel dimension to construct a joint feature vector; and step 604, inputting the joint feature vector into a fully connected classifier.

5. The fundus image classification method of claim 1, wherein, The operation of performing element-wise feature response attenuation on the original feature map by using the base suppression mask in step 105 is performed based on a negative correlation mapping relationship between the numerical values of each element in the base suppression mask and the feature response intensity at the corresponding position in the structural feature map, and the negative correlation mapping relationship maps the high response anatomical structure feature to a weight coefficient close to zero; the width and height dimensions of the first resolution input tensor are respectively one quarter of the width and height dimensions of the second resolution input tensor; The structure perception channel and the classification feature channel are independent in network architecture, and there is only one-way data flow injection in the feature orthogonal correction operation stage. The structure perception channel does not receive the back propagation from the classification feature channel.

6. The fundus image classification method of claim 4, wherein, The global structure description vector contains geometric topological feature data representing the overall tortuosity and branch density of the fundus vascular network; the stitching processing in step 603 encodes the geometric topological feature data as a global context constraint, and jointly encodes the local texture abnormal feature in the lesion feature vector.

7. The fundus image classification method of claim 1, wherein, The classification feature channel includes a deep convolutional neural network; the classification probability vector output in step 106 is generated by the full connection classifier through the Softmax function mapping, and each dimension of the classification probability vector corresponds to the confidence value of a preset retinal lesion grading category.

8. An apparatus for classifying a fundus image, characterized by comprising: It comprises: A multi-scale sampling construction module is configured to acquire fundus image data to be processed, perform multi-scale sampling operation on the fundus image data, and construct a first resolution input tensor and a second resolution input tensor respectively, the spatial dimension of the first resolution being lower than that of the second resolution; A structure feature extraction module is configured to input the first resolution input tensor into a preset structure perception channel to extract a structure feature map representing the inherent anatomical topology of the fundus by convolution operation; a texture entropy calculation module is configured to perform local sliding window scanning on the structure feature map to calculate the local Shannon entropy of the feature response value in each window and generate a texture complexity map; A mask generation module is configured to generate a base suppression mask for anatomical structure based on the structure feature map, and extract the region with a value higher than a preset complexity threshold in the texture complexity map to generate an exemption compensation mask; A feature orthogonal correction module is configured to input the second resolution input tensor into a classification feature channel to extract an original feature map, and perform feature orthogonal correction operation, which includes performing element-wise feature response attenuation on the original feature map by using the base suppression mask, and performing weighted back-annotation on the attenuated feature map by using the exemption compensation mask to generate a lesion residual feature map; A classification decision output module is configured to input the lesion residual feature map into a full connection classifier to calculate and output a classification probability vector of the fundus image data.

9. An apparatus for classifying a fundus image, characterized by comprising: The computer program is executed by the processor to implement the fundus image classification method of claim 1.

10. A storage medium of fundus image classification, on which a computer program is stored, characterized in that, The computer program is executed by the processor to implement the fundus image classification method of claim 1. The computer program is executed by the processor to implement the fundus image classification method of claim 1.

Images