Nuclear cataract grading classification method based on AS-OCT image

Abstract

The invention discloses a nuclear cataract grading classification method based on an AS-OCT image, and relates to the technical field of nuclear cataract grading. The nuclear cataract grading classification method based on the AS-OCT image comprises the steps of construction of a CDRFF-Net grading network based on the AS-OCT image, image segmentation and model input, feature extraction, construction of a supplementary discrimination region extractor module, construction of an image strong feature fusion module and nuclear cataract grading. The invention provides a CDRFF-Net classification network for nuclear cataract classification, classification precision is improved by identifying distinguishing region features, a complementary distinguishing region extractor module and an image enhancement feature fusion module are designed, rich distinguishing region feature extraction and efficient fusion are realized, and the classification accuracy is improved. The ordinal number information of the severity of the nuclear cataract is coded by using a soft label, and ordinal number category information is taken as priori knowledge to be incorporated into a classification model.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear cataract grading, and in particular to a classification method for nuclear cataract grading based on AS-OCT images. Background Art

[0002] Nuclear cataract is one of the leading causes of vision loss worldwide, and its incidence is rising as the population ages. It mainly affects the nuclear area of ​​the lens, where proteins gradually degenerate with age, causing the lens to harden and become opaque, impairing vision. The progression of nuclear cataracts varies from person to person and may gradually worsen over months to years. Therefore, early diagnosis and timely intervention are essential to prevent further deterioration of vision. In this context, accurate grading of nuclear cataracts becomes crucial as it can help doctors develop more appropriate treatment plans, while the current grading of nuclear cataracts relies on the doctor's subjective judgment, which requires a certain amount of experience.

[0003] Anterior segment optical coherence tomography (AS-OCT) is an advanced, non-invasive, non-contact optical coherence tomography imaging technology that provides high-resolution, high-contrast real-time imaging capabilities, is easy to operate, and has a faster scanning speed. In recent years, ophthalmologists have increasingly adopted AS-OCT to diagnose ocular surface-related diseases. Compared with traditional retinoscope imaging and slit lamp imaging, AS-OCT shows significant advantages. Retinoscope imaging mainly captures cataract opacity information, while slit lamp imaging is greatly affected by external lighting conditions and is highly dependent on the doctor's experience and lacks precise quantification capabilities. In contrast, AS-OCT images can clearly capture the lens nucleus and quantitatively measure the degree of opacity, such as the attached Figure 2 This is essential to help doctors diagnose and evaluate the pathological changes of nuclear cataracts. Therefore, a method that can automatically identify the degree of such opacity is needed.

[0004] Despite the good performance of these methods, some shortcomings were observed in this application: existing nuclear cataract grading research methods usually use coarse-grained classification methods, which are not sensitive enough to small differences due to subtle differences between adjacent nuclear levels. These methods usually rely on overall features for classification and ignore the fine-grained changes in different image regions. For example, the severity of the nucleus is positively correlated with the opacity of the nucleus, while the distribution of opacity in the lens nucleus is uneven, which means that coarse-grained methods may mask the differences in opacity between different regions, making it difficult to detect small opacity changes within the nucleus and accurately distinguish different pathological levels.

[0005] In fine-grained visual classification, the features of distinguishing regions play a key role. However, although most attention-based fine-grained classification models are able to locate distinguishing regions, they still have shortcomings when dealing with fine-grained tasks. Specifically, they may ignore less obvious but still important regional features. In addition, extracting features from different regions in isolation will ignore the interrelationships and dependencies between them. In amorphous optical coherence tomography images, opacity is widely distributed in the lens nucleus, and it is difficult to capture all significant feature changes by focusing only on the region. Features from different regions are usually associated, and failure to capture these regional features and correlations will affect the accuracy and robustness of the classification model.

[0006] Aphakic grading has an ordinal prior, indicating that there is a natural order relationship between different levels. However, most nuclear cataract grading studies use one-hot label encoding and fail to take advantage of this ordinal feature. One-hot encoding treats each level as an independent category and fails to reflect the continuity between levels, ignoring their order information. This encoding method limits the performance of the model when processing ordinal data, thus affecting the classification results. Summary of the invention

[0007] In view of the deficiencies of the prior art, the present invention provides a classification method for nuclear cataract grading based on AS-OCT images, which solves the problems raised in the above-mentioned background technology.

[0008] To achieve the above objectives, the present invention is implemented by the following technical scheme: a classification method for nuclear cataract grading based on AS-OCT images, the method comprising the following specific steps:

[0009] S1: Image segmentation and model input; the lens nucleus region of the AS-OCT image is segmented according to the pre-trained deep segmentation network, and the AS-COT cropped image is used as the input of the grading model;

[0010] S2: Feature extraction: The hierarchical model extracts features from the input image based on the dynamic Swin Transformer, and introduces a mask generator using a dynamic sparsification mechanism to divide the feature area into information-rich areas and information-sparse areas, and performs data processing based on a dual-channel setting of slow and fast paths;

[0011] S3: construct a supplementary discriminative region extractor module; the supplementary discriminative region extractor module consists of a feature enhancement and suppression module and a discriminative feature extraction module. The hierarchical model obtains the feature map of attention enhancement and the feature map of attention suppression through the feature enhancement and suppression module, and combines the mask generated by S2 with the obtained feature map through the discriminative feature extraction module to complete the further extraction of discriminative region features;

[0012] S4: Construct an image strong feature fusion module; the image strong feature fusion module completes the effective aggregation of features in different regions of the feature map, and uses the graph attention mechanism to focus on different feature subspaces to complete the fusion of features in different regions;

[0013] S5: Nuclear cataract grading; the data labels are converted into soft probability distribution and the cross entropy loss is used. The grading model can accurately capture and realize the grading of nuclear cataract.

[0014] A further improvement of the technical solution of the present invention is that the size of the AS-OCT image in S1 after segmentation is 224×224.

[0015] A further improvement of the technical solution of the present invention is that the dynamic Swin Transformer in S2 processes the standard data in the following specific steps:

[0016] F′ i-1 =PatchMerging(F i-1 );

[0017] F i =Swins(F' i-1 )#(1);

[0018] where F' i-1 is the output feature of the previous stage, PatchMerging(·) represents the image block merging layer, and Swins(·)r represents the dynamic Swin Transformer in each stage.

[0019] A further improvement of the technical solution of the present invention is that the mask generator in S2 is composed of a normalization layer, a linear layer and a logsoftmax layer, wherein the mask generator generates a score map of two channels, the first channel represents the probability that each image block belongs to an information-rich area, and the second channel represents the probability that each image block belongs to an information-sparse area.

[0020] A further improvement of the technical solution of the present invention is that the data processing according to the dual-channel setting of the slow path and the fast path in S2 includes the following specific steps:

[0021]

[0022] Where Split(·) represents the split operation, Slow(·) represents the slow path operation, Fast(·) represents the fast path operation, Reassemble(·) represents the reassemble operation, and F′ 3,j is the output feature of the previous section.

[0023] A further improvement of the technical solution of the present invention is that the feature enhancement and suppression module and the discriminant feature extraction module in S3 include the following specific steps: extracting features and 4C1 represent the dimension and number of channels of the feature respectively;

[0024] S3.1: Evenly divide F2 features, the size of each group of features is expressed as Where 1≤i≤M, then the 1×1 and ReLU operations are determined by calculating the importance of each group;

[0025] S3.2: Normalize using mean and softmax operations, including,

[0026]

[0027] in represents the normalized importance factor of all groups;

[0028] S3.3: F2 is reweighted to obtain, in is the feature factor, and sets the mask vector S i ∈R M×1 Represents the importance factors of all groups, setting the values ​​of the feature factors to 0 and other numbers to 1 by multiplying the mask vector S i and characteristic factors Completion feature factors Inhibition of

[0029]

[0030] in represents the normalized importance factor of all groups.

[0031] A further improvement of the technical solution of the present invention is that in S5, different classification labels of nuclear cataracts are corresponding to the regression labels according to the distribution sequence, and the beneficial effect of soft label encoding is achieved based on the prior information of the category order.

[0032] Compared with the prior art, the beneficial effect of the present invention is that a CDRFF-Net grading network for grading nuclear cataracts is proposed, the classification accuracy is improved by identifying discriminative regional features, a complementary discriminant region extractor module and a graph enhancement feature fusion module are designed to achieve rich discriminant region feature extraction and efficient fusion, and soft labels are used to encode the ordinal information of the severity of nuclear cataracts, and the ordinal category information is incorporated as prior knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the process of grading nuclear cataract of the present invention;

[0034] Figure 2 The overall framework of this application in CDRFF-Net classification

[0035] Figure 3 a is a schematic diagram of the mask generator, and b is a schematic diagram of the dynamic Swin Transformer module;

[0036] Figure 4 a is a schematic diagram of the FES module, and b is a schematic diagram of the DFE module;

[0037] Figure 5 This is the NC severity distribution map of the NC_AS-OCT_ dataset. DETAILED DESCRIPTION

[0038] Various exemplary embodiments, features and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0039] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0040] In addition, in order to better illustrate the present application, numerous specific details are provided in the specific embodiments below. It should be understood by those skilled in the art that the present application can also be implemented without certain specific details. In some examples, methods, means, and elements well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present application.

[0041] The present invention provides a classification method for nuclear cataract grading based on AS-OCT images, the method comprising constructing a CDRFF-Net grading network based on AS-OCT images, specifically comprising:

[0042] S1: Image segmentation and model input; the lens nucleus region of the AS-OCT image is segmented according to the pre-trained deep segmentation network, and the AS-COT cropped image is used as the input of the grading model;

[0043] The size of the AS-OCT image in S1 after segmentation is 224×224.

[0044] The output training set The image containing the kernel region is used as input data, X m ∈R H×W Corresponding label set Y m={Normal, Mild, Moderate, Severe}, Normal, Mild, Moderate and Severe represent normal, mild, moderate and severe respectively. For evaluation in D train The reasoning ability of the trained classification model;

[0045] S2: Feature extraction: The hierarchical model extracts features from the input image based on the dynamic Swin Transformer, and introduces a mask generator using a dynamic sparsification mechanism to divide the feature area into information-rich areas and information-sparse areas, and performs data processing based on a dual-channel setting of slow and fast paths;

[0046] The dynamic Swin Transformer in S2 processes standard data in the following specific steps:

[0047] F′ i-1 =PatchMerging(F i-1 );

[0048] F i =Swins(F' i-1 )#(1);

[0049] where F' i-1 is the output feature of the previous stage, PatchMerging(·) represents the image block merging layer, and Swins(·)r represents the dynamic Swin Transformer in each stage.

[0050] The mask generator in S2 is composed of a normalization layer, a linear layer and a logsoftmax layer, wherein the mask generator generates a score map of two channels, the first channel represents the probability that each image block belongs to an information-rich region, and the second channel represents the probability that each image block belongs to an information-sparse region.

[0051] The data processing according to the dual-channel setting of the slow path and the fast path in S2 includes the following specific steps:

[0052]

[0053]

[0054] Where Split(·) represents the split operation, Slow(·) represents the slow path operation, Fast(·) represents the fast path operation, Reassemble(·) represents the reassemble operation, and F′ 3,j is the output feature of the previous section.

[0055] S3: construct a supplementary discriminative region extractor module; the supplementary discriminative region extractor module consists of a feature enhancement and suppression module and a discriminative feature extraction module. The hierarchical model obtains the feature map of attention enhancement and the feature map of attention suppression through the feature enhancement and suppression module, and combines the mask generated by S2 with the obtained feature map through the discriminative feature extraction module to complete the further extraction of discriminative region features;

[0056] The feature enhancement and suppression module and the discriminant feature extraction module in S3 include the following specific steps: extracting features and 4C1 represent the dimension and number of channels of the feature respectively;

[0057] S3.1: Evenly divide F2 features, the size of each group of features is expressed as Where 1≤i≤M, then the 1×1 and ReLU operations are determined by calculating the importance of each group;

[0058] S3.2: Normalize using mean and softmax operations, including,

[0059]

[0060] in represents the normalized importance factor of all groups;

[0061] S3.3: F2 is reweighted to obtain, in is the feature factor, and sets the mask vector S i ∈R M×1 Represents the importance factors of all groups, setting the value of the most important feature factor to 0 and the others to 1, by multiplying the mask vector S i and characteristic factors Completion feature factors Inhibition of

[0062]

[0063] in represents the normalized importance factor of all groups.

[0064] S4: Construct an image strong feature fusion module; the image strong feature fusion module completes the effective aggregation of features in different regions of the feature map, and uses the graph attention mechanism to focus on different feature subspaces to complete the fusion of features in different regions;

[0065] S5: Nuclear cataract grading; the data labels are converted into soft probability distribution and the cross entropy loss is used. The grading model can accurately capture and realize the grading of nuclear cataract.

[0066] In S5, different classification labels of nuclear cataracts are corresponding to the regression labels according to the distribution sequence, and soft label encoding is implemented based on the category order prior information.

[0067] This application proposes a soft label ordinal regression method, which not only retains the ordered information between categories, but also incorporates this information into the loss function, by converting the data labels into soft probability distributions and using cross entropy loss, so that the model can capture the subtle differences between input instances and different categories, rather than hard classifying them into specific categories. To achieve this, this application first assigns ordinal regression labels 1, 2, 3, and 4 corresponding to the nuclear cataract grading labels "normal, mild, moderate, and severe", which naturally include clear order relationships between categories.

[0068] Specifically, Compute the encoding vector y of the true label for a given label level r k , where y = {r1, r2, …, r C}, where C is the number, r k is the true sequence label, φ(r k ,r i ) is the selected distance metric loss function, where the simple absolute error |ij| is selected as the penalty function, penalizing the distance r between the true metric values k and r j , for example, the true label is 2. After soft label encoding, it is represented as [0.1966, 0.5344, 0.1966, 0.0723];

[0069] Furthermore, the cross entropy loss function is widely used in classification tasks, but the labels used are encoded with soft labels, specifically:

[0070]

[0071] where y i is the soft label encoding y of the true label, p i is the predicted probability of the label.

[0072] In order to evaluate whether the technical solution of this application is feasible, this application conducted experiments on two datasets, including a local AS-OCT dataset, a public ophthalmic OCT dataset, and a public fundus image dataset;

[0073] The AS-OCT dataset includes data from 448 patients, of which 386 provided right eye data, 382 provided left eye data, and 320 provided binocular data, totaling 12,288 AS-OCT images: normal (3008), mild (3056), moderate (4128), and severe (2096). AS-OCT images cover the entire lens structure, while only the nuclear area is associated with nuclear cataract (nuclear cataract) lesions. This application uses a pre-trained deep segmentation network to automatically crop the core area. Given the lack of a standard nuclear cataract grading system based on AS-OCT images, the nuclear cataract labels of AS-OCT images are mapped from slit lamp images using LOCSIII. This application divides the participants' AS-OCT images into mutually exclusive training and test sets, ensuring that each participant's images are assigned to only one of the datasets. The training set includes images from 314 participants, and the test set includes 3,611 images from 134 participants. In addition, an independent validation set consisting of images from 44 participants randomly selected from the training set was constructed to evaluate model performance.

[0074] The public ophthalmology OCT dataset contains 24,000 images divided into eight categories: age-related macular degeneration (AMD), choroidal neovascularization (CNV), diabetic macular edema (DME), drusen, macular hole (MH), diabetic retinopathy (DR), central serous chorioretinopathy (CSC), and healthy. In this dataset, 25,600 images are used for training, 2,800 for validation, and 2,800 for testing. Each category includes 3,200 images for training, 350 for validation, and 350 for testing.

[0075] The public fundus image dataset is a DR fundus image dataset provided by the Kaggle blindness detection competition. The dataset contains 3.662 images, of which 2.929 are used for training and 733 are used for testing. The images are divided into five levels: (1) no obvious retinal lesions (level 0), (2) mild NPDR (level 1), (3) moderate NPDR (level 2), (4) severe NPDR (level 3), and (5) PDR (level 4).

[0076] In order to comprehensively evaluate the effectiveness of the proposed method in nuclear cataract grading, this application selected accuracy (ACC), sensitivity (Sen), F1 score, precision (PRE) and Kappa as evaluation indicators. ACC, Sen, F1, Kappa and PRE are commonly used indicators for evaluating nuclear cataracts based on AS-OCT images. The definitions of these indicators are as follows:

[0077]

[0078] Where TP, FP, TN and FN represent the number of true positives, false positives, true negatives and false negatives, respectively.

[0079] The proposed model is implemented on the PyTorch platform. The mask generator in the CDRE module is placed after the 6th dynamic Swin Transformer and the retention rate t is set to 0.5.

[0080] In order to verify the effectiveness of the fine-grained grading network for nuclear cataract severity assessment proposed in this application, this application selected a group of representative baseline models for comparative analysis. These models include existing CNNs, attention-based methods, methods specifically designed for fine-grained classification, and existing methods focused on nuclear cataract grading. In this study, this application compared the proposed CDRFF-Net with a variety of advanced models, including traditional CNN models (such as Vgg16 and ResNet34), attention mechanism models (such as ECA-Net34 and SE-Net34), fine-grained classification models (such as FBSD, TransFG, and FET-FGVC), and nuclear cataract grading models (such as RIRN et and Ranking-MFCNet), through the use of five key performance indicators: accuracy, precision, sensitivity, F1 score and Kappa coefficient for systematic evaluation, CDRFF-Net outperforms the comparison models in all key performance indicators. For example, compared with ResNet34, CDRFF-Net's accuracy is improved by 7.02% (from 83.56% to 90.58%), and the Kappa coefficient is improved by 9.62% (from 77.49% to 87.11%), demonstrating its high efficiency in category differentiation;

[0081] In addition, this application compares CDRFF-Net with models specifically designed for fine-grained tasks, such as FBSD, TransFG, and FET-FGVC. CDRFF-Net performs well in handling subtle differences between categories. For example, compared with FBSD, CDRFF-Net's accuracy is improved by 3.96% (from 86.62% to 90.58%); compared with TransFG, CDRFF-Net's accuracy is improved by 2.94% (from 87.64% to 90.58%); compared with FET-FGVC, CDRFF-Net improves its accuracy by 1.65% (from 88.93% to 90.58%), confirming the effectiveness of the CDRE module and the GFF module in achieving rich distinguishable regional feature extraction and efficient fusion.

[0082] At the same time, compared with the nuclear cataract grading model, CDRFF-Net outperforms RIRNet and Ranking-MFCNet in terms of accuracy, which is improved by 4.28% (from 86.3% to 90.58%) and 3.14% (from 87.44% to 90.58%), respectively, and also improves in Kappa coefficient, which is improved by 5.87% (from 81.24% to 87.11%) and 4.32% (from 82.79% to 87.11%), respectively.

[0083] In order to verify the effectiveness of the fine-grained grading network for nuclear cataract severity assessment proposed in this application; This application verifies the effectiveness of treating nuclear cataract grading as a fine-grained classification task, and also demonstrates the complementary discriminant region extraction of the CDRE block, the effective feature fusion capability of the GFF module, and the advantages of adding the order information between categories to the cross entropy loss function. These results further emphasize the importance of fine-grained methods in medical image analysis, especially when dealing with diseases such as nuclear cataracts that have progressive and continuous turbidity changes.

[0084] This application further uses soft label encoding and ordinal priors to compare the nuclear cataract grading performance of different backbone networks. This application selected ResNet34, ECA-Net34, RIRNet and CDRFF-Net (containing only CDRE blocks and GFF modules) as backbone networks, and added soft label encoding and ordinal priors in the experiment, using SLE to improve the nuclear cataract scoring performance of all backbone networks. Specifically, after using SLE, the ACC of ResNet34 increased from 83.56% to 85.21%, an increase of 1.65%; the ACC of ECA-Net34 increased from 84.56% to 86.23%; an increase of 1.67%; the ACC of RIRNet increased from 86.30% to 88.02%; an increase of 1.72%; the ACC of CDRFF-Net (CDRE+GFF) increased from 88.9% to 90.58%, an increase of 1.68%.

[0085] These results show that soft label encoding with ordinal priors improves the nuclear cataract scoring performance of all backbone networks. It is worth noting that CDRFF-Net combined with SLE achieves the highest classification accuracy, verifying the effectiveness and superiority of the designed module in this application.

[0086] In order to further verify the effectiveness of the method proposed in this application in the task of medical image classification, this application applies CDRFF-Net to two public datasets: the OCT-Eye dataset and the fundus image dataset. The image categories in the OCT-Eve dataset do not have ordinal features, so the CDRFF-Net of this application uses cross-entropy loss without soft label encoding. In contrast, the fundus image dataset is an ordered category dataset. In the OCT-Eye dataset, the CDRE module in CDRFF-Net is able to extract discriminative features from different categories, while the GFF module effectively fuses these features, leading to more accurate classification results. For the Fundus-Eye dataset, CDRFF-Net not only focuses on discriminative regional features, but also utilizes prior knowledge of category order. By using cross-entropy loss with soft label encoding, it achieves higher classification accuracy. Therefore, these results further demonstrate the generalization ability of the proposed method.

[0087] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0088] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A classification method for nuclear cataract grading based on AS-OCT images, characterized in that: The method includes constructing a CDRFF-Net hierarchical network based on AS-OCT images, specifically including: S1: Image segmentation and model input; the lens nucleus region of the AS-OCT image is segmented according to the pre-trained deep segmentation network, and the AS-COT cropped image is used as the input of the grading model; S2: Feature extraction: The hierarchical model extracts features from the input image based on the dynamic Swin Transformer, and introduces a mask generator using a dynamic sparsification mechanism to divide the feature area into information-rich areas and information-sparse areas, and performs data processing based on a dual-channel setting of slow and fast paths; S3: construct a supplementary discriminative region extractor module; the supplementary discriminative region extractor module consists of a feature enhancement and suppression module and a discriminative feature extraction module. The hierarchical model obtains the feature map of attention enhancement and the feature map of attention suppression through the feature enhancement and suppression module, and combines the mask generated by S2 with the obtained feature map through the discriminative feature extraction module to complete the further extraction of discriminative region features; S4: Construct an image strong feature fusion module; the image strong feature fusion module completes the effective aggregation of features in different regions of the feature map, and uses the graph attention mechanism to focus on different feature subspaces to complete the fusion of features in different regions; S5: Nuclear cataract grading; the data labels are converted into soft probability distribution and the cross entropy loss is used. The grading model can accurately capture and realize the grading of nuclear cataract.

2. The classification method for nuclear cataract grading based on AS-OCT images according to claim 1, characterized in that: The size of the AS-OCT image in S1 after segmentation is 224×224.

3. The classification method for nuclear cataract grading based on AS-OCT images according to claim 1, characterized in that: The dynamic Swin Transformer in S2 processes standard data in the following specific steps: F i-1 =PatchMerging(F i-1 ); F i =Swins(F' i-1 )#(1); where F' i-1 is the output feature of the previous stage, PatchMerging(·) represents the image block merging layer, Swins(·) represents the Swin Transfomer module, and r represents the dynamic Swin Transformer in each stage.

4. The classification method for nuclear cataract grading based on AS-OCT images according to claim 1, characterized in that: The mask generator in S2 is composed of a normalization layer, a linear layer and a logsoftmax layer, wherein the mask generator generates a score map of two channels, the first channel represents the probability that each image block belongs to an information-rich region, and the second channel represents the probability that each image block belongs to an information-sparse region.

5. The classification method for nuclear cataract grading based on AS-OCT images according to claim 1, characterized in that: The data processing according to the dual-channel setting in S2 includes the following specific steps: Where Split(·) indicates a split operation, Slow(·) indicates a slow path operation, Fast(·) indicates a fast path operation, and Reassemble(·) indicates a reassemble operation. is the output feature of the previous section.

6. The classification method for nuclear cataract grading based on AS-OCT images according to claim 1, characterized in that: The feature enhancement and suppression module and the discriminant feature extraction module in S3 include the following specific steps: extracting features and 4C1 represent the dimension and number of channels of the feature respectively; S3.1: Evenly divide F2 features, the size of each group of features is expressed as Where 1≤i≤M, then the 1×1 and ReLU operations are determined by calculating the importance of each group; S3.2: Normalize using mean and softmax operations, including, in represents the normalized importance factor of all groups; S3.3: F2 is reweighted to obtain, in is the feature factor, and sets the mask vector S i ∈R M×1 Represents the importance factors of all groups, sets the values ​​of important feature factors to 0 and the values ​​of other groups to 1, and multiplies them by the mask vector S i and characteristic factors Completion feature factors Inhibition, including in represents the normalized importance factor of all groups.

7. The classification method for nuclear cataract grading based on AS-OCT images according to claim 1, characterized in that: In S5, different classification labels of nuclear cataracts are corresponding to the regression labels according to the distribution sequence, and soft label encoding is implemented based on the category order prior information.

Images