Slit lamp image-based open-angle glaucoma deep learning identification method and system

Abstract

The invention discloses a slit lamp image-based open-angle glaucoma deep learning identification method and system, and relates to the technical field of artificial intelligence and medical diagnosis crossing. The method comprises the steps of preprocessing a slit lamp image and basic clinical data of a target object, processing a standardized image by adopting a pre-trained deep learning model, and performing feature fusion on image depth features and standardized data by adopting a multi-modal feature fusion network. According to the method, automatic classification of two main glaucoma types is achieved, and the problems that in the prior art, distinguishing angle-separating glaucoma from angle-closing glaucoma mainly depends on doctor experience, subjectivity is high, efficiency is not high, traditional chamber angle microscopy examination is high in contact performance, and an existing AI method usually depends on OCT and other special equipment are solved. And objective and efficient auxiliary decision making is provided for clinical diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and medical diagnosis, in particular to a glaucoma depth learning identification method and system based on a slit lamp image. BACKGROUND

[0002] As the leading irreversible blinding eye disease worldwide, the key to effective control of glaucoma lies in early detection, early diagnosis and accurate treatment. Clinically, glaucoma is mainly divided into primary open-angle glaucoma (POAG) and primary angle-closure glaucoma (PACG). The pathophysiological mechanisms, clinical progression and treatment strategies of these two types are different, so it is crucial to accurately classify them before developing a treatment plan.

[0003] Currently, the gold standard for clinically distinguishing POAG and PACG is gonioscopy. This method requires an experienced ophthalmologist to use a specially designed gonioscope to directly contact the patient's cornea under a slit lamp microscope to observe and evaluate the open state of the angle. However, this method has significant limitations: First, gonioscopy is a contact examination that requires placing the examination lens directly on the patient's eyeball, which is extremely unfriendly to patients with poor cooperation and can easily cause patient discomfort and even corneal abrasion and other complications.

[0004] Second, gonioscopy has very high technical requirements for doctors, not only requiring skilled operation skills to correctly place and adjust the gonioscope angle, but also requiring rich clinical experience to accurately interpret the morphological features of the angle. In primary medical institutions, ophthalmologists with such professional skills are often scarce, resulting in uneven diagnostic accuracy.

[0005] Third, gonioscopy is highly subjective, different doctors may have different judgments on the same patient's angle state, especially when the angle is in a critical state, the consistency of the diagnosis result is difficult to guarantee.

[0006] Fourth, gonioscopy takes a long time, from lens disinfection, patient cooperation, examination operation to result interpretation, the examination requires experienced ophthalmologists to operate, has strong operator dependence, subjective result interpretation, and long examination time, etc., it is difficult to popularize and apply in the case of large outpatient volume.

[0007] Although non-contact imaging technologies such as ultrasound biomicroscopy and anterior segment optical coherence tomography can visualize the angle structure, their equipment costs are high and have not been popularized in primary medical institutions. Existing artificial intelligence assisted diagnosis methods mostly focus on analyzing optic nerve damage using fundus photography for early screening of glaucoma, and less on accurate differentiation of glaucoma types. In addition, some methods rely on complex parameter measurement and statistical modeling of OCT images, with a complex process.

[0008] Anterior segment slit lamp examination is the most basic and most popular examination method in ophthalmology, and its image contains rich information of anterior chamber angle. However, how to objectively and efficiently utilize the information for intelligent classification is still a challenge in the field. SUMMARY

[0009] The purpose of the present application is to provide a slit lamp image-based open-angle glaucoma deep learning identification method and system, to solve the problem that the existing technology mainly relies on the experience of doctors to distinguish open-angle glaucoma and closed-angle glaucoma, which is subjective and inefficient, and the traditional gonioscopy is highly contactive and the existing AI method usually relies on special equipment such as OCT, and to provide objective and efficient auxiliary decision-making for clinical diagnosis.

[0010] To achieve the above purpose, the present application provides the following scheme: In a first aspect, the present application provides a slit lamp image-based open-angle glaucoma deep learning identification method, which comprises: preprocessing the slit lamp image and the basic clinical data of the target object to obtain a standardized image of the slit lamp image and standardized data of the basic clinical data; processing the standardized image using a pre-trained deep learning model to obtain image deep features of the slit lamp image, wherein: the deep learning model comprises an initial convolutional layer, a plurality of stackingly arranged residual modules, and a pooling layer, the output of the initial convolutional layer being used as the input of the first residual module, the output of the previous residual module being used as the input of the next residual module, and the output of the last residual module being used as the input of the pooling layer, wherein: the initial convolutional layer is used for preliminary feature extraction of the standardized image to obtain preliminary features; the residual module comprises an attention module and a plurality of stackingly arranged convolutional blocks, the plurality of stackingly arranged convolutional blocks being used for multiple times of feature extraction of the preliminary features; and the attention module is used for feature enhancement of the preliminary features and adding the preliminary features after feature enhancement to the preliminary features after multiple times of feature extraction through residual connection to obtain deep features; the pooling layer is used for global average pooling operation of the deep features to obtain the image deep features; a multi-modal feature fusion network is used for feature fusion of the image deep features and the standardized data to obtain fusion features; outputting the glaucoma classification result of the target object according to the fusion features.

[0011] In a second aspect, the application provides a glaucoma depth learning identification system based on a slit lamp image, the glaucoma depth learning identification system based on the slit lamp image comprising: a preprocessing module configured to preprocess a slit lamp image of a target object and basic clinical data to obtain a standardized image of the slit lamp image and standardized data of the basic clinical data; a feature extraction module configured to process the standardized image by using a pre-trained deep learning model to obtain an image depth feature of the slit lamp image, wherein: the deep learning model comprises an initial convolutional layer, a plurality of stacked residual modules, and a pooling layer, an output of the initial convolutional layer is used as an input of a first residual module, an output of a previous residual module is used as an input of a next residual module, and an output of a last residual module is used as an input of the pooling layer, wherein: the initial convolutional layer is configured to perform preliminary feature extraction on the standardized image to obtain preliminary features; the residual module comprises an attention module and a plurality of stacked convolutional blocks, the plurality of stacked convolutional blocks are configured to perform multiple times of feature extraction on the preliminary features, and the attention module is configured to perform feature enhancement on the preliminary features and add the preliminary features after the feature enhancement to the preliminary features after the multiple times of feature extraction through a residual connection to obtain a depth feature; the pooling layer is configured to perform a global average pooling operation on the depth feature to obtain the image depth feature; a multi-modal fusion module configured to perform feature fusion on the image depth feature and the standardized data by using a multi-modal feature fusion network to obtain a fusion feature; a classification and prediction module configured to output a glaucoma classification result of the target object according to the fusion feature.

[0012] In a third aspect, the application provides a computer device, comprising a memory, a processor, a computer program stored in the memory and executable on the processor, and the processor executes the computer program to implement the steps of the glaucoma depth learning identification method based on a slit lamp image according to any one of the above.

[0013] In a fourth aspect, the application provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the steps of the glaucoma depth learning identification method based on a slit lamp image according to any one of the above.

[0014] In a fifth aspect, the present application provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the glaucoma deep learning identification method based on a slit lamp image according to any one of the above.

[0015] According to the specific embodiments provided in the present application, the following technical effects are disclosed: The present application provides a glaucoma deep learning identification method and system based on a slit lamp image. First, the slit lamp image of the patient and the clinical data are standardized. The slit lamp image is obtained based on conventional anterior segment photography, without the need for contact examination equipment such as an angle lens, thereby improving the comfort and safety of the patient. Second, a deep learning model is used to extract features from the standardized image of the slit lamp image, accurately obtaining various morphological features required for glaucoma classification, eliminating subjective differences in manual interpretation, and ensuring the consistency and repeatability of the classification results. Third, the slit lamp image features and clinical basic data are combined, and a multi-modal fusion network is used to fully utilize information of different dimensions, thereby improving the accuracy and robustness of the classification. Finally, the glaucoma classification results are output, realizing the automatic classification of the two main types of glaucoma, solving the problem that the existing technology mainly relies on the experience of doctors to distinguish between open-angle and closed-angle glaucoma, which is subjective and inefficient, and the traditional gonioscopy examination is highly contact and the existing AI method usually relies on special equipment such as OCT, thereby providing an objective and efficient auxiliary decision for clinical diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments 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.

[0017] Figure 1 A flowchart of a glaucoma deep learning identification method based on a slit lamp image according to an embodiment of the present application; Figure 2 A flowchart of a glaucoma deep learning identification method based on a slit lamp image according to an embodiment of the present application; Figure 3 A structural block diagram of a residual module according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative work are within the protection scope of the present application.

[0019] The above purposes, features and advantages of the present application will be more apparent and understandable. The present application will be further described in detail with reference to the drawings and specific embodiments.

[0020] As shown in the drawings, Figure 1 The present application provides a kind of based on slit lamp image open angle glaucoma deep learning identification method, including the following steps S101 to step S104. Wherein: Step S101, the slit lamp image of target object and basic clinical data are preprocessed, obtain the standardization image of the slit lamp image and the standardization data of the basic clinical data, wherein: The slit lamp image of target object is preprocessed, and the standardization image of the slit lamp image specifically includes the following steps S201 to step S203. Wherein: Step S201, the slit lamp image of target object is obtained, wherein the slit lamp image includes the clear imaging of the eyeball structure of the target object, and the eyeball structure includes cornea, iris, pupil and anterior chamber structure. The slit lamp image is obtained by standard ocular fundus photography equipment, and after ocular fundus photography equipment is collected, original image data transmission is carried out. Because different models of ocular fundus photography equipment exist differences in image format, resolution and the like, the image source is identified and processed by reading the metadata information of the image, and the metadata information includes image resolution, color space, acquisition device model and the like.

[0021] Step S202, the slit lamp image is quality evaluated, and slit lamp image meeting the evaluation standard is screened, and qualified image with clearness up to standard, no occlusion and uniform illumination is screened out; As an optional implementation, the quality evaluation of the slit lamp image includes image clarity evaluation, illumination uniformity evaluation, structure integrity evaluation and non-occlusion evaluation, wherein: The method for evaluating image definition is to calculate the Laplacian variance value of the key region of the slit lamp image as the definition score, and compare it with a definition threshold value, which is determined by ROC curve analysis on a historical clinical image library, and can best distinguish the diagnostic availability; the method for evaluating illumination uniformity is to calculate the brightness standard deviation in the key region of the slit lamp image as the illumination uniformity score, and compare it with an illumination standard deviation value range, which is determined based on the statistical distribution of the "diagnosable" image library annotated by clinical experts; the method for evaluating structural integrity is to ensure that key structures such as the corneal limbus and the pupil boundary are clearly visible; the method for evaluating non-occlusion is to exclude images with heavy occlusion of eyelashes and eyelids.

[0022] In step S203, the slit lamp image meeting the evaluation criteria is subjected to image standardization processing, which includes size unification, image enhancement processing, data enhancement processing, gray scale normalization and region of interest (ROI) extraction, to obtain the standardized image; wherein: The size unification method is to adjust the slit lamp image meeting the evaluation criteria to 224x224 pixels, specifically by using a bilinear interpolation or bicubic interpolation algorithm to scale the image to meet the input standard of the deep learning model, and ensuring that the image detail information is not lost during the image scaling process; The gray scale normalization is to convert the color of the slit lamp image subjected to the scale unification processing, so that the slit lamp image remains an RGB three-channel color image or is converted into a gray scale image, and the pixel value of the slit lamp image is normalized to the interval [0, 1]; it should be noted that the key information of the slit lamp image is mainly reflected in the structural features rather than the color features, and the gray scale normalization can reduce the computational complexity after converting the slit lamp image into a gray scale image, but in some cases, preserving color information can help identify iris texture and other detailed features, so the present application provides a flexible gray scale normalization processing strategy.

[0023] The image enhancement processing method includes contrast enhancement, histogram equalization and noise suppression, wherein the contrast enhancement includes using the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm to enhance the local contrast and highlight the key anatomical structures of the anterior segment, such as the cornea, iris and anterior chamber; the noise suppression includes using a bilateral filter to reduce noise while preserving edge information, eliminating random noise generated during the acquisition process, and improving image quality.

[0024] The data enhancement processing method includes random rotation, translation, scaling and flipping, and the purpose of data enhancement processing of the slit lamp image is to expand the training data set of the deep learning model.

[0025] The preprocessing of the basic clinical data of the target object obtains standardized data of the basic clinical data, and specifically includes the following steps S301 to S302. Wherein: Step S301, collecting the basic clinical data of the target object, the basic clinical data is obtained in the clinical routine examination of the target object, and the basic clinical data includes the baseline information of the target object, such as age, gender, intraocular pressure (IOP), eye information (left / right eye) and family history; wherein, the age is a continuous variable, the unit is year; the gender is a classification variable, including male / female; the intraocular pressure value is a continuous variable, the unit is mmHg; the eye is a classification variable, including left eye / right eye. It is worth noting that the clinical data is an important reference index for glaucoma diagnosis, and different types of glaucoma show different statistical characteristics on these indicators, for example, the patients with angle-closure glaucoma usually have shallow anterior chamber depth and narrow angle.

[0026] Step S302, data standardization processing is performed on the basic clinical data to obtain standardized data of uniform scale, the data standardization processing includes Z-score standardization method and Min-Max normalization method, the numerical range of different clinical indicators is quite different, the standardization processing can eliminate the influence of dimension, so that each index has equivalent weight when the characteristics are fused.

[0027] Step S102, using a pre-trained deep learning model to process the standardized image to obtain image depth features of the slit lamp image, wherein: The deep learning model includes an initial convolution layer, a plurality of stacked residual modules and a pooling layer, the output of the initial convolution layer is used as the input of the first residual module, the output of the previous residual module is used as the input of the next residual module, and the output of the last residual module is used as the input of the pooling layer, wherein: The initial convolution layer is used for preliminary feature extraction of the standardized image to obtain preliminary features; The residual module includes an attention module and a plurality of stacked convolution blocks, the plurality of stacked convolution blocks are used for multiple times of feature extraction of the preliminary features; the attention module is used for feature enhancement of the preliminary features and adding the preliminary features after feature enhancement and the preliminary features after multiple times of feature extraction through residual connection to obtain depth features; The pooling layer is used for global average pooling operation of the depth features to obtain the image depth features, the image depth features contain high-level semantic information of the slit lamp image, and are an important basis for open-angle classification; the global average pooling operation can convert a two-dimensional feature map into a one-dimensional feature vector while retaining global semantic information, thereby providing compact feature representation for subsequent feature fusion and classification.

[0028] In the embodiment, the image depth features include the morphological features of the iris-corneal junction region, the depth features of the anterior chamber space, the iris features, and the overall visual features of the anterior chamber angle opening state. Specifically, the step S102 is implemented to extract the depth features from the standardized image of the slit lamp image. The curvature, width, contrast, and other visual features of the iris-corneal junction region are automatically learned through multi-layer convolution without explicitly measuring the geometric parameters. The apparent features of the relative depth of the anterior chamber are extracted from the standardized image of the slit lamp image to obtain the depth visual features of the anterior chamber. The depth discrimination ability is obtained by comparing with a large number of labeled samples. The morphological representations of the iris root insertion position, the peripheral iris elevation degree, and the iris posterior convexity in the standardized image of the slit lamp image are extracted to obtain the iris root morphological features. The local features are integrated into the semantic expression of the overall anterior chamber angle opening state through the global average pooling operation to obtain the global feature of the anterior chamber angle opening degree.

[0029] As an optional implementation, in the embodiment of the application, the deep learning model takes the RestNet-18 model as the network backbone, integrates the attention module in the residual module, and automatically learns the important features of the attention angle region by using the channel attention mechanism and the spatial attention mechanism. The weights of the RestNet-18 model come from the ImageNet data set IMAGENET1K_V1, the optimizer is Adam, the learning rate is set to 1e-4, the loss function uses the cross entropy with class weights, and a highly abstract feature map tensor is output. The tensor retains the key semantic information of the input image, and its spatial dimension is significantly reduced compared with the original image. The initial convolution layer uses a 7x7 convolution kernel with a step of 2 to extract low-level features of the standardized image, such as edge, texture, and other basic visual information. The RestNet model includes a plurality of stacked residual modules, as shown in Figure 3 For any residual module ResBlock, it includes a main path and a residual connection path. The main path includes two stacked convolution layers Conv1 and Conv2, and the residual connection path is provided with an attention module AttentionBlock. The input of the residual module is respectively passed through the main path and the residual connection path. The main path extracts depth features from the input image features, the residual connection path directly adds the input to the output through a skip connection to relieve the gradient vanishing problem, and the attention module AttentionBlock is introduced in the transmission process to enhance the feature extraction ability of the network to the key region. Finally, the output of the main path and the output of the residual connection path are added to obtain the output of the residual module.

[0030] It should be further explained that the attention module AttentionBlock includes a channel attention module and a spatial attention module. The channel attention module can adaptively adjust the weight of different feature channels to highlight important feature channels related to the chamber state. The spatial attention module focuses on key spatial positions in the image, such as the chamber area, the anterior chamber depth measurement area, etc.

[0031] It can be understood that the residual module is the core structure of the RestNet model. By stacking multiple residual modules, higher-level abstract features are extracted layer by layer, so that the RestNet model can learn hierarchical representation from low-level features to high-level semantic features. In the embodiment, the RestNet model is preferably a RestNet-18 model, and the total number of layers of the RestNet-18 model is 18 layers, wherein the residual modules are divided into 4 stages (Stage1~Stage4), and each stage contains 2 residual modules ResBlock.

[0032] In step S103, the image depth feature and the standardized data are fused by using a multi-modal feature fusion network to obtain a fusion feature.

[0033] As an optional implementation, step S103 specifically includes steps S401 to S402. Wherein: In step S401, the image depth feature and the standardized data are dimensionally converted by using a first fully connected layer, so that the image depth feature and the standardized data have the same dimension. It should be noted that the dimension of the image depth feature output by the deep learning model is 512, and the dimension of the standardized data is 4. After the image depth feature and the standardized data are dimensionally converted by the first fully connected layer, the dimension of the image depth feature and the standardized data is converted to 256. In the embodiment, the first fully connected layer includes 2 to 3 layers, and the number of neurons in each layer decreases in turn. In addition, the first fully connected layer can also learn the nonlinear relationship between the indicators of the clinical data and extract the clinical feature representation related to the type of glaucoma.

[0034] In step S402, the image depth feature and the standardized data after dimension conversion are fused by using a feature fusion strategy to obtain the fusion feature.

[0035] In step S402 of the embodiment of the present application, the feature fusion strategy includes simple concatenation, weighted summation, or attention dynamic fusion mechanism. Wherein, the image depth feature and the standardized data after dimension conversion are fused by using the attention dynamic fusion mechanism, including steps S501 to S503. Wherein: Step S501, taking the image depth feature and the standardized data as inputs of a multi-head attention module, calculating outputs of each head, splicing and linearly projecting the outputs of each head to obtain an output of the multi-head attention module; Step S502, taking the output of the multi-head attention module as an input of an adaptive weight generation network, generating a weight of the image depth feature and a weight of the standardized data through the adaptive weight generation network.

[0036] As an optional implementation, in step S502, the generating the weight of the image depth feature and the weight of the standardized data through the adaptive weight generation network comprises: generating the weight of the image depth feature and the weight of the standardized data by using a gating mechanism; wherein an expression of the gating mechanism is: ; In the formula, is a fusion weight, ; is an activation function; is a weight matrix of the image depth feature; is the image depth feature; is a weight matrix of the standardized data; is the standardized data; is a bias term.

[0037] Step S503, generating the fusion feature according to the weight of the image depth feature and the weight of the standardized data.

[0038] As an optional implementation, in step S503, an expression of the generating the fusion feature according to the weight of the image depth feature and the weight of the standardized data is: ; In the formula, is the fusion feature.

[0039] As an optional implementation, the feature fusion of the image depth feature and the standardized data by using the multi-modal feature fusion network to obtain the fusion feature further comprises: performing a Dropout regularization processing on the fusion feature, and a ratio of the Dropout regularization processing is set to 0.2-0.5. Dropout is an effective regularization technique, which enhances the generalization ability of the model by randomly discarding part of the neurons, avoids overfitting of the model on the training data, and specifically, in the training process, different Dropout ratios are applied to the image depth feature extraction stage, the standardized data extraction stage and the feature fusion stage to cope with the overfitting risk of different modalities. For the image depth feature extraction stage, due to its numerous parameters and complex feature hierarchy, the highest risk of overfitting exists, so a higher Dropout ratio (such as 0.5) is adopted to force learning of more robust spatial features.

[0040] For the standardized data extraction stage, due to its low feature dimension and clear semantics, a lower Dropout ratio (such as 0.2) is adopted to protect its key structured information from being excessively destroyed.

[0041] For the feature fusion stage, a medium Dropout ratio (such as 0.3) is adopted to promote the robust integration of information from different modalities.

[0042] Step S104, outputting the glaucoma classification result of the target object according to the fusion feature, specifically comprising the following steps S601 to S602. Wherein: Step S601, inputting the fusion feature into a second fully connected layer, applying an activation function to the output of the second fully connected layer, and converting the output of the second fully connected layer into a probability distribution, the probability distribution containing an open-angle glaucoma probability value and a closed-angle glaucoma probability value, the sum of the open-angle glaucoma probability value and the closed-angle glaucoma probability value being 1; the second fully connected layer containing a 128-dimensional or 256-dimensional hidden layer, the second fully connected layer further integrating comprehensive features and learning high-level semantic representations for final classification judgment.

[0043] Step S602, judging whether the open-angle glaucoma probability value is greater than the closed-angle glaucoma probability value, if the result is yes, selecting open-angle glaucoma as the glaucoma classification result; otherwise, selecting closed-angle glaucoma as the glaucoma classification result.

[0044] In step S104 of the embodiments of the present application, the outputting the glaucoma classification result of the target object according to the fusion feature further comprises the following steps S701 to S702. Wherein: Step S701, taking the probability value corresponding to the classification result as the confidence, the confidence reflecting the determination degree of the model on the classification result, high confidence indicating that the model is more reliable, and low confidence indicating that manual review may be needed; Step S702, judging whether the confidence is lower than a preset threshold, if the result is yes, marking the classification result and sending the classification result to a manual review queue, for boundary cases with low confidence, automatically triggering a manual review mechanism, and finally determining by a professional ophthalmologist, to ensure the accuracy and safety of the diagnosis.

[0045] Further, the step S104 of outputting the glaucoma classification result of the target object according to the fusion feature further includes generating a diagnosis report containing the classification result, confidence, and key feature visualization information, which not only provides the classification result, but also displays the key areas concerned by the model through visualization techniques such as heat maps, thereby enhancing the interpretability of the result and helping doctors understand the basis for the judgment of the model.

[0046] The steps S101-S104 are implemented as follows. First, the slit lamp images and clinical data of the patient are standardized, which are obtained based on conventional anterior segment photography without the need for contact examination equipment such as gonioscopes, thereby improving the comfort and safety of the patient during examination. Second, a deep learning model is used to extract features from the standardized slit lamp images, accurately obtaining various morphological features required for glaucoma classification, eliminating subjective differences in manual interpretation, and ensuring consistency and repeatability of the classification result. Third, the slit lamp image features are combined with the clinical basic data, and different dimensions of information are fully utilized through a multi-modal fusion network to improve the accuracy and robustness of the classification. Finally, the glaucoma classification result is output, realizing the automatic classification of the two main types of glaucoma and solving the problems in the prior art, such as the strong subjectivity and low efficiency of the classification of open-angle and closed-angle glaucoma mainly relying on the experience of doctors, the strong contact of traditional gonioscopy examination, and the dependence of existing AI methods on special equipment such as OCT, thereby providing an objective and efficient auxiliary decision for clinical diagnosis.

[0047] Further, the open-angle glaucoma deep learning recognition method based on slit lamp images further includes a cross-device domain adaptation method, which comprises the following steps S801-S804. Wherein: Step S801, constructing a device feature extractor, which is used to identify the imaging characteristics of different anterior segment photography devices and extract the imaging feature differences of different anterior segment photography devices based on the imaging characteristics; the different anterior segment photography devices are different types / brands of slit lamp image acquisition devices, such as slit lamp cameras of different brands.

[0048] Step S802, using an adversarial domain adaptation technique to eliminate the imaging feature differences at the feature space level.

[0049] Step S803, obtaining the slit lamp images collected by the anterior segment photography devices with the imaging feature differences eliminated.

[0050] Step S804, taking the slit lamp images as source domain images and converting them into target domain styles through a style transfer network to enhance the cross-device generalization ability of the model.

[0051] Further, in the embodiments of the present application, the cross-device domain adaptive method further comprises establishing a device calibration database for recording characteristic distribution statistics of each device.

[0052] Based on the same inventive concept, the embodiments of the present application also provide a slit lamp image-based open-angle glaucoma deep learning identification system. The slit lamp image-based open-angle glaucoma deep learning identification system comprises a preprocessing module, a feature extraction module, a multi-modal fusion module, and a classification prediction module. Wherein: The preprocessing module is configured to preprocess the slit lamp image and the basic clinical data of the target object to obtain a standardized image of the slit lamp image and standardized data of the basic clinical data. The feature extraction module is configured to process the standardized image by using a pre-trained deep learning model to obtain image deep features of the slit lamp image, wherein: The deep learning model comprises an initial convolution layer, a plurality of stacked residual modules, and a pooling layer. The output of the initial convolution layer is used as the input of the first residual module, the output of the previous residual module is used as the input of the next residual module, and the output of the last residual module is used as the input of the pooling layer, wherein: The initial convolution layer is configured to perform preliminary feature extraction on the standardized image to obtain preliminary features. The residual module comprises an attention module and a plurality of stacked convolution blocks. The plurality of stacked convolution blocks are configured to perform multiple feature extractions on the preliminary features. The attention module is configured to perform feature enhancement on the preliminary features and add the preliminary features after feature enhancement to the preliminary features after multiple feature extractions through residual connection to obtain deep features. The pooling layer is configured to perform global average pooling operation on the deep features to obtain the image deep features. The multi-modal fusion module is configured to perform feature fusion on the image deep features and the standardized data by using a multi-modal feature fusion network to obtain fusion features. The classification prediction module is configured to output the glaucoma classification result of the target object according to the fusion features.

[0053] In addition, the slit lamp image-based open-angle glaucoma deep learning identification system further comprises a cross-device domain adaptive module. The cross-device domain adaptive module comprises a feature extractor construction unit, an imaging feature difference elimination unit, an image acquisition unit, and an image conversion unit. Wherein: The feature extractor construction unit is configured to construct a device feature extractor for identifying imaging characteristics of different anterior segment photography devices and extracting imaging feature differences of different anterior segment photography devices based on the imaging characteristics. The different anterior segment photography devices are different types / brands of slit lamp image acquisition devices, such as slit lamp cameras of different brands.

[0054] The imaging feature difference elimination unit is configured to eliminate the imaging feature differences at a feature space level by using an adversarial domain adaptation technique. The image acquisition unit is configured to acquire slit lamp images acquired by the anterior segment photography devices after the imaging feature differences are eliminated. The image conversion unit is configured to convert the slit lamp images as source domain images into target domain styles by using a style transfer network, thereby enhancing the cross-device generalization capability of the model.

[0055] Further, in the embodiments of the present application, the cross-device domain adaptation module further includes a database construction unit configured to establish a device calibration database to record feature distribution statistics of each device.

[0056] In an exemplary embodiment, a computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor implementing the steps in the above-mentioned method embodiments when executing the computer program.

[0057] In an exemplary embodiment, a computer readable storage medium is provided, storing a computer program, which, when executed by a processor, implements the steps in the above-mentioned method embodiments.

[0058] In an exemplary embodiment, a computer program product is provided, including a computer program, which, when executed by a processor, implements the steps in the above-mentioned method embodiments.

[0059] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0060] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to a memory, a database or other medium used in the embodiments provided in the present application can include at least one of a non-volatile and a volatile memory. The non-volatile memory can include a read-only memory (ROM), a magnetic tape, a floppy disk, a flash memory, an optical storage, a high-density embedded non-volatile memory, a resistive random access memory (ReRAM), a magnetoresistive random access memory (MRAM), a ferroelectric random access memory (FRAM), a phase change memory (PCM), a graphene memory, etc. The volatile memory can include a random access memory (RAM) or an external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as a static random access memory (SRAM) or a dynamic random access memory (DRAM), etc.

[0061] The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0062] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0063] The principles and implementation modes of the present application are described by applying specific examples in the present application. The above-mentioned embodiments are only used to help understand the method and its core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range can be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A kind of opening angle glaucoma deep learning identification method based on slit lamp image, it is characterized in that, The open-angle glaucoma deep learning identification method based on a slit lamp image comprises the following steps: Preprocessing a slit lamp image and basic clinical data of a target object to obtain a standardized image of the slit lamp image and standardized data of the basic clinical data; Processing the standardized image by using a pre-trained deep learning model to obtain image deep features of the slit lamp image, wherein: The deep learning model comprises an initial convolutional layer, a plurality of stacked residual modules, and a pooling layer, the output of the initial convolutional layer serving as the input of the first residual module, the output of the previous residual module serving as the input of the next residual module, and the output of the last residual module serving as the input of the pooling layer, wherein: The initial convolutional layer is used for performing preliminary feature extraction on the standardized image to obtain preliminary features; The residual module comprises an attention module and a plurality of stacked convolutional blocks, the plurality of stacked convolutional blocks being used for performing multiple times of feature extraction on the preliminary features; the attention module is used for performing feature enhancement on the preliminary features and adding the preliminary features after feature enhancement to the preliminary features after multiple times of feature extraction through residual connection to obtain deep features; The pooling layer is used for performing a global average pooling operation on the deep features to obtain the image deep features; Using a multi-modal feature fusion network to perform feature fusion on the image deep features and the standardized data to obtain fusion features; Outputting a glaucoma classification result of the target object according to the fusion features. 2.The glaucoma detection method based on a deep learning algorithm using a slit lamp image according to claim 1, wherein, The preprocessing of the slit lamp image and the basic clinical data of the target object to obtain the standardized image of the slit lamp image and the standardized data of the basic clinical data comprises the following steps: Obtaining a slit lamp image of a target object; the slit lamp image comprises clear imaging of eyeball structures of the target object, the eyeball structures comprising a cornea, an iris, a pupil, and an anterior chamber structure; Performing quality assessment on the slit lamp image to screen slit lamp images meeting the assessment criteria; Performing image standardization processing on the slit lamp images meeting the assessment criteria to obtain the standardized image; the image standardization processing comprises size unification, image enhancement processing, data enhancement processing, gray scale normalization, and region of interest (ROI) extraction; Collecting basic clinical data of the target object; the basic clinical data comprises the age, gender, intraocular pressure value, and eye information of the target object; Performing data standardization processing on the basic clinical data to obtain standardized data of a unified scale, the data standardization processing comprising a Z-score standardization method and a Min-Max normalization method. 3.The glaucoma detection method based on a deep learning algorithm using a slit lamp image according to claim 1, wherein, The feature fusion of the image deep features and the standardized data by using the multi-modal feature fusion network to obtain the fusion features comprises the following steps: Converting the dimensions of the image deep features and the standardized data by using a first fully connected layer so that the image deep features and the standardized data have the same dimensions; Performing feature fusion on the image deep features and the standardized data after dimension conversion by using a feature fusion strategy to obtain the fusion features. 4.The glaucoma detection method based on a deep learning algorithm using a slit lamp image according to claim 3, wherein, The feature fusion strategy includes simple splicing, weighted summation, or an attention dynamic fusion mechanism; wherein the feature fusion of the image depth feature and the standardized data after dimension conversion by adopting the attention dynamic fusion mechanism includes: taking the image depth feature and the standardized data as inputs of a multi-head attention module, calculating the outputs of each head, splicing and linearly projecting the outputs of each head to obtain the output of the multi-head attention module; taking the output of the multi-head attention module as an input of an adaptive weight generation network, generating the weight of the image depth feature and the weight of the standardized data by the adaptive weight generation network; generating the fusion feature according to the weight of the image depth feature and the weight of the standardized data. 5.The glaucoma detection method based on a deep learning algorithm using a slit lamp image according to claim 4, wherein, The generation of the weight of the image depth feature and the weight of the standardized data by the adaptive weight generation network includes: generating the weight of the image depth feature and the weight of the standardized data by a gating mechanism; wherein the expression of the gating mechanism is: ； wherein: is a fusion weight, ; is an activation function; is a weight matrix of image depth features; is an image depth feature; is a weight matrix of normalized data; is normalized data; is a bias term; The expression for generating the fusion feature according to the weight of the image depth feature and the weight of the standardized data is: ； In the formula, is a fusion feature. 6.The glaucoma detection method based on a deep learning algorithm using a slit lamp image according to claim 1, wherein, The output of the fusion feature includes: inputting the fusion feature into a second fully connected layer, applying an activation function to the output of the second fully connected layer, converting the output of the second fully connected layer into a probability distribution, and the probability distribution contains an open-angle glaucoma probability value and a closed-angle glaucoma probability value; determining whether the open-angle glaucoma probability value is greater than the closed-angle glaucoma probability value, if the determination result is yes, selecting open-angle glaucoma as the glaucoma classification result; otherwise, selecting closed-angle glaucoma as the glaucoma classification result.

7. The glaucoma detection method according to claim 6, wherein, The output of the fusion feature also includes: taking the probability value corresponding to the classification result as a confidence degree; determining whether the confidence degree is lower than a preset threshold, if the determination result is yes, marking the classification result, and sending the classification result to an artificial review queue. 8.The glaucoma detection method based on a deep learning algorithm using a slit lamp image according to claim 1, wherein, The feature fusion of the image depth feature and the standardized data by the multi-modal feature fusion network to obtain the fusion feature also includes: performing Dropout regularization processing on the fusion feature, and the ratio of the Dropout regularization processing is set to 0.2-0.

5.

9. The glaucoma identification method based on a slit lamp image according to any one of claims 1-8, characterized in that, The image depth feature includes the morphological feature of the iris-corneal junction area, the depth feature of the anterior chamber space, the iris feature, and the overall visual feature of the chamber angle opening state. 10.A system for glaucoma detection based on deep learning of open angle glaucoma from slit lamp images, characterized in that, The open-angle glaucoma deep learning recognition system based on a slit lamp image includes: a preprocessing module for preprocessing the slit lamp image of a target object and the basic clinical data to obtain a standardized image of the slit lamp image and standardized data of the basic clinical data; a feature extraction module for processing the standardized image by a pre-trained deep learning model to obtain an image depth feature of the slit lamp image, wherein: The deep learning model comprises an initial convolutional layer, a plurality of stacked residual modules, and a pooling layer, an output of the initial convolutional layer serving as an input of a first residual module, an output of a previous residual module serving as an input of a next residual module, and an output of a last residual module serving as an input of the pooling layer, wherein: an initial convolutional layer, configured to perform preliminary feature extraction on the standardized image to obtain preliminary features; a residual module, comprising an attention module and a plurality of stacked convolutional blocks, the plurality of stacked convolutional blocks being configured to perform multiple times of feature extraction on the preliminary features; the attention module being configured to perform feature enhancement on the preliminary features and add the preliminary features after feature enhancement and the preliminary features after multiple times of feature extraction through a residual connection to obtain deep features; a pooling layer, configured to perform a global average pooling operation on the deep features to obtain the image deep features; a multi-modal fusion module, configured to perform feature fusion on the image deep features and the standardized data by using a multi-modal feature fusion network to obtain fused features; a classification and prediction module, configured to output a glaucoma classification result of the target object according to the fused features.

Images