Eye OCT image analysis method and system based on deep learning

Abstract

The invention discloses an eye OCT image analysis method and system based on deep learning, and relates to the related field of OCT image analysis, and the method comprises the steps: collecting multi-angle eye OCT images; the collected eye OCT image is preprocessed; segmenting the preprocessed eye OCT image by adopting a segmentation model based on deep learning; and according to a segmentation result, calculating eye structure parameters for the femtosecond laser cataract surgery. Through the method, the problem of insufficient segmentation precision of a traditional image processing algorithm is solved. In addition, according to the eye structure parameters obtained through calculation, a cutting track of the femtosecond laser cataract surgery is automatically generated; a deep learning-based cataract detection and grading model is adopted to analyze the preprocessed eye OCT image, the opacification degree of the cataract is judged, and the optimal presplitting nuclear energy is given, so that the risk of planning a femtosecond laser cataract surgery by subjective experience is reduced, and the safety of the surgery is improved.

Description

Technical Field

[0001] This application relates to the field of OCT image analysis, and particularly to an ophthalmic OCT image analysis method and system based on deep learning. Background Art

[0002] In the process of ophthalmic diagnosis and treatment, OCT image analysis plays a crucial role. Compared with other commonly used medical imaging technologies, OCT technology has significant advantages such as non-invasiveness, high resolution at the micron level, and relatively low cost. This technology not only plays a key role in the diagnosis of fundus diseases, but is also widely used in the diagnosis and treatment of glaucoma and cataracts. It provides visual insights for ophthalmologists and greatly improves the level of diagnosis and treatment of ophthalmic diseases.

[0003] Traditional image processing algorithms, such as threshold segmentation and active contour models, have many deficiencies in analyzing OCT images. First, the segmentation accuracy of these algorithms for low-contrast structures such as the posterior surface of the lens is poor, and key structural boundaries cannot be extracted. Second, their processing speed is slow and it is difficult to meet the needs of clinical real-time diagnosis. Moreover, these algorithms cannot effectively process the artifacts generated by the strong reflection of intraocular lenses after surgery, affecting the accuracy of the analysis results. In addition, there are also defects in the existing technology in cataract surgery planning and laser energy setting. The traditional method requires doctors to manually mark the degree of lens opacity, which is inefficient and difficult to quantify the full-thickness opacity depth. At the same time, the planning of the surgical cutting trajectory and the setting of laser energy also rely on the subjective experience of doctors, with a certain risk of human error, affecting the surgical accuracy and safety. Summary of the Invention

[0004] The present invention aims to solve the deficiencies of the existing technology and proposes an ophthalmic OCT image analysis method and system based on deep learning. By using a deep learning model to analyze ophthalmic OCT images, precise segmentation of key eye structures such as the cornea, aqueous humor, and lens in ophthalmic OCT images is achieved, effectively solving the problem of insufficient segmentation accuracy of traditional image processing algorithms for low-contrast structures. At the same time, based on the output results of the cataract detection and grading model, the femtosecond laser cataract surgery is reasonably planned to improve the surgical safety.

[0005] This application provides an ophthalmic OCT image analysis method based on deep learning, including the following steps:

[0006] (1) Collect ophthalmic OCT images from multiple angles;

[0007] (2) Perform preprocessing on the collected ophthalmic OCT images;

[0008] (3) Segment the preprocessed ophthalmic OCT images using a segmentation model based on deep learning;

[0009] (4) Calculate the eye structure parameters for femtosecond laser cataract surgery based on the segmentation results;

[0010] Furthermore, use an optical coherence tomography (OCT) device to collect multi-angle eye OCT images. During the collection process, select an appropriate scanning mode according to the patient's eye condition, and adjust the scanning range, depth, and resolution. A beam of light is emitted during the operation of the scanning device, passes through the eye tissue and reflects back. The detector receives these reflected lights and generates eye OCT images.

[0011] Furthermore, perform preprocessing on the collected eye OCT images, including cropping and normalization. To obtain a more abundant dataset for training the deep learning model, apply various data augmentation methods to expand the data volume, including random rotation, random flipping, affine transformation, and noise addition. These methods can simulate images from different perspectives in real scenarios and increase the diversity of the data.

[0012] Furthermore, the specific steps for segmenting the preprocessed eye OCT images using a deep learning-based segmentation model are as follows:

[0013] Construct a deep learning-based segmentation model, and train the model based on the annotated data to obtain the network parameters of the optimal model as the pre-trained segmentation model;

[0014] Input the preprocessed eye OCT images into the pre-trained segmentation model;

[0015] Use the segmentation model to segment the input eye OCT images and output the segmentation results.

[0016] Furthermore, the deep learning-based segmentation model is U-net++. This model is an improved U-net segmentation network, which has a symmetric encoder part and decoder part, and connects the encoder and decoder levels through skip connections. The encoder part is responsible for extracting the high-level feature representations of the image, including multiple convolutional and downsampling operations. The decoder part has the same structure as the encoder part, maps these features back to the original image size through convolutional and upsampling operations, and generates the segmentation results. U-net++ enhances the reusability of features through more fine-grained feature fusion and dense skip connection mechanisms, enabling the network to have stronger segmentation ability for target regions with complex boundaries, and has extremely strong robustness and generalization ability.

[0017] Furthermore, when training a deep learning-based segmentation model, a deep supervision training strategy is adopted. This mechanism introduces multiple output layers and corresponding loss functions at different stages of the decoder, enabling the model to perform supervised learning at multiple levels. This approach helps the model learn features at different levels, accelerates the training convergence speed of the network, and at the same time, helps improve the generalization ability of the model, reduces the risk of overfitting, and further improves the segmentation accuracy.

[0018] Furthermore, after the deep learning-based segmentation model segments the ocular OCT images, the cornea, aqueous humor, and lens boundaries can be clearly distinguished from the images; in addition, the curvatures of the upper and lower surfaces of the cornea, the curvatures of the upper and lower surfaces of the lens, the central thickness of the lens, and the pupil diameter are accurately extracted from the segmented images. Based on these data, the corneal radius of curvature, lens volume, tilt angle of the intraocular lens, and axial position parameters are calculated.

[0019] Furthermore, according to the above-mentioned ocular structure parameters obtained, the cutting trajectory of the femtosecond laser cataract surgery is automatically generated, reducing the risk of planning cataract surgery based on subjective experience.

[0020] Furthermore, to reduce the laser loss in femtosecond laser cataract surgery, a deep learning-based cataract detection and grading model is used to analyze the preprocessed ocular OCT images, judge the degree of cataract turbidity, and give the optimal pre-cracking nucleus energy;

[0021] Furthermore, using a deep learning-based cataract detection and grading model to analyze the preprocessed ocular OCT images, the specific steps include:

[0022] Construct a deep learning-based cataract detection and grading model, train it based on cataract image data, and retain the network parameters of the optimal model as the pre-trained cataract detection and grading model;

[0023] Input the preprocessed ocular OCT images into the pre-trained cataract detection and grading model;

[0024] Use the cataract detection and grading model to analyze the input ocular OCT images, output the degree of cataract turbidity, and select the optimal pre-cracking nucleus energy value.

[0025] Furthermore, the cataract detection and grading model adopts the ConvNeXt structure, which combines the local receptive field of the convolutional neural network and the Transformer structure design, and realizes the accurate grading of the degree of cataract turbidity through depthwise separable large kernel convolution. Among them, the cataract turbidity level is divided into levels 1-20. Through comprehensive evaluation by analyzing the turbidity degree of different positions of the lens output by the model, the optimal pre-cracking nucleus energy is given, reducing the laser loss and improving the safety of the surgery.

[0026] Furthermore, ConvNeXt consists of multiple ConvNeXt blocks. The ConvNeXt block adopts an inverted bottleneck structure, including three convolutional layers, all of which use depthwise separable convolutions. The inverted bottleneck structure expands and then compresses the channel dimension of the features, avoiding information loss caused by the compressed dimension in the bottleneck structure when information is transformed between different dimensional feature spaces. The ConvNeXt block introduces a large kernel convolutional layer, which is consistent with the mechanism of multi-head self-attention in Transformer and can capture long-range context information, improving the model's ability to understand long-range dependencies in images.

[0027] This application also provides an eye OCT image analysis system based on deep learning, including:

[0028] An eye OCT image acquisition module for acquiring eye OCT scan images from multiple angles;

[0029] An eye OCT image preprocessing module for preprocessing the acquired eye OCT images;

[0030] A deep learning segmentation module that uses a deep learning-based segmentation model to segment the preprocessed eye OCT images;

[0031] A parameter quantization module for calculating eye structure parameters for femtosecond laser cataract surgery according to the segmentation results;

[0032] A femtosecond laser cutting trajectory automatic generation module for automatically generating the cutting trajectory of femtosecond laser cataract surgery according to the calculated parameters;

[0033] An energy automatic setting module that uses a deep learning-based cataract detection and grading model to analyze the preprocessed eye OCT images, judge the turbidity degree of the cataract, and give the optimal pre-cracking nucleus energy.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] The eye OCT image analysis system based on deep learning provided by this application of the present invention realizes precise segmentation of key structures such as the cornea, aqueous humor, and lens in eye OCT images through deep learning methods. In addition, the system also provides a cataract detection and grading model, which can automatically judge the turbidity degree of the cataract and accurately give the optimal pre-cracking nucleus energy according to the detection results, effectively reducing laser damage and significantly improving surgical safety. At the same time, the system can automatically generate the cutting trajectory of cataract surgery, simplifying the doctor's operation process, reducing the surgical risk caused by human errors, and further improving the surgical cutting accuracy.

[0036] In terms of postoperative evaluation, the system can accurately quantify key parameters such as the corneal radius of curvature, lens volume, tilt angle and axial position of the intraocular lens, providing high-precision quantitative support for doctors. The application of deep learning methods not only greatly improves the image processing speed, meets the needs of clinical real-time diagnosis, but also effectively solves the problem of insufficient accuracy of traditional image processing algorithms in segmenting low-contrast structures, significantly improving the segmentation accuracy and robustness. Brief Description of the Drawings

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings of the embodiments of the present invention will be briefly introduced below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the operations in the front or below do not necessarily need to be executed precisely in sequence. On the contrary, according to needs, various steps can be executed in reverse order or simultaneously. At the same time, other operations can also be added to these processes, or one or several steps can be removed from these processes.

[0038] Figure 1 It is a schematic flowchart of a method for analyzing ocular OCT images based on deep learning provided by an embodiment of this application.

[0039] Figure 2 It is a detailed structural diagram of a segmentation model based on deep learning provided by an embodiment of this application.

[0040] Figure 3 It is a model training flowchart of the segmentation model provided by an embodiment of this application.

[0041] Figure 4 It is a detailed structural diagram of a cataract detection and grading model based on deep learning provided by an embodiment of this application.

[0042] Figure 5 It is a schematic structural diagram of an ocular OCT image analysis system based on deep learning provided by an embodiment of this application.

[0043] Figure 6 It is the original ocular OCT image.

[0044] Figure 7 It is a quantitative graph of the ocular OCT image processed by the ocular OCT image analysis system. Detailed Embodiments

[0045] The above description is only an overview of the technical solutions of this application. In order to be able to understand the technical means of this application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of this application more obvious and understandable, the following specifically gives the detailed embodiments of this application.

[0046] To make the objectives, technical solutions and advantages of this application clearer, the following will further describe this application in detail with reference to the accompanying drawings. The described embodiments should not be regarded as limitations of this application. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this application.

[0047] In the following description, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or server that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application.

[0048] Embodiment 1. The embodiment of this application provides an eye OCT image analysis method based on deep learning, as Figure 1 shown:

[0049] Step S1: Collect eye OCT images from multiple angles. Use an optical coherence tomography (OCT) device to collect multi-angle scan images of the anterior segment of the eye. During the collection process, select an appropriate scan mode according to the patient's eye condition, and adjust the scan range, depth, and resolution. A beam of light is emitted during the operation of the scanning device, passes through the eye tissue and reflects back. The detector receives these reflected lights and generates an eye OCT image.

[0050] Step S2: Perform preprocessing on the collected eye OCT images. Crop the collected eye OCT images to a size of 224×224 to meet the size requirements of the input features of the deep learning model. Normalize each pixel point in the image so that its pixel value range is within [0, 255]. To obtain a more abundant dataset for training the deep learning model, apply multiple data augmentation methods to expand the data volume:

[0051] Flip the cropped eye OCT images along the horizontal and vertical directions respectively as new training samples; rotate the cropped eye OCT images 5°, 15°, and 20° clockwise and counterclockwise respectively and add them to the existing dataset to simulate the dynamic changes of images in reality; at the same time, perform affine transformation on the cropped eye OCT images using rotation, translation, and flipping operations to increase the complexity of the images; in addition, add noise to the cropped eye OCT images to simulate the interference of noise in the real imaging process. The introduction of these complex samples helps to improve the adaptability of the segmentation model to different changes in eye OCT images, thereby significantly enhancing the generalization performance of the model.

[0052] Step S3: Segment the preprocessed eye OCT image using a deep learning-based segmentation model. The specific steps are as follows:

[0053] Construct a deep learning-based segmentation model and train the model based on the annotated data to obtain the network parameters of the optimal model as the pre-trained segmentation model. The deep learning-based segmentation model used is U-net++. The detailed structure is as Figure 2 shown:

[0054] This model is an improved U-net segmentation network. The structure of this model is divided into an encoder part and a decoder part. The encoder part is responsible for extracting the high-level feature representation of the image, including multiple convolutional and downsampling operations. The dimension of the input eye OCT image to the encoder is 224×224×3 (where 224×224 is the image size and 3 is the number of channels). After being processed by the encoder, high-dimensional intermediate layer features are obtained. The output features of each convolutional layer are fused through dense skip connections to achieve multi-scale information fusion, which helps the model learn richer and more representative features. The dense skip connection mechanism retains the feature information from each stage of the encoder, reducing the loss of information during propagation. The decoder part has the same structure as the encoder part. Through convolutional and upsampling operations, these feature maps are mapped back to the original image size and the segmentation result is generated. U-net++ enhances the reusability of features through fine-grained feature fusion and a dense skip connection mechanism, enabling the network to have accurate segmentation ability for target regions with complex boundaries and having extremely strong robustness and generalization ability.

[0055] When training the deep learning-based segmentation model, a deep supervision training strategy is adopted. This mechanism introduces multiple output layers at different stages of the decoder, calculates the corresponding loss function, and enables the model to perform supervised learning at multiple levels. The detailed training process of the segmentation model is as Figure 3 shown:

[0056] For the target eye OCT image, that is, the eye OCT image to be recognized and processed, manual marking is first performed. For each input eye OCT image, a class label (the classes include background, cornea, aqueous humor, and lens) is assigned to each pixel to generate a segmentation map. The value of each pixel in the segmentation map is no longer the color information of the original image, but represents the class information (class value is an integer) to which the pixel belongs, thereby obtaining the theoretical boundary. Secondly, the target image is input into the deep learning-based segmentation model, and the model outputs a predicted segmentation map. Calculate the deviation between the predicted segmentation map and the true segmentation map, that is, the loss function. In this example, Dice Loss is used as the loss function. The calculation formula of Dice Loss is as follows:

[0057]

[0058] Among them, Dice Loss represents the value of the loss function, p is the predicted segmentation map (multi-class probability map) of the model, g is the ground truth segmentation map, and ∈ is a very small parameter used to avoid the denominator being zero. This example involves a multi-classification problem, and the Dice Loss of all classes is averaged to obtain the final loss function:

[0059]

[0060] Among them, f(x) represents the total loss function, K is the total number of classes, and Dice Loss k is the value of the loss function for the k-th class. The advantage of Dice Loss is that it can effectively handle the class imbalance problem because it focuses on the overlapping part between two sets rather than simply calculating the proportion of correctly classified pixels.

[0061] Secondly, according to the value of the loss function, the model parameters are dynamically updated. In this embodiment, a deep supervision training strategy is adopted. The decoder outputs multiple predicted segmentation maps at different stages, and the corresponding loss functions are calculated respectively; for the predicted segmentation maps output at each level, a total of 4 loss functions of the predicted results are calculated in this embodiment, and the weighted average of these 4 loss functions is used as the final loss function value. This deep supervision training method helps the model learn features at different levels, accelerates the training convergence speed of the model. At the same time, it helps to improve the generalization ability of the model, reduce the risk of overfitting, and further improve the segmentation accuracy.

[0062] When the model training converges to the loss function, the training ends; adjust the hyperparameter settings of the model, conduct multiple trainings, and retain the model parameters with the smallest loss function at convergence as the final pre-trained segmentation model. The stochastic gradient descent method is adopted during the training process, and the Adam optimizer is selected with an initial learning rate of 3e -4 .

[0063] Input the preprocessed eye OCT image into the pre-trained segmentation model.

[0064] Use the segmentation model to segment the input eye OCT image and output the segmentation result.

[0065] Step S4, according to the segmentation result, calculate the eye structure parameters for femtosecond laser cataract surgery. After segmenting the eye OCT image with the deep learning-based segmentation model, the cornea, aqueous humor, and lens boundaries can be clearly distinguished from the image; Figure 6 is an OCT image of the anterior segment of the eye. Use the segmentation model to accurately segment this image, and after quantifying the image, obtain Figure 7, in the figure, the red part represents the cornea, and the green part represents the lens. The upper and lower curvatures p1 and p2 of the cornea, the upper and lower curvatures p3 and p4 of the lens, the central thicknesses h1 and h2 of the cornea and the lens, and the pupil diameter d are accurately extracted therefrom; according to the above-mentioned eye structure parameters, the corneal curvature radius, the lens volume, the tilt angle of the intraocular lens, and the axial position parameters are calculated. A segmentation model based on deep learning is used to achieve accurate segmentation of the eye OCT image, and further calculate the eye structure parameters. The accurate measurement of these parameters provides strong support for the personalized planning of femtosecond laser cataract surgery.

[0066] Example 2. The embodiment of the present application provides a method for analyzing eye OCT images based on deep learning. Based on the parameters calculated in Example 1, the cutting trajectory of femtosecond laser cataract surgery can be automatically generated. The automatically generated cataract surgery cutting trajectory includes a pre-cracked nucleus of the cataract, a capsular ring incision, and a corneal side incision; automatically generating the surgery cutting trajectory can reduce the surgical risk caused by subjective experience.

[0067] To further improve the safety of the surgery, a cataract detection and grading model based on deep learning is used to analyze the preprocessed eye OCT image. The specific steps include:

[0068] Construct a cataract detection and grading model based on deep learning, and train it based on cataract image data, and retain the network parameters of the optimal model as the pre-trained cataract detection and grading model;

[0069] Input the preprocessed eye OCT image into the pre-trained cataract detection and grading model;

[0070] Use the cataract detection and grading model to analyze the input eye OCT image, output the cataract turbidity degree, and select the optimal pre-cracked nucleus energy value.

[0071] The cataract detection and grading model adopts the ConvNeXt structure, as Figure 4 shown. This structure combines the local receptive field of the convolutional neural network and the Transformer structure design, and uses depthwise separable convolution to implement a lightweight model for deployment on computer devices. ConvNeXt consists of multiple ConvNeXt blocks. The ConvNeXt block adopts an inverted bottleneck structure, including three convolutional layers, all of which use depthwise separable convolution; the inverted bottleneck structure expands and then compresses the channel dimension of the feature, so that when information is converted between different dimensional feature spaces, the information loss caused by compressing the dimension in the bottleneck structure can be avoided. The large kernel convolutional layer introduced by the ConvNeXt block is consistent with the mechanism of multi-head self-attention in the Transformer, and can capture long-range context information and improve the model's ability to understand long-range dependence relationships in images.

[0072] The training method adopted by the cataract detection and grading model is the same as that of the segmentation model. The cataract grades in the ocular OCT images are extracted manually as training labels. Among them, the cataract turbidity grades are divided into levels 1-20. The prediction results of the model are compared with the label data, and the loss function is calculated. The cross-entropy function is selected as the loss function for this model, and the calculation formula of this function is as follows:

[0073]

[0074] where N represents the number of samples, C represents the number of categories, that is, the number of levels of cataract turbidity, y i,c is the true label of the i-th sample in the c-th category, and p i,c is the probability of the i-th sample in the c-th category predicted by the model. The model parameters are updated dynamically according to the loss function value until the model converges, and the optimal model parameters are retained as the pre-trained cataract detection and grading model.

[0075] The preprocessed ocular OCT images are input into the pre-trained cataract detection and grading model. The detailed process of the model processing the ocular OCT images includes the following steps:

[0076] First, the preprocessed ocular OCT images undergo a layer of convolution for preliminary feature extraction of the images, reducing noise interference and generating primary feature maps;

[0077] Secondly, the primary feature maps are sent into 3 consecutive ConvNeXt blocks 1 for processing, increasing the number of channels of the feature maps with a size of 224×224×3 to 96, and the size of the feature maps becomes 56×56. The ConvNeXt block consists of a depthwise separable convolutional layer, layer normalization, and a GELU (Gaussian Error Linear Unit) activation function, and applies channel scaling and random dropout strategies. The depthwise separable convolutional layer uses a large kernel convolution with a kernel size of 7, and the normalization method is the same layer normalization as that of the Transformer structure. Compared with the batch normalization commonly used in convolution, this normalization method can achieve better network performance. The calculation formula of the GELU activation function adopted is expressed as:

[0078] GELU(x) = xΦ(x)

[0079]

[0080] Among them, \(x\) represents the input, \(\varPhi(x)\) is the cumulative function of the standard normal distribution, and \(erf(\cdot)\) is the error function. The GELU activation function can not only retain the advantages of the ReLU function and solve the problem of gradient disappearance, but also simulate the natural distribution characteristics of the data, help the model learn complex feature representations, and improve the prediction ability and generalization ability of the model. The channel scaling strategy adjusts the number of channels or the weight values of each channel during the training process to reduce the model complexity; the random dropout strategy is a regularization technique that randomly discards the outputs of some neurons during the training process to break the excessive dependence between neurons and prevent the model from overfitting.

[0081] Then, the feature map is fed into the Downsample module to downsample the features. The size of the feature map becomes half of the original. The downsampled feature map is then fed into 3 stacked ConvNeXt blocks 2 for further processing. At this time, the size of the feature map is \(28\times28\times192\).

[0082] Next, repeat the above operations. The feature map passes through multiple Downsample modules and ConvNeXt blocks to obtain richer feature information and generate high-level feature representations. The size of the high-level feature map output by ConvNeXt block 4 is \(7\times7\times768\).

[0083] Finally, the high-level feature map undergoes global pooling and linear layer mapping to output the prediction results, realizing the grading of the cataract turbidity degree. Through analyzing the turbidity degrees at different positions of the lens output by the model, a comprehensive evaluation is made to give the optimal pre-cracking nucleus energy, reduce the laser loss, and improve the safety of the surgery.

[0084] Embodiment 3. An eye OCT image analysis system based on deep learning, as Figure 5 shown, includes:

[0085] An eye OCT image acquisition module 10 for acquiring eye OCT images from multiple angles;

[0086] An eye OCT image preprocessing module 20 for performing preprocessing on the acquired eye OCT images;

[0087] A deep learning segmentation module 30 that uses a deep learning-based segmentation model to segment the preprocessed eye OCT images;

[0088] A parameter quantization module 40 for calculating the eye structure parameters for femtosecond laser cataract surgery according to the segmentation results;

[0089] A femtosecond laser cutting trajectory automatic generation module 50 for automatically generating the cutting trajectory of femtosecond laser cataract surgery according to the calculated parameters. This module includes a cataract pre-cracking nucleus module 51, a capsulotomy module 52, and a corneal side-cut module 53;

[0090] The energy automatic setting module 60 analyzes the preprocessed eye OCT image by using a cataract detection and grading model based on deep learning, determines the turbidity degree of the cataract, and gives the optimal pre-cracking nucleus energy.

[0091] The eye OCT image analysis system based on deep learning provided by the embodiment of the present invention can execute the eye OCT image analysis method based on deep learning provided by any embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution of the method.

[0092] Although various references are made to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The various units and modules included are only divided according to functional logic, but are not limited to the above division as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of mutual distinction and do not limit the protection scope of the present invention.

[0093] The above specific embodiments do not constitute a limitation to the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principle of the present application shall be included within the protection scope of the present application. In some cases, the actions or steps recorded in the present application can be executed in a different order from that in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. An eye OCT image analysis method based on deep learning, characterized in that, The method includes: (1) Collecting eye OCT images from multiple angles; (2) Performing preprocessing on the collected eye OCT images; (3) Segmenting the preprocessed eye OCT images using a deep learning-based segmentation model; (4) Calculating eye structure parameters for femtosecond laser cataract surgery based on the segmentation results.

2. The method for analyzing ocular OCT images based on deep learning according to claim 1, characterized in that, In step (1), an optical coherence tomography imaging device is used to collect eye OCT images from multiple angles. During the collection process, an appropriate scanning mode needs to be selected according to the patient's eye condition, and the scanning range, depth, and resolution are adjusted. A beam of light is emitted during the operation of the scanning device, passes through the eye tissue and reflects back, and the detector receives these reflected lights and generates eye OCT images.

3. A method for analyzing eye OCT images based on deep learning according to claim 1, characterized in that, In step (3), the specific steps of segmenting the preprocessed eye OCT images using a deep learning-based segmentation model include: Constructing a deep learning-based segmentation model and training the model based on labeled data to obtain the network parameters of the optimal model as the pre-trained segmentation model; Inputting the preprocessed eye OCT images into the pre-trained segmentation model; Using the segmentation model to segment the input eye OCT images and output the segmentation results.

4. The method for analyzing eye OCT images based on deep learning according to claim 1, wherein In step (3), the deep learning-based segmentation model is U-net++. This model is an improved U-net segmentation network with a symmetric encoder part and decoder part, and the encoder and decoder levels are connected through skip connections. The encoder part is responsible for extracting high-level feature representations of the image, including multiple convolutional and downsampling operations. The decoder part has the same structure as the encoder part, maps these features back to the original image size through upsampling and convolutional operations, and generates the segmentation results.

5. The method for analyzing eye OCT images based on deep learning according to claim 1, characterized in that In step (4), after segmenting the eye OCT images using the deep learning-based segmentation model, the cornea, aqueous humor, and lens boundaries are distinguished from the image; and the upper and lower surface curvatures of the cornea, the upper and lower surface curvatures of the lens, the central thickness of the lens, and the pupil diameter are extracted from the image. Based on these data, the corneal curvature radius, lens volume, intraocular lens tilt angle, and axial position parameters are calculated.

6. The method for analyzing eye OCT images based on deep learning according to claim 5, characterized in that, According to the calculated eye structure parameters, the cutting trajectory of femtosecond laser cataract surgery is automatically generated.

7. The method for analyzing eye OCT images based on deep learning according to claim 1, characterized in that, In step (2), a deep learning-based cataract detection and grading model is used to analyze the preprocessed eye OCT images, judge the cataract turbidity degree, and give the optimal pre-cracking nucleus energy.

8. The method for analyzing eye OCT images based on deep learning according to claim 7, characterized in that, The specific steps of using the deep learning-based cataract detection and grading model to analyze the preprocessed eye OCT images include: Constructing a deep learning-based cataract detection and grading model and training it based on cataract image data, and retaining the network parameters of the optimal model as the pre-trained cataract detection and grading model; Inputting the preprocessed eye OCT images into the pre-trained cataract detection and grading model; Using the cataract detection and grading model to analyze the input eye OCT images, output the cataract turbidity degree, and select the optimal pre-cracking nucleus energy value.

9. The method for analyzing eye OCT images based on deep learning according to claim 8, wherein, The cataract detection and grading model adopts the ConvNeXt structure, which combines the local receptive field of the convolutional neural network and the Transformer structure design to achieve accurate grading of the cataract turbidity degree. Among them, ConvNeXt consists of multiple ConvNeXt blocks. The ConvNeXt block adopts an inverted bottleneck structure, including three convolutional layers, all of which use depthwise separable convolutions. The large kernel convolutional layer introduced in the ConvNeXt block is consistent with the mechanism of multi-head self-attention in the Transformer to capture long-range context information.

10. An eye OCT image analysis system based on deep learning, characterized in that, The system is used to implement a method for analyzing ocular OCT images based on deep learning according to any one of claims 1-9. The system includes: An ocular OCT image acquisition module for acquiring ocular OCT images from multiple angles; An ocular OCT image preprocessing module for performing preprocessing on the acquired ocular OCT images; A deep learning segmentation module that uses a segmentation model based on deep learning to segment the preprocessed ocular OCT images; A parameter quantization module for calculating the ocular structure parameters for femtosecond laser cataract surgery according to the segmentation results; A femtosecond laser cutting trajectory automatic generation module for automatically generating the cutting trajectory of femtosecond laser cataract surgery according to the calculated parameters; An energy automatic setting module that uses a cataract detection and grading model based on deep learning to analyze the preprocessed ocular OCT images, judge the cataract turbidity degree, and give the optimal pre-cracking nucleus energy.

Images