AI model-based optic nerve image segmentation and identification method and system

Abstract

The invention discloses an optic nerve image segmentation and recognition method and system based on an AI model. Semantic segmentation processing is carried out on an optic nerve image of a patient object by adopting an image processing technology based on deep learning. Firstly, contrast enhancement, Gaussian filtering and image size adjustment are performed on an optic nerve image of a patient object, so that image quality is improved, and optic nerve structural features are highlighted. And then, semantic feature extraction and pixel-level classification are performed on the preprocessed optic nerve image by using a neural network model, so that accurate segmentation and recognition of the optic nerve image are realized. Particularly, in the Gaussian filtering process of the optic nerve image, a mode of dynamically adjusting the standard deviation of the Gaussian filter is adopted to cope with the quality difference of different images, so that the noise reduction processing effect of the image is optimized. Therefore, the accuracy of optic nerve image segmentation can be effectively improved, the lesion area of optic nerves can be identified, and more accurate auxiliary information can be provided for clinical diagnosis of eye disease patients.

Description

Technical Field

[0001] This application relates to the field of image segmentation, and more specifically, to a method and system for optic nerve image segmentation and recognition based on an AI model. Background Art

[0002] In the field of medical image analysis, especially in the diagnosis of ophthalmic diseases, the accurate segmentation and recognition of optic nerve images play a crucial role. By accurately extracting the information of the optic nerve structure, doctors can better evaluate the health status of the optic nerve, detect the development of diseases (such as glaucoma, optic neuritis, etc.), and make effective treatment plans.

[0003] Traditionally, the analysis of optic nerve images relied on manual or semi-automatic methods, which usually required a large amount of manual intervention, not only time-consuming and laborious, but also easily affected by the subjective factors of operators, resulting in insufficient consistency and accuracy of the results. With the development of computer vision technology and artificial intelligence, automated methods have gradually been introduced into the processing of optic nerve images, aiming to improve efficiency and reduce human errors.

[0004] However, the existing AI-based optic nerve image segmentation methods still face some challenges. For example, various noises and artifacts that may be introduced during the acquisition of optic nerve images will interfere with the clear presentation of the optic nerve structure and affect the final segmentation effect. The conventional image preprocessing method, Gaussian filtering, although can smooth the image and reduce noise to a certain extent, due to the significant quality differences in optic nerve images obtained by different devices, including differences in resolution, contrast, etc., it brings difficulties to image processing under a unified standard.

[0005] Therefore, an optimized method and system for optic nerve image segmentation and recognition based on an AI model are expected. Summary of the Invention

[0006] To solve the above technical problems, this application is proposed. Embodiments of this application provide a method and system for optic nerve image segmentation and recognition based on an AI model, which perform semantic segmentation processing on the optic nerve images of patient objects by using image processing technology based on deep learning. First, the contrast of the optic nerve images of patient objects is enhanced, Gaussian filtering is performed, and the image size is adjusted to improve the image quality and highlight the structural features of the optic nerve. Then, a neural network model is used to extract semantic features and perform pixel-level classification on the preprocessed optic nerve images, thereby achieving accurate segmentation and recognition of the optic nerve images. In particular, during the process of performing Gaussian filtering on the optic nerve images, the standard deviation of the Gaussian filter is dynamically adjusted to cope with the quality differences of different images, thereby optimizing the noise reduction effect of the images. In this way, the accuracy of optic nerve image segmentation can be effectively improved, the diseased areas of the optic nerve can be identified, and more accurate auxiliary information can be provided for the clinical diagnosis of eye disease patients.

[0007] According to one aspect of this application, a method for optic nerve image segmentation and recognition based on an AI model is provided, which includes:

[0008] Retrieve the optic nerve images of the target patient object;

[0009] Preprocess the optic nerve images of the target patient object to obtain preprocessed optic nerve images;

[0010] Input the preprocessed optic nerve images into an optic nerve image feature extractor based on a dynamic convolution tagging model to obtain an optic nerve image semantic encoding feature matrix;

[0011] Input the optic nerve image semantic encoding feature matrix into an image semantic segmenter based on the Softmax function to obtain an image semantic segmentation result.

[0012] According to another aspect of this application, a system for optic nerve image segmentation and recognition based on an AI model is provided, which includes:

[0013] An image retrieval module for retrieving the optic nerve images of the target patient object;

[0014] An image preprocessing module for preprocessing the optic nerve images of the target patient object to obtain preprocessed optic nerve images;

[0015] An optic nerve image feature extraction module for inputting the preprocessed optic nerve images into an optic nerve image feature extractor based on a dynamic convolution tagging model to obtain an optic nerve image semantic encoding feature matrix;

[0016] An image semantic segmentation module, which is used to input the optic nerve image semantic encoding feature matrix into an image semantic segmenter based on the Softmax function to obtain an image semantic segmentation result.

[0017] Compared with the prior art, a method and system for optic nerve image segmentation and recognition based on an AI model provided by the present application performs semantic segmentation processing on the optic nerve image of a patient object by adopting an image processing technology based on deep learning. First, the contrast of the optic nerve image of the patient object is enhanced, Gaussian filtering is performed, and the image size is adjusted to improve the image quality and highlight the optic nerve structure features. Then, a neural network model is used to extract semantic features and perform pixel-level classification on the preprocessed optic nerve image, so as to achieve accurate segmentation and recognition of the optic nerve image. In particular, during the process of performing Gaussian filtering on the optic nerve image, a method of dynamically adjusting the standard deviation of the Gaussian filter is adopted to cope with the quality differences of different images, thereby optimizing the noise reduction effect of the image. In this way, the accuracy of optic nerve image segmentation can be effectively improved, the diseased area of the optic nerve can be identified, and more accurate auxiliary information can be provided for the clinical diagnosis of eye disease patients. Description of the Drawings

[0018] By describing the embodiments of the present application in more detail with reference to the drawings, the above and other objects, features, and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 It is a flowchart of a method for optic nerve image segmentation and recognition based on an AI model according to an embodiment of the present application;

[0020] Figure 2 It is a schematic diagram of data flow of a method for optic nerve image segmentation and recognition based on an AI model according to an embodiment of the present application;

[0021] Figure 3 It is a flowchart of sub-step S2 of a method for optic nerve image segmentation and recognition based on an AI model according to an embodiment of the present application;

[0022] Figure 4 It is a block diagram of a system for optic nerve image segmentation and recognition based on an AI model according to an embodiment of the present application. Detailed Description of the Embodiments

[0023] Next, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the example embodiments described herein.

[0024] As used in this application and the claims, unless the context clearly indicates otherwise, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0025] Although this application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or the server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0026] Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the operations before or below do not necessarily need to be executed precisely in order. On the contrary, as needed, 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.

[0027] Next, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the example embodiments described herein.

[0028] In the technical solution of this application, an optic nerve image segmentation and recognition method based on an AI model is proposed. Figure 1 FIG. is a flowchart of an optic nerve image segmentation and recognition method based on an AI model according to an embodiment of the present application. Figure 2 FIG. is a schematic diagram of data flow of an optic nerve image segmentation and recognition method based on an AI model according to an embodiment of the present application. As Figure 1 and Figure 2As shown, the AI model-based optic nerve image segmentation and recognition method according to an embodiment of the present application includes the steps of: S1, retrieving the optic nerve image of a target patient object; S2, preprocessing the optic nerve image of the target patient object to obtain a preprocessed optic nerve image; S3, inputting the preprocessed optic nerve image into an optic nerve image feature extractor based on a dynamic convolution labeling model to obtain an optic nerve image semantic coding feature matrix; S4, inputting the optic nerve image semantic coding feature matrix into an image semantic segmenter based on the Softmax function to obtain an image semantic segmentation result.

[0029] Specifically, in S1, the optic nerve image of the target patient object is retrieved. This process generally involves a Hospital Information System (HIS), a Picture Archiving and Communication System (PACS), and specialized ophthalmic imaging equipment. Before accessing any patient data, it is necessary to ensure that the operator has legal permission and goes through the necessary authentication processes, such as using a username / password, biometric technology, etc., and at the same time confirm that the operator has obtained the patient's informed consent form allowing them to access and process personal health information. After logging into the hospital's information management system or a dedicated ophthalmic diagnosis and treatment platform, the operator inputs or selects the identifier of a specific patient (such as a medical record number, ID number, or other unique identification code) as needed to locate the patient's Electronic Medical Record (EMR). Once the correct patient file is determined, the operator can then browse the patient's examination history to find ophthalmology-related diagnosis records or appointment records. If the patient has had optic nerve images collected, these images can be found in the corresponding reports or image lists. This is usually done through the PACS system integrated in the HIS, which is responsible for storing, managing, and transmitting medical images. These images may be generated by different ophthalmic imaging devices, such as Optical Coherence Tomography (OCT) scanners, fundus cameras, etc., and thus may have different formats and resolutions. After downloading, the images are imported into a designated software application for further preprocessing and analysis. Throughout the process, strict privacy policies and regulatory requirements must be followed to ensure the security and confidentiality of patient data, using encryption technology and access control mechanisms to prevent unauthorized data leakage or tampering. Each retrieval action should be detailedly recorded, including the timestamp, operator information, the type and scope of the accessed data, etc., for future audit and tracking.

[0030] Specifically, in S2, the optic nerve image of the target patient object is preprocessed to obtain a preprocessed optic nerve image. Considering that the optic nerve of the patient object and the surrounding structures may have problems such as noise, low contrast, and blurred edges in the original image due to various reasons (such as lighting conditions, imaging device quality, etc.), which may affect the subsequent image segmentation effect. Therefore, in this application, the optic nerve image is first preprocessed to enhance the image quality, making it easier for the model to learn useful features. In a specific example of this application, as Figure 3 shown, S2 includes: S21, enhancing the image contrast of the optic nerve image to obtain a contrast-enhanced optic nerve image; S22, using a Gaussian filter to perform image noise reduction on the contrast-enhanced optic nerve image to obtain a noise-reduced optic nerve image; S23, adjusting the image size of the noise-reduced optic nerve image to obtain the preprocessed optic nerve image.

[0031] Specifically, in S21, the image contrast of the optic nerve image is enhanced to obtain a contrast-enhanced optic nerve image. Considering that in the image segmentation task, accurate edge detection is the key to successful segmentation. Therefore, in this application, the image contrast of the optic nerve image is first enhanced, making the tiny structures and detailed information in the image clearer and more prominent, while effectively strengthening the edge information in the optic nerve image, enabling the image segmentation algorithm to more accurately identify the boundary of the optic nerve structure, thereby improving the accuracy and precision of image segmentation. In the embodiment of this application, the histogram equalization technique is used to enhance the contrast of the optic nerve image to obtain a contrast-enhanced optic nerve image.

[0032] In a specific example, considering that the optic nerve of the patient object and the surrounding structures may have problems such as noise, low contrast, and blurred edges in the original image due to various reasons (such as lighting conditions, imaging device quality, etc.), which in turn affect the effect of subsequent image segmentation. Therefore, the present application first performs image preprocessing on the optic nerve image to enhance the image quality, making it easier for the model to learn useful features. One of the specific preprocessing steps is to perform image contrast enhancement on the optic nerve image to obtain a contrast-enhanced optic nerve image. Contrast enhancement is an important preprocessing step to improve image quality and feature recognition. It aims to improve the distribution of different gray levels in the image, making the difference between the optic nerve structure and the surrounding tissues more obvious, thus contributing to subsequent segmentation and recognition tasks. In the field of medical image analysis, especially when diagnosing ophthalmic diseases, the original optic nerve image may exhibit low contrast due to factors such as the acquisition device, lighting conditions, or the patient's eye condition. Low contrast makes it difficult to accurately capture the tiny structures in the image, affecting the doctor's judgment of the lesion area. Therefore, through appropriate contrast enhancement techniques, the boundaries and details of the optic nerve and its surrounding tissues can be effectively highlighted, providing a basis for more accurate image segmentation.

[0033] Commonly used contrast enhancement methods include Histogram Equalization, Adaptive Histogram Equalization (AHE), local contrast adjustment, etc. These methods have their own characteristics, and the specific implementation method needs to consider the specific situation of the image and the expected effect. Histogram Equalization is a classic and effective global contrast enhancement method. It expands the dynamic range of the image by redistributing the probability density function of the image pixel values, making the overall brightness distribution of the image more uniform, thereby increasing the visual contrast. However, simply applying global histogram equalization may cause some areas to be too bright or too dark, especially when there are large gray level changes in the image. Therefore, in actual operation, it is often necessary to combine other strategies or use improved versions of techniques such as CLAHE (Contrast Limited Adaptive Histogram Equalization) to avoid the negative effects brought by over-enhancement.

[0034] Adaptive Histogram Equalization (AHE) divides an image into multiple small blocks and performs histogram equalization independently for each block. This method can better preserve the local characteristics of the image and is especially suitable for images with complex textures or uneven illumination conditions. However, if directly applied to the entire image, it may lead to unnatural transitions between adjacent blocks, resulting in the so-called "seam effect". To address this, some smoothing mechanisms are usually introduced, such as using weighted averaging to fuse the results of adjacent blocks to ensure the quality of the final output image. In addition to the above histogram-based methods, operations based on the spatial domain can also be used for local contrast adjustment. Such methods mainly operate on the neighborhood around each pixel point, such as calculating its mean, variance, or other statistics, and then adjusting the brightness value of the pixel point according to the set rules. This way can preserve the spatial information of the image to a certain extent while achieving a better contrast enhancement effect. Furthermore, multi-scale analysis or multi-resolution representation can be combined to further optimize the effect of local contrast adjustment. The specific implementation process is as follows: Load the optic nerve image data to be processed from the PACS system or other storage media; If the original image is a color image, it is first converted to a single-channel grayscale image because contrast enhancement operations are usually performed on grayscale values; Count the frequency distribution of grayscale values of all pixel points in the image to form an initial grayscale histogram; Based on the grayscale histogram, calculate the cumulative distribution function CDF (Cumulative Distribution Function) corresponding to each grayscale level, which represents the proportion of all pixel numbers less than or equal to the current grayscale level to the total pixel number; Use the CDF to establish a new grayscale mapping relationship, that is, replace the grayscale value of each pixel point in the original image with its corresponding new grayscale value. This can make the dark areas in the image brighter and the bright areas darker, but overall increase the distinction between different grayscale levels; To avoid artificial artifacts caused by extreme brightness changes, a reasonable contrast enhancement upper limit can be set according to clinical needs to prevent image distortion; Finally, apply the grayscale values after contrast enhancement back to the original image to generate the optic nerve image with enhanced contrast.

[0035] To ensure that the enhancement effect is both significant and not overly exaggerated, a parameter adjustment mechanism can also be introduced to allow users to flexibly adjust the enhancement intensity according to the actual situation. For example, in some cases, doctors may want to retain more information of the original image, and in this case, a milder enhancement mode can be selected; while in other cases, a stronger contrast enhancement may be required to highlight specific structures. In short, by carefully designing and reasonably configuring the contrast enhancement algorithm, not only can the quality of optic nerve images be effectively improved, but also higher-quality input data can be provided for subsequent AI model training, thereby improving the performance of the entire optic nerve image segmentation system. During this process, the authenticity and clinical applicability of the images should always be concerned to ensure that the enhanced images still faithfully reflect the actual situation of the patients.

[0036] Specifically, in S22, a Gaussian filter is used to perform image denoising on the contrast-enhanced optic nerve image to obtain a denoised optic nerve image. In a specific example of the present application, first, based on the contrast-enhanced optic nerve image, the recommended standard deviation of the Gaussian filter is determined. Considering that random factors in the imaging process may introduce unwanted high-frequency information or noise points, therefore, the present application further uses a Gaussian filter to perform denoising processing on the contrast-enhanced optic nerve image. The Gaussian filter is a low-pass filter that can effectively smooth the image and reduce the influence of high-frequency noise in the image. Compared with other types of filters, the Gaussian filter can better maintain the important edges and structural information in the image without blurring while removing noise. In particular, considering that in the process of applying the Gaussian filter to remove image noise, the selection of the standard deviation of the Gaussian filter has a decisive influence on the final image processing effect. If the standard deviation is set too small, the noise may not be effectively removed; on the contrary, if the standard deviation is set too large, the image may be overly smoothed and important detail information may be lost. Therefore, when setting the standard deviation of the Gaussian filter, the noise level and information density of the image need to be comprehensively considered. Based on this, the present application proposes a Gaussian filter parameter optimization method, which intelligently recommends the standard deviation of the Gaussian filter by analyzing the image quality of the optic nerve image, so as to achieve adaptive denoising processing of optic nerve images with different qualities.

[0037] Specifically, first, the contrast-enhanced optic nerve image is subjected to image block processing to obtain a sequence of contrast-enhanced optic nerve image patches. In the technical solution of the present application, in order to more finely capture the quality characteristics of the contrast-enhanced optic nerve image, the contrast-enhanced optic nerve image is subjected to image block processing to refine the granularity of image quality analysis. In this way, by dividing the complete image into a sequence of contrast-enhanced optic nerve image patches, it is helpful to more deeply capture the detail characteristics of the local area of the image, thereby providing richer and more accurate information for subsequent standard deviation recommendation.

[0038] Next, image feature extraction is performed on each of the contrast-enhanced optic nerve image patches in the sequence of contrast-enhanced optic nerve image patches to obtain a sequence of semantic coding feature vectors of the local optic nerve region images. Those of ordinary skill in the art should be aware that MobileNet is a lightweight deep learning model that can effectively reduce the number of model parameters and computational complexity while maintaining a high feature extraction accuracy by using depthwise separable convolutions for image processing. Based on this, in the technical solution of this application, the MobileNet model is used to extract the visual patterns and structural information of each contrast-enhanced optic nerve image patch, which can effectively reduce the consumption of computing resources and achieve an initial characterization of the quality of each local region image.

[0039] In a specific example, MobileNet decomposes the traditional standard convolution into two steps by introducing the depthwise separable convolution technique: first, a per-channel convolution, and then a pointwise convolution. This decomposition significantly reduces the number of parameters and computational complexity, enabling MobileNet to significantly reduce the demand for computing resources while maintaining a high level of accuracy. For application scenarios that need to run on different devices and have performance requirements, such as optic nerve image analysis, MobileNet is a very ideal choice.

[0040] To better capture the detailed features of different regions in the optic nerve image, the system divides the original image into blocks. That is, the entire image is divided into multiple smaller rectangular regions or "patches", and each patch represents a local part of the image. This can refine the granularity of the image quality analysis and provide richer and more accurate information for subsequent standard deviation recommendations. At the same time, by processing these small patches separately, the computational burden caused by the large image size can be alleviated to a certain extent. Next, each contrast-enhanced optic nerve image patch is fed into the MobileNet model as input. Before this, some normalization processes may be required for the patches, such as resizing to meet the input requirements of MobileNet or normalizing the pixel values to ensure a consistent numerical range. Such preprocessing helps to improve the generalization ability and prediction accuracy of the model.

[0041] After entering MobileNet, the patch data will sequentially pass through a series of depthwise separable convolutional layers. These layers are responsible for extracting rich visual patterns and structural information from the input data. Specifically, the first step of depthwise separable convolution is to perform channel-wise convolution, that is, applying a small filter independently to each input channel; the second step is to combine the results of all channels through pointwise convolution to form a new feature map. Compared with traditional convolution, this method greatly reduces the computational amount and memory occupancy, while retaining sufficient expressive power to capture important image features. After each round of convolution, a non-linear activation function (such as ReLU) is usually added to introduce non-linearity into the model. This step is crucial for breaking linear relationships and enhancing the expressiveness of the model. Non-linear activation enables the network to learn more complex feature representations, thus more accurately reflecting the subtle changes in the optic nerve images.

[0042] In addition to convolutional layers, MobileNet also includes pooling layers. Pooling operations are used to reduce the spatial dimension of the feature map, further reducing the computational cost and helping the model focus on the most important information. Common pooling methods include max pooling and average pooling. Max pooling selects the maximum value within each window as the output, while average pooling takes the average value. Both methods can effectively reduce redundant information and improve the robustness and generalization ability of the model. After multiple rounds of convolution and pooling operations, the feature maps generated by the last few convolutional layers are flattened and fed into fully connected layers. The role of fully connected layers is to integrate all local features and generate a compact but information-rich feature vector. This feature vector comprehensively represents the main visual patterns and structural information of each optic nerve image patch, constituting a preliminary characterization of the image quality of this local area. To adapt to different application scenarios, the number and dimension size of fully connected layers can also be adjusted as needed.

[0043] Finally, the MobileNet model will generate a corresponding feature vector for each input optic nerve image patch. These feature vectors are then collected to form a complete sequence, which is used to describe the different local characteristics of the entire optic nerve image. Based on this feature vector sequence, the overall properties of the image can be further analyzed through various advanced algorithms (such as clustering, classification, etc.), or directly used for subsequent tasks, such as Gaussian filter parameter optimization, semantic segmentation, etc.

[0044] Benefiting from the design of depthwise separable convolution, MobileNet can significantly reduce the consumption of computing resources while ensuring good performance. In addition, by simply modifying the network architecture or parameter settings, MobileNet can easily adapt to different types and scales of datasets to meet diverse application requirements. Despite its relatively simple structure, MobileNet still maintains a high feature extraction accuracy and can effectively capture the key details in optic nerve images, providing a solid foundation for subsequent processing.

[0045] Subsequently, perform feature aggregation based on kinetic analysis on the sequence of semantic encoding feature vectors of the local region image of the optic nerve to obtain an initial image quality significantly aggregated encoding vector. By performing feature aggregation on the sequence of semantic encoding feature vectors of the local region image of the optic nerve, the overall quality of the contrast-enhanced optic nerve image can be comprehensively characterized. In particular, in order to enhance the understanding of the overall internal feature structure of the contrast-enhanced optic nerve image, the present application proposes a feature aggregation method based on kinetic analysis, which evaluates the distribution of each local image patch feature in the high-dimensional feature space by simulating the kinetic principle of the physical world, learns the feature association pattern between each local image patch, and thereby performs feature integration to achieve a comprehensive representation of the quality of the entire optic nerve image.

[0046] In an embodiment of the present application, the specific steps for performing feature aggregation based on kinetic analysis on the sequence of semantic coding feature vectors of the local region image of the optic nerve include: First, based on the feature static potential energy distribution of the sequence of semantic coding feature vectors of the local region image of the optic nerve, calculate the pseudo-anchoring aggregation center representation vector of the local region image of the optic nerve. That is, first input each semantic coding feature vector of the local region image of the optic nerve in the sequence of semantic coding feature vectors of the local region image of the optic nerve into the feature static potential energy measurement network to obtain a sequence of feature static potential energy measurement coefficients of the local region image of the optic nerve; here, the feature static potential energy measurement network is used to evaluate the potential energy level of each semantic coding feature vector of the local region image of the optic nerve under static conditions, so as to reveal the importance and influence of each local region in the current global feature distribution of the optic nerve image, as well as the internal connection between each local region, thereby obtaining a sequence of feature static potential energy measurement coefficients of the local region image of the optic nerve. Then, input the sequence of feature static potential energy measurement coefficients of the local region image of the optic nerve into the feature energy level screening gating unit to obtain a sequence of feature static potential energy weight factors of the local region image of the optic nerve; that is, perform gating masking processing on the obtained sequence of feature static potential energy measurement coefficients of the local region image of the optic nerve through the feature energy level screening gating unit to screen out the local region features that contribute more to the overall image quality representation, ensure that the local patch features with higher potential energy values, that is, larger potential influences, are retained, and assign corresponding weight factors, thereby reducing the risk of overfitting. Furthermore, based on the sequence of feature static potential energy weight factors of the local region image of the optic nerve, calculate the weighted sum of the sequence of semantic coding feature vectors of the local region image of the optic nerve to obtain the pseudo-anchoring aggregation center representation vector of the local region image of the optic nerve; that is, perform weighted aggregation on the sequence of semantic coding feature vectors of the local region image of the optic nerve based on the obtained weight factors to reveal the feature distribution trend center of each local patch of the optic nerve image, and generate the pseudo-anchoring aggregation center representation vector of the local region image of the optic nerve. In this way, the overall quality level of the optic nerve image can be accurately reflected, thereby providing effective guidance for the subsequent local feature aggregation process.

[0047] In a specific example, based on the feature static potential energy distribution of the sequence of semantic coding feature vectors of the local region image of the optic nerve, calculate the pseudo-anchoring aggregation center representation vector of the local region image of the optic nerve with the following formula; where the formula is:

[0048] V = {v 1 , v 2 ,..., v i ,..., v n}

[0049]

[0050]

[0051]

[0052]

[0053] Among them, V represents the sequence of semantic coding feature vectors of the local region image of the optic nerve, and v 1 、v 2 、v i and v n respectively represent the first, second, ith, and nth semantic coding feature vectors of the local region image of the optic nerve in the sequence of semantic coding feature vectors of the local region image of the optic nerve. n is the number of feature vectors in the sequence of semantic coding feature vectors of the local region image of the optic nerve. represents the eigenvalue at the ith position in the ith semantic coding feature vector of the local region image of the optic nerve, μ i and σ i respectively represent the feature mean and feature variance of the ith semantic coding feature vector of the local region image of the optic nerve. L is the feature scale value of the semantic coding feature vector of the local region image of the optic nerve, and e i is the static potential energy metric coefficient of the ith semantic coding feature vector of the local region image of the optic nerve, w i represents the static potential energy metric coefficient of the normalized local region image feature of the ith optic nerve, and mask(·) represents the gating mask function. represents the gating threshold, w si represents the static potential energy weight factor of the local region image feature of the ith optic nerve, and v a represents the pseudo-anchoring aggregation center representation vector of the local region image of the optic nerve.

[0054] Next, calculate the aggregation movement direction of each optic nerve local region image semantic coding feature vector in the sequence of optic nerve local region image semantic coding feature vectors relative to the optic nerve local region image pseudo-anchored aggregation center representation vector to obtain a sequence of optic nerve local region image semantic aggregation movement directions. Here, by calculating the aggregation movement direction of each optic nerve local region image semantic coding feature vector relative to the optic nerve local region image pseudo-anchored aggregation center representation vector, the change trend of each local patch feature relative to the center point is quantified, revealing the dynamic relationship and its evolution path within the optic nerve image feature space, so as to facilitate the identification of the relative relationship between features and the group behavior pattern. More specifically, the specific process of calculating the aggregation movement direction of each optic nerve local region image semantic coding feature vector in the sequence of optic nerve local region image semantic coding feature vectors relative to the optic nerve local region image pseudo-anchored aggregation center representation vector includes: calculating the value of the inverse cosine function between each optic nerve local region image semantic coding feature vector in the sequence of optic nerve local region image semantic coding feature vectors and the optic nerve local region image pseudo-anchored aggregation center representation vector as the aggregation movement direction to obtain the sequence of optic nerve local region image semantic aggregation movement directions.

[0055] In a specific example, calculate the aggregation movement direction of each optic nerve local region image semantic coding feature vector in the sequence of optic nerve local region image semantic coding feature vectors relative to the optic nerve local region image pseudo-anchored aggregation center representation vector with the following formula to obtain a sequence of optic nerve local region image semantic aggregation movement directions; where the formula is:

[0056]

[0057] where ||·|| represents the norm of the vector, arccos(·) is the inverse cosine function, and θ i represents the aggregation movement direction of the i-th optic nerve local region image semantic coding feature vector relative to the optic nerve local region image pseudo-anchored aggregation center representation vector.

[0058] Subsequently, based on the sequence of the semantic aggregation running directions of the local optic nerve region images, the sequence of the semantic encoding feature vectors of the local optic nerve region images is dynamically aggregated towards the pseudo-anchored aggregation center representation vector of the local optic nerve region images to obtain the initial image quality significantly aggregated encoding vector. That is, based on the aggregation running directions and the feature static potential energy metric coefficients of each local optic nerve region image semantic encoding feature vector, each local optic nerve region image semantic encoding feature vector is dynamically aggregated towards the pseudo-anchored aggregation center representation vector of the local optic nerve region images to obtain the initial image quality significantly aggregated encoding vector, realizing the dynamic characterization of the overall quality of the optic nerve images. In this way, not only can the feature information of each local region of the optic nerve image be effectively integrated, but also the mutual correlations and overall structure among the feature of each local region image can be revealed through the dynamic aggregation process, thereby providing more comprehensive and in-depth data support for the quality analysis of the optic nerve image and the recommendation of the standard deviation of the Gaussian filter.

[0059] In a specific example, based on the sequence of the semantic aggregation running directions of the local optic nerve region images, the sequence of the semantic encoding feature vectors of the local optic nerve region images is dynamically aggregated towards the pseudo-anchored aggregation center representation vector of the local optic nerve region images by the following formula to obtain the initial image quality significantly aggregated encoding vector; where the formula is:

[0060]

[0061] where sin(·) represents the sine function, W i and b i respectively represent the weight matrix and the bias term, and X represents the initial image quality significantly aggregated encoding vector.

[0062] Finally, based on the initial image quality, the encoding vectors are significantly aggregated to generate the recommended standard deviation of the Gaussian filter. That is, in a specific example of the present application, the initial image quality significantly aggregated encoding vector is input into the Gaussian filter parameter mapping module based on the decoder to obtain the recommended standard deviation of the Gaussian filter. In the technical solution of the present application, the decoder adopts a multi-layer perceptron (MLP) structure, and through the training process, it learns the complex mapping relationship between the image features and the standard deviation of the Gaussian filter, so as to be able to provide adaptive filtering parameters for different optic nerve image qualities to achieve the best noise reduction effect. Based on this, after receiving the initial image quality significantly aggregated encoding vector, the decoder captures the deep image quality feature information through multi-layer non-linear transformation of it, and combines the mapping relationship learned in the training process to map it into the standard deviation parameter space of the Gaussian filter, thereby outputting the recommended standard deviation value for the current optic nerve image quality, so as to guide the Gaussian filter to effectively remove noise while retaining the key detail information in the image. In this way, high-quality noise reduction processing of the optic nerve image can be realized, thus providing clearer and more accurate image data for subsequent medical diagnosis and analysis.

[0063] In the case where each optic nerve local region image semantic encoding feature vector in the sequence of optic nerve local region image semantic encoding feature vectors respectively represents the image semantic encoding features of the local region determined by image block processing of the contrast-enhanced optic nerve image, when performing feature aggregation based on feature dynamics analysis on it, the initial image quality significantly aggregated encoding vector will also have distribution diversity of feature aggregation due to the image representation difference of the local region of the source image. Therefore, when the initial image quality significantly aggregated encoding vector performs decoding probability regression through the Gaussian filter parameter mapping module based on the decoder, it will affect the accuracy of the decoding result.

[0064] Based on this, when performing decoding probability regression on the initial image quality significantly aggregated encoding vector through the Gaussian filter parameter mapping module based on the decoder, first optimize the initial image quality significantly aggregated encoding vector, and the optimization process specifically includes:

[0065] Calculate the sum of the absolute values w of all eigenvalues of the initial image quality significantly aggregated encoding vector 1 , and the square root w of the sum of the squares of all eigenvalues of the initial image quality significantly aggregated encoding vector 2 , and divide w 2 by w 1 to obtain the initial image quality significantly aggregated encoding probability value p, that is:

[0066] w 1 = ∑ i |vi |

[0067]

[0068]

[0069] wherein, v i represents the i-th eigenvalue of the initial image quality significant aggregation coding vector;

[0070] Perform maximum normalization on each eigenvalue in the initial image quality significant aggregation coding vector V to obtain the initial image quality significant aggregation coding probability vector V':

[0071]

[0072] wherein, v max represents the maximum value in the initial image quality significant aggregation coding vector, and v' i represents the i-th eigenvalue of the initial image quality significant aggregation coding probability vector;

[0073] After subtracting the initial image quality significant aggregation coding probability vector from the initial image quality significant aggregation coding probability value pointwise, further multiply pointwise by the reciprocal of the initial image quality significant aggregation coding probability value to obtain the initial image quality significant aggregation coding domain gradient vector wherein, represents pointwise subtraction, and ⊙ represents pointwise multiplication;

[0074] After subtracting the initial image quality significant aggregation coding probability vector from the unit eigenvector pointwise, perform pointwise division with the initial image quality significant aggregation coding domain gradient vector to obtain the initial image quality significant aggregation coding differential vector, that is:

[0075]

[0076] wherein, V I represents the unit eigenvector, V 1 ⊙-1 represents the element-wise reciprocal of the initial image quality significant aggregation coding domain gradient vector, that is, calculate the reciprocal of each eigenvalue of the initial image quality significant aggregation coding domain gradient vector, V 2 represents the initial image quality significant aggregation coding differential vector;

[0077] Calculate the power function of the initial image quality significant aggregation coding domain gradient vector with the initial image quality significant aggregation coding probability value as the exponent V 1 ⊙p, and perform a weighted sum with the exponential function exp(V 2 ) of the initial image quality significant aggregation coding differential vector with the base of the natural constant to obtain an optimized initial image quality significant aggregation coding vector, that is:

[0078]

[0079] where α and β represent weighted hyperparameters, represents point addition, and V o represents the optimized initial image quality significant aggregation coding vector.

[0080] That is, through the probability density domain gradient discrete differential corresponding to the initial image quality significant aggregation coding vector, construct the mutually exclusive generalization representation behavior unit of the initial image quality significant aggregation coding vector, and based on the different behavior unit organization spaces under the non-uniform topological structure of the initial image quality significant aggregation coding vector, enhance the static response characteristics of the short-range correlation mesoscopic information configuration of the initial image quality significant aggregation coding vector for the generative probability generalization representation behavior, so as to ensure the steady-state convergence of the optimization process when mapping between the generation target and the extracted features in the feature space - generation probability, and improve the accuracy of the recommended standard deviation of the Gaussian filter obtained through the Gaussian filter parameter mapping module based on the decoder.

[0081] Furthermore, after obtaining the recommended standard deviation, update the current standard deviation of the Gaussian filter to the recommended standard deviation to obtain an updated Gaussian filter; and use the updated Gaussian filter to perform Gaussian filtering on the contrast-enhanced optic nerve image to obtain the denoised optic nerve image. Specifically, replace the standard deviation parameter of the previously used Gaussian filter with the newly calculated recommended standard deviation. This means that the Gaussian filter will now run with a new parameter configuration to more precisely match the specific requirements of the current image. Perform Gaussian filtering on the previously contrast-enhanced optic nerve image using the updated Gaussian filter. At this stage, the Gaussian filter will traverse each pixel point in the image and calculate a new pixel value based on the weighted average of the pixel values in its neighborhood. The weights are defined by the Gaussian function, with its center at the currently processed pixel position, and pixels farther away from the center contribute less. In this way, the high-frequency noise components in the image are effectively suppressed, but the main structure and boundary information of the image are retained. After completing Gaussian filtering, the system will output a denoised optic nerve image. This image has better visual quality and lower noise level compared to the original image, which is beneficial for subsequent image segmentation and recognition tasks. In addition, due to the method of dynamically adjusting the standard deviation, even for optic nerve images from different sources and qualities, the system can provide consistent and high-quality denoising results, improving the robustness and generalization ability of the overall processing flow.

[0082] Specifically, in S23, the image size of the denoised optic nerve image is adjusted to obtain the preprocessed optic nerve image. Considering that when processing optic nerve images acquired from different sources or devices, the sizes of the original images may vary. Therefore, in this application, further through an image scaling algorithm, the denoised optic nerve image is adjusted to a unified size standard to ensure the consistency and compatibility of the images under different devices and application scenarios, thus completing the preprocessing process of the optic nerve image. In the embodiments of this application, the image scaling algorithm adopts bilinear interpolation technology.

[0083] Particularly, in S3, the preprocessed optic nerve image is input into an optic nerve image feature extractor based on a dynamic convolutional token model to obtain an optic nerve image semantic encoding feature matrix. That is, in the technical solution of this application, image feature extraction is performed on the preprocessed optic nerve image for subsequent image segmentation and recognition tasks. In the technical solution of this application, in order to take into account both the global and local information of the preprocessed optic nerve image, an optic nerve image feature extractor based on a dynamic convolutional token model is used to process the preprocessed optic nerve image to obtain an optic nerve image semantic encoding feature matrix. Specifically, the Dynamic Convolutional Transformer Model (DynamicViT) is a new type of vision Transformer model that combines a convolutional neural network (CNN) and a Transformer model. First, the preprocessed optic nerve image is divided into multiple local blocks, and image feature extraction is performed on each local block using a CNN model. Then, the image features of each local block are combined into a sequence form, and a Transformer model is used to capture the long-range dependence relationships between different local features for global information interaction processing. Thus, it can comprehensively utilize the local feature extraction ability of the CNN model and the global information processing ability of the Transformer model to effectively capture the detailed features of each local patch in the preprocessed optic nerve image and the global context semantic information of the image, providing a richer and more comprehensive image feature representation for subsequent image segmentation and recognition.

[0084] In particular, in step S4, the semantic encoding feature matrix of the optic nerve image is input into an image semantic segmenter based on the Softmax function to obtain an image semantic segmentation result. In particular, the Softmax function can convert each element in the semantic encoding feature matrix of the optic nerve image into a probability value, indicating the likelihood that the element belongs to a specific semantic category. The image semantic segmenter uses these probability values to determine the semantic category of each pixel point, thereby achieving pixel-level semantic segmentation of the entire optic nerve image. In this way, different tissue structures and lesion areas in the optic nerve image can be clearly distinguished, providing more intuitive and detailed diagnostic information for doctors.

[0085] In summary, the method for segmenting and recognizing optic nerve images based on an AI model according to an embodiment of the present application is elucidated. It performs semantic segmentation processing on the optic nerve images of patient objects by using image processing techniques based on deep learning. First, the contrast of the optic nerve images of patient objects is enhanced, Gaussian filtering is performed, and the image size is adjusted to improve the image quality and highlight the optic nerve structure features. Then, a neural network model is used to extract semantic features and perform pixel-level classification on the preprocessed optic nerve images, thereby achieving accurate segmentation and recognition of the optic nerve images. In particular, during the Gaussian filtering of the optic nerve images, the standard deviation of the Gaussian filter is dynamically adjusted to cope with the quality differences of different images, thereby optimizing the noise reduction effect of the images. In this way, the accuracy of optic nerve image segmentation can be effectively improved, and the lesion areas of the optic nerve can be identified, providing more accurate auxiliary information for the clinical diagnosis of eye disease patients.

[0086] Furthermore, a system for segmenting and recognizing optic nerve images based on an AI model is also provided.

[0087] Figure 4 FIG. is a block diagram of a system for segmenting and recognizing optic nerve images based on an AI model according to an embodiment of the present application. As Figure 4 shown, the system 300 for segmenting and recognizing optic nerve images based on an AI model according to an embodiment of the present application includes: an image retrieval module 310 for retrieving the optic nerve images of a target patient object; an image preprocessing module 320 for preprocessing the optic nerve images of the target patient object to obtain preprocessed optic nerve images; an optic nerve image feature extraction module 330 for inputting the preprocessed optic nerve images into an optic nerve image feature extractor based on a dynamic convolution labeling model to obtain a semantic encoding feature matrix of the optic nerve image; and an image semantic segmentation module 340 for inputting the semantic encoding feature matrix of the optic nerve image into an image semantic segmenter based on the Softmax function to obtain an image semantic segmentation result.

[0088] As described above, the optic nerve image segmentation and recognition system 300 based on an AI model according to an embodiment of the present application can be implemented in various wireless terminals, such as a server with an optic nerve image segmentation and recognition algorithm based on an AI model. In a possible implementation, the optic nerve image segmentation and recognition system 300 based on an AI model according to an embodiment of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the optic nerve image segmentation and recognition system 300 based on an AI model can be a software module in the operating system of the wireless terminal, or can be an application developed for the wireless terminal; of course, the optic nerve image segmentation and recognition system 300 based on an AI model can also be one of the many hardware modules of the wireless terminal.

[0089] Alternatively, in another example, the optic nerve image segmentation and recognition system 300 based on an AI model and the wireless terminal can also be separate devices, and the optic nerve image segmentation and recognition system 300 based on an AI model can be connected to the wireless terminal through a wired and / or wireless network and transmit interaction information in accordance with a predefined data format.

[0090] The embodiments of the present disclosure have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of the technology in the market, or to enable other ordinary skill in the art in the technical field to understand the embodiments disclosed herein.

Claims

1. A optic nerve image segmentation and recognition method based on an AI model, characterized in that: include: Retrieving an optic nerve image of a target patient; Preprocessing the optic nerve image of the target patient to obtain a preprocessed optic nerve image; Inputting the preprocessed optic nerve image into a optic nerve image feature extractor based on a dynamic convolutional labeling model to obtain a optic nerve image semantic encoding feature matrix; The optic nerve image semantic encoding feature matrix is ​​input into an image semantic segmentation device based on a Softmax function to obtain an image semantic segmentation result.

2. The optic nerve image segmentation and recognition method based on the AI ​​model according to claim 1, characterized in that: Preprocessing the optic nerve image of the target patient to obtain a preprocessed optic nerve image includes: performing image contrast enhancement on the optic nerve image to obtain a contrast-enhanced optic nerve image; Using a Gaussian filter to perform image noise reduction on the contrast-enhanced optic nerve image to obtain a noise-reduced optic nerve image; The image size of the noise-reduced optic nerve image is adjusted to obtain the preprocessed optic nerve image.

3. The optic nerve image segmentation and recognition method based on the AI ​​model according to claim 2 is characterized in that: Using a Gaussian filter to perform image noise reduction on the contrast-enhanced optic nerve image to obtain a noise-reduced optic nerve image, comprising: determining a recommended standard deviation of the Gaussian filter based on the contrast-enhanced optic nerve image; Updating the current standard deviation of the Gaussian filter to the recommended standard deviation to obtain an updated Gaussian filter; The updated Gaussian filter is used to perform Gaussian filtering on the contrast-enhanced optic nerve image to obtain the noise-reduced optic nerve image.

4. The optic nerve image segmentation and recognition method based on the AI ​​model according to claim 3 is characterized in that: Determining a recommended standard deviation of the Gaussian filter based on the contrast-enhanced optic nerve image includes: Performing image block processing on the contrast-enhanced optic nerve image to obtain a sequence of contrast-enhanced optic nerve image blocks; Performing image feature extraction on each contrast-enhanced optic nerve image block in the sequence of contrast-enhanced optic nerve image blocks to obtain a sequence of optic nerve local area image semantic coding feature vectors; Performing feature aggregation based on dynamic analysis on the sequence of semantic coding feature vectors of the local area image of the optic nerve to obtain a significant aggregated coding vector of initial image quality; A recommended standard deviation of the Gaussian filter is generated based on the initial image quality significant aggregation code vector.

5. The optic nerve image segmentation and recognition method based on the AI ​​model according to claim 4 is characterized in that: Performing image feature extraction on each contrast-enhanced optic nerve image block in the sequence of contrast-enhanced optic nerve image blocks to obtain a sequence of optic nerve local area image semantic coding feature vectors, including: Each contrast-enhanced optic nerve image block in the sequence of contrast-enhanced optic nerve image blocks is input into a optic nerve image feature extractor based on MobileNet to obtain a sequence of semantic encoding feature vectors of the optic nerve local area image.

6. The optic nerve image segmentation and recognition method based on the AI ​​model according to claim 5, characterized in that: Performing feature aggregation based on dynamic analysis on the sequence of semantic coding feature vectors of the local area image of the optic nerve to obtain an initial image quality significant aggregation coding vector, including: Calculating a pseudo-anchored aggregation center representation vector of the local area image of the optic nerve based on a characteristic static potential energy distribution of a sequence of semantically encoded feature vectors of the local area image of the optic nerve; Calculating the aggregation movement direction of each optic nerve local area image semantic encoding feature vector in the sequence of the optic nerve local area image semantic encoding feature vector relative to the optic nerve local area image pseudo anchor aggregation center representation vector to obtain a sequence of optic nerve local area image semantic aggregation running directions; Based on the sequence of the semantic aggregation running directions of the local area image of the optic nerve, the sequence of the semantic coding feature vectors of the local area image of the optic nerve is dynamically aggregated toward the pseudo-anchored aggregation center representation vector of the local area image of the optic nerve to obtain the initial image quality significant aggregation coding vector.

7. The optic nerve image segmentation and recognition method based on the AI ​​model according to claim 6, characterized in that: Based on the characteristic static potential energy distribution of the sequence of semantic encoding feature vectors of the local area image of the optic nerve, a pseudo-anchored aggregation center representation vector of the local area image of the optic nerve is calculated, including: Inputting each optic nerve local area image semantic encoding feature vector in the sequence of optic nerve local area image semantic encoding feature vectors into a feature static potential energy measurement network to obtain a sequence of optic nerve local area image feature static potential energy measurement coefficients; Inputting the sequence of static potential energy measurement coefficients of the local area image features of the optic nerve into a feature energy level screening gating unit to obtain a sequence of static potential energy weight factors of the local area image features of the optic nerve; Based on the sequence of static potential weight factors of the optic nerve local area image features, the weighted sum of the sequence of semantic encoding feature vectors of the optic nerve local area image is calculated to obtain the pseudo-anchored aggregation center representation vector of the optic nerve local area image.

8. The method for segmenting and recognizing optic nerve images based on an AI model according to claim 7, characterized in that: Calculating the aggregation movement direction of each optic nerve local area image semantic encoding feature vector in the sequence of the optic nerve local area image semantic encoding feature vector relative to the optic nerve local area image pseudo anchor aggregation center representation vector to obtain a sequence of optic nerve local area image semantic aggregation running directions, including: The arc cosine function value between each semantic coding feature vector of the local area image of the optic nerve in the sequence of the semantic coding feature vector of the local area image of the optic nerve and the pseudo-anchor aggregation center representation vector of the local area image of the optic nerve is calculated as the aggregation movement direction to obtain the sequence of the semantic aggregation running directions of the local area image of the optic nerve.

9. The method for segmenting and recognizing optic nerve images based on an AI model according to claim 8, characterized in that: Generating a recommended standard deviation of the Gaussian filter based on the initial image quality significant aggregation coding vector, including: The initial image quality significant aggregation coding vector is input into a decoder-based Gaussian filter parameter mapping module to obtain a recommended standard deviation of the Gaussian filter.

10. A optic nerve image segmentation and recognition system based on an AI model, characterized in that: include: An image acquisition module, used to acquire an optic nerve image of a target patient; An image preprocessing module, used for preprocessing the optic nerve image of the target patient to obtain a preprocessed optic nerve image; A optic nerve image feature extraction module, used for inputting the preprocessed optic nerve image into a optic nerve image feature extractor based on a dynamic convolutional labeling model to obtain a optic nerve image semantic encoding feature matrix; The image semantic segmentation module is used to input the optic nerve image semantic encoding feature matrix into an image semantic segmentor based on a Softmax function to obtain an image semantic segmentation result.

Images