Retinochoroidal disease course structure change monitoring system based on artificial intelligence

Abstract

The invention relates to the technical field of medical image processing and artificial intelligence, and discloses an artificial intelligence-based retinochoroid disease course structure change monitoring system, which comprises an image acquisition unit, a feature extraction unit, a lesion segmentation unit and the like. The image acquisition unit acquires retina optical coherence tomography sequence image data and fundus color image data; the feature extraction unit extracts a retina choroidal structure feature tensor through a three-dimensional convolutional neural network; the lesion segmentation unit outputs a lesion area probability distribution diagram by using a U-Net segmentation network; the time sequence alignment unit is used for registering the multi-time-point images to generate a displacement change matrix; the dynamic analysis unit extracts related indexes; an abnormal scoring unit constructs a disease course progress score; the decision grading unit outputs disease course stage classification labels; the multi-modal fusion unit fuses the multi-modal features; the report generation unit generates a structured disease course monitoring report. And favorable support is provided for diagnosis and treatment of ophthalmic diseases and illness state tracking.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing and artificial intelligence technology, and in particular to an artificial intelligence-based retinal choroidal disease course structure change monitoring system. Background Art

[0002] Since 2013, central serous chorioretinopathy (CSC) has been redefined as choroidal vascular / Bruch's membrane disease or choroidal disease in authoritative fundus disease literature and is considered a common manifestation of pachychoroidopathy. The pathogenesis of CSC involves dilation, stasis, remodeling of choroidal vessels, and abnormal connections between vortex veins, leading to increased choriocapillary venous pressure, which in turn damages blood vessels, retinal pigment epithelium, and outer retinal barrier, causing fluid leakage in the macular area. However, the specific risk factors or causes of choroidal vascular abnormalities have not yet been clarified.

[0003] CSC is common in young and middle-aged people (especially men), who are the backbone of families in contemporary society. CSC is one of the common causes of vision loss in young and middle-aged people, especially in high-pressure occupational groups (such as white-collar workers, IT industry personnel, etc.). Patients often have limited work ability due to visual distortion and decreased dark adaptation, and even cause anxiety or depression. As the course of the disease prolongs, chronic and recurrent CSC requires repeated treatment, increasing the medical burden. Long-term treatment of the disease may increase the social, psychological and economic burden on patients and their families. Early intervention through visualization research methods to reduce social productivity losses is crucial to protecting the vision health of the working population.

[0004] CSC often has no obvious symptoms in the early stages. When patients experience obvious vision loss and other symptoms, the disease may have already developed to a more serious stage, making treatment extremely difficult. Therefore, early and accurate monitoring of structural changes in the retinal choroidal disease course is crucial.

[0005] Traditional monitoring methods for retinal and choroidal diseases have many limitations. In terms of image acquisition, the imaging resolution of early devices was limited, making it difficult to clearly capture the subtle structural changes in the retina and choroid. For example, ordinary fundus cameras can only obtain the general morphology of the fundus, and it is difficult to detect the changes in microstructures such as the retinal pigment epithelium layer, choroidal capillary layer, and Bruch's membrane. In terms of analysis methods, it mainly relies on the subjective experience of doctors for judgment. Doctors observe the images and evaluate the condition based on personal experience. However, this method is greatly affected by factors such as the doctor's professional level and fatigue, and there may be significant differences in the diagnostic results among different doctors, lacking objectivity and accuracy. Moreover, traditional methods are difficult to quantitatively analyze the development trend of diseases and cannot accurately predict the progression speed and severity of diseases.

[0006] With the development of technology, optical coherence tomography (OCT) and fundus color imaging technology have been gradually applied to ophthalmic clinics. OCT can provide high-resolution tomographic images of the retina and choroid, clearly showing the structures of each layer; fundus color images can directly present the morphology and distribution of fundus blood vessels. However, a large amount of image data generated by these technologies requires efficient analysis methods. Manually analyzing these image data is not only time-consuming and laborious but also difficult to discover the potential laws in the data.

[0007] The rise of artificial intelligence technology has brought new opportunities for the monitoring of retinal and choroidal diseases. However, the existing artificial intelligence-based monitoring systems still have deficiencies. In terms of feature extraction, some algorithms cannot fully mine the key information in the images, resulting in incomplete feature extraction and affecting subsequent lesion judgment and disease course analysis. The accuracy of lesion segmentation also needs to be improved. Some systems cannot accurately distinguish the lesion area from normal tissues, especially the detection ability for micro-lesions such as microvascular lesions is weak. In addition, the existing monitoring systems lack in comprehensively analyzing multi-modal image data, failing to fully utilize the complementary advantages of OCT and fundus color images and unable to provide comprehensive and accurate basis for clinical diagnosis. Therefore, it is of great clinical significance and application value to develop a more efficient, accurate, and intelligent monitoring system for the structural changes of the retinal and choroidal disease course. Summary of the Invention

[0008] The purpose of the present invention is to provide a monitoring system for the structural changes of the retinal and choroidal disease course based on artificial intelligence to solve the problems raised in the above background technology.

[0009] To achieve the above purpose, the present invention provides the following technical solution: A monitoring system for the structural changes of the retinal and choroidal disease course based on artificial intelligence, the system includes: An image acquisition unit, which is used to acquire the optical coherence tomography sequence image data and fundus color image data of the target patient; A feature extraction unit, which is used to perform hierarchical feature extraction on the retinal optical coherence tomography sequence image data through a three-dimensional convolutional neural network to obtain a retinal choroid structure feature tensor; A lesion segmentation unit, which is used to input the retinal choroid structure feature tensor into a pre-trained U-Net segmentation network and output a probability distribution map of the retinal choroid lesion area; A temporal alignment unit, which is used to perform dynamic registration on multi-time point images in the retinal optical coherence tomography sequence image data to generate a spatio-temporally aligned retinal interlayer displacement change matrix; A dynamic analysis unit, which is used to perform temporal modeling on the spatio-temporally aligned retinal interlayer displacement change matrix based on a long short-term memory network to extract the choroid thickness change rate, blood vessel density attenuation gradient, and lesion area diffusion coefficient; An abnormal score unit, which is used to construct a disease progression score through a weighted fusion algorithm according to the choroid thickness change rate, the blood vessel density attenuation gradient, and the lesion area diffusion coefficient; A decision classification unit, which is used to input the disease progression score into a gradient boosting decision tree model and output a classification label for the retinal choroid disease stage; A multi-modal fusion unit, which is used to perform feature concatenation on the vascular texture features extracted from the fundus color image data by a residual network and the retinal choroid structure feature tensor to generate a multi-modal fusion feature vector; A report generation unit, which is used to integrate the disease stage classification label, the lesion area probability distribution map, and the multi-modal fusion feature vector to generate a structured disease process monitoring report.

[0010] Preferably, the execution steps of the feature extraction unit include: Perform interlayer boundary enhancement processing on the retinal optical coherence tomography sequence image data, and use anisotropic diffusion filtering to eliminate noise interference; Extract local spatial features of the retinal pigment epithelium layer, choroid capillary layer, and Bruch's membrane in the axial, transverse, and temporal dimensions through a three-dimensional convolution kernel; Combine multi-scale feature maps with an attention mechanism module to generate a retinal choroid structure feature tensor containing spatial weight distribution.

[0011] Preferably, the execution steps of the lesion segmentation unit include: Construct a cascaded U-Net network architecture, where the primary network segments the large choroid blood vessel area, and the secondary network focuses on microvascular lesion detection; Introduce a deformable convolution module in the skip connection layer to adapt to the retinal curvature variations of different patients; Adopt a hybrid loss function, combining Dice coefficient loss and edge perception loss to optimize the segmentation boundary accuracy.

[0012] Preferably, the execution steps of the temporal alignment unit include: Calculate the interlayer displacement vector field of adjacent time-point images based on the optical flow field algorithm; Perform rigid registration on multi-time-point images through a Lie group transformation matrix to eliminate eye movement artifacts; Use a non-rigid registration algorithm to correct local retinal deformations and construct a spatio-temporal aligned interlayer displacement change matrix.

[0013] Preferably, the execution steps of the dynamic analysis unit include: Input the interlayer displacement change matrix sliced by time window into a bidirectional long short-term memory network; Extract the standard deviation change rate of choroidal thickness within each time window as a dynamic stability index; Model the topological relationship of the vascular network through a graph convolutional network and calculate the density attenuation gradient of vascular branch points.

[0014] Preferably, the execution steps of the anomaly scoring unit include: Establish a dynamic weight allocation model based on the entropy weight method, and the entropy weight method automatically adjusts the weight coefficient according to the historical variability of each index; Perform Z-score normalization on the choroidal thickness change rate to eliminate the baseline differences between individuals; Adopt fuzzy logic rules to fuse the standardized indexes and generate a disease progression score in the range of 0-1.

[0015] Preferably, the execution steps of the decision-making classification unit include: Construct a gradient boosting decision tree model containing a Shapley value interpretation module, and the module is used to visualize the feature contribution degree; Adopt a stratified sampling strategy to balance the distribution of samples in different disease stages; Introduce KL divergence constraints at the leaf nodes to ensure the interpretability of the classification boundary between adjacent disease stages.

[0016] Preferably, the execution steps of the multi-modal fusion unit include: Perform vascular skeletonization on the fundus color image data to generate a topological connectivity graph; Extract vascular fractal dimension, curvature and intersection point density features through a graph attention network; Adopt a gated fusion mechanism to dynamically adjust the fusion ratio of optical coherence tomography features and vascular texture features.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: In terms of image processing and feature extraction, the system demonstrates powerful capabilities. The feature extraction unit performs interlayer boundary enhancement processing and noise elimination on the retinal optical coherence tomography sequence image data, extracts local spatial features in multiple dimensions using a three-dimensional convolutional kernel, and generates a structural feature tensor in combination with an attention mechanism, enabling more accurate and comprehensive acquisition of the structural features of the retinal choroid. Compared with traditional methods, it greatly improves the accuracy and integrity of feature extraction, providing a solid data foundation for subsequent lesion analysis.

[0018] The lesion segmentation unit adopts a cascaded U-Net network architecture and a deformable convolution module, combined with a hybrid loss function, significantly improving the accuracy of lesion segmentation. It can not only accurately segment the large choroidal vascular region but also focus on the detection of microvascular lesions, adapt to the retinal curvature variations of different patients, optimize the accuracy of the segmentation boundary, effectively avoid the missed detection and misjudgment of the lesion area, and help doctors observe the lesion range and morphology more clearly.

[0019] The temporal alignment unit uses the optical flow field algorithm, Lie group transformation matrix, and non-rigid registration algorithm to achieve the dynamic registration of multi-time-point images, generating a spatio-temporal aligned retinal interlayer displacement change matrix. This process effectively eliminates the influence of eye movement artifacts and local retinal deformations, making the analysis of the changes in the retinal choroid structure over time more accurate and reliable, and enabling the timely discovery of subtle structural change trends.

[0020] The dynamic analysis unit, based on the long short-term memory network and graph convolutional network, extracts indicators such as the choroidal thickness change rate, vascular density attenuation gradient, and lesion area diffusion coefficient, quantifying the disease development process from multiple dimensions. These quantitative indicators provide objective data support for doctors, helping to more accurately evaluate the severity and development trend of the condition, and providing a scientific basis for formulating personalized treatment plans.

[0021] The abnormal scoring unit automatically adjusts the weight coefficients through the entropy weight method, performs standardization processing on the indicators and fuses fuzzy logic rules. The constructed disease course progression score can more reasonably reflect the disease progression. Compared with the traditional subjective scoring method, it has higher objectivity and accuracy, facilitating doctors to uniformly evaluate and compare the conditions of different patients.

[0022] The decision classification unit adopts a gradient boosting decision tree model containing a Shapley value interpretation module, combined with a stratified sampling strategy and KL divergence constraint, not only improving the accuracy of disease course stage classification but also visualizing the feature contribution degree to ensure the interpretability of the classification boundary. Doctors can intuitively understand the impact of each feature on the judgment of the disease course stage, enhancing the credibility and persuasiveness of the diagnosis results.

[0023] The multimodal fusion unit fuses the vascular texture features of the fundus color image data with the tensor of the retinal choroid structure features, giving full play to the complementary advantages of multimodal data. The generated multimodal fusion feature vector contains richer information, provides a more comprehensive perspective for diagnosis, and helps to discover lesion features that are difficult to detect in single-modal images.

[0024] The report generation unit integrates the disease course stage classification label, the probability distribution map of the lesion area, and the multimodal fusion feature vector to generate a structured disease course monitoring report. This report form is standardized and comprehensive in information, facilitating doctors to quickly understand the patient's condition, and also facilitating the storage, management, and sharing of medical data, improving the efficiency and quality of medical work. Description of the Drawings

[0025] Figure 1 It is the working principle diagram of the artificial intelligence-based retinal choroid disease course structure change monitoring system described in the present invention; Figure 2 It is the working flow chart of the lesion segmentation unit; Figure 3 It is the working flow chart of the abnormal scoring unit. Detailed Embodiments

[0026] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0027] Please refer to Figures 1-3 , the present invention provides a technical solution: based on an artificial intelligence-based retinal choroid disease course structure change monitoring system, the system includes: Image acquisition unit: Using professional ophthalmic image acquisition equipment, retinal optical coherence tomography (OCT) sequence image data and fundus color image data of the target patient are obtained. These image data contain the structural information of the retinal choroid at different time points and various information such as fundus blood vessels, providing the original data basis for subsequent analysis.

[0028] Feature extraction unit: Through a three-dimensional convolutional neural network, hierarchical feature extraction is performed on the obtained retinal optical coherence tomography sequence image data to obtain the tensor of retinal choroid structure features. This feature tensor contains the key structural features of the retinal choroid and is an important basis for subsequent lesion analysis.

[0029] Lesion segmentation unit: Input the extracted retinal choroid structural feature tensor into a pre - trained U - Net segmentation network. After the network's operation and processing, output the probability distribution map of the retinal choroid lesion area. This distribution map intuitively shows the distribution possibility of the lesion area in the retinal choroid, which helps doctors judge the lesion location and scope.

[0030] Temporal alignment unit: Dynamically register multi - time - point images in the retinal optical coherence tomography sequence image data. Through a series of algorithmic processes, generate a spatio - temporal aligned retinal inter - layer displacement change matrix, which reflects the displacement changes of each retinal layer at different time points and provides dynamic data support for analyzing the disease course development.

[0031] Dynamic analysis unit: Based on the Long Short - Term Memory (LSTM) network, perform temporal modeling on the spatio - temporal aligned retinal inter - layer displacement change matrix. Through modeling, extract key indicators such as the choroid thickness change rate, blood vessel density attenuation gradient, and lesion area diffusion coefficient, which can quantitatively reflect the lesion development trend of the retinal choroid.

[0032] Abnormal score unit: According to the extracted choroid thickness change rate, blood vessel density attenuation gradient, and lesion area diffusion coefficient, construct a disease course progress score through a weighted fusion algorithm. This score comprehensively evaluates the progress degree of the disease course in the form of a numerical value, which is convenient for doctors to intuitively understand the severity of the condition.

[0033] Decision - making classification unit: Input the constructed disease course progress score into a gradient - boosting decision tree model. After the model's analysis and judgment, output the classification labels of the retinal choroid disease course stages. These labels divide the disease course into different stages, which helps doctors formulate targeted treatment plans.

[0034] Multi - modal fusion unit: Cascade the vascular texture features extracted from fundus color image data through a residual network with the retinal choroid structural feature tensor to generate a multi - modal fusion feature vector. The multi - modal fusion feature vector combines the feature information of different - modality images, making subsequent analysis more comprehensive and accurate.

[0035] Report generation unit: Integrate the disease course stage classification labels, the lesion area probability distribution map, and the multi - modal fusion feature vector to generate a structured disease course monitoring report. This report presents the lesion situation of the patient's retinal choroid in a structured form, which is convenient for doctors to consult and diagnose.

[0036] The present invention will be further described below in conjunction with Examples 1 to 6: Example 1: This embodiment mainly elaborates on the specific implementation steps of the feature extraction unit. Its unique function is to extract the structural features of the retinal choroid more accurately and comprehensively, providing high-quality feature data for subsequent lesion analysis.

[0037] ① Perform interlayer boundary enhancement processing on the retinal optical coherence tomography sequence image data, and use anisotropic diffusion filtering to eliminate noise interference. In retinal OCT images, the accurate identification of interlayer boundaries is crucial for feature extraction. Anisotropic diffusion filtering is a non-linear filtering method, and its formula is: where, represents the image data, is the time variable, is the divergence operator, is the gradient operator, is the diffusion coefficient, which is a function of the image gradient . This filtering method can retain the edge information of the image while smoothing the noise, making the boundaries between the retinal layers clearer and providing a better basis for subsequent feature extraction.

[0038] ② Extract the local spatial features of the retinal pigment epithelium layer, choroid capillary layer, and Bruch's membrane in the axial, transverse, and temporal dimensions through a three-dimensional convolution kernel. The three-dimensional convolution kernel performs convolution operations on the image data in three dimensions, capable of capturing local features in both space and time. Slide the three-dimensional convolution kernel in the axial, transverse, and temporal dimensions to perform convolution operations on the image regions of the retinal pigment epithelium layer, choroid capillary layer, and Bruch's membrane, thereby extracting the local spatial features of these layers, which contain the structural information and change trends of each layer.

[0039] ③ Combine the multi-scale feature maps with the attention mechanism module to generate a retinal choroid structural feature tensor containing the spatial weight distribution. The multi-scale feature maps can reflect the structural features of the retinal choroid from different resolution perspectives, while the attention mechanism module can assign different weights to different regions according to the importance of the features. By combining the two, the generated retinal choroid structural feature tensor can highlight important structural features, suppress unimportant information, and more accurately reflect the structural characteristics of the retinal choroid.

[0040] Example 2: This embodiment details the specific operation steps of the lesion segmentation unit. Its unique function is to improve the accuracy and adaptability of lesion segmentation, and it can better identify the retinal choroid lesion regions of different patients.

[0041] ①Construct a cascaded U-Net network architecture, where the primary network segments the large choroidal blood vessel region and the secondary network focuses on microvascular lesion detection. The cascaded U-Net network architecture combines the advantages of two U-Net networks. The primary network uses its larger receptive field to first segment the large choroidal blood vessel region, providing a basic framework for subsequent microvascular lesion detection. Based on the primary network, the secondary network conducts more detailed detection of microvascular lesions, being able to more comprehensively discover the lesion regions.

[0042] ②Introduce a deformable convolution module in the skip connection layer to adapt to the retinal curvature variations of different patients. There are differences in the retinal curvature of different patients, which will affect the accuracy of lesion segmentation. The deformable convolution module can adaptively adjust the sampling positions of the convolution kernels according to the curvature of the retina, and its formula is: Among them, is the value of the output feature map at position , is the input feature map, is the sampling region, is the convolution kernel weight, is the sampling offset, is the learnable offset. By introducing the deformable convolution module, the network can better adapt to the retinal curvature of different patients and improve the accuracy of lesion segmentation.

[0043] ③Adopt a hybrid loss function, combining the Dice coefficient loss and the edge-aware loss to optimize the accuracy of the segmentation boundary. The Dice coefficient loss is used to measure the overlap degree between the prediction result and the ground truth label, and the formula is: Among them, is the predicted lesion region, is the ground truth lesion region. The edge-aware loss pays more attention to the boundary of the lesion region. By combining these two loss functions, it can optimize the accuracy of the segmentation boundary while ensuring the accuracy of the segmented region, making the segmentation result more in line with the actual lesion situation.

[0044] Example 3: This example mainly elaborates on the specific implementation process of the temporal alignment unit, and its unique role is to achieve precise registration of multi-time point images, providing an accurate data basis for subsequent dynamic analysis.

[0045] ①Calculate the inter-layer displacement vector field of adjacent time point images based on the optical flow field algorithm. The optical flow field algorithm calculates the motion of pixels in adjacent time point images to obtain the inter-layer displacement vector field, and its formula is: Among them, is the image data, , are the spatial coordinates, is the time, , are respectively and the optical flow components in the directions. This vector field reflects the displacement of each layer of the retina between adjacent time points and is an important basis for subsequent registration.

[0046] ② Rigidly register the multi-time-point images through the Lie group transformation matrix to eliminate eye movement artifacts. Eye movement may cause artifacts in the images, affecting the analysis results. The Lie group transformation matrix can perform rigid transformations on multi-time-point images, including translation and rotation operations. Its matrix form is: Among them, is the rotation matrix, is the translation vector. Rigidly registering the images through the Lie group transformation matrix can eliminate the overall displacement and rotation caused by eye movement and preliminarily align the images in space.

[0047] ③ Use the non-rigid registration algorithm to correct the local deformation of the retina and construct an inter-layer displacement change matrix for spatio-temporal alignment. The non-rigid registration algorithm can further correct the local deformation of the retina, taking into account factors such as the elastic deformation of each layer of the retina. By processing the corrected images, an inter-layer displacement change matrix for spatio-temporal alignment is constructed. This matrix accurately reflects the displacement changes of each layer of the retina at different time points and provides reliable data support for dynamic analysis.

[0048] Example 4: This example covers the specific implementation steps of the dynamic analysis unit and the abnormal scoring unit. Its unique role is to accurately extract the key indicators of the disease course development and construct a reasonable disease course progress score to provide a quantitative basis for disease assessment.

[0049] ① Dynamic analysis unit: Slice the inter-layer displacement change matrix according to the time window and input it into the bidirectional long short-term memory network. The bidirectional long short-term memory network can simultaneously learn the forward and backward time series information and is very effective for analyzing the change trend of the inter-layer displacement change matrix over time. Slice the inter-layer displacement change matrix according to a certain time window and input each slice into the bidirectional long short-term memory network in turn. The network can learn the displacement change characteristics within different time windows.

[0050] Extract the standard deviation change rate of choroidal thickness within each time window as the dynamic stability index. The change in choroidal thickness is one of the important indicators for evaluating retinopathy. Calculate the standard deviation change rate of choroidal thickness within each time window, and the formula is: This index can reflect the change stability of choroidal thickness within different time windows. The larger the change rate, the greater the fluctuation of choroidal thickness, and the more unstable the condition may be.

[0051] Model the topological relationship of the vascular network through a graph convolutional network, and calculate the density attenuation gradient of vascular branch points. The graph convolutional network can effectively process data with topological structures. By modeling the vascular network, represent the vascular network in the form of a graph, where nodes represent vascular branch points and edges represent vascular connection relationships. Calculate the density attenuation gradient of vascular branch points, which can reflect the change trend of vascular density over time and is of great significance for evaluating the development of vascular lesions.

[0052] ② Abnormal scoring unit: Establish a dynamic weight allocation model based on the entropy weight method. The entropy weight method automatically adjusts the weight coefficients according to the historical variability of each indicator. The entropy weight method is an objective weight assignment method, and its formula is: Among them, is the weight of the th indicator, is the entropy value of the th indicator, is the total number of indicators. Through the entropy weight method, according to the historical variability of indicators such as the change rate of choroidal thickness, the density attenuation gradient of blood vessels, and the diffusion coefficient of the lesion area, weights are automatically assigned to each indicator to make the scoring results more objective and accurate.

[0053] Perform Z - score standardization on the change rate of choroidal thickness to eliminate the baseline differences between individuals. The Z - score standardization formula is: Among them, is the standardized value, is the original data, is the mean of the data, is the standard deviation of the data. By performing standardization on the change rate of choroidal thickness, the baseline differences caused by individual differences between different patients can be eliminated, making the data of different patients comparable.

[0054] The standardized indicators are fused using fuzzy logic rules to generate a disease progression score in the range of 0 - 1. Fuzzy logic rules can handle uncertain and fuzzy information. Based on the standardized indicators such as the change rate of choroidal thickness, the attenuation gradient of vascular density, and the diffusion coefficient of the lesion area, corresponding fuzzy logic rules are formulated. Through these rules, multiple indicators are fused to generate a disease progression score in the range of 0 - 1. The closer the score is to 1, the more severe the disease progression.

[0055] Example 5: This example details the specific implementation process of the decision - grading unit and the multi - modal fusion unit. Its unique role is to accurately divide the disease course stage and provide more comprehensive feature information to assist doctors in making more accurate diagnoses.

[0056] ② Decision - grading unit: Construct a gradient - boosting decision tree model that includes a Shapley value interpretation module, which is used to visualize the feature contribution degree. The Shapley value can measure the contribution degree of each feature to the model output. The formula is: Where, is the Shapley value of the th feature, is the feature subset, is the set of all features, is the model output value corresponding to the feature subset By constructing a gradient - boosting decision tree model that includes a Shapley value interpretation module, it is possible to intuitively understand the contribution of each indicator in the disease progression score to the classification of the disease course stage, which helps doctors understand the decision - making process of the model.

[0057] Adopt a stratified sampling strategy to balance the distribution of samples in different disease course stages. The number of samples in different disease course stages may be uneven, which will affect the training effect of the model. The stratified sampling strategy divides the samples into different layers according to the disease course stage, and samples are taken in each layer so that the samples of each disease course stage have a reasonable proportion in the training set, improving the generalization ability of the model.

[0058] Introduce the KL - divergence constraint at the leaf nodes to ensure the interpretability of the classification boundary between adjacent disease course stages. The KL - divergence is used to measure the difference between two probability distributions. The formula is: Where, and They are two probability distributions. Introducing the KL divergence constraint at the leaf nodes of the gradient boosting decision tree model can make the classification boundary between adjacent disease stages clearer, more interpretable, and facilitate doctors to accurately judge the disease stage based on the output of the model.

[0059] ② Multimodal fusion unit: Perform vascular skeletonization on the fundus color image data to generate a topological connectivity graph. Vascular skeletonization can extract the centerlines of the fundus blood vessels and generate a topological connectivity graph, which reflects the connection relationship and topological structure of the blood vessels and provides a basis for subsequent feature extraction.

[0060] Extract the vascular fractal dimension, curvature, and intersection density features through a graph attention network. The graph attention network can automatically assign attention weights according to the relationship between nodes, process the topological connectivity graph, and extract features such as the vascular fractal dimension, curvature, and intersection density. These features can reflect the morphological and structural characteristics of the fundus blood vessels from different angles and, combined with the retinal choroid structure feature tensor, provide more comprehensive information.

[0061] Adopt a gated fusion mechanism to dynamically adjust the fusion ratio of optical coherence tomography features and vascular texture features. The gated fusion mechanism dynamically adjusts the fusion ratio of optical coherence tomography features and vascular texture features through a gating function, and the formula is: Among them, is the fused feature vector, is the optical coherence tomography feature, is the vascular texture feature, is the gating value, and its value range is between 0 and 1. Through the gated fusion mechanism, the fusion ratio of the two features can be adaptively adjusted according to different situations to generate a more effective multimodal fusion feature vector.

[0062] Example 6: This example focuses on the report generation unit, whose unique role is to integrate the key information obtained from the previous analysis and present it to medical staff in a structured and easy-to-read manner, providing strong support for clinical diagnosis and treatment.

[0063] In an actual application scenario, assume that the system monitors a patient with a retinal choroid disease. After the processing of each previous unit, the disease stage classification label, the probability distribution map of the lesion area, and the multimodal fusion feature vector have been obtained.

[0064] The report generation unit will first create a structured document framework, for example, adopting the format of a common medical report template, including main sections such as patient basic information, examination items, and result analysis.

[0065] In the patient's basic information section, basic information such as the patient's name, age, gender, medical record number, etc. will be extracted from the system's patient information database and filled in the corresponding positions of the report.

[0066] For the examination item section, relevant information on optical coherence tomography (OCT) of the retina and fundus color imaging examinations involved in this monitoring will be listed in detail, including the examination time, examination equipment, etc.

[0067] In the result analysis section, key content related to the course of the retina and choroid disease will be highlighted. The classification labels of the disease course stages will be presented in clear text, such as "early lesion stage", "progressive stage", etc., so that doctors can understand the current disease course stage of the patient at a glance.

[0068] The probability distribution map of the lesion area will be embedded in the report in the form of an image. To help doctors better understand, a brief text description will be added next to the image, marking the areas with higher lesion probabilities and comparing them with the normal areas. For example, "As shown in the figure, near the macula of the retina, the probability distribution map of the lesion area shows that the lesion probability in this area reaches 70%, which is higher than the probability level of the normal area, indicating that there may be a relatively serious lesion here." For the multi-modal fusion feature vector, since it contains rich but relatively abstract information, it will be analyzed and interpreted to some extent before being presented in the report. For example, through the analysis of the multi-modal fusion feature vector, it is found that the blood vessel texture features are significantly different from the normal samples in some dimensions. The report will describe it as "The analysis of the multi-modal fusion feature vector shows that the blood vessel texture features of the fundus deviate from the normal standards in terms of curvature and intersection density. Combining other indicators, this may reflect that the lesion situation of the blood vessels is closely related to the progression of the disease course."

[0069] In addition to the above information, the report generation unit will also add some summary content. For example, based on all the analysis results, a general condition assessment will be given, "According to the monitoring results of the structural changes in the course of the retina and choroid disease, the patient is currently in the progressive stage of the disease course. The lesion area is mainly concentrated near the macula, and abnormal changes have also occurred in the blood vessel texture. It is recommended to further conduct relevant examinations and formulate a personalized treatment plan in combination with the clinical symptoms."

[0070] Finally, the report generation unit will save the generated structured disease course monitoring report in a common document format, such as PDF, which is convenient for doctors to consult in the electronic medical record system and is also convenient for printing for patients' referral or consultation.

[0071] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0072] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based monitoring system for the structural changes in the course of retinopathy and choroidopathy, characterized in that, Comprising: An image acquisition unit, which is used to acquire retinal optical coherence tomography sequence image data and fundus color image data of a target patient; A feature extraction unit, which is used to perform hierarchical feature extraction on the retinal optical coherence tomography sequence image data through a three-dimensional convolutional neural network to obtain a retinal choroid structure feature tensor; A lesion segmentation unit, which is used to input the retinal choroid structure feature tensor into a pre-trained U-Net segmentation network and output a probability distribution map of the retinal choroid lesion area; A temporal alignment unit, which is used to perform dynamic registration on multi-time point images in the retinal optical coherence tomography sequence image data to generate a spatio-temporally aligned retinal interlayer displacement change matrix; A dynamic analysis unit, which is used to perform temporal modeling on the spatio-temporally aligned retinal interlayer displacement change matrix based on a long short-term memory network and extract the choroid thickness change rate, blood vessel density attenuation gradient, and lesion area diffusion coefficient; An abnormal scoring unit, which is used to construct a disease course progression score through a weighted fusion algorithm according to the choroid thickness change rate, the blood vessel density attenuation gradient, and the lesion area diffusion coefficient; A decision classification unit, which is used to input the disease course progression score into a gradient boosting decision tree model and output a classification label for the retinal choroid disease course stage; A multi-modal fusion unit, which is used to perform feature concatenation on the blood vessel texture features extracted from the fundus color image data by a residual network and the retinal choroid structure feature tensor to generate a multi-modal fusion feature vector; A report generation unit, which is used to integrate the disease course stage classification label, the lesion area probability distribution map, and the multi-modal fusion feature vector to generate a structured disease course monitoring report.

2. The system according to claim 1, characterized in that, The execution steps of the feature extraction unit include: Performing interlayer boundary enhancement processing on the retinal optical coherence tomography sequence image data, and using anisotropic diffusion filtering to eliminate noise interference; Extracting local spatial features of the retinal pigment epithelium layer, choroid capillary layer, and Bruch's membrane in the axial, transverse, and temporal dimensions through a three-dimensional convolution kernel; Combining multi-scale feature maps with an attention mechanism module to generate a retinal choroid structure feature tensor containing spatial weight distribution.

3. The system according to claim 1, wherein The execution steps of the lesion segmentation unit include: Constructing a cascaded U-Net network architecture, where the primary network segments the large choroidal blood vessel area, and the secondary network focuses on microvascular lesion detection; Introducing a deformable convolution module in the skip connection layer to adapt to the retinal curvature variation of different patients; Adopting a hybrid loss function and combining the Dice coefficient loss and the edge perception loss to optimize the segmentation boundary accuracy.

4. The system according to claim 1, characterized in that The execution steps of the temporal alignment unit include: Calculating the interlayer displacement vector field of adjacent time point images based on the optical flow field algorithm; Performing rigid registration on multi-time point images through a Lie group transformation matrix to eliminate eye movement artifacts; Correct the local deformation of the retina using a non-rigid registration algorithm and construct an interlayer displacement change matrix for spatio-temporal alignment.

5. The system according to claim 1, characterized in that, The execution steps of the dynamic analysis unit include: Slice the interlayer displacement change matrix by time window and input it into a bidirectional long short-term memory network; Extract the standard deviation change rate of the choroidal thickness within each time window as the dynamic stability index; Model the topological relationship of the vascular network through a graph convolutional network and calculate the density attenuation gradient of vascular branch points.

6. The system according to claim 1, characterized in that, The execution steps of the anomaly scoring unit include: Establish a dynamic weight allocation model based on the entropy weight method, and the entropy weight method automatically adjusts the weight coefficient according to the historical variability of each index; Perform Z-score normalization on the choroidal thickness change rate to eliminate the baseline differences between individuals; Adopt fuzzy logic rules to fuse the normalized indicators and generate a disease progression score in the range of 0-1.

7. The system according to claim 1, wherein The execution steps of the decision classification unit include: Construct a gradient boosting decision tree model containing a Shapley value interpretation module, and the module is used to visualize the feature contribution degree; Adopt a stratified sampling strategy to balance the distribution of samples at different disease stages; Introduce KL divergence constraints at the leaf nodes to ensure the interpretability of the classification boundary between adjacent disease stages.

8. The system according to claim 1, wherein The execution steps of the multimodal fusion unit include: Perform vascular skeletonization on the fundus color image data to generate a topological connectivity graph; Extract the vascular fractal dimension, curvature, and intersection point density features through a graph attention network; Adopt a gated fusion mechanism to dynamically adjust the fusion ratio of optical coherence tomography features and vascular texture features.

Images