A method for detecting pseudo-lesions of fundus

Abstract

The present application provides a method for detecting pseudo-lesions of the fundus, comprising: performing image feature analysis of the lesion area according to preliminary lesion markings, locating and evaluating the cell density and the degree of vascular leakage, and determining the risk of disease aggravation; predicting the tortuosity and vascular fragility of new blood vessels and evaluating the probability of complications in patients according to morphological processing and feature analysis of multi-phase vascular images; predicting the risk of serious complications according to the lesion markings and vascular feature analysis of the patient, combined with blood test results, and re-identifying and excluding pseudo-lesion images that are mistaken for lesion images; comparing the feature changes of pseudo-lesions and real lesion areas according to time series analysis of fundus disease images, evaluating the disease development path, and performing early intervention on the disease; segmenting the new blood vessel area based on the use of sensitive index filtering technology, and evaluating the risk of disease aggravation through vascular feature statistics.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for detecting pseudo-lesions of fundus. Background Art

[0002] In the field of ophthalmological diagnosis, especially in the diagnosis and treatment of fundus diseases, accurate image analysis is crucial. Retinal fundus diseases, such as diabetic retinopathy and macular degeneration, pose a serious threat to patients' vision. With the advancement of medical imaging technology, especially the application of optical coherence tomography (OCT), fundus photography, and fluorescent fundus angiography, doctors are able to observe the structure and vascular conditions of the retina in more detail. However, fundus imaging diagnosis still faces some key challenges. One major challenge is the identification of pseudo-lesions, which refer to image features that appear in fundus images and are similar to real lesions, but are not actually caused by any fundus disease. These pseudo-lesions may be caused by various factors, such as optical artifacts, noise in image processing, vascular shadows, retinal pigment deposition, etc. The presence of pseudo-lesions not only increases the complexity of diagnosis, but may also lead to misdiagnosis or overdiagnosis, thereby affecting the formulation of treatment decisions. Traditional pseudo-lesion identification relies on the doctor's experience and subjective judgment, which to some extent limits the accuracy and efficiency of diagnosis. In addition, with the increasing amount of medical imaging data, it is becoming increasingly difficult to analyze and interpret these data based solely on the doctor's experience and expertise. Therefore, there is an urgent need for a method that can automatically and accurately identify and classify pseudo-lesions. In addition, the dynamic monitoring of pseudo-lesions is also an important issue. Since the development of fundus diseases may be accompanied by changes in retinal image characteristics, how to monitor these changes in real time and adjust the treatment strategy accordingly is the key to improving the treatment effect. Existing methods are usually unable to effectively monitor the dynamic changes of pseudo-lesions, which limits the flexibility of doctors in making treatment decisions. In short, the existing fundus imaging diagnostic methods have many limitations in the identification, classification and dynamic monitoring of pseudo-lesions, and new technical solutions are urgently needed to improve the accuracy, efficiency and adaptability of diagnosis. Summary of the invention

[0003] The present invention provides a method for detecting pseudo-lesions of the fundus, which mainly comprises:

[0004] After collecting multi-phase and multi-modal fundus lesion image sequences, a convolutional neural network based on the position attention mechanism is used to identify pseudo-lesions and predict the change rules of boundary heterogeneity to obtain preliminary lesion markings.

[0005] Based on the preliminary lesion marking, image feature analysis of the lesion area is performed to locate and evaluate cell density and vascular leakage, and determine the risk of disease aggravation;

[0006] Based on the morphological processing and feature analysis of multi-phase vascular images, the tortuosity and fragility of new blood vessels are predicted, and the probability of complications in patients is assessed;

[0007] Based on the patient's lesion markers and vascular characteristics analysis, combined with blood test results, the risk of serious complications is predicted, and pseudo-lesion images that are mistaken for lesion images are re-identified and excluded;

[0008] Through time series analysis of fundus disease images, the characteristic changes of pseudo-lesions and real lesion areas are compared to evaluate the progression of the disease and conduct early intervention on the disease.

[0009] The sensitivity index filtering technology is used to segment the neovascularization area and the risk of disease progression is assessed through vascular feature statistics.

[0010] Based on OCT and OCTA images, the status of vascular lesions after treatment can be monitored, the association between different treatment strategies and high-risk complications can be identified, the treatment effect can be evaluated, and the selection of personalized treatment strategies can be guided.

[0011] In some embodiments, after acquiring a multi-phase multi-modal fundus lesion image sequence, a convolutional neural network based on a position attention mechanism is used to identify pseudo lesions and predict the change law of boundary heterogeneity to obtain a preliminary lesion mark, including:

[0012] Multi-phase fundus images are acquired using multimodal imaging technology, including optical coherence tomography and fluorescent fundus angiography. The images are preprocessed, including noise reduction, contrast enhancement and color correction. Based on the fundus images containing pseudo-lesion area annotations, a convolutional neural network model enhanced with position attention mechanism is trained to identify key areas in multi-phase multimodal fundus images, where the key areas are pseudo-lesion areas. The brightness or color value of each pixel in the image is compared with its adjacent pixels, and pixel-level difference calculation is performed to obtain the difference value between each pixel and its neighborhood, thereby highlighting the pseudo-lesion area. If there is no pseudo-lesion area in the fundus image, the Canny edge detection algorithm is used to separate the lesion area based on the preprocessed multi-phase multimodal fundus image. The segmented lesion image is subjected to boundary integrity check. If the lesion boundary is incomplete, the Canny edge detection algorithm is used. The Snake algorithm is used for image completion. For the completed lesion image, the morphological features are extracted to determine whether the morphology is stable. If the morphology is stable, the lesion mark is recorded. If the morphology is unstable, the lesion morphological changes are monitored and recorded. According to the recorded lesion morphological changes, images with enhanced lesion heterogeneity are screened, and the SLIC algorithm is used to segment the images with enhanced lesion heterogeneity to obtain preliminary lesion marks. Heterogeneity enhancement refers to the situation where the difference between the lesion area and the surrounding normal tissue in terms of morphology, color, and texture is greater than the preset difference threshold. By comparing the changes of the lesion areas divided in the continuous image sequence over time, the area increase rate and boundary expansion speed of the lesion area are obtained to evaluate the regional growth rate. If the growth rate of the heterogeneous area is less than the preset rate threshold, it is judged to be a benign lesion. If the growth rate of the heterogeneous area is greater than the preset rate threshold, it is judged to be a malignant lesion.

[0013] In some embodiments, the method of performing image feature analysis of the lesion area based on the preliminary lesion marking, locating and evaluating the cell density and the degree of vascular leakage, and determining the risk of disease aggravation includes:

[0014] The image of the lesion area is obtained, and the image is processed using a convolutional neural network. The multi-scale features of the image, including shape, texture, and color, are extracted through multi-level convolution and pooling operations; the clarity of the extracted features at each scale is determined. If the clarity is lower than the preset first clarity threshold, the pyramid scene parsing network is used to enhance the feature expression; the enhanced features are subjected to image pyramid parsing to obtain feature maps at different resolutions; the Canny edge detection algorithm is used on each resolution feature map to locate the tissue structure area; the located tissue structure area is included in the sample library, and the cell distribution density of the tissue structure area is annotated; based on the annotated sample images, the convolutional neural network algorithm is used for model training, a cell density assessment model is constructed, the cell density distribution is predicted, and the cell density is determined uniformity; if the predicted result is that the uniformity of cell distribution density is less than the preset uniformity threshold, it is judged that the range of cell infiltration is expanded; if the predicted result is that the uniformity of cell distribution density is greater than the preset uniformity threshold, the original assessment is maintained; based on the labeled sample images, the convolutional neural network algorithm is used for model training to identify the vascular area in the image; by calculating the ratio of the area of ​​the dilated vascular area to the area of ​​the normal vascular area, the area difference between the segmented vascular area and the normal vascular area is compared, and the degree of vascular leakage is assessed according to the area ratio; if the degree of vascular leakage is greater than the preset degree threshold, it is judged that the vascular leakage is aggravated; if the degree of vascular leakage is less than the preset degree threshold, the vascular state is normal; based on the judgment results of cell infiltration and vascular leakage, the risk of aggravation of the disease is assessed and an assessment report is output.

[0015] In some embodiments, the prediction of the tortuosity and fragility of new blood vessels and the evaluation of the probability of complications in patients based on morphological processing and feature analysis of multi-phase vascular images includes:

[0016] Through the fundus image library, multi-phase vascular images are obtained, and the vascular structure morphology is enhanced by binarization and morphological processing of dilation and erosion. If the clarity of the vascular contour in the enhanced image is less than the preset second clarity threshold, the Gaussian blur technology is used to smooth the noise and irregular edges in the image to obtain a vascular image with a clarity greater than the preset second clarity threshold. For the vascular image with a clarity greater than the preset second clarity threshold, a thinning algorithm based on distance transformation is used to extract the single-pixel vascular centerline. The number of turning angles on the vascular centerline is calculated and accumulated to determine the total degree of tortuosity of the new blood vessels. The tortuosity values ​​at different phases are obtained and formed into a time series in chronological order. Based on the historical time series of the tortuosity values, the ARIMA algorithm is used to train the model and predict the changing trend of the tortuosity values. The histological characteristics of the blood vessels, including the integrity of the vessel wall and the distortion of the lumen, are extracted. Based on the histological characteristics of the blood vessels, the support vector machine algorithm is used to train the model to construct a vascular fragility assessment model to determine the risk of vascular rupture. Based on the changing trend of the tortuosity and the judgment of vascular fragility, the probability of complications in patients is assessed.

[0017] In some embodiments, the prediction of the risk of serious complications based on the analysis of the patient's lesion markers and vascular characteristics in combination with blood test results, and re-identification and exclusion of pseudo-lesion images that are mistaken for lesion images, include:

[0018] Obtain lesion markers, vascular characteristics and clinical indicator data sets, construct sample feature vectors, which include lesion size, lesion shape, lesion location information, vascular structural characteristics, and clinical indicators, where the clinical indicators are blood test results; use the Pearson correlation coefficient method to perform correlation analysis on the sample vectors, and identify feature sets with a correlation with worsening of the disease that is higher than the preset correlation threshold, as early warning signals, including lesion size, lesion shape, vascular structural characteristics, and clinical indicators; based on the early warning signals, use the support vector machine algorithm to train the model, build a severe complication risk prediction model, and assess the patient's severe complication risk; based on the historical data of the early warning signals, use the long short-term memory network algorithm to train the model, build a severe complication association model, and predict each feature of the early warning signal Change trends over time; monitor the time series changes of warning signals in new samples. If the characteristics of the warning signals are detected to be abnormal, obtain the abnormal patterns of the characteristics of the warning signals, which include sudden changes in characteristics, periodic fluctuations in characteristics, trend breaks, and abnormal ranges of characteristic fluctuations; use the support vector machine algorithm to train the model based on the abnormal patterns of the characteristics of the warning signals, re-identify pseudo-lesions, and identify and exclude pseudo-lesion images that are mistaken for lesion images; use the severe complication risk prediction model to predict the risk of severe complications based on the warning signals of the re-screened lesion images; if the predicted risk of severe complications is greater than the preset risk threshold, combine the patient's clinical data and imaging analysis results to predict the development path of the disease, and the clinical data include age, health status, and medical history;

[0019] The method also includes classifying the pseudo lesions according to the pattern of changes in the characteristics of the pseudo lesions.

[0020] The pseudo lesions are classified according to the pattern of pseudo lesion feature changes, specifically including:

[0021] Obtain retinal fundus images from multiple time points, including annotated real lesion and pseudo-lesion images, and preprocess the images, including denoising, color correction, and contrast enhancement. Extract key features of pseudo-lesions, including shape, size, edge sharpness, color consistency, and texture pattern. Based on the key features of historical pseudo-lesions, use the ARIMA algorithm to train the model, predict the feature changes of pseudo-lesions at different time points, and identify the pattern of pseudo-lesion feature changes, including size growth, color change, and texture evolution. Design a pseudo-lesion classification formula based on the pattern of pseudo-lesion feature changes

[0022] C=α·F shape +β·F color +γ·F texture +δ·ΔF time , evaluate the category of pseudo lesions, where C is the pseudo lesion classification evaluation index, and F shape , F color, F texture Represent the metrics of shape, color and texture respectively, ΔF time Represents the amount of change in the time series, and α, β, γ, and δ are weights determined based on historical feature data. According to the calculation results of the pseudo-lesion classification evaluation index and the preset index threshold, the pseudo-lesions are classified to determine the types of pseudo-lesions, including optical artifacts, pixel noise, vascular shadows, retinal pigment deposition, and vitreous floaters. Use the labeled data set to train the pseudo-lesion classification formula, optimize the parameters, and improve the classification accuracy. Verify the performance of the pseudo-lesion classification formula on an independent test data set to evaluate the accuracy and reliability of the pseudo-lesion classification formula in actual clinical applications. Adjust the pseudo-lesion classification formula parameters and the weight of the classification formula according to the test results. Integrate the pseudo-lesion classification formula into existing ophthalmic diagnostic tools.

[0023] In some embodiments, the time series analysis of fundus disease images, comparing the characteristic changes of pseudo-lesions and real lesion areas, evaluating the disease progression path, and performing early intervention on the disease include:

[0024] Obtain fundus disease image time series samples, including fundus disease image samples in the early, middle and late stages of the lesion; evaluate the information density and degree of change in each frame of the image through the hill climbing algorithm, and select key frames with information density greater than a preset density threshold and a degree of change greater than a preset change threshold; use the convolutional neural network-based segmentation algorithm U-Net to annotate the selected key frames for lesions; extract the significant features of the lesions based on the annotated results, including morphology, size, color, and texture; compare the changes in the significant features of the same lesion at different time points to determine the disease progression path, which includes worsening, improvement, and stabilization; process the fundus image data in the time series through the convolutional neural network, learn the development pattern of the disease, establish a disease course dynamics change map, and display the progression stage and change trend of the disease in a visual way; compare the information and feature change patterns in the disease course dynamics change map with those mistakenly identified as lesion maps The characteristic change patterns of pseudo-lesion images are compared to obtain the changing trends of pseudo-lesions and real lesion areas with treatment and disease progression; the early change patterns of pseudo-lesion characteristics of different categories are obtained, and the correlation between the characteristic change patterns of pseudo-lesions of different categories and the disease stage is calculated by the Pearson correlation coefficient method to evaluate whether early intervention of the disease is needed. The change patterns include morphological changes, color changes and evolution of texture patterns. The disease stage includes early, middle and late stages. If the correlation between the pseudo-lesion characteristic change pattern and the disease stage is greater than the preset correlation threshold, early intervention of the disease is required. According to the information and characteristic change patterns in the disease course dynamics change map, the risk of disease deterioration at different stages is evaluated. The characteristic change patterns include the growth of lesion size, changes in morphology or deepening of color. Combined with the patient's clinical data and image analysis results, the disease development path is predicted. The clinical data include age, health status and medical history.

[0025] It also includes: monitoring the dynamic changes of pseudo-lesions based on characteristic analysis of fundus disease images, and adjusting early intervention strategies in real time.

[0026] The method of analyzing the characteristics of fundus disease images, monitoring the dynamic changes of pseudo-lesions, and adjusting the early intervention strategy in real time specifically includes:

[0027] Obtain fundus disease images, apply convolutional neural networks to model training, and extract pseudo-lesion features, including morphological details, color depth, and texture accuracy. Based on retinal fundus images, use dynamic Bayesian networks to train models, monitor changes in pseudo-lesion features over time, and predict future trends in pseudo-lesion features. Based on pseudo-lesion features, establish an early intervention scoring formula.

[0028] G=ω*F detail +θ*ΔF time +λ*P trend, predicting the timing of early intervention, where G represents the intervention index,

[0029] F detail Indicates the feature detail score of pseudo-lesions, ΔF time Indicates the rate of change of the feature over time, P trend represents the predicted future trend score, and ω, θ, and λ are weight parameters based on historical pseudo-lesion feature data. Based on the real-time retinal fundus image monitoring results, the changes in pseudo-lesion characteristics are continuously monitored, and the intervention index is adjusted according to the latest data to update the intervention recommendations in real time. The early intervention scoring formula is tested in clinical trials to obtain feedback from doctors and patients on the accuracy of the intervention timing. The parameters of the early intervention scoring formula are adjusted and optimized based on clinical results. The early intervention scoring formula is integrated into the existing ophthalmic diagnosis and decision support system to provide doctors with comprehensive treatment recommendations.

[0030] In some embodiments, segmenting the neovascularization area by using a sensitivity index filtering technique and evaluating the risk of disease progression by using vascular feature statistics include:

[0031] Obtain fundus images, use convolutional neural networks for model training, build a diseased vascular region recognition model, extract multi-scale vascular features, and locate the diseased vascular region; intercept the diseased vascular image, and use sensitive index filtering to enhance the details of the inner wall of the lumen; based on the similarity of the feature vectors of the diseased vascular image, retrieve cases from the sample library whose similarity difference with the feature vector of the current diseased vascular image is less than the preset difference threshold, and the feature vector includes the size, shape, and color of the lesion area; the cases retrieved from the sample library contain annotated diseased vascular areas, and the annotated similar cases are prior knowledge; by using the annotated similar cases as additional training data and integrating them into the diseased vascular region recognition model, accurate segmentation of the diseased vessels is achieved; the segmentation results using the diseased vascular region recognition model are post-processed, and the post-processing includes removing small noise areas and smoothing edges; the vascular characteristics of the identified new blood vessels are statistically analyzed, including distribution range, density, and degree of curvature; according to the vascular characteristics of the new blood vessels, a decision tree algorithm is used for model training to evaluate the risk of disease deterioration based on the current vascular characteristics.

[0032] In some embodiments, the OCT and OCTA images are used to monitor the vascular lesions after treatment, identify the association between different treatment strategies and high-risk complications, evaluate the treatment effect and guide the selection of personalized treatment strategies, including:

[0033] Obtain fundus optical coherence tomography images after treatment, and use exponential smoothing denoising technology to improve image clarity and visibility of details; based on the fundus optical coherence tomography images after treatment, use the Canny edge detection algorithm to segment the contours of each layer of retinal tissue, identify and mark each retinal layer of the retina, including the retinal neuroepithelial layer, photoreceptor layer, outer nuclear layer, outer limiting membrane, connecting layer between the outer nuclear layer and the inner nuclear layer, inner nuclear layer, and nerve fiber layer; measure the distance between the retinal layers to determine whether the retinal thickness is normal; detect the grating distortion in the image, and use the Canny edge detection algorithm to locate the grating area in the image to determine whether the grating morphology is distorted; determine whether the retinal thickness and grating morphology have returned to the ideal state. If not, use OCT angiography imaging to compare with conventional imaging to determine whether the vascular revascularization is normal. The ideal state is determined by the preset ideal state threshold of retinal thickness and the preset ideal state threshold of grating morphology; in OCTA images, the Canny edge detection algorithm is used to identify new blood vessels, and the number, diameter, and tortuosity of the new blood vessels are counted; according to the number, diameter, and tortuosity of the new blood vessels, the decision tree algorithm is used to train the model, build a decision tree model, and judge high-risk vascular complications; obtain data on different treatment strategies, including treatment type, duration, and dosage; obtain fundus images, vascular lesions, and recorded complications of patients after treatment; according to the data of different treatment strategies, patient age, disease course, and treatment response, the support vector machine algorithm is used to train the model to evaluate the risk of complications of different treatment strategies; according to the risk of complications of different treatment strategies, guide clinical decision-making and select the treatment plan that best suits the patient.

[0034] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0035] The present invention discloses a method for detecting fundus pseudo-lesions. The present invention can accurately identify pseudo-lesions through high-precision image analysis, significantly improve the accuracy and efficiency of diagnosis, and thus reduce the risk of misdiagnosis or overdiagnosis. Through detailed analysis of the lesion area, the present invention can assess the risk of worsening of the disease, including cell density and degree of vascular leakage, and provide doctors with more accurate disease assessment and treatment decisions. The invention supports dynamic monitoring of fundus vascular images, helps predict the occurrence of complications, and provides key information for early intervention and disease management. By continuously tracking the progression of the disease, it can help doctors evaluate the treatment effect and adjust the treatment strategy in time to adapt to changes in the disease. Overall, the solution of the present invention provides doctors with a comprehensive, accurate and dynamic disease management tool, which helps to improve the treatment effect, reduce the burden on patients, and optimize the allocation of medical resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1The present invention is a flow chart of a method for detecting pseudo-lesions of the fundus.

[0037] Figure 2 It is a schematic diagram of a method for detecting fundus pseudo-lesions of the present invention.

[0038] Figure 3 It is another schematic diagram of a method for detecting fundus pseudo-lesions of the present invention. DETAILED DESCRIPTION

[0039] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0040] like Figure 1-3 In this embodiment, a method for detecting pseudo-lesions of the fundus may specifically include:

[0041] S101. After collecting multi-phase and multi-modal fundus lesion image sequences, a convolutional neural network based on a position attention mechanism is used to identify pseudo-lesions and predict the change pattern of boundary heterogeneity to obtain preliminary lesion markings.

[0042] Use multimodal imaging technology to obtain multi-phase fundus images. Multimodal imaging technology includes optical coherence tomography and fluorescent fundus angiography. Preprocess the image, including noise reduction, contrast enhancement and color correction; train a convolutional neural network model enhanced with position attention mechanism based on the fundus image containing pseudo-lesion area annotations to identify the key areas in the multi-phase multimodal fundus image, which are pseudo-lesion areas; compare the brightness or color value of each pixel in the image with its adjacent pixels, perform pixel-level difference calculations, obtain the difference value between each pixel and its neighborhood, and highlight the pseudo-lesion area. If there is no pseudo-lesion area in the fundus image, use the Canny edge detection algorithm to separate the lesion area based on the preprocessed multi-phase multimodal fundus image; perform a boundary integrity check on the segmented lesion image. If the lesion boundary is incomplete, use the Snake algorithm to complete the image. For the completed lesion image, extract morphological features to determine whether the morphology is stable. If the morphology is stable, record the lesion mark.

[0043] If the morphology is unstable, monitor and record the changes in lesion morphology. According to the recorded changes in lesion morphology, screen images with enhanced lesion heterogeneity, and segment the images with enhanced lesion heterogeneity using the SLIC algorithm to obtain preliminary lesion markings. Heterogeneity enhancement refers to the situation where the difference between the lesion area and the surrounding normal tissue in terms of morphology, color, and texture is greater than the preset difference threshold; by comparing the changes in the lesion area divided in the continuous image sequence over time, obtain the area increase rate and boundary expansion speed of the lesion area, and evaluate the regional growth rate. If the growth rate of the heterogeneous area is less than the preset rate threshold, it is judged as a benign lesion; if the growth rate of the heterogeneous area is greater than the preset rate threshold, it is judged as a malignant lesion.

[0044] For example, fundus images of patients were obtained using optical coherence tomography and fluorescent fundus angiography, and a series of fundus images containing 20 time points were obtained, each time point including images of two modalities, OCT and fluorescent fundus angiography. A Gaussian filter was applied to all images, with a kernel size of 5x5 pixels. The contrast enhancement operation equalized the histogram of the image, and color correction was performed to eliminate illumination unevenness. Images containing pseudo-lesion area annotations were used as training data to train the convolutional neural network model, and annotated images containing 100 pseudo-lesions were used as training sets. For each pixel, the difference in its brightness or color value was compared with the surrounding 8 neighboring pixels, and the average color difference between the central pixel and its neighboring pixels was calculated. If there was no pseudo-lesion in the image, the Canny edge detection algorithm was applied to separate the lesion and non-lesion areas, and the Snake algorithm was used to complete the incomplete lesion boundary, and the number of iterations was set to 50. The morphological features of the lesion after completion, such as area and perimeter, were extracted, and the changes in the lesion at 5 consecutive time points were monitored to record whether the morphology was stable. The superpixel segmentation algorithm was applied to the heterogeneity-enhanced lesion image, and the SLIC algorithm was used to compare the area of ​​the lesion region in the continuous images, and the lesion area from time point T1 to T5 increased from 2 square millimeters to 3 square millimeters. If the area increase rate of the lesion area is less than the preset threshold, for example, less than 10% at each time point, it is judged as a benign lesion. If the area increase rate is greater than 10%, it is judged as a malignant lesion.

[0045] S102. Based on the preliminary lesion marking, perform image feature analysis of the lesion area, locate and evaluate the cell density and degree of vascular leakage, and determine the risk of disease aggravation.

[0046] The image of the lesion area is obtained, and the image is processed by a convolutional neural network. The multi-scale features of the image, including shape, texture and color, are extracted through multi-level convolution and pooling operations. The clarity of the extracted features of each scale is determined. If the clarity is lower than the preset first clarity threshold, the pyramid scene parsing network is used to enhance the feature expression. The enhanced features are parsed by an image pyramid to obtain feature maps at different resolutions. On each resolution feature map, the Canny edge detection algorithm is used to locate the tissue structure area. The located tissue structure area is included in the sample library, and the cell distribution density of the tissue structure area is annotated. Based on the annotated sample images, the convolutional neural network algorithm is used for model training, a cell density evaluation model is constructed, the cell density distribution is predicted, and the uniformity of the cell density is determined. If the predicted result is that the uniformity of the cell distribution density is less than the preset uniformity threshold, it is determined that the cell infiltration range is expanded. If the predicted result is that the uniformity of the cell distribution density is greater than the preset uniformity threshold, the original evaluation is maintained. Based on the annotated sample images, the convolutional neural network algorithm is used for model training to identify the vascular area in the image. By calculating the ratio of the area of ​​the dilated blood vessel area to the area of ​​the normal blood vessel area, the area difference between the segmented blood vessel area and the normal blood vessel area is compared, and the degree of vascular leakage is evaluated based on the area ratio. If the degree of vascular leakage is greater than the preset threshold, it is judged that the vascular leakage has aggravated. If the degree of vascular leakage is less than the preset threshold, the vascular state is normal. Based on the judgment results of cell infiltration and vascular leakage, the risk of aggravation of the disease is assessed and an evaluation report is output.

[0047] For example, a series of fundus images including high-resolution images of the lesion area are acquired using optical coherence tomography. A standard convolutional neural network, such as VGG-16, is applied to process the images and extract multi-scale features, including shape, texture, and color. The texture features identified by the model in the first layer may not be clear enough, and the clarity score is lower than the preset threshold of 0.5. Due to the lack of clarity of some features, a pyramid scene parsing network, such as PSPNet, is applied to enhance the feature expression. The enhanced features are processed using image pyramid parsing technology to obtain feature maps at different resolutions. On the feature maps of each resolution, the Canny edge detection algorithm is applied to locate the cell and vascular structure areas. On the located cell structure area, the cell distribution density is annotated, and the cell density in the area is obtained as 200 cells per square millimeter. A convolutional neural network model is trained using the annotated sample images to predict the cell density distribution. On the test data set, the cell density distribution predicted by the model shows a uniformity of 0.6, which is less than the preset uniformity threshold of 0.7, indicating that the range of cell infiltration has expanded. The convolutional neural network model is also used to identify the vascular area in the image. The vascular leakage is evaluated by calculating the ratio of the area of ​​the dilated vascular area to the area of ​​the normal vascular area. The area of ​​the dilated vascular area is 150% of the area of ​​the normal vascular area, which exceeds the preset leakage threshold of 120%, indicating that the vascular leakage has worsened. The evaluation results of cell infiltration and vascular leakage are combined to output an assessment report on the risk of worsening of the disease. The report points out that due to uneven cell density and vascular leakage exceeding the threshold, there is a risk of worsening of the disease.

[0048] S103. Based on the morphological processing and feature analysis of multi-phase vascular images, the tortuosity and fragility of the new blood vessels are predicted, and the probability of complications in patients is evaluated.

[0049] Through the fundus image library, multi-phase vascular images are obtained, and the vascular structure morphology is enhanced by binarization and morphological processing of dilation and erosion. If the clarity of the vascular contour in the enhanced image is less than the preset second clarity threshold, the Gaussian blur technique is used to smooth the noise and irregular edges in the image to obtain a vascular image with a clarity greater than the preset second clarity threshold. For the vascular image with a clarity greater than the preset second clarity threshold, a thinning algorithm based on distance transformation is used to extract the single-pixel vascular centerline. The number of turning angles on the vascular centerline is calculated, and the number of turning angles is accumulated to determine the total degree of tortuosity of the new blood vessels. The tortuosity values ​​at different phases are obtained and a time series is formed in chronological order. According to the historical time series of the tortuosity values, the ARI MA algorithm is used for model training to predict the change trend of the tortuosity values. The histological characteristics of the blood vessels are extracted, including the integrity of the vessel wall and the distortion of the lumen. According to the histological characteristics of the blood vessels, the support vector machine algorithm is used for model training to construct a vascular fragility assessment model to determine the risk of vascular rupture. According to the change trend of the tortuosity and the judgment of vascular fragility, the probability of complications in patients is assessed.

[0050] For example, a series of vascular images at different time points were obtained from the fundus image library. A total of 30 images were obtained, covering a period of 6 months, with 5 images per month. Morphological processing, including dilation and erosion operations, was applied to these images to enhance the morphology of vascular structures. The threshold binarization method was applied to convert the images into binary images, and the threshold was set to a grayscale value of 150. For those images with vascular contour clarity lower than the preset threshold of 0.6, Gaussian blurring techniques were applied with a kernel size of 5x5 to smooth the noise and irregular edges in the images. A single-pixel vascular centerline was extracted using a distance transform-based thinning algorithm. In one image, the length of the extracted centerline was 200 pixels. The number of turning angles was calculated for the vascular centerline of each image, and an average of 10 turning angles were detected in each image. These tortuosity values ​​were arranged in chronological order to form a time series, which was then analyzed and trended using the ARI MA model. The histological characteristics of the vessels were extracted from the images, including vessel wall integrity and lumen distortion, and it was found that some vascular regions showed vessel wall incompleteness. The support vector machine algorithm was used to train the vascular fragility assessment model. The input features included the integrity of the vessel wall and the degree of lumen distortion. In the test data set, the model predicted a high risk of vascular rupture. Combining the trend of tortuosity and vascular fragility assessment, the probability of complications in patients was predicted. The model predicted that areas with high tortuosity and vessel wall fragility had a 40% probability of developing severe complications.

[0051] S104. Based on the analysis of the patient's lesion markers and vascular characteristics and in combination with the blood test results, the risk of serious complications is predicted, and pseudo-lesion images that are mistaken for lesion images are re-identified and excluded.

[0052] Obtain lesion markers, vascular features, and clinical indicator data sets, and construct sample feature vectors. The feature vectors include lesion size, lesion shape, lesion location information, vascular structural features, and clinical indicators. Clinical indicators are blood test results. Use the Pearson correlation coefficient method to perform correlation analysis on the sample vectors, and identify feature sets with a correlation with worsening of the disease that is higher than the preset correlation threshold, as early warning signals, including lesion size, lesion shape, vascular structural features, and clinical indicators. Based on the early warning signals, use the support vector machine algorithm for model training, build a severe complication risk prediction model, and assess the patient's risk of severe complications. Based on the historical data of the early warning signals, use the long short-term memory network algorithm for model training, build a severe complication association model, and predict the change trend of each feature of the early warning signal over time. Monitor the time series changes of the warning signal in the new sample. If the characteristics of the warning signal are abnormal, obtain the abnormal pattern of the characteristics of the warning signal, which includes sudden changes in characteristics, periodic fluctuations in characteristics, trend breaks, and abnormal fluctuation ranges of characteristics; according to the abnormal pattern of the characteristics of the warning signal, use the support vector machine algorithm to train the model, re-identify pseudo-lesions, identify and exclude pseudo-lesion images that are mistaken for lesion images; according to the warning signals of the re-screened lesion images, use the severe complication risk prediction model to predict the risk of severe complications. If the predicted risk of severe complications is greater than the preset risk threshold, the development path of the disease is predicted in combination with the patient's clinical data and image analysis results. The clinical data includes age, health status, and medical history.

[0053] For example, a dataset of 100 samples is obtained from the fundus lesion image database, each of which contains lesion markers and vascular feature information, as well as clinical indicators such as hemoglobin level and white blood cell count. A feature vector is constructed, including lesion size with an average area of ​​2 square millimeters, lesion shape with irregular or round shape, vascular structure features with vascular density and dilation, and blood test results with hemoglobin level of 13g / dL and white blood cell count of 7,000 / microliter. Correlation analysis is performed using the Pearson correlation coefficient method, and it is found that lesion size and vascular dilation are highly correlated with worsening of the disease. The Pearson correlation coefficient is greater than the preset threshold of 0.7, and these features are used as early warning signals. The support vector machine algorithm is used to train a severe complication risk prediction model based on the early warning signal. The model predicts that 45% of the samples have a risk of severe complications on the test set. The long short-term memory network algorithm is used to train the model on historical data to construct a severe complication association model. The model predicts that the growth trend of lesion size is highly correlated with the risk of severe complications in the next 3 months. By monitoring the time series changes of the warning signals in the data of another 50 patients in the new sample, the abnormal patterns of the characteristics of the warning signals are obtained. The abnormal patterns include sudden changes in characteristics, periodic fluctuations in characteristics, trend breaks, and abnormal ranges of characteristic fluctuations. For example, the rapid growth of lesion size in 10 samples is monitored, indicating an abnormal pattern. The support vector machine algorithm is used to re-identify pseudo-lesions for these 10 samples to exclude images misdiagnosed as lesions. The severe complication risk prediction model is applied to the screened samples, and 5 samples are predicted to have a risk of 80%, which is greater than the preset risk threshold of 70%, and requires immediate medical intervention. New medical imaging data of these 5 high-risk samples are obtained and analyzed using convolutional neural networks. It is found that 3 of the samples have neovascularization, indicating potential risk of severe complications.

[0054] Pseudo-lesions are classified according to the pattern of changes in their characteristics.

[0055] Obtain retinal fundus images from multiple time points, including annotated real lesion and pseudo-lesion images, and preprocess the images, including denoising, color correction, and contrast enhancement. Extract key features of pseudo-lesions, including shape, size, edge sharpness, color consistency, and texture pattern. Based on the key features of historical pseudo-lesions, use the ARIMA algorithm to train the model, predict the feature changes of pseudo-lesions at different time points, and identify the pattern of pseudo-lesion feature changes, including size growth, color change, and texture evolution. Design a pseudo-lesion classification formula based on the pattern of pseudo-lesion feature changes

[0056] C=α·F shape +β·F color +γ·F texture +δ·ΔF time, evaluate the category of pseudo lesions, where C is the pseudo lesion classification evaluation index, and F shape , F color , F texture Represent the metrics of shape, color and texture respectively, ΔF time Represents the amount of change in the time series, and α, β, γ, and δ are weights determined based on historical feature data. According to the calculation results of the pseudo-lesion classification evaluation index and the preset index threshold, the pseudo-lesions are classified to determine the types of pseudo-lesions, including optical artifacts, pixel noise, vascular shadows, retinal pigment deposition, and vitreous floaters. Use the labeled data set to train the pseudo-lesion classification formula, optimize the parameters, and improve the classification accuracy. Verify the performance of the pseudo-lesion classification formula on an independent test data set to evaluate the accuracy and reliability of the pseudo-lesion classification formula in actual clinical applications. Adjust the pseudo-lesion classification formula parameters and the weight of the classification formula according to the test results. Integrate the pseudo-lesion classification formula into existing ophthalmic diagnostic tools.

[0057] For example, there is a series of retinal fundus images from the same patient, taken in January, March, and May. In these images, there is a specific pseudo lesion. The images from January, March, and May are denoised, color corrected, and contrast enhanced, and the contrast is enhanced using adaptive histogram equalization. The features of the pseudo lesions are extracted at each time point. The size of the pseudo lesions is 2mm×2mm in January, 2.5mm×2.5mm in March, and 3mm×3mm in May. The color changes from light red to dark red, and the texture changes from smooth to slightly rough. The ARIMA algorithm is used to analyze the features of the pseudo lesions, the changes in size, color, and texture over time. The model predicts that by July, the size of the pseudo lesions may grow to 3.5mm×3.5mm, the color will be darker red, and the texture will be rougher. Apply the formula C=α·F shape +β·F color +γ·F texture +δ·ΔF time , where α, β, γ, and δ are obtained through historical data training, and the determined weights are

[0058] α=0.3,β=0.4,γ=0.2,δ=0.1, calculate the C value and get

[0059] C = 0.3 × 2 + 0.4 × 3 + 0.2 × 2 + 0.1 × 1 = 2.3. If the preset classification threshold is 2.0, since the C value is 2.3, which exceeds the threshold, pseudo-lesions may be classified as categories with specific characteristics, such as pigmentation. Use a large number of annotated data sets to train the model and adjust the weights in the formula to improve classification accuracy. Validate the model on an independent test data set to ensure accuracy and reliability. Adjust the model parameters based on the test results and integrate the optimized model into the ophthalmic diagnostic tool.

[0060] S105. Based on the time series analysis of fundus disease images, the characteristic changes of pseudo-lesions and real lesion areas are compared to evaluate the disease progression path and conduct early intervention on the disease.

[0061] Obtain fundus disease image time series samples, including fundus disease image samples in the early, middle and late stages of the disease. Use the hill climbing algorithm to evaluate the information density and degree of change in each frame of the image, and select key frames with information density greater than the preset density threshold and degree of change greater than the preset change threshold. For the selected key frames, use the convolutional neural network-based segmentation algorithm U-Net to annotate the lesions. According to the annotation results, extract the significant features of the lesions, including morphology, size, color, and texture. Compare the changes in the significant features of the same lesion at different time points to determine the disease progression path, which includes disease progression, disease improvement, and disease stabilization. Use the convolutional neural network to process the fundus image data in the time series, learn the disease progression pattern, establish the disease course dynamics change map, and display the progression stage and change trend of the disease in a visual way. Compare the information and feature change rules in the disease course dynamics change map with the feature change rules of the pseudo-lesion images that are mistaken for lesion images, and obtain the change trends of the pseudo-lesions and real lesion areas with treatment and disease progression. Obtain the early change patterns of different types of pseudo-lesion features, calculate the correlation between the change patterns of different types of pseudo-lesion features and the stage of the disease through the Pearson correlation coefficient method, and evaluate whether early intervention of the disease is needed. The change patterns include morphological changes, color changes, and the evolution of texture patterns. The disease stages include early, middle, and late stages. If the correlation between the change pattern of pseudo-lesion features and the stage of the disease is greater than the preset correlation threshold, early intervention of the disease is required; based on the information and feature change rules in the disease course dynamics change map, evaluate the risk of disease deterioration at different stages. Feature change rules include the growth of lesion size, changes in morphology, or deepening of color. Combine the patient's clinical data and imaging analysis results to predict the disease development path. Clinical data include age, health status, and medical history.

[0062] For example, 150 samples of time series image data were collected from the fundus lesion database, including fundus images of the early, middle and late stages of the lesions. Each sample contained one image taken every month during a period of 6 months, for a total of 6 images. The hill climbing algorithm was applied to analyze the image sequence of each sample, and the information density threshold was set to 0.7 and the change degree threshold was set to 0.6. The key frames that met the conditions were selected. For a certain sample, the images of the 3rd and 5th months were selected as key frames. The U-Net convolutional neural network was used to automatically annotate the key frames with lesions, and the significant features of the lesions were extracted, such as irregular morphology, average size of 2 square millimeters, dark red color and rough texture. By processing the image data in the time series, the development pattern of the disease was learned, and a dynamic change map of the disease course was established to visualize the progression stage and change trend of the disease. By comparing the real lesions in the dynamic change map of the disease course with the pseudo-lesion images that were mistaken for lesions, it was found that the real lesion area increased significantly in the 5th month, while the pseudo-lesion did not change much in size. The Pearson correlation coefficient method was used to analyze the correlation between the characteristic change patterns of different types of pseudo-lesions and the disease stage. It was found that the correlation coefficient between the morphological change pattern of the vascular shadow pseudo-lesion and the mid-stage of the disease was 0.75, which exceeded the preset threshold of 0.7. Since the characteristic change pattern of the vascular shadow pseudo-lesion is highly correlated with the disease stage, it is judged that the sample needs early intervention. Based on the information in the disease course dynamics change map, the risk of disease deterioration is assessed, and combined with clinical data, the patient is 45 years old and has a history of diabetes, the development path of the disease is predicted.

[0063] Based on the characteristic analysis of fundus disease images, the dynamic changes of pseudo-lesions are monitored and early intervention strategies are adjusted in real time.

[0064] Obtain fundus disease images, apply convolutional neural networks to model training, and extract pseudo-lesion features, including morphological details, color depth, and texture accuracy. Based on retinal fundus images, use dynamic Bayesian networks to train models, monitor changes in pseudo-lesion features over time, and predict future trends in pseudo-lesion features. Based on pseudo-lesion features, establish an early intervention scoring formula.

[0065] G=ω*F detail +θ*ΔF time +λ*P trend , predicting the timing of early intervention, where G represents the intervention index,

[0066] F detail Indicates the feature detail score of pseudo-lesions, ΔF time Indicates the rate of change of the feature over time, P trendrepresents the predicted future trend score, and ω, θ, and λ are weight parameters based on historical pseudo-lesion feature data. Based on the real-time retinal fundus image monitoring results, the changes in pseudo-lesion characteristics are continuously monitored, and the intervention index is adjusted according to the latest data to update the intervention recommendations in real time. The early intervention scoring formula is tested in clinical trials to obtain feedback from doctors and patients on the accuracy of the intervention timing. The parameters of the early intervention scoring formula are adjusted and optimized based on clinical results. The early intervention scoring formula is integrated into the existing ophthalmic diagnosis and decision support system to provide doctors with comprehensive treatment recommendations.

[0067] For example, a convolutional neural network is used to train a model for fundus disease images. 1,000 fundus images can be used as a training set, which includes images of normal fundus and fundus diseases. After training, the characteristics of fundus diseases can be automatically identified. When evaluating the characteristics of fundus diseases, three indicators, morphological detail, color depth, and texture accuracy, can be used for scoring. The score of each indicator is represented by a range of 0-100. For a fundus image, it can be calculated that its morphological detail score is 80, the color depth score is 70, and the texture accuracy score is 90. In order to monitor the changes in pseudo-lesion features over time and predict future trends, a dynamic Bayesian network can be used for model training. There are fundus image data at 100 time points, which can be used to train the model and obtain a model that can predict the trend of feature changes. According to the pseudo-lesion characteristics, an early intervention scoring formula G=ω*F can be established detail +θ*ΔF time +λ*P trend The weight parameters ω = 0.5, θ = 0.3, λ = 0.2 are selected. Based on the previous example, the feature detail score F of the pseudo-lesion can be calculated. detail =80, the rate of change of the feature over time ΔF time =10, predicted future trend score P trend =70. Substituting these values ​​into the formula G=0.5*80+0.3*10+0.2*70, we can get the intervention index G=57. Based on the real-time retinal fundus image monitoring results, the changes in pseudo-lesion characteristics can be continuously monitored and the intervention index can be adjusted based on the latest data. The latest fundus image data shows the feature detail score F detail =85, the rate of change of the feature over time ΔF time =15, predicted future trend score P trend=75. Substituting these values ​​into the formula G=0.5*85+0.3*15+0.2*75, the adjusted intervention index G=62 can be obtained. When testing the early intervention scoring formula in clinical trials, doctors and patients can be invited to participate and let them evaluate the accuracy of the intervention timing based on their own experience and feelings. In the trial, doctors and patients agreed that the accuracy of the intervention timing was high. When adjusting and optimizing the parameters of the early intervention scoring formula based on clinical results, adjustments can be made based on the feedback from doctors and patients. For example, if they think that the weight of the feature detail score is high, the value of the parameter ω can be increased. Finally, the early intervention scoring formula is integrated into the existing ophthalmic diagnosis and decision support system to provide doctors with comprehensive treatment recommendations. For example, when the doctor inputs the fundus image, the system can automatically calculate the intervention index and give corresponding intervention recommendations to help the doctor decide whether to take early intervention measures.

[0068] S106. Segment the neovascularization area based on the sensitivity index filtering technology, and evaluate the risk of disease progression through vascular feature statistics.

[0069] Obtain fundus images, use convolutional neural networks for model training, build a diseased vascular region recognition model, extract multi-scale vascular features, and locate the diseased vascular region. Intercept the diseased vascular image, and use sensitive index filtering to enhance the details of the inner wall of the lumen. Based on the similarity of the feature vectors of the diseased vascular image, retrieve cases from the sample library whose similarity difference with the feature vector of the current diseased vascular image is less than the preset difference threshold. The feature vector includes the size, shape, and color of the lesion area. The cases retrieved from the sample library contain annotated diseased vascular regions, and the annotated similar cases are prior knowledge. By using the annotated similar cases as additional training data and integrating them into the diseased vascular region recognition model, accurate segmentation of the diseased vessels is achieved. The segmentation results using the diseased vascular region recognition model are post-processed, including removing small noise areas and smoothing edges. The vascular features of the identified new blood vessels are statistically analyzed, including distribution range, density, and curvature. According to the vascular features of the new blood vessels, the decision tree algorithm is used for model training to evaluate the risk of disease deterioration based on the current vascular features.

[0070] For example, 50 fundus images are collected from the medical database, each image shows the retinal condition of different patients. The pre-trained convolutional neural network model VGG-16 is used to process the images, extract multi-scale vascular features, and locate the diseased vascular area. The diseased vascular area is intercepted from each fundus image, and the size of each intercepted area is 200x200 pixels. The sensitive index filtering technology is used to enhance the details of the inner wall of the lumen so that the structure of the blood vessel can be observed more clearly. Based on the intercepted diseased vascular image, a feature vector is constructed, including the size of the lesion area is 20 square millimeters, the morphology is irregular, and the color is dark red. The similarity difference threshold is set to 0.3, and cases with a similarity difference of less than 0.3 with the current diseased vascular image feature vector are retrieved from the sample library. The retrieved similar cases, including the annotated diseased vascular area, are used as additional training data and fused into the diseased vascular area recognition model, and the updated model is used to accurately segment the diseased vascular area. The segmentation results are post-processed, including removing noise areas less than 5 square millimeters and smoothing edges. The characteristics of the new blood vessels identified by statistics, such as the distribution range of 50 square millimeters, the density of 10 blood vessels per square millimeter, and the average curvature of 0.5. The decision tree algorithm is used to train the model to assess the risk of disease progression based on the current blood vessel characteristics. In the test data set, the model predicts that 30% of the samples have a high risk of disease progression.

[0071] S107. Based on OCT and OCTA images, monitor the status of vascular lesions after treatment, identify the association between different treatment strategies and high-risk complications, evaluate the treatment effect and guide the selection of personalized treatment strategies.

[0072] Obtain fundus optical coherence tomography images after treatment, and use exponential smoothing denoising technology to improve image clarity and visibility of details. According to fundus optical coherence tomography images after treatment, use the Canny edge detection algorithm to segment the contours of each layer of retinal tissue, identify and mark each retinal layer of the retina, including the retinal neuroepithelial layer, photoreceptor layer, outer nuclear layer, outer limiting membrane, connecting layer between the outer nuclear layer and the inner nuclear layer, inner nuclear layer, and nerve fiber layer. Measure the distance between the retinal layers to determine whether the retinal thickness is normal. Detect the grating distortion in the image, and locate the grating area in the image using the Canny edge detection algorithm to determine whether the grating morphology is distorted. Determine whether the retinal thickness and grating morphology have returned to the ideal state. If not, use OCT angiography imaging to compare with conventional imaging to determine the vascular regeneration. The ideal state is determined by the preset retinal thickness ideal state threshold and the preset grating morphology ideal state threshold. In the OCTA image, the Canny edge detection algorithm is used to identify new blood vessels, and the number, diameter, and tortuosity of new blood vessels are counted. According to the number, diameter, and tortuosity of new blood vessels, the decision tree algorithm is used for model training, and a decision tree model is constructed to determine high-risk vascular complications. Data on different treatment strategies are obtained, including treatment type, duration, and dosage. Fundus images, vascular lesions, and recorded complications of patients after treatment are obtained. According to the data of different treatment strategies, patient age, course of disease, and treatment response, the support vector machine algorithm is used for model training to evaluate the risk of complications of different treatment strategies; based on the risk of complications of different treatment strategies, clinical decision-making is guided to select the treatment plan that best suits the patient.

[0073] For example, 30 post-treatment fundus OCT images were collected from the hospital database, and exponential smoothing denoising technology was applied to each image to improve image clarity and visibility of details. The Canny edge detection algorithm was applied, and the edge detection threshold was set to 100 and 200 to segment the contours of the retinal tissue layers, and the various layers of the retina were identified and marked, including the retinal neuroepithelium, photoreceptor layer, outer nuclear layer, outer limiting membrane, connection layer between the outer nuclear layer and the inner nuclear layer, inner nuclear layer, and nerve fiber layer. The distance between the retinal layers was measured, and the distance from the retinal neuroepithelium to the photoreceptor layer was 120 microns. The grating area was located by the Canny algorithm, and its morphology was judged. It was found that the grating morphology was slightly distorted in some areas. The ideal state threshold of retinal thickness and grating morphology was set. If the retinal thickness and grating morphology did not return to the ideal state, the OCTA image was used to analyze the vascular regeneration. The Canny edge detection algorithm was used to identify new blood vessels, and the number of new blood vessels was counted as 5, the average diameter was 50 microns, and the average curvature was 0.3. The decision tree algorithm was used to train the model, and the risk of high-risk vascular complications was assessed based on the characteristics of neovascularization. The model predicted that 20% of the samples were at high risk. Data on different treatment strategies were collected, such as anti-VEGF injection, duration of 6 months, and dosage of 2 mg each time. The support vector machine algorithm was used to train the model to evaluate the risk of complications of different treatment strategies and compare them with the patient's age of 50 years, course of disease of 3 years, and treatment response. Based on the results of the model output, doctors are provided with treatment recommendations that best suit the patient.

[0074] The above description is merely a preferred embodiment of one or more embodiments of the present specification and is not intended to limit one or more embodiments of the present specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present specification shall be included in the scope of protection of one or more embodiments of the present specification.

Claims

1. A method for detecting pseudo-lesions of the fundus, characterized in that: The method comprises: After collecting multi-phase and multi-modal fundus lesion image sequences, a convolutional neural network based on the position attention mechanism is used to identify pseudo-lesions and predict the change rules of boundary heterogeneity to obtain preliminary lesion markings. Based on the preliminary lesion marking, image feature analysis of the lesion area is performed to locate and evaluate cell density and vascular leakage, and determine the risk of disease aggravation; Based on the morphological processing and feature analysis of multi-phase vascular images, the tortuosity and fragility of new blood vessels are predicted, and the probability of complications in patients is assessed; Based on the patient's lesion markers and vascular characteristics analysis, combined with blood test results, the risk of serious complications is predicted, and pseudo-lesion images that are mistaken for lesion images are re-identified and excluded; Through time series analysis of fundus disease images, the characteristic changes of pseudo-lesions and real lesion areas are compared to evaluate the progression of the disease and conduct early intervention on the disease. The sensitivity index filtering technology is used to segment the neovascularization area and the risk of disease progression is assessed through vascular feature statistics. Based on OCT and OCTA images, monitor the status of vascular lesions after treatment, identify the association between different treatment strategies and high-risk complications, evaluate treatment effects and guide the selection of personalized treatment strategies; The method predicts the tortuosity and fragility of new blood vessels and assesses the probability of complications in patients based on morphological processing and feature analysis of multi-phase vascular images, including: Through the fundus image library, multi-phase vascular images are obtained, and binarization, dilation and erosion morphological processing are used to enhance the vascular structure morphology; If the clarity of the blood vessel contour in the enhanced image is less than a preset second clarity threshold, a Gaussian blur technique is applied to smooth the noise and irregular edges in the image to obtain a blood vessel image with a clarity greater than the preset second clarity threshold; For the blood vessel images whose clarity is greater than a preset second clarity threshold, a thinning algorithm based on distance transformation is used to extract the single-pixel blood vessel centerline; Calculate the number of turning angles on the midline of the blood vessel, accumulate the number of turning angles, and determine the overall degree of tortuosity of the new blood vessels; Obtain the tortuosity values ​​at different phases and construct a time series in chronological order; According to the historical time series of tortuosity values, the ARIMA algorithm is used to train the model and predict the changing trend of tortuosity values; Extract the histological characteristics of blood vessels, including vessel wall integrity and lumen distortion; Based on the histological characteristics of blood vessels, the support vector machine algorithm is used for model training to build a vascular fragility assessment model to determine the risk of vascular rupture; The probability of complications in patients is assessed based on the changing trend of tortuosity and vascular fragility.

2. The method according to claim 1, wherein: After collecting the multi-phase multi-modal fundus lesion image sequence, the pseudo-lesions are identified and the boundary heterogeneity change law is predicted by a convolutional neural network based on the position attention mechanism to obtain a preliminary lesion mark, including: Use multimodal imaging technology to obtain multi-phase fundus images, including optical coherence tomography and fundus fluorescence angiography; Preprocess the images, including noise reduction, contrast enhancement, and color correction; Based on the fundus images containing pseudo-lesion area annotations, a convolutional neural network model enhanced with position attention mechanism is trained to identify key areas in multi-phase and multi-modal fundus images, where the key areas are pseudo-lesion areas. Compare the brightness or color value of each pixel in the image with its adjacent pixels, perform pixel-level difference calculation, obtain the difference value between each pixel and its neighborhood, and highlight the pseudo-lesion area; If there is no pseudo lesion area in the fundus image, the lesion area is separated using the Canny edge detection algorithm based on the preprocessed multi-phase and multi-modal fundus image; The segmented lesion image is checked for boundary integrity. If the lesion boundary is incomplete, the Snake algorithm is used to complete the image. For the completed lesion image, extract the morphological features and judge whether the morphology is stable. If the morphology is stable, record the lesion mark; if the morphology is unstable, monitor and record the lesion morphological changes; According to the recorded changes in lesion morphology, images with enhanced lesion heterogeneity were screened, and the SLIC algorithm was used to segment the images with enhanced lesion heterogeneity to obtain preliminary lesion markings. Heterogeneity enhancement refers to the situation where the difference between the lesion area and the surrounding normal tissue in morphology, color and texture is greater than the preset difference threshold; By comparing the changes of the lesion areas divided in the continuous image sequence over time, the area increase rate and boundary expansion speed of the lesion area are obtained to evaluate the regional growth rate; If the growth rate of the heterogeneous region is less than the preset rate threshold, it is judged as a benign lesion; if the growth rate of the heterogeneous region is greater than the preset rate threshold, it is judged as a malignant lesion.

3. The method according to claim 1, wherein: The method includes: performing image feature analysis of the lesion area based on the preliminary lesion marking, locating and evaluating the cell density and vascular leakage, and determining the risk of disease aggravation, including: Obtain images of the lesion area, process the images using a convolutional neural network, and extract multi-scale features of the images, including shape, texture, and color, through multi-level convolution and pooling operations; Determine the clarity of the extracted features at each scale, and if the clarity is lower than a preset first clarity threshold, use a pyramid scene parsing network to enhance feature expression; Perform image pyramid analysis on the enhanced features to obtain feature maps at different resolutions; On each resolution feature map, the Canny edge detection algorithm is used to locate the tissue structure area; The located tissue structure area is included in the sample library, and the cell distribution density of the tissue structure area is marked; Based on the labeled sample images, the convolutional neural network algorithm is used for model training to build a cell density assessment model, predict cell density distribution, and determine the uniformity of cell density; If the predicted result is that the uniformity of the cell distribution density is less than the preset uniformity threshold, it is determined that the cell infiltration range is expanded; if the predicted result is that the uniformity of the cell distribution density is greater than the preset uniformity threshold, the original evaluation is maintained; Based on the labeled sample images, a convolutional neural network algorithm is used to train the model and identify the blood vessel areas in the image; By calculating the ratio of the area of ​​the dilated blood vessel area to the area of ​​the normal blood vessel area, the area difference between the segmented blood vessel area and the normal blood vessel area was compared, and the degree of vascular leakage was evaluated based on the area ratio; If the degree of vascular leakage is greater than the preset threshold, it is judged that the vascular leakage is aggravated, and if the degree of vascular leakage is less than the preset threshold, the vascular state is normal; Based on the results of the cell infiltration and vascular leakage levels, the risk of disease aggravation is assessed and an evaluation report is output.

4. The method according to claim 1, wherein: The method of predicting the risk of serious complications based on the patient's lesion markers and vascular characteristics analysis combined with blood test results, and re-identifying and excluding pseudo-lesion images that are mistaken for lesion images, includes: Obtain lesion markers, vascular characteristics and clinical index data sets, and construct sample feature vectors. The feature vectors include lesion size, lesion shape, lesion location information, vascular structural characteristics and clinical indicators. The clinical indicators are blood test results. Using the Pearson correlation coefficient method, we conduct correlation analysis on sample vectors and identify feature sets with a correlation with disease aggravation higher than a preset correlation threshold as early warning signals, including lesion size, lesion shape, structural characteristics of blood vessels, and clinical indicators; Based on the early warning signals, the support vector machine algorithm is used for model training to build a severe complication risk prediction model to assess the patient's severe complication risk; Based on the historical data of early warning signals, the long short-term memory network algorithm is used for model training to build a severe complication association model and predict the changing trend of each feature of the early warning signal over time; Monitor the time series changes of the warning signal in the new sample. If the characteristics of the warning signal are abnormal, obtain the abnormal pattern of the characteristics of the warning signal. The abnormal pattern includes sudden changes in characteristics, periodic fluctuations in characteristics, trend breaks, and abnormal fluctuation ranges of characteristics. According to the abnormal pattern of the characteristics of the warning signal, the support vector machine algorithm is used to train the model, re-identify the pseudo-lesion, and identify and exclude the pseudo-lesion images that are mistaken for lesion images; According to the early warning signals of the re-screened lesion images, the risk of serious complications is predicted using the serious complication risk prediction model; If the predicted risk of serious complications is greater than the preset risk threshold, the progression of the disease is predicted based on the patient’s clinical data and imaging analysis results. The clinical data includes age, health status, and medical history. It also includes: classifying pseudo-lesions according to the pattern of changes in the characteristics of pseudo-lesions, specifically including: obtaining retinal fundus images from multiple time points, including annotated real lesions and pseudo-lesion images, and preprocessing the images, including denoising, color correction and contrast enhancement; extracting key features of pseudo-lesions, including shape, size, edge sharpness, color consistency and texture pattern; using the ARIMA algorithm to train the model based on the key features of historical pseudo-lesions, predicting the characteristic changes of pseudo-lesions at different time points, and identifying the pattern of changes in the characteristics of pseudo-lesions, including size growth, color change and texture evolution; designing a pseudo-lesion classification formula based on the pattern of changes in the characteristics of pseudo-lesions C=α·F shape +β·F color +γ·F texture +δ·ΔF time , evaluate the category of pseudo lesions, where C is the pseudo lesion classification evaluation index, and F shape , F color , F texture Represent the metrics of shape, color and texture respectively, ΔF time represents the amount of change in the time series, and α, β, γ, and δ are weights determined based on historical feature data; based on the calculation results of the pseudo-lesion classification evaluation index and the preset index threshold, the pseudo-lesions are classified to determine the types of pseudo-lesions, including optical artifacts, pixel noise, vascular shadows, retinal pigment deposition, and vitreous floaters; the pseudo-lesion classification formula is trained using annotated data sets, the parameters are optimized, and the classification accuracy is improved; the performance of the pseudo-lesion classification formula is verified on an independent test data set, and the accuracy and reliability of the pseudo-lesion classification formula in actual clinical applications are evaluated; the pseudo-lesion classification formula parameters and the classification formula weights are adjusted based on the test results; the pseudo-lesion classification formula is integrated into existing ophthalmic diagnostic tools.

5. The method according to claim 1, wherein: The method of comparing the characteristic changes of pseudo-lesions and real lesion areas based on the time series analysis of fundus disease images, evaluating the disease progression path, and conducting early intervention on the disease includes: Obtain time series samples of fundus disease images, including fundus disease image samples in the early, middle and late stages of the lesion; The information density and degree of change in each frame of the image are evaluated by a hill climbing algorithm, and key frames with information density greater than a preset density threshold and a degree of change greater than a preset change threshold are selected; For the selected key frames, the convolutional neural network-based segmentation algorithm U-Net is used to mark the lesions; According to the annotation results, extract the significant features of the lesion, including shape, size, color and texture; Compare the changes in significant features of the same lesion at different time points to determine the disease progression path, which includes disease progression, disease improvement, and disease stabilization; Through convolutional neural networks, the fundus image data in time series is processed to learn the development pattern of the disease, establish a dynamic change map of the disease course, and display the progression stage and change trend of the disease in a visual way; Compare the information and feature change patterns in the disease course dynamics change map with the feature change patterns of pseudo-lesion images that are mistaken for lesion images, and obtain the change trends of pseudo-lesions and real lesion areas with treatment and disease progression; Obtain the early change patterns of different types of pseudo-lesion features, calculate the correlation between the change patterns of different types of pseudo-lesion features and the disease stage by using the Pearson correlation coefficient method, and evaluate whether early intervention of the disease is needed. The change patterns include morphological changes, color changes, and the evolution of texture patterns. The disease stages include early, middle, and late stages. If the correlation between the pseudo-lesion feature change pattern and the disease stage is greater than the preset correlation threshold, early intervention of the disease is required; The risk of disease progression at different stages is assessed based on the information and characteristic change patterns in the disease course dynamics map. The characteristic change patterns include the growth of lesion size and the change of shape or the deepening of color. Combine the patient's clinical data and imaging analysis results to predict the disease progression path. Clinical data includes age, health status and medical history. It also includes: based on the characteristic analysis of fundus disease images, monitoring the dynamic changes of pseudo-lesions and adjusting the early intervention strategy in real time, specifically including: obtaining fundus disease images, applying convolutional neural networks for model training, extracting pseudo-lesion features, including morphological details, color depth, and texture accuracy; based on retinal fundus images, using dynamic Bayesian networks for model training, monitoring the changes of pseudo-lesion features over time, and predicting the future trend of pseudo-lesion features; based on pseudo-lesion features, establishing an early intervention scoring formula, G = ω*F detail +θ*ΔF time +λ*P trend , predicting the timing of early intervention, where G represents the intervention index, F detail Indicates the feature detail score of pseudo-lesions, ΔF time Indicates the rate of change of the feature over time, P trend represents the predicted future trend score, ω, θ, λ are weight parameters based on historical pseudo-lesion feature data; based on the real-time retinal fundus image monitoring results, the changes in pseudo-lesion characteristics are continuously monitored, and the intervention index is adjusted according to the latest data, and the intervention recommendations are updated in real time; the early intervention scoring formula is tested in clinical trials to obtain feedback from doctors and patients on the accuracy of the intervention timing; the parameters of the early intervention scoring formula are adjusted and optimized according to clinical results; the early intervention scoring formula is integrated into the existing ophthalmic diagnosis and decision support system to provide doctors with comprehensive treatment recommendations.

6. The method according to claim 1, wherein: The method of segmenting the neovascularization area by using the sensitivity index filtering technology and evaluating the risk of disease progression by using vascular feature statistics includes: Obtain fundus images, use convolutional neural networks to train models, build diseased vascular area recognition models, extract multi-scale vascular features, and locate diseased vascular areas; Capture the diseased blood vessel image and use sensitive index filtering to enhance the details of the inner wall of the lumen; Based on the similarity of the feature vector of the lesion blood vessel image, the case whose similarity difference with the feature vector of the current lesion blood vessel image is less than a preset difference threshold is retrieved from the sample library, and the feature vector includes the size, shape and color of the lesion area; The cases retrieved from the sample library contain annotated diseased vascular regions, and the annotated similar cases are prior knowledge; By using annotated similar cases as additional training data and integrating them into the diseased vessel region recognition model, accurate segmentation of diseased vessels can be achieved; Post-processing the segmentation results using the diseased vessel region recognition model, including removing small noise areas and smoothing edges; Statistically identify the vascular characteristics of new blood vessels, including distribution range, density, and tortuosity; According to the vascular characteristics of the new blood vessels, a decision tree algorithm is used to train the model and evaluate the risk of disease progression based on the current vascular characteristics.

7. The method according to claim 1, wherein: Based on OCT and OCTA images, the system monitors the vascular lesions after treatment, identifies the association between different treatment strategies and high-risk complications, evaluates the treatment effect and guides the selection of personalized treatment strategies, including: Obtain fundus optical coherence tomography images after treatment, and use exponential smoothing denoising technology to improve image clarity and visibility of details; According to the fundus optical coherence tomography images after treatment, the Canny edge detection algorithm was used to segment the contours of the various layers of retinal tissue, identify and mark the various retinal layers of the retina, including the retinal neuroepithelial layer, photoreceptor layer, outer nuclear layer, outer limiting membrane, connecting layer between the outer nuclear layer and the inner nuclear layer, inner nuclear layer, and nerve fiber layer; Measure the distance between the layers of the retina to determine whether the retinal thickness is normal; Detect the grating distortion in the image, and locate the grating area in the image through the Canny edge detection algorithm to determine whether the grating shape is distorted; Determine whether the retinal thickness and grating morphology have returned to the ideal state. If not, use OCT angiography imaging to compare with conventional imaging to determine the state of vascular regeneration. The ideal state is determined by the preset retinal thickness ideal state threshold and the preset grating morphology ideal state threshold; In OCTA images, the Canny edge detection algorithm was used to identify new blood vessels and count the number, diameter, and tortuosity of the new blood vessels; According to the number, diameter and tortuosity of new blood vessels, the decision tree algorithm is used to train the model and build a decision tree model to judge high-risk vascular complications; Obtain data on different treatment strategies, including treatment type, duration, and dosage; Obtain fundus images, vascular lesions, and documented complications of patients after treatment; Based on the data of different treatment strategies, patient age, disease course and treatment response, the support vector machine algorithm was used to train the model and evaluate the risk of complications of different treatment strategies; Guide clinical decision-making based on the risk of complications of different treatment strategies and select the treatment plan that best suits the patient.

Images