Training method of intelligent medical decision model based on ophthalmic surgery

Abstract

The application belongs to the technical field of intelligent medical decision support systems, and discloses a training method of an intelligent medical decision model based on ophthalmic surgery, which comprises the following steps: obtaining original ophthalmic data of a patient and performing preliminary data processing to obtain an ophthalmic data set of the patient; analyzing image data in the ophthalmic data set of the patient, detecting and extracting an eye feature vector, and then preliminarily classifying the eye condition of the patient and marking an eye disease; extracting a clinical feature vector of non-image data in the ophthalmic data set of the patient, performing multi-source data fusion with the eye feature vector, and obtaining an influence feature vector; taking the influence feature vector as input to train an intelligent medical decision model for ophthalmic surgery; and explaining the prediction result of the intelligent medical decision model for ophthalmic surgery by using an explainability tool to generate a model explanation and a surgery evaluation report, thereby improving the accuracy of diagnosis and enhancing the explainability of the model and the ability of individualized treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent medical decision support systems, more specifically, the present application relates to a training method of an intelligent medical decision model based on ophthalmic surgery. BACKGROUND

[0002] In the modern medical field, intelligent medical decision models have gradually become an important tool for assisting doctors in diagnosis and treatment planning. However, traditionally, these information is often scattered and lacks an effective integration mechanism to fully utilize all available information resources, and the analysis method of single data source limits the comprehensive understanding of the patient's condition, affecting the quality of medical decision.

[0003] Although many existing medical decision support systems have begun to try to use multiple types of data for analysis and prediction, and try to apply artificial intelligence to the research in the medical field, the existing technology still has significant deficiencies, first, the data integration capability is limited, and the advantages of multi-source heterogeneous data are not fully utilized, resulting in information loss, second, the feature engineering efficiency is low, and there is a lack of automated and intelligent feature discovery and optimization process, which limits the model performance, finally, the model explanation is poor, and the complex model structure makes it difficult to understand and verify, reducing the user's acceptance, these problems together lead to the limitations and challenges of existing models in practical application. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a training method of an intelligent medical decision model based on ophthalmic surgery, comprising:

[0005] S1, obtaining the original ophthalmic data of the patient and performing preliminary data processing to obtain the ophthalmic data set of the patient;

[0006] S2, analyzing the image data in the ophthalmic data set of the patient, detecting and extracting the eye feature vector, including smoothing the image data, repeatedly reducing the size of the smoothed image to obtain different scale images, using an improved active contour algorithm to refine the contour curve of the first scale image, and refining the remaining scale images through the contour curve of the first scale image, using curve wave transformation to extract feature information at different scales, and integrating to obtain a comprehensive image feature vector;

[0007] Using the comprehensive image feature vector to construct a network graph, setting an initial filtering threshold, gradually increasing the filtering threshold, at each increase, traversing the network graph, identifying all edge sets and constructing topological features to obtain a topological feature vector, and fusing the comprehensive image feature vector and the topological feature vector to obtain an eye feature vector;

[0008] S3. Extract clinical feature vectors from non-imaging data in the patient's ophthalmology dataset, perform multi-source data fusion on the ocular feature vectors and the clinical feature vectors, and identify the influencing feature vectors that affect the patient's ophthalmic surgery outcome;

[0009] S4. Using the influencing feature vector as input, training an intelligent medical decision-making model for ophthalmic surgery is performed. Based on the trained intelligent medical decision-making model for ophthalmic surgery, the predicted success probability of the ophthalmic surgery and the predicted risk level of adverse postoperative outcomes are output;

[0010] S5. Based on the predicted success probability and predicted risk level, the prediction results of the ophthalmic surgery intelligent medical decision-making model are explained through interpretability tools to generate a model explanation and surgical evaluation report.

[0011] Furthermore, the patient ophthalmological dataset is obtained in the following manner:

[0012] Raw ophthalmology data include imaging data, electronic health records, clinical examination data, and wearable device and self-reported data;

[0013] Perform preliminary data processing on the patient's original ophthalmic data, including using data cleaning technology to clean the original ophthalmic data, standardizing and formatting the non-imaging data based on the original ophthalmic data after data cleaning, and encrypting the patient's personal identity information, and then integrating all the original ophthalmic data after preliminary data processing into the patient's ophthalmic data set.

[0014] Furthermore, the eye feature vector is extracted by:

[0015] Step 31: performing denoising, contrast adjustment, and size normalization on the image data according to the image data in the patient's ophthalmological dataset;

[0016] Step 32: For the processed image data, use a Gaussian filter to smooth the image;

[0017] Step 33: Reduce the size of the smoothed image to half of its original size to form a low-resolution image. Repeat the reduction process until the maximum number of repetitions is reached to obtain images of different scales, which are recorded as scaled images.

[0018] Step 34: Refine the contour curve of each scale image, and extract feature information at different scales using curvelet transform based on all refined scale images, and then integrate the feature information at different scales to obtain a comprehensive image feature vector;

[0019] Step 35: Convert the comprehensive image feature vector into a network graph, and process the network graph using topological data analysis technology to obtain a topological feature vector;

[0020] Step 36: Perform feature filling and normalization processing on the obtained comprehensive image feature vector and topological feature vector, and use the feature weighted summation method to fuse the processed comprehensive image feature vector and topological features to obtain the eye feature vector.

[0021] Furthermore, the method of obtaining the comprehensive image feature vector includes:

[0022] Arrange all scale images in descending order of scale size to obtain a scale image sequence;

[0023] Select the first scale image in the scale image sequence and set the initial contour curve of the first scale image, including:

[0024] For the first scale image, the local threshold segmentation technique is used to perform threshold segmentation on the scale image to obtain the foreground area and the background area;

[0025] For the scaled image after threshold segmentation, the edge detection algorithm is applied to generate a binary edge map of each scaled image. The points with a value of 1 in the edge map are edge points, and the points with a value of 0 are non-edge points.

[0026] For each foreground region on the scaled image, the ratio of the sum of all edge points in the foreground region to the area of ​​the foreground region is taken as the edge density of the foreground region; and the average value of the gradient intensity in the foreground region is taken as the average gradient intensity of the foreground region;

[0027] Set the edge density threshold and average gradient threshold. For any foreground area, if the edge density or average gradient intensity in the foreground area is greater than or equal to the set edge density threshold or average gradient threshold, it is determined that there is a clear edge in the foreground area. The foreground area is recorded as the retinal vessel area, and its edge is used as the initial contour curve of the retinal vessel:

[0028] If the edge density and average gradient intensity in the foreground area are both less than the set edge density threshold and average gradient threshold, the foreground area is judged to be a suspected eye lesion area, and the foreground area is recorded as the lesion area, and its boundary is used as the initial contour curve of the lesion area:

[0029] Integrate the obtained initial contour curves of all retinal blood vessels and the initial contour curves of the lesion area into the first scale image according to the coordinate positions to serve as the initial contour curves of the first scale image;

[0030] The initial contour curve of the first scale image is refined using the improved active contour algorithm to obtain the refined contour curve of the first scale image. Based on the refined contour curve of the first scale image, all remaining scale images in the scale image sequence are refined to obtain the refined contour curves of all scale images.

[0031] According to all scale images after the contour curve is refined, each scale image is transformed by the curvelet transform method to generate the curvelet coefficients of each scale image, which are horizontally spliced ​​into the initial image feature vector;

[0032] The local threshold method is used to denoise the initial image feature vector. The principal component analysis method is used to reduce the dimensionality of the denoised initial image feature vector to obtain a low-dimensional feature vector as the comprehensive image feature vector.

[0033] Furthermore, the method of obtaining the refined contour curves of all scale images includes:

[0034] Step 51: Based on the initial contour curve of the first scale image, a refinement function of the contour curve is defined. The precise position of the contour curve is adjusted and refined by the refinement function. The value of the refinement function is the sum of the fitting function value and the shift function value. The value of the fitting function is the weighted sum of the smoothness value and the curvature value of the contour curve.

[0035] The smoothing value of the contour curve is defined as: PH CC =∫(α0+α jl ×jl(CC(cs),edge))×|CC 1 (cs)| 2 dcs;

[0036] Among them, PH CC Indicates the smoothing value of the contour curve CC, CC 1 (cs) represents the first-order derivative of the contour curve CC at point cs, CC(cs) represents the coordinate position of point cs on the contour curve CC, CC(cs)=(x(cs),y(cs)), x(cs) and y(cs) represent the coordinates of point cs on the x-axis and y-axis, ∫dcs represents the integral along the curve CC, α0 represents the basic smoothing value weight factor, α jl represents the adjustment coefficient of distance jl, (α0+α jl ×jl(CC(cs),edge)) represents the total weight value in the formula, and jl(CC(cs),edge) represents the distance between CC(cs) and the nearest edge edge;

[0037] The curvature value of the contour curve is defined as: WQ CC =∫(β0+β jl ×jl(CC(cs),edge))×|CC 2 (cs)| 2 dcs; among them, CC 2 (cs) represents the second-order derivative of the contour curve CC at point cs, β0 represents the basic bending value weight, β jlan adjustment coefficient representing a distance;

[0038] The moving function value is defined as:

[0039]

[0040] wherein qw cd represents the weight of the cdth scale, At the scale cd, the scale image I cd represents the gradient intensity at the coordinate position CC(cs), and represents the gradient direction of the scale image at the coordinate position CC(cs), and represents the direction angle of the contour curve CC at CC(cs), represents the cosine value between the contour curve direction and the scale gradient direction;

[0041] Step 53: initializing a0, a jl , b0, b jl , and qw cd as a parameter combination, performing parameter tuning using a grid search, evaluating the cross-validation scores under different parameter combinations using k-fold cross-validation, and selecting an optimal parameter combination;

[0042] Step 54: iteratively updating the thinning function using the gradient descent method until the absolute difference between the thinning function value in the current iteration and the thinning function value in the last iteration is less than a preset thinning function difference threshold or the maximum number of iterations is reached, to obtain the thinned contour curve of the first scale image;

[0043] Step 55: taking the thinned contour curve of the cdth scale image as the initial contour curve of the (cd+1)th scale image in the scale image sequence;

[0044] Step 56: repeating steps 51 and 55 until the thinned contour curve of the last scale image in the scale image sequence is obtained.

[0045] Further, the manner of obtaining the topological feature vector comprises:

[0046] According to the comprehensive image feature vector, each coefficient in the comprehensive image feature vector is taken as a node, the cosine similarity between each two nodes is calculated, and a similarity threshold is set;

[0047] When the cosine similarity between two nodes is greater than or equal to the similarity threshold, an edge is established between the two nodes; and when the cosine similarity between two nodes is less than the similarity threshold, no edge is established between the two nodes;

[0048] According to the nodes and edges, a network graph is constructed;

[0049] Traverse the network graph, define the number of edges connected to a node in the network graph as the weight of the node;

[0050] Set an initial filtering threshold, increase the filtering threshold step by step, and check all node pairs each time the filtering threshold is increased. If the weights of the two nodes are both less than or equal to the current filtering threshold, connect the two nodes, and stop increasing the filtering threshold until there are edges between all nodes.

[0051] At each filtering threshold, construct the current simplicial complex according to the current edge set;

[0052] Wherein, the current edge set represents the set of all original edges in the network graph and all newly added edges under the current filtering threshold, and the simplicial complex represents a set of multiple simplices, including vertices, edges and high-dimensional simplices. Vertices are 0-dimensional simplices, representing a node, edges are 1-dimensional simplices, and high-dimensional simplices represent higher-dimensional simplices.

[0053] For each simplicial complex at each filtering threshold, the topological features in the simplicial complex are counted respectively, including the number of connected components, loops and voids. The definition of connected component is: a combination of nodes in the simplicial complex, at least one edge connecting any two nodes in the combination, and the combination cannot be expanded to a larger connected set by adding more nodes. Loop represents a closed path in the simplicial complex with the same start and end point, and void represents an area surrounded by one or more loops without other nodes inside the area.

[0054] For different filtering thresholds, starting from the initial filtering threshold, record the filtering threshold at which each topological feature first appears and the filtering threshold at which it first disappears.

[0055] The filtering threshold at which each topological feature first appears and the filtering threshold at which it first disappears are used as the interval of each topological feature.

[0056] Count the total number of topological features, define a topological network of tp x tp, and cover the interval between the initial filtering threshold and the maximum filtering threshold, and tp is greater than the total number of topological features.

[0057] For each topological feature, use a Gaussian kernel function as the weight function to calculate the weight value of each cell in the topological network Where QZH(q, p) represents the weight value, u r and v r are the filtering threshold at which the rth topological feature first appears and the filtering threshold at which it first disappears, q and p are the coordinates of the cells in the topological network, and σ represents the standard deviation of the Gaussian kernel.

[0058] For each topological network cell, the total weight value of all topological features in the cell is accumulated, and the total weight value of each cell on the topological network is horizontally spliced into a topological feature vector in the order of moving from the top left corner to the bottom right corner.

[0059] Further, the manner of identifying the influence feature vector affecting the effect of the patient's ophthalmic surgery includes:

[0060] According to the non-image data in the patient's ophthalmic data set, the natural language processing technology is used to convert the text type data in the non-image data into an embedding vector, and the embedding vector and the numerical type data in the non-image data are used as features to horizontally splice into a clinical feature vector, which is integrated with the eye feature vector into a comprehensive feature set;

[0061] The random forest classifier is selected as the base model, the comprehensive feature set is used as the input, the result of the ophthalmic surgery is defined as the target variable of the random forest model, the random forest model is trained, and the influence score of each feature on the ophthalmic surgery is output;

[0062] According to the influence score, one feature with the lowest influence score is removed from all features;

[0063] The remaining features are used as input to retrain the random forest model, calculate new influence scores, and remove the feature with the lowest new influence score, repeat the iteration until a predetermined number of features is reached, stop the iteration, and the remaining all features are used as the influence features affecting the effect of the patient's ophthalmic surgery, and are horizontally spliced into an influence feature vector.

[0064] Further, the training manner of the ophthalmic surgery intelligent medical decision model includes:

[0065] The deep neural network is used to construct the ophthalmic surgery intelligent medical decision model, the architecture of the ophthalmic surgery intelligent medical decision model is set as a shared layer and a specific task layer, and the specific task layer is set as a surgery success rate branch layer and a postoperative ideal recovery state branch layer;

[0066] The influence feature vector is input into the shared layer, and the output of the shared layer is input into the surgery success rate branch layer and the postoperative ideal recovery state branch layer, respectively;

[0067] The surgery success rate branch layer extracts features related to surgery success from the shared layer output through a fully connected layer and outputs a predicted success probability;

[0068] The postoperative ideal recovery state branch layer extracts features related to postoperative recovery from the shared layer output and outputs a predicted risk degree of not reaching an ideal recovery state;

[0069] collecting a sample set, including an influence feature vector of each patient and a corresponding label of whether the surgery is successful and a label of whether the patient achieves an ideal recovery state after the surgery;

[0070] proportionally dividing the sample set into a training set and a validation set, defining a loss function of the surgery success rate branch layer as a weighted binary cross-entropy loss, defining a loss function of the postoperative ideal recovery state branch layer as a binary cross-entropy loss, and defining a total loss function as a weighted sum of the weighted binary cross-entropy loss value and the binary cross-entropy loss value;

[0071] performing forward propagation on the model using the training set, calculating the loss values of the surgery success rate branch layer and the postoperative ideal recovery state branch layer using the corresponding loss functions respectively, and then calculating the total loss function, and performing back propagation through the total loss function, and updating the model parameters including the parameters of the shared layer and the two branch layers using the Adam optimization algorithm in each iteration of the model;

[0072] For each iteration, using AUC as an evaluation indicator, calculating the AUC value on the validation set;

[0073] According to the AUC value on the validation set, calculating the difference between the AUC value after the current iteration and the AUC value of the last iteration, denoted as an iteration difference value;

[0074] Setting an iteration difference value threshold, if the iteration difference value is greater than the iteration difference value threshold, it is determined that the performance of the model is improved;

[0075] If the iteration difference value is less than or equal to the iteration difference value threshold, it is determined that the performance of the model is not improved;

[0076] If the performance of the model on the validation set does not improve in continuous DC iterations, the training is stopped, and a trained ophthalmic surgery intelligent medical decision-making model is obtained.

[0077] Further, the way of interpreting the prediction result of the ophthalmic surgery intelligent medical decision-making model by the interpretability tool includes:

[0078] According to the original ophthalmic data of a new patient, predicting the predicted success probability and the predicted risk degree of the patient by the ophthalmic surgery intelligent medical decision-making model;

[0079] Using the SHAP algorithm, calculating the SHAP value of each feature in the influence feature vector according to the predicted success probability and the predicted risk degree of the patient respectively, as the contribution value of each feature to the prediction result of the ophthalmic surgery intelligent medical decision-making model, and obtaining the contribution value of each feature to the predicted success probability and the predicted risk degree respectively;

[0080] Setting a contribution threshold, if the contribution value of the feature to the prediction result is greater than the contribution threshold, it is determined that the feature will have an impact on the prediction result of the corresponding patient, and the feature is marked.

[0081] If the contribution value of the feature to the prediction result is less than or equal to the contribution threshold value, it is determined that the feature will not affect the prediction result of the corresponding patient, and the feature is not marked.

[0082] Further, the model explanation and operation evaluation report is generated in the following manner:

[0083] According to the prediction success probability, the prediction risk degree and the contribution value of each feature to the prediction success probability and the prediction risk degree, a model explanation and operation evaluation report is generated, comprising:

[0084] The patient information summary, the predicted success probability of the operation, the predicted risk degree after the operation and the marked features; the patient information summary includes the basic personal information, medical history and clinical examination data of the patient.

[0085] The technical effects and advantages of the training method of the intelligent medical decision model based on ophthalmic surgery are as follows:

[0086] The present application provides more comprehensive information for the model by multi-source data fusion of image data, electronic health records, clinical examination data and wearable device and self-reported data, which helps to improve the prediction accuracy; secondly, in the process of extracting the eye feature vector, an improved active contour algorithm is used to refine the image contour curve, which improves the accuracy of complex eye structure recognition; at the same time, by using the image after refining the contour curve by using the improved active contour algorithm to construct the network graph and applying topological data analysis, the deep structure information of the eye features can be captured, which is difficult to achieve in traditional methods; then, the random forest classifier is used for feature selection, and the feature with the lowest impact score is removed by iteration to ensure that the final feature set can best reflect the influencing factors of the patient's ophthalmic surgery effect; then, the design of shared layer plus specific task layer is adopted, and the success rate of operation and postoperative recovery state are predicted respectively, which makes the model pay attention to multiple targets at the same time, improves the effectiveness and pertinence of prediction; finally, the SHAP value is used to quantify the contribution of each feature to the model output, which enhances the transparency of the model, facilitates the doctor to understand and trust the prediction result of the model, and improves the performance and practicality of the intelligent medical decision model based on ophthalmic surgery. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 The present application is a schematic diagram of the training method of the intelligent medical decision model based on ophthalmic surgery;

[0088] Figure 2 The present application is a schematic diagram of the training system of the intelligent medical decision model based on ophthalmic surgery;

[0089] Figure 3Schematic diagram of the method for extracting eye feature vectors in the training method of the intelligent medical decision-making model based on ophthalmic surgery of the present invention. DETAILED DESCRIPTION

[0090] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0091] Example 1

[0092] See also Figure 1 and Figure 3 As shown, the training method of the intelligent medical decision-making model based on ophthalmic surgery described in this embodiment includes:

[0093] S1. Obtain the patient's original ophthalmological data and perform preliminary data processing to obtain the patient's ophthalmological dataset;

[0094] S2. Analyze the image data in the patient's ophthalmology dataset to detect and extract eye feature vectors;

[0095] S3. Extract clinical feature vectors of non-imaging data in the patient's ophthalmology dataset, perform multi-source data fusion on the ocular feature vectors and the clinical feature vectors, and identify the influencing feature vectors that affect the patient's ophthalmic surgery results;

[0096] S4. Using the influencing feature vector as input, training an intelligent medical decision-making model for ophthalmic surgery is performed. Based on the trained intelligent medical decision-making model for ophthalmic surgery, the predicted success probability of the ophthalmic surgery and the predicted risk level of adverse postoperative outcomes are output;

[0097] S5. Based on the predicted success probability and predicted risk level, interpret the prediction results of the ophthalmic surgery intelligent medical decision-making model using interpretability tools to generate a model explanation and surgical evaluation report;

[0098] Ways to obtain patient ophthalmology datasets include:

[0099] Original ophthalmic data include imaging data (obtained through methods such as retinal scans and optical coherence tomography (OCT), etc.), electronic health records (including the patient's basic personal information, medical history (including ophthalmic problems in the past five years, surgical history, and family medical history), and medication use records (including records of medications currently being used or used in the past six months by the patient), obtained through EHR), clinical examination data (including vision test results, intraocular pressure measurement data, and refractive status data), wearable devices (obtained through physiological data provided by smart watches or fitness trackers), and self-reported data (obtained through questionnaires or oral statements by patients);

[0100] Perform preliminary data processing on the patient's original ophthalmological data, including using data cleaning technology to clean the original ophthalmological data (including deleting duplicate records, filling missing values, and correcting obvious input errors). Based on the cleaned original ophthalmological data, standardize and unify the format of non-imaging data, encrypt the patient's personal identity information, and then integrate all the original ophthalmological data after preliminary data processing into the patient's ophthalmological data set;

[0101] Methods for detecting and extracting eye feature vectors include:

[0102] Step 31: Based on the image data in the patient's ophthalmological dataset, the image data is subjected to denoising, contrast adjustment, and size normalization processing (to ensure that all images meet standards suitable for analysis);

[0103] Step 32: For the processed image data, use a Gaussian filter to smooth the image (to reduce noise and protect edge information);

[0104] Step 33: Reduce the size of the smoothed image to half of its original size to form a low-resolution image. Repeat the reduction process until the maximum number of repetitions is reached to obtain images of different scales, which are recorded as scaled images.

[0105] Step 34: Refine the contour curve of each scale image. Based on all the refined scale images, extract feature information at different scales using curvelet transform. Then, integrate the feature information at different scales to obtain a comprehensive image feature vector (used to detect subtle lesions at specific scales, such as microaneurysms or minor hemorrhages in the early stages of diabetic retinopathy).

[0106] Step 35: Convert the comprehensive image feature vector into a network graph, and process the network graph using topological data analysis technology to obtain a topological feature vector;

[0107] Step 36: performing feature filling (comparing the lengths of the comprehensive image feature vector and the topological feature vector, and using a filling method to fill the shorter feature vector so that the lengths of the detailed feature vector and the topological feature vector are the same) and normalization processing on the obtained comprehensive image feature vector and the topological feature vector, and then using a feature weighted summation method to fuse the processed comprehensive image feature vector and the topological feature to obtain an eye feature vector;

[0108] The feature weighted summation method takes the detailed features and topological features as two feature sets, assigns a weight to each feature set, and the sum of the two weights is equal to 1;

[0109] For each feature set, multiply it by the corresponding weight, that is, multiply each feature in the feature set by the weight respectively, and then add the two feature sets after multiplication, that is, add the features at corresponding positions in the two feature sets to obtain the final fusion feature vector as the eye feature;

[0110] Ways to obtain the comprehensive image feature vector include:

[0111] Arrange all scale images in descending order of scale size to obtain a scale image sequence;

[0112] Select the first scale image in the scale image sequence and set the initial contour curve of the first scale image, including:

[0113] It should be noted that starting from the image with the lowest resolution (i.e. the largest scale) the processing is done. The structure in the low-resolution image is simpler, and the edges are easier to identify and track, which helps to quickly find the approximate location.

[0114] For the first scale image, use the local threshold segmentation technique (such as the Otsu method, adaptive threshold) to perform threshold segmentation on the scale image to obtain the foreground area and the background area (that is, use the local threshold technique to set a threshold and binarize the image, where the part with a pixel value greater than the threshold is regarded as the foreground (the foreground part is the suspected lesion area), and the rest is regarded as the background);

[0115] For the scaled image after threshold segmentation, an edge detection algorithm (such as Canny, Sobel, Prewitt) is applied to generate a binary edge map of each scaled image. Points with a value of 1 in the edge map are edge points, and points with a value of 0 are non-edge points.

[0116] For each foreground region on the scaled image, the ratio of the sum of all edge points in the foreground region to the area of ​​the foreground region is taken as the edge density of the foreground region; and the average gradient intensity in the foreground region is taken as the average gradient intensity of the foreground region (that is, the gradient intensity of the scaled image at position (x, y) in the foreground region is first calculated, and then the sum of the gradient intensities of all positions in the foreground region of the scaled image is taken as the ratio of the foreground region area to obtain the average gradient intensity).

[0117] Set the edge density threshold and average gradient threshold (based on industry expert experience). For any foreground area, if the edge density or average gradient intensity in the foreground area is greater than or equal to the set edge density threshold or average gradient threshold, it is determined that there is a clear edge in the foreground area. The foreground area is recorded as the retinal vessel area, and its edge is used as the initial contour curve of the retinal vessel:

[0118] If the edge density and average gradient intensity in the foreground area are both less than the set edge density threshold and average gradient threshold, the foreground area is judged to be a suspected eye lesion area, and the foreground area is recorded as the lesion area, and its boundary is used as the initial contour curve of the lesion area:

[0119] Integrate the obtained initial contour curves of all retinal blood vessels and the initial contour curves of the lesion area into the first scale image according to the coordinate positions to serve as the initial contour curves of the first scale image;

[0120] It should be noted that by comprehensively applying edge detection and threshold segmentation techniques, the position of the initial contour curve can be effectively and automatically determined, which is suitable for both accurate delineation of retinal blood vessels and preliminary positioning of lesion areas;

[0121] The initial contour curve of the first scale image is refined using the improved active contour algorithm to obtain the refined contour curve of the first scale image. Based on the refined contour curve of the first scale image, all remaining scale images in the scale image sequence are refined to obtain the refined contour curves of all scale images.

[0122] According to all the scale images after refining the contour curve, the curve wave transform is performed on each scale image through the curve wave transform method to generate the curve wave coefficients of each scale image, which are horizontally spliced into an initial image feature vector (after applying the curve wave transform to each scale and direction image, a group of coefficients is obtained, which constitutes a feature vector. This feature vector can be regarded as a multi-scale and multi-directional representation of the original image, which is called the initial image feature vector. The dimension of the initial image feature vector depends on the number of scales and directions set, that is, the product of the number of scales and the number of directions. For example, if there are 3 scales and 8 directions under each scale, the length of the initial image feature vector is 24. Each coefficient in the initial image feature vector represents the component intensity of the image in a specific scale and direction. Larger coefficients usually correspond to significant structures or features in the image, while smaller coefficients are mainly caused by noise).

[0123] The curve wave transform formula is: where, represents the curve wave coefficient at scale cd and direction , which reflects the component intensity of the image in a specific scale and direction. By calculating these coefficients, multi-scale and multi-directional features in the image can be captured, represents the direction angle index k at scale cd. Different correspond to different directions. The common practice is to divide 360 degrees into multiple directions (such as 8 or 16 directions) to comprehensively cover all possible direction changes, represents the double integration over the entire image plane. The integration range from negative infinity to positive infinity means that all pixel positions in the image are considered. In practice, since the image is usually of finite size, the integration range will be adjusted according to the actual image size. cd (x, y) represents the gray value of the scale image at position (x, y) after refining the contour curve by the improved active contour algorithm. The subscript cd indicates that this scale image is at scale cd. Through multi-scale analysis, image features can be extracted at different scales, represents the curve wave basis function related to scale cd and direction . Curve wave basis functions are usually generated by scaling and translating a series of mother functions (Mother Wavelet). For a given scale cd and direction , the form of the basis function changes to adapt to the feature extraction requirements in different scales and directions. dxdy represents an area element for integration operation. In the continuous case, dxdy represents an infinitesimal area in the two-dimensional plane. In the discrete case, it usually corresponds to a pixel in the image.

[0124] It should be noted that the curved wave transform can provide a powerful tool to analyze and represent the multi-scale and multi-directional features in the image, which is very useful for tasks such as refining the retinal blood vessels of the eye or locating the eye lesion area, it can capture the subtle structure and complex pattern in the image, and provide strong support for subsequent medical diagnosis and analysis, through the method of combining the improved active contour algorithm and the curved wave transform, not only can effectively enhance the key features in the image, but also can provide rich multi-scale and multi-directional feature description, this method is especially suitable for medical image analysis field, especially for the application scene which needs high precision edge detection and complex structure analysis;

[0125] The initial image feature vector is denoised using the local threshold method, and the low-dimensional feature vector obtained by dimension reduction processing of the denoised initial image feature vector using principal component analysis method is used as the comprehensive image feature vector;

[0126] The method for obtaining the refined contour curve of the image of all scales includes:

[0127] Step 51: defining a thinning function of the contour curve according to the initial contour curve of the first scale image, adjusting and refining the accurate position of the contour curve through the thinning function, and the value of the thinning function is the sum of the fitting function value and the moving function value;

[0128] The value of the fitting function is the weighted sum of the smooth value and the bending value of the contour curve;

[0129] The smooth value of the contour curve is defined as: PH CC =∫(α0+α jl ×jl(CC(cs),edge))×|CC 1 (cs)| 2 dcs;

[0130] Wherein, PH CC represents the smooth value of the contour curve CC, CC 1 (cs) represents the first derivative of the contour curve CC at point cs, that is, the change rate of the tangent direction, This quantity describes the direction and slope of the contour curve at point cs, CC(cs) represents the coordinate position of point cs on the contour curve CC, CC(cs)=(x(cs),y(cs)), x(cs) and y(cs) represent the coordinates of point cs on the x-axis and y-axis, and |CC 1 (cs)| 2represents the square modulus of the first derivative, ∫dcs represents the integral along the curve CC, here the integral is along the entire curve CC, that is, the integral is a cumulative operation on each point on the curve, the value of the entire curve is calculated by accumulating the contribution of each point, dcs represents the length element of the small segment along the curve, α0 represents the basic smoothing value weight factor (the initial basic smoothing value weight factor can be determined by random selection method), α jl represents the adjustment coefficient of the distance, (α0+α jl ×jl(CC(cs), edge)) represents the value of the total weight in the formula, jl(CC(cs), edge) represents the distance between CC(cs) and the nearest edge edge;

[0131] It should be noted that the smoothing value formula of the contour curve is improved from the smoothing term formula, and the original smoothing term formula is PH CC =∫α0×|CC 1 (cs)| 2 dcs, by dynamically adjusting the total weight value according to the distance between the current position of the contour curve and the nearest edge, when the contour curve approaches the target edge, d(CC(cs), edge) decreases, allowing more details to be captured, when the contour curve is far away from the target edge, the smoothing value is increased to maintain the smoothness of the overall contour curve, so that the contour curve has different smoothing requirements at different positions;

[0132] The bending value of the contour curve is defined as: WQ CC =∫(β0+β jl ×jl(CC(cs), edge))×|CC 2 (cs)| 2 dcs; wherein, CC 2 (cs) represents the second derivative of the contour curve CC at point cs, β0 represents the basic bending value weight, β jl represents the adjustment coefficient of the distance, when approaching the target edge, the requirement of bending value is reduced, so that the contour curve is more flexible and can better fit the complex boundary; when far away from the target edge, the requirement of bending value is increased to avoid unnecessary bending;

[0133] The moving function value is defined as:

[0134]

[0135] wherein, qw cd represents the weight of the cdth scale, is the gradient intensity of the scale image I cd at the coordinate position CC(cs), specifically, is the gradient vector of the scale image at the point, and its modulus represents the edge strength at this point, a larger gradient value usually corresponds to an edge or boundary in the scale image, represents the gradient direction of the scale image at coordinate position CC(cs), which is obtained by calculating the gradient vector of the image and taking its direction angle, it describes the direction in which the scale image intensity changes most rapidly at this point, θ(CC(cs) represents the direction angle of the contour curve CC at CC(cs), which is obtained by calculating the angle difference between adjacent points on the contour curve, it describes the tangent direction of the contour curve at this point, represents the cosine value between the contour curve direction and the scale gradient direction, which is used to further optimize the direction of the contour curve, so that it is more inclined to move along the actual edge in the image, thereby improving the performance on complex structures (such as retinal blood vessels), when the two directions are completely consistent, the cosine value is 1; when they are perpendicular, the cosine value is 0; when they are opposite, the cosine value is -1;

[0136] Step 53: initialize α0, α jl , β0, β jl and qw cd as parameter combinations, use grid search to optimize parameters, use k-fold cross-validation to evaluate the cross-validation score under different parameter combinations, and select the optimal parameter combination;

[0137] Step 54: use gradient descent method to iteratively update the thinning function, minimize the thinning function, and gradually approach the target edge with the initial contour curve, the position of the contour curve is updated at each iteration until the absolute difference between the thinning function value in the current iteration and the thinning function value in the last iteration is less than the preset thinning function difference threshold or the maximum iteration number is reached, and the thinned contour curve of the first scale image is obtained;

[0138] The update formula is: Δtt represents the step size, represents the gradient of the thinning function with respect to the position of the contour curve;

[0139] Step 55: take the thinned contour curve of the cdth scale image as the initial contour curve of the cd+1th scale image in the sequence of scale images;

[0140] Step 56: repeat steps 51 and 55 until the thinned contour curve of the last scale image in the sequence of scale images is obtained;

[0141] The way to obtain the topological feature vector includes:

[0142] According to the comprehensive image feature vector, each coefficient in the comprehensive image feature vector is taken as a node, the cosine similarity between each two nodes is calculated, and a similarity threshold (set by relevant personnel in the industry according to experience) is set.

[0143] When the cosine similarity between two nodes is greater than or equal to the similarity threshold, an edge is established between the two nodes; when the cosine similarity between two nodes is less than the similarity threshold, no edge is established between the two nodes;

[0144] According to the nodes and edges, a network graph is constructed;

[0145] Traverse the network graph and define the number of connections between the nodes in the network graph as the weight of the node;

[0146] Set an initial filtering threshold and gradually increase the filtering threshold. Check all node pairs each time the threshold is increased. If the weights of both nodes are less than or equal to the current filtering threshold, connect the two nodes until there are edges between all nodes. Then stop increasing the filtering threshold.

[0147] As the threshold increases, originally disconnected parts may begin to connect, forming larger connected components, and even more complex topological features such as rings or holes may appear;

[0148] At each filtering threshold, construct the current simplicial complex based on the current edge set;

[0149] Among them, the current edge set represents the set consisting of all the original edges in the network graph and all the newly added edges under the current filtering threshold. The simplicial complex represents a set consisting of multiple simplices, including vertices, edges and high-dimensional simplices. The vertex is a 0-dimensional simplex, representing a node, the edge is a 1-dimensional simplex, and the high-dimensional simplex represents a higher-dimensional simplex (for example, if three nodes form a triangle (that is, the three nodes are connected by edges), then this triangle can be added to the simplicial complex as a 2-simplex. Similarly, the same is true for higher-dimensional simplices (such as tetrahedron as a 3-simplex)).

[0150] A simplicial complex is a set of simplices such that: if a simplicial belongs to the current simplicial complex, then all its faces also belong to the current simplicial complex; if two simplices both belong to the current simplicial complex, then their intersection is also a simplicial of the current simplicial complex;

[0151] For each simplicial complex under each filtering threshold, the topological features within the simplicial complex are statistically obtained, including the number of connected components, rings and holes, where a connected component is defined as: a set of node combinations in the simplicial complex, in which any two nodes in the node combination are connected by a path consisting of at least one edge, and this set of nodes cannot be expanded into a larger connected set by adding more nodes (for example, a network graph has six nodes (A, B, C, D, E, F). Under a certain filtering threshold, there is an edge between the six nodes (ABCDEF). Then, these six nodes constitute a single connected component, because starting from any node, you can reach any other node through a series of edges. For another example, under a certain filtering threshold, there are two connected components between the six nodes (ABC, DEF). The first connected component includes the node {ABC}, and the second connected component includes the node {DEF}. There is no path connection between these two connected components, so they are independent. It should be noted that in eye images, connected components represent these situations: different anatomical areas of the retina (such as the macula, optic disc), lesions (e.g., exudates and hemorrhages in diabetic retinopathy), areas of retinal detachment, different branches of the vascular network, different structures within the optic disc (e.g., the optic cup), foreign bodies or spaces within the vitreous, or different layers or damaged areas of the cornea); a ring represents a closed path with the same starting and ending points in a simple complex (in ocular images, a ring represents the following: a ring-like structure in the retinal vascular network, a microvascular ring within the macula, a vascular ring around the optic disc, a ring-like path formed by the edge of a retinal break or tear, a choroidal neovascularization (CNV) ), annular structures formed in posterior vitreous detachment (PVD), or annular structures in corneal images (such as the pattern formed by the arrangement of corneal endothelial cells). A hole refers to an area in a simple complex surrounded by one or more rings and without other nodes inside (it should be noted that in ocular image analysis, a hole refers to the following situations: the blank area inside the retinal vascular ring, the avascular area of ​​the fovea, retinal holes or tears, cup-shaped depressions in the optic disc, local spaces in the vitreous cavity, or damaged or missing parts within the lesion area);

[0152] For topological features under different filtering thresholds, starting from the initial filtering threshold, record the filtering threshold when each topological feature first appears, and the filtering threshold when it first disappears (for example, assuming that under a lower threshold, an independent ring appears (that is, it is a new topological feature), then record the filtering threshold when this ring first appears. As the threshold increases, this ring may merge with other rings, or be filled and disappear. In this case, record the filtering threshold when it disappears);

[0153] The filtering threshold at the first appearance and the filtering threshold at the first disappearance are taken as the interval of each topological feature;

[0154] counting the total number of topological features, defining a tp x tp topological network, and covering the interval between the initial filtering threshold and the maximum filtering threshold, and tp is greater than the total number of topological features;

[0155] For each topological feature, using a Gaussian kernel function as a weight function, the weight value of the topological feature to each cell on the topological network is calculated where QZH(q, p) represents the weight value, u r and v r are the filtering threshold when the rth topological feature first appears and the filtering threshold when it first disappears, q and p are the coordinates of the cells on the topological network, and sigma represents the standard deviation of the Gaussian kernel, used to control the width of the influence range;

[0156] For each topological network cell, the total weight value of all topological features in the cell is accumulated, so that the value of each cell represents the "density" of topological features in the region. The total weight value of each cell on the topological network is horizontally spliced in the order of moving from the top left corner to the right bottom corner, forming a topological feature vector (by converting topological features into topological feature vectors, not only the important information of the original topological features is preserved, but also these information can be seamlessly integrated with other types of features, thus improving the performance of the model. In eye image analysis, this method can help better capture the complexity and subtle changes of retinal structures, improving the accuracy of disease diagnosis and treatment monitoring);

[0157] It should be noted that by converting the comprehensive image feature vector into a network graph and using topological data analysis techniques, rich topological features can be extracted from image data. This method not only provides a new perspective on data shape and structure, but also can discover deep patterns and features that traditional methods cannot capture in the diagnosis and analysis of eye diseases;

[0158] The way to identify the influence feature vector that affects the effect of the patient's ophthalmic surgery includes:

[0159] According to the non-image data in the patient's ophthalmic data set, using natural language processing technology to convert the text type data in the non-image data into an embedding vector, and using the embedding vector and the numerical type data in the non-image data as features, horizontally splicing into a clinical feature vector, and then integrating with the eye feature vector into a comprehensive feature set (each value in the eye feature vector is also a feature, combined with all features in the clinical feature vector to form a comprehensive feature set);

[0160] It should be noted that, due to too many features in this comprehensive feature set, if directly used in the subsequent intelligent medical decision model of ophthalmic surgery, it may cause dimension disaster, increase the complexity of calculation, and may introduce noise and irrelevant information, thereby affecting the performance of the model, so it is necessary to filter out the features that can affect the effect of ophthalmic surgery, and then use these filtered features as the final features for the subsequent intelligent medical decision model of ophthalmic surgery;

[0161] The random forest classifier is selected as the basic model because this kind of model not only can provide good prediction performance, but also can give the importance score of each feature;

[0162] The comprehensive feature set is used as input, and the result of ophthalmic surgery (such as represented by binary classification (1 represents successful surgery, 0 represents failed surgery), or more categories (such as excellent, good, general, and poor) according to specific circumstances) is defined as the target variable of the random forest model, and the random forest model is trained to output the influence score of each feature on ophthalmic surgery;

[0163] According to the influence score, all features are sorted, and the feature with the lowest influence score is removed;

[0164] The remaining features are used as input to train the random forest model again, calculate the new influence score, and remove the feature with the lowest new influence score, repeat the iteration until the predetermined number of features is reached, stop the iteration, and the remaining all features are used as the influence features affecting the effect of ophthalmic surgery of the patient, and are horizontally spliced into an influence feature vector;

[0165] The training method of the intelligent medical decision model of ophthalmic surgery includes:

[0166] A deep neural network is used to construct the intelligent medical decision model of ophthalmic surgery, the architecture of the intelligent medical decision model of ophthalmic surgery is set as a shared layer and a specific task layer, and the specific task layer is set as a surgery success rate branch layer and a postoperative ideal recovery state branch layer;

[0167] The influence feature vector is used as input, and the input is input into the shared layer, and the output of the shared layer is input into the surgery success rate branch layer and the postoperative ideal recovery state branch layer, respectively;

[0168] The surgery success rate branch layer extracts features related to surgery success from the shared layer output through a fully connected layer, and outputs a predicted success probability;

[0169] The postoperative ideal recovery state branch layer extracts features related to postoperative recovery from the shared layer output, and outputs a predicted risk degree of not reaching an ideal recovery state;

[0170] Collect a sample set, including each patient's impact feature vector and the corresponding label of whether the surgery was successful and whether the ideal recovery state was achieved after surgery; the label of whether the surgery was successful and whether the ideal recovery state was achieved after surgery are both binary labels, 1 for yes and 0 for no;

[0171] The sample set is divided into a training set and a validation set in proportion. The loss function of the surgical success rate branch layer is defined as the weighted binary cross entropy loss. The loss function of the postoperative ideal recovery state branch layer is defined as the binary cross entropy loss. The total loss function is defined as the weighted sum of the weighted binary cross entropy loss value and the binary cross entropy loss value.

[0172] The model is forward propagated using the training set. The corresponding loss functions are used to calculate the loss values ​​of the surgical success rate branch layer and the ideal postoperative recovery state branch layer. The total loss function is then calculated and backpropagated through the total loss function. The Adam optimization algorithm is used to update the model parameters in each iteration of the model, including the parameters of the shared layer and the two branch layers.

[0173] For each iteration, AUC is used as the evaluation metric and the AUC value is calculated on the validation set;

[0174] According to the AUC value on the validation set, the difference between the AUC value after the current iteration and the AUC value of the previous iteration is calculated and recorded as the iteration difference;

[0175] Set the iteration difference threshold. If the iteration difference is greater than the iteration difference threshold, it is determined that the performance of the model has improved.

[0176] If the iteration difference is less than or equal to the iteration difference threshold, it is determined that the performance of the model has not improved;

[0177] If the performance of the model on the validation set does not improve in consecutive DC iterations, the training is stopped and a trained intelligent medical decision-making model for ophthalmic surgery is obtained;

[0178] Ways to explain the prediction results of the ophthalmic surgery intelligent medical decision-making model through interpretability tools include:

[0179] Based on the original ophthalmic data of new patients, the patient's predicted success probability and predicted risk level are predicted through the ophthalmic surgery intelligent medical decision-making model;

[0180] The SHAP algorithm is used to calculate the SHAP value of each feature in the influencing feature vector based on the patient's predicted success probability and predicted risk level. This value is used as the contribution value of each feature to the prediction results of the intelligent medical decision-making model for ophthalmic surgery, and the contribution value of each feature to the predicted success probability and predicted risk level is obtained respectively.

[0181] Set a contribution threshold. If the contribution value of a feature to the prediction result is greater than the contribution threshold, it is determined that the feature will affect the prediction result of the corresponding patient and the feature is marked.

[0182] If the contribution value of a feature to the prediction result is less than or equal to the contribution threshold, it is determined that the feature will not affect the prediction result of the corresponding patient and the feature will not be marked;

[0183] It should be noted that by analyzing the SHAP values, we can determine which features have the greatest positive or negative impact on the predicted outcome for a particular patient. For example, in an ophthalmic surgery case, if age and preoperative visual acuity are found to have a significant impact on the predicted success rate, then we need to pay special attention to these features and explain in detail in the report how they affect the prediction.

[0184] Methods for generating model interpretation and surgical assessment reports include:

[0185] Based on the predicted success probability, predicted risk level, and the contribution of each feature to the predicted success probability and predicted risk level, a model interpretation and surgical evaluation report is generated, including:

[0186] Patient information summary, surgical success probability, postoperative risk level, and marker features; the patient information summary includes the patient's basic personal information, medical history, and clinical examination data; and AI technology can be used to provide specific medical advice based on the analysis results, including whether surgery is suitable and whether additional measures need to be taken to reduce risks. For example, a 50-year-old female patient with mild cataracts and a preoperative visual acuity of 0.3 was found after model prediction and SHAP analysis to be the main factors affecting the surgical success rate and postoperative recovery. Therefore, in the generated model explanation and surgical evaluation report, in addition to listing the basic patient information summary, surgical success probability, postoperative risk level, and marker features, AI tools can also emphasize the importance of these two factors and generate corresponding medical advice accordingly, such as "Considering the patient's age and current visual condition, although the surgical success rate is high, attention should still be paid to postoperative care to reduce the risk of complications." and "Because the patient is older and has lower preoperative visual acuity, this may affect the success rate of the surgery."

[0187] The embodiment integrates various types of data (such as image data, electronic health records, clinical examination data, and wearable device data), eliminates the limitations of single-dimensional data monitoring, improves the accuracy and reliability of diagnosis, and provides high-quality input data for subsequent analysis by using an improved active contour algorithm to refine the image contour curve and obtain comprehensive image features. This step can more accurately identify complex eye structures (such as retinal blood vessel regions or lesion regions), thereby providing additional insights by capturing deep structure information of eye features and revealing hidden patterns and relationships in the data. Then, the random forest classifier is used for feature selection, and the feature with the lowest impact score is removed iteratively to determine the set of factors that best reflect the effectiveness of the patient's ophthalmic surgery, ensuring the quality and relevance of the model input data. Then, the design of shared layers and specific task layers is adopted to predict the success rate of surgery and postoperative recovery status, allowing the model to focus on multiple targets simultaneously and optimizing overall performance. Finally, the SHAP value is used to quantify the contribution of each feature to the model output, enhancing the transparency of the model and helping doctors understand the working principle and prediction basis of the model, promoting effective communication between doctors and patients. The performance of the intelligent medical decision-making model based on ophthalmic surgery is significantly improved, not only improving the accuracy of diagnosis, but also enhancing the interpretability of the model and the ability of personalized treatment.

[0188] Embodiment Two

[0189] Please refer to Figure 2 The embodiment does not describe some parts in detail, which are described in Embodiment 1. The training system of the intelligent medical decision-making model based on ophthalmic surgery includes:

[0190] Data acquisition module: acquire the original ophthalmic data of the patient and perform preliminary data processing to obtain the ophthalmic data set of the patient;

[0191] Eye feature extraction module: analyze the image data in the ophthalmic data set of the patient, detect and extract the eye feature vector;

[0192] Influence identification module: extract the clinical feature vector of the non-image data in the ophthalmic data set of the patient, perform multi-source data fusion on the eye feature vector and the clinical feature vector, and identify the influence feature vector that affects the effectiveness of the ophthalmic surgery of the patient;

[0193] Model construction module: take the influence feature vector as input, train to obtain an intelligent medical decision-making model for ophthalmic surgery, and output the predicted success probability of ophthalmic surgery and the predicted risk degree of adverse outcomes after ophthalmic surgery based on the trained intelligent medical decision-making model for ophthalmic surgery;

[0194] Model explanation module: according to the prediction success probability and the prediction risk degree, the prediction result of the intelligent medical decision model of ophthalmic surgery is explained by an explainable tool, and a model explanation and surgery evaluation report is generated.

[0195] Embodiment three

[0196] The embodiment provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the operation mode of the training method of the intelligent medical decision model based on ophthalmic surgery when executing the computer program.

[0197] Since the electronic device introduced in the embodiment is used to implement the training method of the intelligent medical decision model based on ophthalmic surgery in the embodiment, the specific implementation mode of the electronic device and various changes thereof can be understood by those skilled in the art based on the training method of the intelligent medical decision model based on ophthalmic surgery in the embodiment, so the implementation of the electronic device in the method in the embodiment will not be introduced in detail. As long as the electronic device used to implement the training method of the intelligent medical decision model based on ophthalmic surgery in the embodiment is implemented by those skilled in the art, it belongs to the scope of protection of the present application.

[0198] The above formulas are dimensionless numerical calculations, the formulas are obtained by collecting a large amount of data to simulate the latest real situation, and the preset parameters and threshold values in the formulas are set by those skilled in the art according to the actual situation.

[0199] The above is only the preferred embodiment of the present application, the protection scope of the present application is not limited to the above-mentioned embodiments, any technical solution belonging to the idea of the present application is also within the protection scope of the present application. It should be noted that for ordinary technical users in the technical field, some improvements and decorations without departing from the principle of the present application are also considered as the protection scope of the present application.

Claims

1. A training method for an intelligent medical decision-making model based on ophthalmic surgery, characterized in that: include: S1. Obtain the patient's original ophthalmological data and perform preliminary data processing to obtain the patient's ophthalmological data set; S2. Analyze the image data in the patient's ophthalmology dataset to detect and extract eye feature vectors, including smoothing the image data, repeatedly reducing the size of the smoothed image to obtain images of different scales, using an improved active contour algorithm to refine the contour curve of the first scale image, and using the contour curve of the first scale image to refine the remaining scale images, using curvelet transform to extract feature information at different scales, and integrating them to obtain a comprehensive image feature vector; Use the comprehensive image feature vector to construct a network graph, set an initial filtering threshold, and gradually increase the filtering threshold. At each increase, traverse the network graph, identify all edge sets and construct topological features to obtain a topological feature vector. Fuse the comprehensive image feature vector with the topological feature vector to obtain an eye feature vector. S3. Extract clinical feature vectors from non-imaging data in the patient's ophthalmology dataset, perform multi-source data fusion on the ocular feature vectors and the clinical feature vectors, and identify the influencing feature vectors that affect the patient's ophthalmic surgery outcome; S4. Using the influencing feature vector as input, training an intelligent medical decision-making model for ophthalmic surgery is performed. Based on the trained intelligent medical decision-making model for ophthalmic surgery, the predicted success probability of the ophthalmic surgery and the predicted risk level of adverse postoperative outcomes are output; S5. Based on the predicted success probability and predicted risk level, the prediction results of the ophthalmic surgery intelligent medical decision-making model are explained through interpretability tools to generate a model explanation and surgical evaluation report.

2. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 1, characterized in that: The patient ophthalmology dataset is obtained by: Raw ophthalmology data include imaging data, electronic health records, clinical examination data, and wearable device and self-reported data; Perform preliminary data processing on the patient's original ophthalmic data, including using data cleaning technology to clean the original ophthalmic data, standardizing and formatting the non-imaging data based on the original ophthalmic data after data cleaning, and encrypting the patient's personal identity information, and then integrating all the original ophthalmic data after preliminary data processing into the patient's ophthalmic data set.

3. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 2, characterized in that: The eye feature vector extraction method includes: Step 31: performing denoising, contrast adjustment, and size normalization on the image data according to the image data in the patient's ophthalmological dataset; Step 32: For the processed image data, use a Gaussian filter to smooth the image; Step 33: Reduce the size of the smoothed image to half of its original size to form a low-resolution image. Repeat the reduction process until the maximum number of repetitions is reached to obtain images of different scales, which are recorded as scaled images. Step 34: Refine the contour curve of each scale image, and extract feature information at different scales using curvelet transform based on all refined scale images, and then integrate the feature information at different scales to obtain a comprehensive image feature vector; Step 35: Convert the comprehensive image feature vector into a network graph, and process the network graph using topological data analysis technology to obtain a topological feature vector; Step 36: Perform feature filling and normalization processing on the obtained comprehensive image feature vector and topological feature vector, and use the feature weighted summation method to fuse the processed comprehensive image feature vector and topological features to obtain the eye feature vector.

4. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 3, characterized in that: The method of obtaining the comprehensive image feature vector includes: Arrange all scale images in descending order of scale size to obtain a scale image sequence; Select the first scale image in the scale image sequence and set the initial contour curve of the first scale image, including: For the first scale image, the local threshold segmentation technique is used to perform threshold segmentation on the scale image to obtain the foreground area and the background area; For the scaled image after threshold segmentation, the edge detection algorithm is applied to generate a binary edge map of each scaled image. The points with a value of 1 in the edge map are edge points, and the points with a value of 0 are non-edge points. For each foreground region on the scaled image, the ratio of the sum of all edge points in the foreground region to the area of ​​the foreground region is taken as the edge density of the foreground region; and the average value of the gradient intensity in the foreground region is taken as the average gradient intensity of the foreground region; Set the edge density threshold and average gradient threshold. For any foreground area, if the edge density or average gradient intensity in the foreground area is greater than or equal to the set edge density threshold or average gradient threshold, it is determined that there is a clear edge in the foreground area. The foreground area is recorded as the retinal vessel area, and its edge is used as the initial contour curve of the retinal vessel: If the edge density and average gradient intensity in the foreground area are both less than the set edge density threshold and average gradient threshold, the foreground area is judged to be a suspected eye lesion area, and the foreground area is recorded as the lesion area, and its boundary is used as the initial contour curve of the lesion area: Integrate the obtained initial contour curves of all retinal blood vessels and the initial contour curves of the lesion area into the first scale image according to the coordinate positions to serve as the initial contour curves of the first scale image; The initial contour curve of the first scale image is refined using the improved active contour algorithm to obtain the refined contour curve of the first scale image. Based on the refined contour curve of the first scale image, all remaining scale images in the scale image sequence are refined to obtain the refined contour curves of all scale images. According to all scale images after the contour curve is refined, each scale image is transformed by the curvelet transform method to generate the curvelet coefficients of each scale image, which are horizontally spliced ​​into the initial image feature vector; The local threshold method is used to denoise the initial image feature vector. The principal component analysis method is used to reduce the dimensionality of the denoised initial image feature vector to obtain a low-dimensional feature vector as the comprehensive image feature vector.

5. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 4, characterized in that: The method of obtaining the refined contour curves of all scale images includes: Step 51: Based on the initial contour curve of the first scale image, a refinement function of the contour curve is defined. The precise position of the contour curve is adjusted and refined by the refinement function. The value of the refinement function is the sum of the fitting function value and the shift function value. The value of the fitting function is the weighted sum of the smoothness value and the curvature value of the contour curve. The smoothing value of the contour curve is defined as: PH CC =∫(α0+α jl ×jl(CC(cs),edge))×|CC 1 (cs)| 2 dcs; Among them, PH CC Indicates the smoothing value of the contour curve CC, CC 1 (cs) represents the first-order derivative of the contour curve CC at point cs, CC(cs) represents the coordinate position of point cs on the contour curve CC, CC(cs)=(x(cs),y(cs)), x(cs) and y(cs) represent the coordinates of point cs on the x-axis and y-axis, ∫dcs represents the integral along the curve CC, α0 represents the basic smoothing value weight factor, α jl represents the adjustment coefficient of distance jl, (α0+α jl ×jl(CC(cs),edge)) represents the total weight value in the formula, and jl(CC(cs),edge) represents the distance between CC(cs) and the nearest edge edge; The curvature value of the contour curve is defined as: WQ CC =∫(β0+β jl ×jl(CC(cs),edge))×|CC 2 (cs)| 2 dcs; among them, CC 2 (cs) represents the second-order derivative of the contour curve CC at point cs, β0 represents the basic bending value weight, β jl Indicates the adjustment factor of distance; The moving function value is defined as: Among them, qw cd represents the weight of the cd-th scale, At scale cd, the scale image I cd The gradient strength at the coordinate position CC(cs), represents the gradient direction of the scale image at the coordinate position CC(cs), θ(CC(cs) represents the direction angle of the contour curve CC at CC(cs), Represents the cosine value between the contour curve direction and the scale gradient direction; Step 53: Set α0, α jl , β0, β jl and qw cd Initialize as a parameter combination, use grid search for parameter tuning, use k-fold cross validation to evaluate the cross validation scores under different parameter combinations, and select the optimal parameter combination; Step 54: Iteratively update the refinement function using the gradient descent method until the absolute difference between the refinement function value in the current iteration and the refinement function value in the previous iteration is less than a preset refinement function difference threshold or the maximum number of iterations is reached, thereby obtaining a refinement contour curve of the first scale image. Step 55: Using the refined contour curve of the cd-th scale image as the initial contour curve of the cd+1-th scale image in the scale image sequence; Step 56: Repeat steps 51 and 55 until the refined contour curve of the last scale image in the scale image sequence is obtained.

6. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 5, characterized in that: The topological feature vector is obtained in the following manner: According to the comprehensive image feature vector, each coefficient in the comprehensive image feature vector is regarded as a node, the cosine similarity between every two nodes is calculated, and the similarity threshold is set; When the cosine similarity between two nodes is greater than or equal to the similarity threshold, an edge is established between the two nodes; when the cosine similarity between two nodes is less than the similarity threshold, no edge is established between the two nodes; According to the nodes and edges, a network graph is constructed; Traverse the network graph and define the number of connections between the nodes in the network graph as the weight of the node; Set an initial filtering threshold and gradually increase the filtering threshold. Check all node pairs each time the threshold is increased. If the weights of both nodes are less than or equal to the current filtering threshold, connect the two nodes until there are edges between all nodes. Then stop increasing the filtering threshold. At each filtering threshold, construct the current simplicial complex based on the current edge set; Among them, the current edge set represents the set of all original edges in the network graph and all newly added edges under the current filtering threshold. The simplicial complex represents a set composed of multiple simplices, including vertices, edges and high-dimensional simplices. The vertex is a 0-dimensional simplex, representing a node, the edge is a 1-dimensional simplex, and the high-dimensional simplex represents a higher-dimensional simplex. For each simplicial complex under each filtering threshold, the topological features within the simplicial complex are statistically analyzed, including the number of connected components, loops, and holes. A connected component is defined as a set of node combinations in the simplicial complex where any two nodes are connected by a path consisting of at least one edge, and the node combination cannot be expanded to a larger connected set by adding more nodes. A loop is a closed path in the simplicial complex with the same starting and ending points. A hole is an area in the simplicial complex that is surrounded by one or more loops and has no other nodes inside. For topological features under different filtering thresholds, starting from the initial filtering threshold, record the filtering threshold when each topological feature first appears and the filtering threshold when it first disappears; The filtering threshold at the first appearance and the filtering threshold at the first disappearance are taken as the interval of each topological feature; Count the total number of topological features and define a tp×tp topological network that covers the interval between the initial filtering threshold and the maximum filtering threshold, and tp is greater than the total number of topological features; For each topological feature, a Gaussian kernel function is used as the weight function to calculate the weight value of the topological feature for each cell on the topological network. Among them, QZH(q,p) represents the weight value, u r and v r are the filtering thresholds when the rth topological feature first appears and disappears, respectively. q and p are the coordinates of the cells on the topological network. σ represents the standard deviation of the Gaussian kernel. For each topological network cell, the total weight value of all topological features in the cell is accumulated, and the total weight value of each cell on the topological network is horizontally spliced ​​into a topological feature vector in the order of moving from the upper left corner to the right to the lower right corner.

7. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 6, characterized in that: The method of identifying the influencing feature vector that affects the effect of the patient's ophthalmic surgery includes: Based on the non-imaging data in the patient ophthalmology dataset, natural language processing technology is used to convert the text-type data in the non-imaging data into embedded vectors. The embedded vectors and the numerical data in the non-imaging data are used as features to horizontally splice into clinical feature vectors, which are then integrated with the ocular feature vectors to form a comprehensive feature set. The random forest classifier was selected as the basic model, the comprehensive feature set was used as input, the outcome of ophthalmic surgery was defined as the target variable of the random forest model, and the random forest model was trained to output the impact score of each feature on ophthalmic surgery. Sort all features by their impact scores and remove the feature with the lowest impact score; The random forest model is trained again using the remaining features as input to calculate new influence scores, and the features with the lowest new influence scores are removed. The iteration is repeated until the predetermined number of features is reached. The iteration is stopped, and all the remaining features are used as influencing features that affect the patient's ophthalmic surgery results and horizontally spliced ​​into an influence feature vector.

8. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 7, characterized in that: The training method of the ophthalmic surgery intelligent medical decision-making model includes: Use deep neural networks to build an intelligent medical decision-making model for ophthalmic surgery. Set the architecture of the intelligent medical decision-making model for ophthalmic surgery to be a shared layer and a specific task layer. The specific task layer is set as the surgical success rate branch layer and the ideal postoperative recovery state branch layer. The impact feature vector is used as input and input into the shared layer. The output of the shared layer is input into the surgery success rate branch layer and the ideal postoperative recovery state branch layer respectively. The surgical success rate branch layer extracts features related to surgical success from the shared layer output through the fully connected layer and outputs the predicted success probability; The ideal postoperative recovery state branch layer extracts features related to postoperative recovery from the shared layer output and outputs the predicted risk level of not reaching the ideal recovery state; Collect a sample set, including the impact feature vector of each patient and the corresponding label of whether the surgery is successful and whether the ideal recovery state is achieved after surgery; The sample set is divided into a training set and a validation set in proportion. The loss function of the surgical success rate branch layer is defined as the weighted binary cross entropy loss. The loss function of the postoperative ideal recovery state branch layer is defined as the binary cross entropy loss. The total loss function is defined as the weighted sum of the weighted binary cross entropy loss value and the binary cross entropy loss value. The model is forward propagated using the training set. The corresponding loss functions are used to calculate the loss values ​​of the surgical success rate branch layer and the ideal postoperative recovery state branch layer. The total loss function is then calculated and backpropagated through the total loss function. The Adam optimization algorithm is used to update the model parameters in each iteration of the model, including the parameters of the shared layer and the two branch layers. For each iteration, AUC is used as the evaluation metric and the AUC value is calculated on the validation set; According to the AUC value on the validation set, the difference between the AUC value after the current iteration and the AUC value of the previous iteration is calculated and recorded as the iteration difference; Set the iteration difference threshold. If the iteration difference is greater than the iteration difference threshold, it is determined that the performance of the model has improved. If the iteration difference is less than or equal to the iteration difference threshold, it is determined that the performance of the model has not improved; If the performance of the model on the validation set does not improve in consecutive DC iterations, the training is stopped and a trained intelligent medical decision-making model for ophthalmic surgery is obtained.

9. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 8, characterized in that: The method of explaining the prediction results of the ophthalmic surgery intelligent medical decision-making model by using the explainability tool includes: Based on the original ophthalmic data of new patients, the patient's predicted success probability and predicted risk level are predicted through the ophthalmic surgery intelligent medical decision-making model; The SHAP algorithm is used to calculate the SHAP value of each feature in the influencing feature vector based on the patient's predicted success probability and predicted risk level. This value is used as the contribution value of each feature to the prediction results of the intelligent medical decision-making model for ophthalmic surgery, and the contribution value of each feature to the predicted success probability and predicted risk level is obtained respectively. Set a contribution threshold. If the contribution value of a feature to the prediction result is greater than the contribution threshold, it is determined that the feature will affect the prediction result of the corresponding patient and the feature is marked. If the contribution value of a feature to the prediction result is less than or equal to the contribution threshold, it is determined that the feature will not affect the prediction result of the corresponding patient and the feature will not be marked.

10. The training method of the intelligent medical decision-making model based on ophthalmic surgery according to claim 9, characterized in that: The model interpretation and surgical evaluation report generation methods include: Based on the predicted success probability, predicted risk level, and the contribution of each feature to the predicted success probability and predicted risk level, a model interpretation and surgical evaluation report is generated, including: Patient information summary, predicted success probability of surgery, predicted risk level after surgery and marker characteristics; Patient information summary includes the patient's basic personal information, medical history and clinical examination data.

Images