A feature image recognition system
By using a feature image recognition system and combining encoder-decoder and data dimensionality reduction techniques with machine learning methods, a feature prediction model for fundus ultrasound images is constructed. This solves the problem of imprecise feature extraction in existing technologies, achieving higher recognition accuracy and model interpretability, and assisting doctors in developing more scientific treatment plans.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING HOSPITAL
- Filing Date
- 2023-05-05
- Publication Date
- 2026-06-19
AI Technical Summary
Current technologies for disease diagnosis using fundus ultrasound images suffer from insufficient feature extraction precision, resulting in low model accuracy and difficulty in effectively assisting doctors in developing treatment plans.
A feature image recognition system is used to establish the mapping relationship of retinal vascular structure feature parameters through the encoder-decoder method. Combined with data dimensionality reduction, machine learning and model ensemble techniques, a feature prediction model based on fundus ultrasound images is constructed to achieve accurate analysis and feature recognition of fundus ultrasound images.
It improves the accuracy of feature recognition in fundus ultrasound images and the interpretability of the models, enhances the scientific basis for medical workers' diagnosis and treatment planning of fundus diseases, and improves work efficiency.
Smart Images
Figure CN116468718B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a feature image recognition system. Background Technology
[0002] Fundus photography is a common ophthalmic examination method that examines the morphological changes of the entire retina. It works by using a specialized camera to record the images seen through an ophthalmoscope. Fundus photography can reveal the morphology of the retina, optic disc, macula, and retinal vessels, as well as any abnormalities such as hemorrhages, exudates, aneurysms, retinal degeneration areas, retinal tears, neovascularization, atrophic spots, and pigmentary disorders. Through fundus examination, doctors can observe the optic disc, retinal arteries and veins, macula, and retina.
[0003] Regarding the application of fundus ultrasound images, a study by a Google research team titled "Predicting Cardiovascular Risk Factors from Retinal Fundus Images Using Deep Learning" was the first to demonstrate that fundus ultrasound images can effectively predict cardiovascular disease risk factors, including age, sex, smoking history, BMI, systolic blood pressure, diastolic blood pressure, and glycated hemoglobin. The study also found that the retinal vascular region was the primary area identified by the predictive model, suggesting the predictive role of retinal vascular changes in cardiovascular disease. This study also predicted the occurrence of major cardiovascular and cerebrovascular events over 5 years, but the model's AUC was only 70%. Since the majority of participants in this dataset were healthy individuals, with only 631 cases (1.3%) having experienced major cardiac events, this is one reason for the model's low accuracy. Furthermore, other studies have constructed models using fundus ultrasound images to predict indicators obtained from coronary computed tomography (CT) scans and cardiac magnetic resonance imaging. Son et al. used fundus ultrasound images from 20,130 cases to intelligently assess coronary artery calcium score (CACS), achieving an AUC of 82.3%-83.2% for their predictive model. It is worth noting that the indicators predicted by this model are not the actual CACS values, but rather groups with high CACS and low CACS. Diaz-Pinto et al. constructed a model to predict left ventricular end-diastolic volume (LVEDV) and left ventricular mass index (LVM) using fundus photography. They further constructed a model to predict myocardial infarction using demographic information and LVEDV / LVM (predicted by fundus photography). They found that the accuracy of the model combining demographic information and LVEDV / LVM (predicted by fundus photography) in predicting myocardial infarction (74%) was significantly higher than that of the model using only demographic information (66%).
[0004] Fundus ultrasound images, as an intermediate step in the diagnosis of related diseases, involve the acquisition and analysis of specific features. While not directly involved in the diagnosis itself, they provide doctors with a feasible basis for analyzing symptoms and developing treatment intervention plans, thus making their judgments more scientific. Therefore, effectively acquiring and refining the characteristics of fundus ultrasound images is a crucial issue that urgently needs to be addressed. Summary of the Invention
[0005] This invention provides a feature image recognition system that can achieve accurate analysis and feature recognition of fundus color ultrasound images.
[0006] According to one aspect of the present invention, a feature image recognition system is provided, comprising:
[0007] The feature quantization unit is used to acquire fundus color ultrasound images and establish a mapping relationship with retinal vascular structure morphological feature parameters through an encoder-decoder method to perform feature quantization.
[0008] The data dimensionality reduction unit is used to perform data dimensionality reduction by analyzing the structure and distribution of the quantized feature data obtained by the feature quantization, and to obtain the feature space with the greatest difference.
[0009] The prediction model building unit is used to perform spatial dimensionality reduction based on the feature space using several random projection matrices to obtain corresponding low-dimensional subspaces; and to build a low-dimensional feature parameter prediction model for each low-dimensional subspace using machine learning methods.
[0010] The model integration unit is used to integrate the low-dimensional feature parameter prediction model using a majority voting method to obtain an integrated model corresponding to the low-dimensional subspace.
[0011] The final model unit is used to perform final integration of the integrated models using a model stacking method to obtain the final prediction model;
[0012] The model prediction unit is used to identify selected features in the acquired fundus ultrasound image based on the final prediction model.
[0013] The system also includes:
[0014] The image acquisition unit is used to acquire fundus ultrasound images from different users and generate fundus ultrasound images.
[0015] The system also includes:
[0016] The invalid dimension removal unit is used to identify the correlation between population characteristics and the quantitative characteristic data, as well as the correlation between the dimensions of the factors, through multi-factor correlation analysis. Combined with the correlation analysis between the factors, it removes the dimension indicators with strong autocorrelation or weak autocorrelation.
[0017] The data dimensionality reduction unit is also used for:
[0018] The mapping relationship between the feature group with a sample size of n and the non-feature group with a sample size of m is obtained by using the empirical optimal transfer estimation method, and the mapping matrix W between the two groups of data is calculated; the (i,j)th element in matrix W represents the weight of the i-th data in the feature group to the j-th data in the non-feature group, denoted as w_ij;
[0019] The displacement vector of the feature group data is calculated based on the mapping matrix W: the i-th data of the feature group x_i is denoted as {z_1, z_2, …, z_m}, and the non-feature group data is denoted as {z_1, z_2, …, z_m}. Then the displacement vector corresponding to x_i is (z_1*w_i1 + z_2*w_i2 + … + z_m*w_im – x_i), denoted as y_i.
[0020] Construct a nonparametric regression model between the displacement vector y_i and the data x_i: y_i = f(x_i) + e_i; where e_i is the model error.
[0021] By estimating the regression function f, the transport map is smoothed to obtain the Wasserstein distance.
[0022] The prediction model building unit is specifically used for:
[0023] On the feature space, spatial dimensionality reduction is performed using several random projection matrices P_1, …, P_k to obtain k different low-dimensional subspaces B_1, …, B_k;
[0024] For each low-dimensional subspace, support vector regression, random forest, and k-nearest neighbor regression machine learning methods are used to establish a prediction model for low-dimensional feature parameters and selected features.
[0025] The model integration unit is specifically used for:
[0026] The machine learning prediction model built on the j-th low-dimensional subspace B_j obtained by dimensionality reduction is integrated using the majority voting method to obtain the integrated model g_j corresponding to the low-dimensional subspace B_j.
[0027] The final model unit is specifically used for:
[0028] For the k low-dimensional subspace ensemble models {g_1, …, g_k}, the model stacking method is used to perform the final ensemble of the k models to obtain the final prediction model.
[0029] The system also includes a data grouping unit, used to group selected features and corresponding fundus ultrasound images into feature groups and non-feature groups;
[0030] The final prediction model is trained and optimized based on the feature groups and non-feature groups respectively.
[0031] The data grouping unit is further configured to divide the selected features and corresponding fundus ultrasound images into a training set and a validation set, which are used to train the final prediction model and validate the final prediction model, respectively.
[0032] The model prediction unit is specifically used for:
[0033] The system acquires a color Doppler ultrasound image of the user's fundus from an image acquisition device, inputs it into the final prediction model of the final model unit, and compares it with pre-set selected features to obtain the correspondence between the fundus ultrasound image and the selected features.
[0034] This invention proposes a feature image recognition system, comprising multiple implementation devices including a feature quantization unit, a data dimensionality reduction unit, a prediction model building unit, a model integration unit, a final model unit, and a model prediction unit. It can establish a selected feature prediction model based on the structural features of the retinal blood vessels in the fundus, accurately identify selected features in fundus ultrasound images, and quantify and compare specific features identified by medical professionals based on experience through instruments and equipment, effectively improving overall work efficiency. It can be used in primary care institutions to assist medical professionals in screening and analyzing specific feature information.
[0035] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0036] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0037] Figure 1 This is a schematic diagram of the feature image recognition system structure in an embodiment of the present invention.
[0038] Figure 2 This is a schematic diagram illustrating the operating principle of the Encoder-Decoder in one embodiment of the present invention.
[0039] Figure 3This is a schematic diagram of the final model acquisition structure in one embodiment of the present invention.
[0040] Figure 4 This is a schematic diagram of the overall system operation structure in one embodiment of the present invention. Detailed Implementation
[0041] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0042] Due to the unique structure of the eye and the transparency of the retina, the retina is the only part of the body where vascular structures can be observed directly and without damage. Previous studies have shown that fundus ultrasound images can effectively reflect many specific physical characteristics, confirming that the fundus is an effective observation window for many physical indicators. In some embodiments of this invention, the greatest breakthrough lies in directly combining fundus ultrasound images with coronary angiography, which can more directly and accurately reflect the actual state of the coronary arteries and the specific state of the cardiovascular system, serving as a characteristic indicator and helping medical professionals obtain specific vital signs more accurately.
[0043] The "black box" effect of existing artificial intelligence prediction models often leads to poor model interpretability and high decision-making risk. Some embodiments of this invention, based on the biomimetic mechanism of human vision, utilize artificial intelligence technologies such as computer vision to quantitatively analyze and process the features of fundus ultrasound images, obtaining information such as blood vessel diameter, curvature, fractal dimension, and angle. By constructing a high-dimensional structural mapping between feature groups and non-feature groups of data, high-dimensional feature difference analysis is performed to extract the key features that constitute the differences between feature groups and non-feature groups. This solves the problem of uninterpretable feature extraction in traditional artificial intelligence prediction models to a certain extent and has high scientific validity.
[0044] Traditional machine learning prediction models are prone to overfitting, resulting in low robustness and poor generalization. In some embodiments of this invention, model ensemble is applied to the extracted key differential feature space to fuse the results of multiple basic prediction models across different dimensions, constructing a selected feature evaluation model based on the quantification of fundus structure features. This addresses, to some extent, the problems of poor stability and inability to generalize in previous models.
[0045] To achieve the specific solution of the present invention described above, the technical solution of the present invention provides a feature image recognition system, such as... Figure 1 As shown, the feature image recognition system includes:
[0046] The feature quantization unit 11 is used to acquire fundus color ultrasound images, establish a mapping relationship with retinal vascular structure morphological feature parameters through an encoder-decoder method, and perform feature quantization.
[0047] The data dimensionality reduction unit 12 is used to perform data dimensionality reduction by analyzing the structure and distribution of the quantized feature data obtained by the feature quantization, and to obtain the feature space with the greatest difference.
[0048] The prediction model building unit 13 is used to perform spatial dimensionality reduction based on the feature space using several random projection matrices to obtain the corresponding low-dimensional subspace; and to build a low-dimensional feature parameter prediction model for each low-dimensional subspace using machine learning methods.
[0049] The model integration unit 14 is used to integrate the low-dimensional feature parameter prediction model using a majority voting method to obtain an integrated model corresponding to the low-dimensional subspace.
[0050] Final model unit 15 is used to perform final integration of the integrated model using a model stacking method to obtain the final prediction model;
[0051] The model prediction unit 16 is used to identify selected features in the acquired fundus ultrasound image based on the final prediction model.
[0052] In one embodiment of the present invention, an assessment model for coronary artery status based on the morphological characteristics of retinal vessels in the fundus can be constructed. Compared with existing methods, this study utilizes artificial intelligence visual computing technology to first digitally analyze the retinal vascular structure in fundus images to obtain relevant morphological feature parameters, and then establishes a statistical prediction model of the relevant morphological feature parameters and coronary artery status, which can increase the understandability and scientific validity of the model. This embodiment of the present invention uses coronary angiography as the outcome indicator for model training, combined with SYNTAX and Gesini scores, to provide a more intuitive and accurate assessment of coronary artery status.
[0053] In one embodiment of the present invention, image feature quantization is first performed. A mapping relationship between fundus ultrasound images and retinal vascular structure morphological feature parameters is established using deep learning methods such as Encoder-Decoder. Encoder-Decoder, commonly referred to as encoder-decoder, is a common model framework in deep learning. The structure of an Encoder-Decoder model includes an encoder and a decoder. The encoder processes the input sequence and then sends the processed vector to the decoder to convert it into the desired output. For the operating principle of Encoder-Decoder, see [link to Encoder-Decoder documentation]. Figure 2 This is existing technical content and will not be elaborated upon here.
[0054] Typically, before processing fundus ultrasound images, it is necessary to first acquire relevant ultrasound images through the system. Specifically, this can be done using a large number of previously acquired fundus ultrasound images. A specific training feature library is then established based on the correspondence between the ultrasound images and the selected features. The feature library can be categorized according to these correspondences; for example, it can be divided into specific groups to distinguish different types of features; it can also be divided into training sets, validation sets, or forward and reverse sets, etc.
[0055] Typically, fundus ultrasound images are acquired using specialized fundus color ultrasound equipment, but other specialized equipment can also be used. For example, fundus ultrasound images can be acquired using various devices such as fundus angiography machines and fundus examination instruments. In this embodiment of the invention, the specific acquisition method is not limited, as long as the fundus ultrasound image contains the necessary features. For example, it needs to contain necessary retinal artery angiography information to facilitate the subsequent selection of retinal artery angiography features as selection criteria.
[0056] The system also includes:
[0057] Image acquisition unit 17 is used to acquire fundus ultrasound images of different users and generate fundus ultrasound images.
[0058] The system also includes:
[0059] The invalid dimension removal unit 18 is used to identify the correlation between population characteristics and the quantitative characteristic data, and the correlation between each dimension of the factors through multi-factor correlation analysis. Combined with the correlation analysis between each factor, it removes dimension indicators with strong autocorrelation or weak autocorrelation.
[0060] Specifically, in one embodiment of the present invention, the correlation between population characteristics and image quantification characteristics is identified by methods such as multi-factor correlation analysis, as well as the correlation between each dimension of the factors and the SYNTAX score and Gesini score. Based on the correlation analysis between each factor, the dimension indicators with strong autocorrelation are removed. Based on the correlation analysis results between each dimension data and the SYNTAX score and Gesini score, the dimension indicators with weak correlation are removed, that is, invalid dimensions are removed.
[0061] The system also includes a data grouping unit 19, used to group selected features and corresponding fundus color ultrasound images into feature groups and non-feature groups;
[0062] The final prediction model is trained and optimized based on the feature groups and non-feature groups respectively.
[0063] Here, the feature group is the group that contains the selected feature, and the non-feature group is the group that does not contain the selected feature.
[0064] The data grouping unit 19 is further configured to divide the selected features and the corresponding fundus ultrasound images into a training set and a validation set, which are used to train the final prediction model and validate the final prediction model, respectively.
[0065] The ratio of the training set to the validation set can be adjusted according to the needs of model training; for example, it can be 9:1.
[0066] The data dimensionality reduction unit 12 is also used for:
[0067] The mapping relationship between the feature group with a sample size of n and the non-feature group with a sample size of m is obtained by using the empirical optimal transfer estimation method, and the mapping matrix W between the two groups of data is calculated; the (i,j)th element in matrix W represents the weight of the i-th data in the feature group to the j-th data in the non-feature group, denoted as w_ij;
[0068] The displacement vector of the feature group data is calculated based on the mapping matrix W: the i-th data of the feature group x_i is denoted as {z_1, z_2, …, z_m}, and the non-feature group data is denoted as {z_1, z_2, …, z_m}. Then the displacement vector corresponding to x_i is (z_1*w_i1 + z_2*w_i2 + … + z_m*w_im – x_i), denoted as y_i.
[0069] Construct a nonparametric regression model between the displacement vector y_i and the data x_i: y_i = f(x_i) + e_i; where e_i is the model error.
[0070] By estimating the regression function f, the transport map is smoothed to obtain the Wasserstein distance.
[0071] The prediction model building unit 13 is specifically used for:
[0072] On the feature space, spatial dimensionality reduction is performed using several random projection matrices P_1, …, P_k to obtain k different low-dimensional subspaces B_1, …, B_k;
[0073] For each low-dimensional subspace, support vector regression, random forest, and k-nearest neighbor regression machine learning methods are used to establish a prediction model for low-dimensional feature parameters and selected features.
[0074] The model integration unit 14 is specifically used for:
[0075] The machine learning prediction model built on the j-th low-dimensional subspace B_j obtained by dimensionality reduction is integrated using the majority voting method to obtain the integrated model g_j corresponding to the low-dimensional subspace B_j.
[0076] The final model unit 15 is specifically used for:
[0077] For the k low-dimensional subspace ensemble models {g_1, …, g_k}, the model stacking method is used to perform the final ensemble of the k models to obtain the final prediction model.
[0078] In one embodiment of the present invention, by analyzing the structure and distribution of the data itself, and using data dimensionality reduction, the feature space in which the differences between eigenvalues and non-eigenvalues corresponding to user data are greatest is identified. We hope that the identified feature subspace satisfies the following properties: a) the difference between eigenvalues and non-eigenvalues is greatest in the identified feature subspace; b) the difference between eigenvalues and non-eigenvalues is small in the complement space of the identified feature subspace. To quantify the differences between different groups of data, we employ the optimal transit Wasserstein distance.
[0079] The mapping relationship between the feature group (sample size n) and the non-feature group (sample size m) data is obtained through the empirical optimal transport estimation method, and the mapping matrix W between the two groups of data is calculated. The (i,j)th element of matrix W represents the weight of the i-th data in the feature group mapped to the j-th data in the non-feature group, denoted as w_ij. The displacement vector of the feature group data is calculated based on the mapping matrix W. Taking the i-th data x_i in the feature group as an example, and denoting the non-feature group data as {z_1, z_2, …, z_m}, the displacement vector corresponding to x_i is (z_1*w_i1 + z_2*w_i2 + … + z_m*w_im – x_i), denoted as y_i. Next, we introduce "smoothing splines" to alleviate the overfitting problem of the empirical optimal transport estimation. Specifically, a nonparametric regression model y_i = f(x_i) + e_i is constructed between the displacement vector y_i and the data x_i, where e_i is the model error. By estimating the regression function f, we can smooth the transport map, thereby obtaining a more accurate Wasserstein distance.
[0080] Using the estimation method described above, we obtain the optimal transport mapping displacement vectors for the characteristic and non-characteristic groups. Then, we perform principal component analysis (PCA) to identify the spatial dimensions with significant differences between the different groups. The number of dimensions to retain is determined by the cumulative explanatory power of the principal components; we select the principal components that achieve a cumulative explanatory power of 80% as the retained dimensions. The subspace spanned by these dimensions is the characteristic subspace we need.
[0081] In this embodiment of the invention, ensemble learning technology is used to integrate the results of multiple basic prediction models in order to improve the accuracy and robustness of the final prediction model.
[0082] In the extracted feature subspace, spatial dimensionality reduction is performed using multiple random projection matrices (denoted as P_1, …, P_k) to obtain k different low-dimensional subspaces, denoted as B_1, …, B_k. For each low-dimensional subspace, machine learning methods such as support vector regression, random forest, and k-nearest neighbor regression are used to establish a predictive model of the low-dimensional feature parameters and the severity of coronary artery lesions.
[0083] For multiple machine learning prediction models built on the j-th low-dimensional subspace B_j obtained by the above dimensionality reduction, the results of the multiple machine learning prediction models are integrated using the "majority vote" method to obtain the integrated model g_j corresponding to the low-dimensional subspace B_j.
[0084] For the k low-dimensional subspace ensemble models {g_1, …, g_k} obtained above, the "model stacking" method is used to perform a final ensemble of the k models to obtain the final prediction model. See [link to documentation] for the specific structure of the final model acquisition. Figure 3 , is the final model obtained by stacking multiple models.
[0085] Specifically, the model prediction unit 16 is used for:
[0086] The system acquires a color Doppler ultrasound image of the user's fundus from an image acquisition device, inputs it into the final prediction model of the final model unit 15, and compares it with a pre-set selected feature to obtain the correspondence between the fundus ultrasound image and the selected feature.
[0087] In one embodiment of the present invention, the entire system operation process includes a data collection section, a fundus color Doppler ultrasound quantitative analysis section, a predictive model establishment section, and a model performance evaluation section. For example... Figure 4 As shown, the data collection section includes multiple data collection and processing processes such as demographic information statistics, coronary angiography scoring, color fundus image processing, and follow-up. Quantitative analysis includes image preprocessing, semantic segmentation networks, optimal network models for vessel segmentation, and digitization of vessel morphological features. The prediction model includes data preparation, data grouping, data training, and model validation. Performance evaluation includes verification processes such as parity-absolute error analysis, median-absolute error analysis, accuracy, precision, recall, and F1 score. In fact, this embodiment uses the correspondence between angiography and cardiovascular features as the basis for training the prediction model.
[0088] In summary, the technical solution of this invention proposes a feature image recognition system, comprising multiple specific implementation devices such as a feature quantization unit, a data dimensionality reduction unit, a prediction model building unit, a model integration unit, a final model unit, and a model prediction unit. This system can establish a selected feature prediction model based on the structural features of the retinal blood vessels in the fundus, accurately identify selected features in fundus ultrasound images, and quantify and compare specific features identified by medical workers based on experience through instruments and equipment, effectively improving overall work efficiency. It can be used in primary care institutions to assist medical workers in screening and analyzing specific feature information.
[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0090] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.
[0093] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A feature image recognition system, characterized by, The system includes: The feature quantization unit is used to acquire fundus color ultrasound images and establish a mapping relationship with retinal vascular structure morphological feature parameters through an encoder-decoder method to perform feature quantization. The data dimensionality reduction unit is used to perform data dimensionality reduction by analyzing the structure and distribution of the quantized feature data obtained by the feature quantization, and to obtain the feature space with the largest Wasserstein distance of the distribution difference of the quantized feature data. The prediction model building unit is used to perform spatial dimensionality reduction based on the feature space using several random projection matrices to obtain the corresponding low-dimensional subspaces; for each low-dimensional subspace, a low-dimensional feature parameter prediction model is built using machine learning methods such as support vector regression, random forest, or k-nearest neighbor regression. The model integration unit is used to integrate the low-dimensional feature parameter prediction model using a majority voting method to obtain an integrated model corresponding to the low-dimensional subspace. The final model unit is used to perform final integration of the integrated models using a model stacking method to obtain the final prediction model; The model prediction unit is used to identify selected features in the acquired fundus ultrasound image according to the final prediction model, compare the selected features with pre-set selected features, and obtain the correspondence between the fundus ultrasound image and the user and the selected features.
2. The feature image recognition system according to claim 1, characterized in that, The system also includes: The image acquisition unit is used to acquire fundus ultrasound images from different users and generate fundus ultrasound images.
3. The feature image recognition system of claim 1, wherein, The data dimensionality reduction unit is also used for: The mapping relationship between the feature group with a sample size of n and the non-feature group with a sample size of m is obtained by using the empirical optimal transfer estimation method, and the mapping matrix W between the two groups of data is calculated; the (i,j)th element in matrix W represents the weight of the i-th data in the feature group to the j-th data in the non-feature group, denoted as w_ij; The displacement vector of the feature group data is calculated based on the mapping matrix W: the i-th data of the feature group x_i is denoted as {z_1, z_2, …, z_m}, and the non-feature group data is denoted as {z_1, z_2, …, z_m}. Then the displacement vector corresponding to x_i is (z_1*w_i1 + z_2*w_i2 + … + z_m*w_im – x_i), denoted as y_i. Construct a nonparametric regression model between the displacement vector y_i and the data x_i: y_i = f(x_i) + e_i; where e_i is the model error. By estimating the regression function f, the transport map is smoothed to obtain the Wasserstein distance.
4. The feature image recognition system of claim 1, wherein, The prediction model building unit is specifically used for: On the feature space, spatial dimensionality reduction is performed using several random projection matrices P_1, …, P_k to obtain k different low-dimensional subspaces B_1, …, B_k; For each low-dimensional subspace, support vector regression, random forest, and k-nearest neighbor regression machine learning methods are used to establish a prediction model for low-dimensional feature parameters and selected features.
5. A feature image recognition system according to claim 4, characterized in that, The model integration unit is specifically used for: The machine learning prediction model built on the j-th low-dimensional subspace B_j obtained by dimensionality reduction is integrated using the majority voting method to obtain the integrated model g_j corresponding to the low-dimensional subspace B_j.
6. A feature image recognition system according to claim 5, characterized in that, The final model unit is specifically used for: For the k low-dimensional subspace ensemble models {g_1, …, g_k}, the model stacking method is used to perform the final ensemble of the k models to obtain the final prediction model.
7. The feature image recognition system according to claim 1, characterized in that, The system also includes a data grouping unit, used to group selected features and corresponding fundus ultrasound images into feature groups and non-feature groups; The final prediction model is trained and optimized based on the feature groups and non-feature groups respectively.
8. A feature image recognition system according to claim 7, characterized in that, The data grouping unit is further configured to divide the selected features and corresponding fundus ultrasound images into a training set and a validation set, which are used to train the final prediction model and validate the final prediction model, respectively.