Self-paced fuzzy clustering method and system for incomplete multi-view medical images

Abstract

The application relates to a self-step fuzzy clustering method and system for incomplete multi-view medical images, and relates to the field of intelligent analysis of medical images. The application solves the problems that existing multi-view clustering technologies cannot effectively process missing partial view data, fuzzy organization structure and lack a progressive dynamic sample selection mechanism. The method comprises the following steps: collecting multi-view medical image data, constructing a view missing label matrix, and performing spatial registration and feature standardization processing; based on a K nearest neighbor neighborhood weighting reconstruction strategy of a feature space, view features with missing data are completed; a standard multi-view fuzzy clustering is adopted, a joint fuzzy clustering objective function fusing a view missing label is constructed, an adaptive view weight distribution strategy and a self-step learning strategy are adopted, and the confidence of a clustering model of fuzzy membership entropy is defined; an alternating optimization strategy is used to update the membership, the clustering center and the view weight of the clustering model, and the self-step fuzzy clustering for incomplete multi-view medical images is completed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image intelligent analysis, and particularly relates to a self-step fuzzy clustering method and system for medical image incomplete multi-view. BACKGROUND

[0002] Medical images are important tools for clinical diagnosis and disease analysis. In clinical practice, multiple views (such as axial, coronal, and sagittal views) of the same lesion or tissue structure are usually collected to achieve more accurate diagnosis and analysis. According to relevant research reports, in the actual clinical data collection process, due to the high radiation risk and cost of some imaging, the scarcity of research subjects, patient privacy, and equipment limitations, etc., there are often cases of missing some view image data, which is referred to as the "incomplete multi-view problem".

[0003] Existing multi-view image analysis methods usually assume complete data. For example, CN110188825A discloses an image clustering method based on discrete multi-view clustering for clustering complete multi-view. However, this method is difficult to apply to actual clinical data with a high view missing rate. Most existing methods for handling missing view data use simple padding (such as mean padding), which is prone to cumulative errors, resulting in insufficient clustering stability and poor accuracy. In addition, the boundaries of tissues in medical images are often fuzzy and gradual, and traditional hard clustering algorithms (such as K-means and spectral clustering) cannot effectively capture this characteristic. Although fuzzy clustering methods can handle fuzzy boundary problems, such as CN103353986A, which discloses a method for segmenting brain MR images using fuzzy clustering, it only applies to single-view data and cannot effectively utilize multi-view information.

[0004] In recent years, Self-Paced Learning (SPL) has been widely applied in image classification and object detection as a progressive learning method from easy to difficult. For example, Chinese patent application publication number CN112598060A discloses a multi-view clustering method based on self-paced learning. This method constructs the graph structure of each view through adaptive graph learning and uses the SPL weight matrix to measure the complexity of the samples, transitioning from fewer "simple" samples to more "complex" samples to improve robustness to noise and outliers. However, this method has the following shortcomings: First, it assumes that the multi-view data is complete, making it difficult to handle image data with high missing rates in clinical settings; second, the clustering process is based on hard segmentation and does not utilize fuzzy membership output, failing to characterize the fuzzy and gradual boundary changes in medical images for clinical data; third, it selects samples only through node-level complexity measurement, lacking a comprehensive evaluation of view-level feature reconstruction errors and clustering confidence. Currently, there is no effective method for applying SPL to the field of incomplete multi-view clustering of medical images. Therefore, how to simultaneously address the issues of missing views and blurred tissue boundaries in medical imaging data, and improve the accuracy and stability of cluster analysis, is a pressing technical challenge in the field. Summary of the Invention

[0005] This invention aims to address the problems of existing multi-view clustering techniques, such as their inability to effectively handle missing view data, ambiguous organizational structures, and the lack of a progressive dynamic sample selection mechanism. To solve these technical problems, this invention achieves the following technical solution:

[0006] Option 1: This invention proposes a self-synchronizing fuzzy clustering method for incomplete multi-view medical images, the method comprising the following steps:

[0007] Step 1: Collect multi-view medical image data, construct a view missing marker matrix based on the collected multi-view medical image data, and perform spatial registration and feature standardization processing;

[0008] Step 2: Based on the K-nearest neighbor weighted reconstruction strategy in the feature space, the missing view features in Step 1 are completed to accurately estimate the feature vector of the missing view.

[0009] Step 3: Based on the features completed in Step 2, standard multi-view fuzzy clustering is used to quickly lock the coarse clustering structure of the data, and the membership update, cluster center update and view weight update of the clustering model are output as the initial state of the self-step learning stage.

[0010] Step 4: Construct a joint fuzzy clustering objective function for missing labels in the fused view, adopt an adaptive view weight allocation strategy, and introduce a local neighborhood structure to preserve the view's weights. Figure One To the point of being compatible;

[0011] Step 5, based on step 4, define the confidence of the clustering model based on fuzzy membership entropy by self-step learning strategy, and optimize the clustering model;

[0012] Step 6, update the membership, clustering center and view weight of the clustering model in step 4 respectively by using alternating optimization strategy until convergence, and complete the self-step fuzzy clustering for medical image incomplete multi-view.

[0013] Further, a preferred embodiment is provided, and the method for spatial registration and feature standardization processing of the collected multi-view medical image data in step 1 is:

[0014] The image size is unified to 256x256 pixels using linear interpolation algorithm;

[0015] The image gray value is normalized to the interval [0, 1] by minimum-maximum normalization;

[0016] The rigid spatial registration is performed by using SimpleITK toolkit to ensure the spatial consistency of different views;

[0017] A view missing mark matrix is constructed, which is , used to represent whether each view data is complete:

[0018] .

[0019] Further, a preferred embodiment is provided, and the K nearest neighbor neighborhood weighting reconstruction strategy based on feature space in step 2 is:

[0020] The method for completing the missing view features in step 1 is:

[0021] Step 2.1, for each sample i, calculate the average vector of the observed view features, denoted as , , wherein represents the view set existing in sample i;

[0022] Step 2.2, local linear regression determines the weight vector ,

[0023] ,

[0024] , wherein , and the closed-form solution is:

[0025] ;

[0026] Step 2.3, complete the features of the missing view v of patient i to obtain the completed vector ,

[0027] .

[0028] Further, a preferred embodiment is provided, and the calculation method of the output clustering model membership degree in step 3 is:

[0029] The membership degree update formula is: ;

[0030] Wherein, ;

[0031] The clustering center update formula is: ;

[0032] The view weight update formula is: ;

[0033] When the number of iterations reaches the preset maximum number or when the update amplitude is extremely small, the iteration is stopped, and the three values obtained in the last round are assigned as initial values 、 、 .

[0034] Further, a preferred embodiment is provided, and the method for constructing the joint fuzzy clustering objective function of the fusion view missing label in step 4 is:

[0035] Step 4.1, define the multi-view fuzzy clustering main objective function :

[0036] ,

[0037] Wherein is a screening mask, which will be given in step 5; the fuzzy factor takes an empirical value of 2;

[0038] Step 4.2, calculate the local neighborhood preservation regularization term : ,

[0039] Wherein , The value is 0.1, which is determined by cross-validation.

[0040] Step 4.3, calculate the view Figure One consistency constraint term : The value is 0.1, which is determined by cross-validation.

[0041] Step 4.4, the total objective function is: .

[0042] Further, a preferred embodiment is provided, in step 5, the confidence of the clustering model based on fuzzy membership entropy is defined by a self-learning strategy, and the method for optimizing the clustering model is:

[0043] Step 5.1, for the current tthiteration, the membership of each sample i to the cluster c is calculated:

[0044]

[0045] wherein, ,

[0046] Step 5.2, the sample clustering confidence is defined: ;

[0047] Step 5.3, sample screening

[0048] The confidence threshold is set for each iteration And the missing view tolerance :

[0049] The initial threshold =0.9, and the missing view tolerance T_0=0;

[0050] Every iteration round, the is reduced by 0.05, is increased by 1;

[0051] The samples meeting the conditions participate in clustering optimization, and the samples not meeting the conditions are temporarily introduced;

[0052] Sample selection mark: ;

[0053] When , the sample i is selected to participate in training in this iteration; when , the sample i is not included in the sample set to participate in training in this iteration.

[0054] Further, a preferred embodiment is provided, in step 6, the method for updating the membership, clustering center and view weight of the clustering model in step 4 respectively by using an alternating optimization strategy is:

[0055] Membership update formula: ;

[0056] Clustering center update formula: ;

[0057] View weight update formula: ;

[0058] The above updates are performed alternately, and the sample set of self-learning is gradually expanded.

[0059] A second aspect of the present application discloses a self-paced fuzzy clustering system for incomplete multi-view medical images, comprising:

[0060] A data standardization module is configured to collect multi-view medical image data, construct a view missing marker matrix based on the collected multi-view medical image data, and perform spatial registration and feature standardization processing;

[0061] A feature completion module is configured to complete the missing view features in the data standardization module based on a K-neighborhood weighted reconstruction strategy in the feature space, so as to accurately estimate the feature vector of the missing view.

[0062] A multi-view fuzzy clustering module is configured to perform standard multi-view fuzzy clustering on the features completed by the feature completion module, quickly lock the rough clustering structure of the data, and output the membership update, cluster center update and view weight update of the clustering model as the initial state of the self-paced learning stage.

[0063] A fuzzy clustering objective function construction module is configured to construct a joint fuzzy clustering objective function fused with the view missing marker, adopt an adaptive view weight distribution strategy, and introduce a local neighborhood structure to maintain the consistency of the views. Figure One

[0064] A clustering model optimization module is configured to define the confidence of the clustering model based on the fuzzy membership entropy and optimize the clustering model by the self-paced learning strategy of the fuzzy clustering objective function construction module.

[0065] A self-paced fuzzy clustering module is configured to update the membership, cluster center and view weight of the clustering model in the fuzzy clustering objective function construction module by an alternating optimization strategy until convergence, thereby completing the self-paced fuzzy clustering for incomplete multi-view medical images.

[0066] A third aspect of the present application discloses a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the steps of the method of the first aspect.

[0067] A fourth aspect of the present application discloses a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program stored in the memory to implement the method of the first aspect.

[0068] The present application has the following advantages:

[0069] The self-paced fuzzy clustering method and system for incomplete multi-view medical images disclosed by the present application adopt a cross-view feature completion strategy combining neighborhood weighting and local linear regression, which can more accurately reconstruct the missing view features in a high missing rate environment, reduce the completion error, and further enhance the stability of the clustering process.​

[0070] The application integrates view weight adaptive optimization, local neighborhood structure preservation and cross-view consistency constraint into the fuzzy clustering objective function, which is more helpful to capture the characteristics of the fuzzy boundary structure in medical images than the traditional hard clustering or single-view fuzzy clustering method. Figure One

[0071] The application combines the confidence measure based on membership entropy with the SPL sample selection mechanism, which can gradually introduce samples in the order of "from easy to difficult", which is theoretically helpful to avoid optimization shock caused by noise or missing data, thereby improving the robustness of the overall clustering.

[0072] The method has clear structure, clear parameters, is easy to implement, and is suitable for brain, lung, liver and other important medical image analysis scenes, and has obvious clinical applicability and promotion value.

[0073] In summary, the method proposed in the application effectively solves the problems of view missing, fuzzy tissue structure and lack of robust optimization mechanism in medical image analysis, improves the accuracy and robustness of medical image clustering analysis, and has important theoretical and clinical application value.

[0074] The application is also suitable for application fields such as medical image processing, pattern recognition and artificial intelligence. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 The flowchart of the self-paced fuzzy clustering method for incomplete multi-view medical images according to the first embodiment. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, not all the embodiments of the application.

[0077] Embodiment one, the embodiment proposes a self-paced fuzzy clustering method for incomplete multi-view medical images, the method comprising the following steps:

[0078] Step 1, collect multi-view medical image data, construct a view missing label matrix based on the collected multi-view medical image data, and perform spatial registration and feature standardization processing;

[0079] Step 2, based on the K nearest neighbor neighborhood weighted reconstruction strategy of the feature space, the missing view features in step 1 are completed for accurate estimation of the feature vector of the missing view;

[0080] ​Step 3, using standard multi-view fuzzy clustering based on the completed features in step 2 to quickly lock the rough clustering structure of the data, and output the membership update, cluster center update and view weight update of the clustering model as the initial state of the subsequent SPL stage;

[0081] Step 4, constructing a joint fuzzy clustering objective function fused with view missing markers, using an adaptive view weight distribution strategy, and introducing a local neighborhood structure to maintain view consistency; Figure One

[0082] Step 5, based on step 4, defining the confidence of the clustering model based on fuzzy membership entropy through a self-step learning strategy, and optimizing the clustering model;

[0083] Step 6, using an alternating optimization strategy to update the membership, cluster center and view weight of the clustering model in step 4 respectively until convergence, completing the self-step fuzzy clustering for medical image incomplete multi-view.

[0084] Embodiment II, this embodiment is a further limitation of the self-step fuzzy clustering method for medical image incomplete multi-view as described in embodiment I, the method for spatial registration and feature standardization processing of the collected multi-view medical image data in step 1 is:

[0085] using linear interpolation algorithm to unify the image size to 256x256 pixels;

[0086] performing minimum-maximum normalization on the image gray value to the interval [0, 1];

[0087] using SimpleITK toolkit for rigid spatial registration to ensure spatial consistency of different views;

[0088] constructing a view missing marker matrix, the matrix is , used to represent whether each view data is complete:

[0089]

[0090] The marker is used to accurately guide the subsequent feature completion.

[0091] Embodiment III, this embodiment is a further limitation of the self-step fuzzy clustering method for medical image incomplete multi-view as described in embodiment I, the method for completing the missing view features in step 1 based on the K-nearest neighbor neighborhood weighting reconstruction strategy in step 2 is:

[0092] Step 2.1, for each sample i, calculate the average vector of its observed view features, denoted as : where ​a set of views in which sample i exists;

[0093] Step 2.2, local linear regression determines weight vector :

[0094] wherein , the closed-form solution of which is: ;

[0095] Step 2.3, feature completion is performed on patient i with missing view v to obtain completion vector : .

[0096] Embodiment Four, the embodiment is a further limitation of the self-step fuzzy clustering method for medical image incomplete multi-view described in Embodiment One, and the calculation method of the membership degree of the output clustering model in Step 3 is:

[0097] Membership degree update formula: ;

[0098] wherein ;

[0099] Clustering center update formula: ;

[0100] When the number of iterations reaches the preset maximum number or when the update amplitude is extremely small, stop iteration, and assign the three values obtained in the last round as initial values , , .

[0101] Embodiment Five, the embodiment is a further limitation of the self-step fuzzy clustering method for medical image incomplete multi-view described in Embodiment One, and the method for constructing the joint fuzzy clustering objective function fused with the missing view label in Step 4 is:

[0102] Step 4.1, define multi-view fuzzy clustering main objective function :

[0103] ,

[0104] wherein is a screening mask, which will be given in Step 5; fuzzy factor take the experience value 2;

[0105] Step 4.2, calculate local neighborhood preservation regularization term ,

[0106] wherein , recommended value is 0.1, determined by cross-validation;

[0107] Step 4.3, calculate the view Figure One consistency constraint term : = 0.1;

[0108] Step 4.4, total objective function: .

[0109] Embodiment six, this embodiment is to provide a preferred embodiment, step 5 through self-step learning strategy, define the confidence of clustering model based on fuzzy membership entropy, and the method for optimizing the clustering model is:

[0110] Step 5.1, for the current tthround of iteration, calculate the membership of each sample i to cluster c:

[0111]

[0112] Wherein,

[0113] Step 5.2, define the sample clustering confidence: ;

[0114] Step 5.3, sample screening

[0115] Set the confidence threshold every iteration And missing view tolerance :

[0116] Initial threshold = 0.9, missing view tolerance T_0=0;

[0117] Every iteration round, will Lower 0.05, Increase 1;

[0118] The sample that meets the conditions participates in clustering optimization, and the sample that does not meet the conditions is temporarily introduced;

[0119] Sample selection mark: ;

[0120] When , sample i is selected to participate in training in this round of iteration; when , sample i is not included in the sample set to participate in training in this round of iteration.

[0121] Embodiment seven, this embodiment is a further limitation of the preferred embodiment of embodiment six, step 6 uses alternating optimization strategy to update the membership, clustering center and view weight of the clustering model in step 4 respectively:

[0122] Membership update formula: ;

[0123] The cluster center update formula is: ;

[0124] The view weight update formula is: ;

[0125] The above updates are alternately performed, and the sample set of self-step learning is gradually expanded.

[0126] Embodiment eight, the embodiment proposes a self-step fuzzy clustering system for incomplete multi-view medical images, the system comprises:

[0127] A data standardization module is configured to collect multi-view medical image data, construct a view missing label matrix based on the collected multi-view medical image data, and perform spatial registration and feature standardization processing;

[0128] A feature completion module is configured to complete the missing view features in step 1 based on the K-nearest neighbor neighborhood weighting reconstruction strategy of the feature space, so as to accurately estimate the feature vector of the missing view;

[0129] A multi-view fuzzy clustering module is configured to use standard multi-view fuzzy clustering on the features completed by the feature completion module, quickly lock the rough clustering structure of the data, and output the membership update, cluster center update and view weight update of the clustering model as the initial state of the subsequent SPL stage;

[0130] A fuzzy clustering objective function construction module is configured to construct a joint fuzzy clustering objective function that fuses the view missing label, adopts an adaptive view weight distribution strategy, and introduces a local neighborhood structure to maintain the consistency of the view; Figure One ;

[0131] A clustering model optimization module is configured to define the confidence of the clustering model based on the fuzzy membership entropy by the self-step learning strategy of the fuzzy clustering objective function construction module, and optimize the clustering model;

[0132] A self-step fuzzy clustering module is configured to update the membership, cluster center and view weight of the clustering model in the fuzzy clustering objective function construction module respectively by using an alternating optimization strategy until convergence, so as to complete the self-step fuzzy clustering for incomplete multi-view medical images.

[0133] Embodiment nine, the embodiment proposes a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by a processor to realize the steps of the method in any one of embodiments one to seven.

[0134] Embodiment ten, the embodiment proposes a computer device, comprising a memory and a processor, the memory has stored therein a computer program, when the processor runs the computer program stored in the memory, the processor executes the method of any one of embodiments one to seven.

[0135] Embodiment eleven, the embodiment proposes an embodiment for explaining embodiments one to ten, which is specifically:

[0136] See Figure 1 The embodiment is explained with reference to Figure 1 It is shown that the embodiment proposes a self-paced fuzzy clustering method for incomplete multi-view medical images. The technical core of the invention is to organically combine the self-paced learning (SPL) strategy with the fuzzy clustering model optimized for multi-view missing data, to realize more efficient and robust automatic clustering analysis of medical image data.

[0137] The specific technical solution of the embodiment includes the following key steps:

[0138] Step one, collect multi-view medical image data, construct a view missing label matrix, clearly label the missing situation of each sample view, and perform spatial registration and feature standardization preprocessing;

[0139] Step two, complete the features of the views with missing data by using the K-nearest neighbor neighborhood weighting reconstruction strategy based on the feature space, to accurately estimate the feature vectors of the missing views;

[0140] Step three, run the standard multi-view fuzzy clustering based on all the completed features, this stage all views exist, no sample screening, aims to quickly lock the rough clustering structure of the data. The membership distribution, cluster center and view weight output in this stage are used as the initial state of the subsequent SPL stage;

[0141] Step four, construct a joint fuzzy clustering objective function that integrates view missing labels, adopt an adaptive view weight allocation strategy, and introduce a local neighborhood structure preservation and view consistency regularization term to improve the accuracy and spatial consistency of clustering analysis; Figure One

[0142] Step five, innovatively introduce the self-paced learning strategy, define the clustering confidence based on fuzzy membership entropy, design a dynamic sample introduction mechanism that gradually increases the difficulty of samples and the tolerance of missing views from easy to difficult, and robustly optimize the clustering model;

[0143] Step six, update the membership, cluster center and view weight of the clustering model respectively by using the alternating optimization strategy until convergence, and obtain the accurate clustering analysis result of medical images.​

[0144] The present application takes brain MRI multi-view medical images as an example, and implements a specific self-step fuzzy clustering method for the case of incomplete multi-view (i.e. partial view missing). In the present embodiment, the following symbol convention is used: the number of iterations is set to t, and the initial value is 0. It is assumed that there are patient samples in total, each patient collects medical image views; for each image, a feature vector with a dimension of is extracted by a convolutional neural network. In addition, the number of clusters in the clustering analysis is preset to , and the number of nearest neighbors K is used in cross-view feature completion. The membership matrix is , , the membership of the sample to each cluster center, and . The cluster center matrix is , the view weight vector is , and . These symbols will remain consistent in all subsequent formula and algorithm descriptions. The specific process and formula derivation are as follows:

[0145] Step one, standardization and missing mark of medical image multi-view data

[0146] First, the following standardization operations are performed on the clinically collected brain MRI images (such as axial, coronal, and sagittal views):

[0147] (1) Use linear interpolation algorithm to unify the image size to 256x256 pixels;

[0148] (2) Normalize the image gray value to the interval [0, 1] by minimum-maximum;

[0149] (3) Use the SimpleITK toolkit for rigid spatial registration to ensure spatial consistency of different views.

[0150] For each patient sample i, a view missing mark matrix is constructed, and the matrix is , which is used to represent whether each view data is complete:

[0151]

[0152] The mark is used to accurately guide the subsequent feature completion.

[0153] Step two, multi-view feature extraction and cross-view feature completion

[0154] Each view uses a convolutional neural network (CNN) to extract features. The network structure is as follows:

[0155] Convolutional layer 1: Convolution kernel (3x3, 32 channels) + BatchNorm + ReLU;

[0156] Convolutional layer 2: Convolution kernel (3x3, 64 channels) + BatchNorm + ReLU;

[0157] Max-pooling layer: 2x2;

[0158] Convolutional layer 3: Convolution kernel (3x3, 128 channels) + BatchNorm + ReLU;

[0159] Global average pooling, output 128-dimensional feature vector .

[0160] In this scheme, in order to improve the accuracy and stability of the missing view feature completion, we additionally introduce a local linear regression (LLR) mechanism on the basis of the classic KNN weighted reconstruction. Specifically, for each sample i, first select K nearest neighbors around the average feature of i, and then automatically learn a set of optimal linear coefficients by least square regression to minimize the reconstruction error in the neighborhood. The regression coefficients are dynamically adjusted according to the local sample distribution, which can more accurately describe the neighborhood structure and improve the stability of the completion.

[0161] First, for each sample i, calculate the average vector of the observed view features of i, denoted as :

[0162] , where denotes the view set of sample i.

[0163] For the missing view v of patient i, based on the average feature vector of the existing views of the sample in the feature space , calculate the Euclidean distance with other sample features, and take the nearest K neighbors, i.e.

[0164] (experience selection K=10) to complete the feature.

[0165] Let , construct the matrix .

[0166] Then use local linear regression to determine the weight vector : ,

[0167] where , and its closed-form solution is: To prevent overfitting of sample features in the local neighborhood, a regularization parameter , the norm of the linear regression coefficient is controlled, The value is small, and the completed feature vector is more suitable But it may overfit, and the completion is unstable; the value is large, and the completed feature vector is more smooth and stable, but the completion error may be larger. Usually = 0.05, further fine-tuning can be done on the validation set.

[0168] Finally, the feature completion of the patient i with the missing view v is performed to obtain the completed vector ;

[0169] Step three, global fuzzy clustering (FCM) initialization

[0170] Before entering the SPL iteration, based on the view that has been completed in step two, run a standard multi-view FCM on all samples once to obtain high-quality initial membership , cluster center and view weight , the update formula of the tthiteration is as follows:

[0171] The membership update formula is: ;

[0172] Wherein, .

[0173] The cluster center update formula is: ;

[0174] The view weight update formula is: ;

[0175] When the number of iterations reaches the preset maximum number or when the update amplitude is very small, stop iteration, and assign the three values obtained in the last round as the initial values , , .

[0176] Step four, self-step multi-view fuzzy clustering model construction

[0177] Construct a multi-view fuzzy clustering objective function that integrates view missing labels, local neighborhood structure preservation, and view Figure One consistency constraints:

[0178] (1) Define the main objective function of multi-view fuzzy clustering :

[0179] ;

[0180] Wherein is a screening mask, which will be given in step five; the fuzzy factor takes an empirical value of 2.

[0181] (2) Local neighborhood preserving regularizer :

[0182] ;

[0183] where , The recommended value is 0.1, which is determined by cross-validation.

[0184] (3) View Figure One consistency constraint :

[0185] ;

[0186] The recommended value is 0.1, which is determined by cross-validation.

[0187] (4) Total objective function: .

[0188] Step five, self-paced learning (SPL) dynamic sample selection mechanism

[0189] First, for the current tthiteration, calculate the membership of each sample i to cluster c:

[0190]

[0191] where

[0192] Define the sample clustering confidence (defined by membership entropy): , set the confidence threshold and the missing view tolerance :

[0193] (1) The initial threshold =0.9, the missing view tolerance T_0=0;

[0194] (2) Each iteration, reduce by 0.05, increase by 1;

[0195] (3) The samples that meet the conditions participate in clustering optimization, and the samples that do not meet the conditions are temporarily suspended.

[0196] Sample selection marker: ;

[0197] When , sample i is selected to participate in training in this iteration; when , sample i is not included in the sample set to participate in training in this iteration.

[0198] Step six, model parameter alternation optimization (iterative update)

[0199] Update in turn using alternation optimization strategy:

[0200] Membership degree update formula: ;

[0201] Cluster center update formula: ;

[0202] View weight update formula: .

[0203] The above updates are alternately performed, and the sample set of self-step learning is gradually expanded.

[0204] Step seven, model optimization termination condition and clustering result output

[0205] When the clustering center moving distance is lower than the set threshold or reaches the set iteration upper limit, the model stops iteration. The final model outputs the clustering category division and fuzzy membership degree of each sample under the multi-view condition.

[0206] In summary, the method proposed in the embodiment effectively solves the problems of view missing, fuzzy tissue structure and lack of robust optimization mechanism in medical image analysis, improves the accuracy and robustness of medical image clustering analysis, and has important theoretical and clinical application value.

[0207] Those skilled in the art can understand that the above description is only preferred embodiments of the present application, and the features described in each embodiment and / or technical solution of the present disclosure can be combined or combined, even if such combination or combination is not explicitly described in the present disclosure. It is not intended to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, and those skilled in the art can modify the technical solutions described in the foregoing embodiments or make equivalent replacement of part of the technical features, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0208] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to these embodiments once they know the basic creative concept. Therefore, the appended technical solutions are intended to include the preferred embodiments and all changes and modifications falling within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and changes of the present application fall within the scope of the technical solutions of the present application and their equivalents, the present application also intends to include these modifications and changes.

Claims

1. A self-paced fuzzy clustering method for incomplete multi-view medical images, characterized in that, The method comprises the following steps: Step 1, collecting multi-view medical image data, constructing a view missing mark matrix based on the collected multi-view medical image data, and performing spatial registration and feature standardization processing; Step 2, based on the K nearest neighbor neighborhood weighted reconstruction strategy of the feature space, the missing view features in step 1 are completed for accurately estimating the feature vector of the missing view; Step 3, based on the completed features in step 2, a standard multi-view fuzzy clustering is used to quickly lock the rough clustering structure of the data, and the membership update, cluster center update and view weight update of the clustering model are output as the initial state of the self-step learning stage; Step 4, a joint fuzzy clustering objective function is constructed by fusing the view missing mark, an adaptive view weight distribution strategy is adopted, and a local neighborhood structure is introduced to maintain view consistency; Step 5, based on step 4, the confidence of the clustering model based on fuzzy membership entropy is defined by a self-step learning strategy, and the clustering model is optimized; Step 6, the membership, cluster center and view weight of the clustering model in step 4 are updated respectively by an alternating optimization strategy until convergence, and the self-step fuzzy clustering for medical image incomplete multi-view is completed; The method for completing the missing view features in step 1 based on the K nearest neighbor neighborhood weighted reconstruction strategy of the feature space in step 2 is: Step 2.

1. For each sample i, compute the average vector of its observed view features, denoted as , wherein, represents a set of views in which the sample i exists, is the vector after feature completion; Step 2.2, Locally linear regression determines weight vector : wherein , set , construct a matrix , The closed-form solution is ; is a regularization parameter, Step 2.

3. Feature completion for patient i with missing views v, resulting in completed vector , 。 2. The self-paced fuzzy clustering method for medical image oriented incomplete multi-view, according to claim 1, wherein, The method for performing spatial registration and feature standardization processing on the collected multi-view medical image data in step 1 is: The image size is unified to 256x256 pixels using a linear interpolation algorithm; The image gray value is normalized to the interval [0, 1] by minimum-maximum normalization; A rigid spatial registration is performed using the SimpleITK toolkit to ensure the spatial consistency of different views; The view missing flag matrix is constructed, which is , to indicate whether each view data is complete: 。 3. The self-paced fuzzy clustering method for medical image incomplete multi-view according to claim 1, characterized in that, The calculation method of the membership of the clustering model output in step 3 is: The membership update formula is: ； wherein ; Cluster center update formula: ; When the iteration number reaches the preset maximum number or when the update amplitude is extremely small, the iteration is stopped, and the three values obtained in the last round are assigned as the initial values , , wherein C is the number of clusters, and m is a fuzzy factor.

4. The self-paced fuzzy clustering method for medical image incomplete multi-view according to claim 1, characterized in that, The method for constructing the joint fuzzy clustering objective function by fusing the view missing mark in step 4 is: Step 4.1, Defining the Multi-view Fuzzy Clustering Main Objective Function : wherein For the screening mask, will be given in step 5; the obscuring factor Take the experience value 2; Step 4.2, computing the local neighborhood preserving regularizer : wherein , with a value of 0.1 determined with cross-validation; Step 4.3, computing view-consistency constraint terms : value of 0.1; Step 4.4, overall objective function: .

5. The self-paced fuzzy clustering method for medical image incomplete multi-view according to claim 1, wherein, The method for defining the confidence of the clustering model based on fuzzy membership entropy by a self-step learning strategy in step 5 is: Step 5.1, for the current tthiteration, the membership of each sample i to the cluster number C is calculated: wherein ; Step 5.2, defining sample cluster confidence: ; Step 5.3, sample screening Setting a confidence threshold for each iteration Tolerance to missing views ; initial threshold value = 0.9, missing view tolerance number T_0 = 0; Each iteration, add decrease 0.05, increase 1; Samples that meet the conditions participate in clustering optimization, and samples that do not meet the conditions are temporarily introduced; Sample selection marker: , When sample i is selected to participate in training in the current iteration; when sample i is not included in the sample set to participate in training in the current iteration.

6. The self-paced fuzzy clustering method for medical image oriented incomplete multi-view, according to claim 1, wherein, The method for updating the membership, cluster center and view weight of the clustering model in step 4 respectively by an alternating optimization strategy in step 6 is: Membership update formula: ; Cluster center update formula: ; View weight update formula: ; The above updates are alternately performed, and the sample set of self-step learning is gradually expanded.

7. A self-paced fuzzy clustering system for medical image incomplete multi-view, characterized in that, The system comprises: A data standardization module for collecting multi-view medical image data, constructing a view missing mark matrix based on the collected multi-view medical image data, and performing spatial registration and feature standardization processing; A feature completion module for completing the missing view features in the data standardization module based on the K nearest neighbor neighborhood weighted reconstruction strategy of the feature space, for accurately estimating the feature vector of the missing view; A feature completion module for completing the missing view features in the data standardization module based on the K nearest neighbor neighborhood weighted reconstruction strategy of the feature space, for accurately estimating the feature vector of the missing view; a multi-view fuzzy clustering module, which is used to adopt standard multi-view fuzzy clustering based on the features completed by the feature completion module, is used to quickly lock the rough clustering structure of the data, and outputs membership update, cluster center update and view weight update of the clustering model as the initial state of the subsequent SPL stage; a fuzzy clustering objective function construction module, which is used to construct a joint fuzzy clustering objective function that fuses view missing labels, adopts an adaptive view weight distribution strategy, and introduces a local neighborhood structure to maintain view consistency; a clustering model optimization module, which is used to define the confidence of the clustering model based on fuzzy membership entropy and optimize the clustering model based on the fuzzy clustering objective function construction module through a self-step learning strategy; a self-step fuzzy clustering module, which is used to update the membership, cluster center and view weight of the clustering model in the fuzzy clustering objective function construction module respectively by using an alternating optimization strategy until convergence, and complete the self-step fuzzy clustering for medical image incomplete multi-view; the method for completing the view features with missing in step 1 by using the K-nearest neighbor neighborhood weighting reconstruction strategy based on the feature space in the feature completion module is as follows: Step 2.

1. For each sample i, compute the average vector of its observed view features, denoted as , wherein, represents a set of views in which the sample i exists, is the vector after feature completion; Step 2.2, Locally linear regression determines weight vector : wherein , set , construct the matrix , the closed-form solution of which is: ; is a regularization parameter, Step 2.

3. Feature completion for patient i with missing views v, resulting in completed vector , 。 8. A computer storage medium having stored thereon a computer program, characterized in that The computer program is executed by a processor to implement the method of any one of claims 1-6.

9. A computer device, comprising: comprise: a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-6.

Images