A deep granular unsupervised learning method for long-tail distribution lung cancer subtype identification

Abstract

The application provides a deep granular unsupervised learning method for long-tailed distribution lung cancer subtype identification, comprising the following steps: S10, preprocessing lung lesion image data set, and generating two different enhanced views for each sample to obtain two sample image data sets; S20, using a pre-trained deep neural network to extract and normalize features of the two sample images to obtain sample feature vectors; S30, training using deep granular unsupervised learning, and unsupervised optimizing features based on the granular structure to improve feature separability under long-tailed distribution; and S40, obtaining final clustering results using the Kmeans method after training. The scheme is suitable for long-tailed distribution situations with significant differences in the number of samples of each subtype in the lung cancer subtype identification task, can improve the deep clustering effect for long-tailed distribution data, and provides reliable classification basis for the auxiliary diagnosis of lung cancer subtype identification.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing, and in particular relates to a deep granulosphere unsupervised learning method for identifying long-tailed lung cancer subtypes. Background Technology

[0002] Lung cancer exhibits high histological and molecular heterogeneity, with significant differences among subtypes in morphological characteristics, molecular pathways, prognosis, and treatment response. With the large-scale acquisition of digital pathological sections, CT images, and multi-omics data, the automated identification and classification of lung cancer subtypes using artificial intelligence has become an important research direction. Compared to supervised learning, which relies on large amounts of high-quality labeled data, unsupervised clustering methods can perform structured grouping and potential subtype discovery of samples in scenarios where accurate annotations are lacking or difficult to obtain. Therefore, they have application value in tasks such as clinical data integration, cohort exploration, and novel subtype discovery.

[0003] However, existing unsupervised clustering techniques for medical imaging still face several challenges in lung cancer subtype identification. First, lung cancer subtype-related features are characterized by high dimensionality, cross-scale complexity, and weak separability. Taking CT scans as an example, information such as nodule morphology, texture, boundary, and internal density exhibits multi-scale characteristics. Direct clustering in the original feature space often fails to yield stable and semantically consistent cluster structures. Second, lung cancer subtype data typically exhibits a significant long-tail distribution, meaning that the number of common subtype samples is far greater than that of rare subtypes, and rare subtypes often have greater clinical value. Although deep clustering methods can improve feature separability by jointly learning feature representations and cluster assignment, existing deep clustering methods generally implicitly assume, to improve training stability and convergence, that cluster distribution is approximately uniform, sample occupancy in each cluster is relatively balanced, and intra-class variance is similar. For example, they suppress cluster collapse through methods such as equalized pseudo-label assignment or entropy regularization. For lung cancer subtype data that naturally exhibit long-tail characteristics, rare subtype samples are scarce and noisier. If forced equal distribution is applied, problems such as head cluster adsorption, tail cluster instability or being swallowed up may occur during training, resulting in biased clustering results and overall performance degradation.

[0004] Therefore, there is an urgent need for an unsupervised learning method for identifying long-tailed lung cancer subtypes. This method should be able to alleviate the problems of head cluster adhesion and cluster collapse, and improve the separability and cluster stability of tail subtypes, without relying on manual annotation, through a cluster representation more suitable for heterogeneous medical data and a robust allocation and update strategy. Such a method can provide technical support for exploring lung cancer subtype structures and assisting in diagnosis. Summary of the Invention

[0005] Objective: This invention addresses the shortcomings of existing technologies by providing a deep particle-sphere unsupervised learning method for identifying long-tailed lung cancer subtypes. This method enables feature representation learning and subtype clustering structure learning of lung cancer samples without manual annotation. Specifically, a deep neural network is first used to extract features from lung cancer-related image data, obtaining a deep representation of the samples. Then, a particle-sphere structure is constructed based on this representation, and clustering features of potential subtypes are iteratively mined through sample-sphere matching and iterative optimization. To address the problems of head clustering, tail subtype instability, and cluster degradation caused by the long-tailed distribution of lung cancer subtype data, a training objective and objective allocation mechanism oriented towards long-tailed distribution is designed to mitigate biased clustering and over-aggregation phenomena that easily occur in existing methods under long-tailed conditions. Simultaneously, an online update method for particle-sphere radius is proposed, incorporating Kalman filtering principles, to adaptively and smoothly update the particle-sphere scale, improving the stability and convergence reliability of the training process. Finally, the clustering allocation results of the samples are output through the deep particle-sphere unsupervised clustering framework, achieving automatic grouping of lung cancer subtypes and discovery of potential subtype structures. This invention provides a robust unsupervised clustering technique for identifying lung cancer subtypes under long-tail distribution conditions, improving the separability of tail subtypes and the reliability of clustering results, and providing a more scientific basis for assisted diagnosis.

[0006] This method specifically includes the following steps:

[0007] Step 1: Preprocess the lung lesion image dataset and generate two enhanced views from the original samples to obtain two sample image datasets.

[0008] Step 2: Input the two sample image datasets obtained in Step 1 into the pre-trained deep neural network for feature extraction, obtain two sample feature sets, and normalize the sample features.

[0009] Step 3: Training is performed based on deep particle unsupervised learning to construct a particle structure characterized by the particle center and radius; during the training process, a matching score is calculated based on the sample features and the particle structure, and the matching score is used to form the predicted distribution of the sample to the particle. Unsupervised optimization is performed based on the consistency constraint between the two enhanced views to update the network parameters and the particle center. Finally, the particle radius is updated using the Kalman filter method.

[0010] Step 4: Use the trained feature extraction network to represent the samples, and use the k-means method to obtain clustering results from the sample representations.

[0011] Step 1 includes the following steps:

[0012] Step 1-1: Perform denoising and intensity normalization on the lung lesion image to obtain a preprocessed image;

[0013] Steps 1-2 involve performing random data augmentation twice on each preprocessed image to generate two augmented view image datasets; let the original number of samples be... The two augmented view datasets are denoted as follows: and :

[0014] (1),

[0015] in, and These represent augmented view datasets respectively. The Middle Augmented view images of the original samples and the augmented view dataset The Middle An enhanced view image of the original sample.

[0016] In steps 1-2, the random data augmentation includes one or more of the following: random flipping, random rotation, random cropping, color transformation, contrast or brightness perturbation, blur perturbation, and adding Gaussian noise.

[0017] Step 2 includes the following steps:

[0018] Let the pre-trained feature extraction network be... The output feature dimension is The two augmented view datasets are input into the feature extraction network respectively to obtain... Features and Features :

[0019] (2),

[0020] in, , ; Represent the space of real numbers;

[0021] L2 normalization is performed on the features to obtain Normalization characteristics and Normalization characteristics :

[0022] (3),

[0023] in, Denotes the Euclidean norm. and This serves as the input feature for step 3.

[0024] Step 3 includes the following steps:

[0025] Step 3-1: Construct the particle-sphere structure based on the input features; let the number of target cluster categories and the number of particles be respectively... and And satisfy:

[0026] (4),

[0027] in These are adjustable coefficients; the normalized features of the two views are combined to form the input feature set. :

[0028] (5),

[0029] For the input feature set Perform K-means clustering to obtain Each cluster center is a set of granular centers. ,in Indicates the first The center of each particle; let the particle be assigned to the center of the first particle. The feature index set of each cluster is , The elements are binary pairs ,in For sample index, For view indexing; obtain the sphere radius based on the average distance within the cluster:

[0030] (6),

[0031] in, Indicates the first The radius of each sphere, Indicate the cardinality of the set; apply a unit norm constraint to the initial sphere centers:

[0032] (7);

[0033] in, This indicates that the left-hand variable is updated to the result of the expression on the right-hand side.

[0034] Step 3-2: Calculate the matching score between the sample and the pellet; index any sample. Index of any view Index of any ball Calculate the squared distance from the sample to the center of the grain. :

[0035] (8),

[0036] And based on squared distance and sphere radius Construct matching scores:

[0037] (9),

[0038] in, For the first The sample at the th The view relative to the first The matching score of each ball. For the power exponent parameter, For smoothing term parameters;

[0039] Step 3-3: Obtain the predicted probability distribution from the matching scores and form the target distribution for training, for any view index. ,based on Calculate the predicted probability distribution:

[0040] (10)

[0041] in, For the first The sample at the th The view belongs to the first The predicted probability of each particle, where exp is the natural exponential function. Temperature parameters for predicting the distribution; based on Calculate intermediate nonnegative quantities And construct the unnormalized target distribution matrix. :

[0042] (11),

[0043] in, The temperature parameters are for the target distribution; for the unnormalized target distribution matrix Perform Sinkhorn-Knopp iterative normalization to obtain the target distribution matrix. Specifically, this includes: introducing row scaling vectors With column scaling vector And order:

[0044] (12)

[0045] in, express A set of 3D non-negative real vectors express A set of 3D non-negative real vectors This represents the diagonalization operator, setting the column boundaries as... And set a smoothing constant. The normalization update formula is:

[0046] (13)

[0047] The row normalization update formula is:

[0048] (14)

[0049] After the iteration ends, by Obtain the target distribution matrix And satisfy:

[0050] (15)

[0051] (16);

[0052] Steps 3-4, based on two views Figure 1 Perform unsupervised optimization and update parameters to ensure consistency;

[0053] Steps 3-5 involve updating the particle radius using logarithmic domain Kalman filtering based on the predicted distribution.

[0054] Steps 3-4 include: the target distribution matrix based on the two views. , With the predicted probability distribution matrix , Constructing Cross-Consistency Loss :

[0055] (17)

[0056] Minimize the loss using gradient descent Update the feature extraction network Parameters and set of granule centers Repeat step 2 to obtain the updated normalized sample features of the two augmented view datasets. and Let the updated set of particle centers be... After each update, apply a unit norm constraint to the center of the sphere:

[0057] (18).

[0058] Steps 3-5 include: indexing any sample Index of any view Generate hard-assigned tags :

[0059] (19)

[0060] Calculate the squared distance from the updated sample to the center of the grain. :

[0061] (20)

[0062] For any number Each ball, collected from the current batch, is hard-allocated to the [number]th [unit]. The set of indices of each sphere :

[0063] (twenty one),

[0064] And order ,exist At that time, in sets Mid-square distance The mean construction of the first Observed square radius of each sphere :

[0065] (twenty two),

[0066] Update the square radius of the sphere before and the updated square radius observations Mapped to logarithmic field state variables respectively and :

[0067] (twenty three),

[0068] (twenty four),

[0069] in, Let be the smoothing constant for the logarithmic field transformation; introduce Representing the state of the logarithm field Uncertainty, process noise This represents the uncertainty increment caused by the slow drift of the radius during training iterations; an uncertainty update is then performed.

[0070] (25)

[0071] Estimating observation noise based on the dispersion of hard-assigned samples : Calculate first exist Within the sample variance :

[0072] (26)

[0073] Again After scaling, we obtain :

[0074] (27)

[0075] Therefore, the Kalman gain is calculated. :

[0076] (28)

[0077] Updating state variables in the logarithmic field With uncertainty :

[0078] (29)

[0079] (30)

[0080] exist Keep constant;;

[0081] Write the updated logarithmic field state back to the sphere radius and apply a lower bound truncation of the radius:

[0082] (31),

[0083] (32),

[0084] in To preset the minimum radius; in Keep Unchanged; repeat steps 3-2 to 3-5 until the preset number of training rounds is reached. .

[0085] In step 4, the normalized features of the two views are added together with equal weights to obtain the sample representation. :

[0086] (33),

[0087] And on Normalization is performed to obtain the normalized sample representation. :

[0088] (34),

[0089] right Perform K-means clustering to obtain the final cluster labels.

[0090] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0091] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0092] The present invention has the following beneficial effects:

[0093] (1) This invention introduces a sphere structure, incorporating the matching relationship between samples and spheres and the sphere occupancy status into the training objective function for joint optimization. By introducing a constraint term oriented towards long-tail distribution into the objective function, different spheres obtain more balanced optimization signals during training, suppressing the dominant role of head spheres in parameter updates, and reducing the probability of tail samples being absorbed or merged by head clusters, thereby alleviating the clustering bias and training degradation caused by long-tail distribution.

[0094] (2) To address the problems of large allocation noise and easy oscillation of the sphere scale caused by the change of sample features with iteration in deep clustering training, this invention draws on the idea of ​​Kalman filtering and proposes an online update mechanism for the sphere radius: by performing temporal smoothing and uncertainty control on the radius update, the radius mutation caused by small batch fluctuations, pseudo-allocation errors or abnormal samples is reduced, so that the sphere radius gradually converges with training, thereby improving the numerical stability and convergence reliability of the clustering process, and reducing the risk of unstable phenomena such as cluster collapse or empty clusters.

[0095] (3) This invention constructs an end-to-end deep unsupervised process from lung cancer sample feature learning to granulocyte clustering structure learning, which can realize automatic clustering and grouping of lung cancer subtypes under the condition of lack of annotation; through the synergistic effect of the expressive power of granulocyte structure and the long-tail-oriented training mechanism, the tail subtype samples obtain clearer clustering boundaries and more stable attribution relationships in the feature space, thereby improving the separability and recognition ability of tail subtypes, and enhancing the semantic consistency and medical practical value of clustering results. Attached Figure Description

[0096] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0097] Figure 1 This is an overall flowchart of the method of the present invention.

[0098] Figure 2 This is a data processing framework diagram of the method of the present invention.

[0099] Figure 3 This is a flowchart of the algorithm of the method of the present invention. Detailed Implementation

[0100] like Figure 1 , Figure 2and Figure 3 As shown, this embodiment of the invention provides a deep granulomatous unsupervised learning method for identifying long-tailed lung cancer subtypes, comprising the following steps:

[0101] Step 1: Preprocess the lung lesion image dataset and generate two different enhanced views for each sample to obtain two sample image datasets;

[0102] Step 2: Use a pre-trained deep neural network to extract and normalize features from the two sample images to obtain sample feature vectors.

[0103] Step 3: Use deep granular unsupervised learning for training, and perform unsupervised optimization of features based on the granular structure to improve the feature separability under long-tailed distribution;

[0104] Step 4: After training is complete, use the K-means method to obtain the final clustering results.

[0105] Step S10 includes the following steps:

[0106] S11. Perform denoising and intensity normalization on the lung lesion image to obtain a preprocessed image;

[0107] S12. Taking the TCGA-LUAD and TCGA-LUSC chest CT datasets publicly available on the TCIA platform as examples, a long-tailed distribution dataset is constructed. Within the effective lung slices, 10 two-dimensional slices are extracted from each case. LUAD (lung adenocarcinoma) is designated as the head class, and 2000 clear two-dimensional images are selected from its corresponding slice sample pool. LUSC (lung squamous cell carcinoma) is designated as the tail class, and 200 two-dimensional images are randomly selected from its corresponding slice sample pool, resulting in a total of 2200 two-dimensional samples and a long-tailed distribution lung cancer dataset with two categories. Two random data augmentations are performed on each preprocessed image to generate two augmented view image datasets; let the original number of samples be... The two augmented view datasets are denoted as follows:

[0108] (1),

[0109] in, and They represent the first Two enhanced view images of the original sample; the random data augmentation includes a combination of random cropping, contrast and brightness perturbation, blur perturbation, and addition of Gaussian noise.

[0110] Step S20 includes the following steps:

[0111] S21. Let the pre-trained feature extraction network be... Its output feature dimension is The two enhanced view sample images are input into the feature extraction network respectively to obtain the features of the two views:

[0112] (2),

[0113] in, , The features are then L2 normalized to obtain the normalized features:

[0114] (3),

[0115] in, Describing the Euclidean norm, the stated and This serves as the input feature for step 3.

[0116] S22. ResNet18 was selected as the pre-trained deep neural network. ResNet18 is a variant of a lightweight deep residual network designed to solve image recognition tasks. To adapt the network input, the two augmented sample images were uniformly scaled to 224×224 pixels.

[0117] Step S30 includes the following steps:

[0118] S31. Construct a particle-sphere structure based on input features. Let the number of target cluster categories and the number of particles be respectively... and And satisfy:

[0119] (4),

[0120] in These are adjustable coefficients used to provide finer-grained granular coverage under long-tailed distribution conditions; the normalized features of the two views are combined to form the input feature set. :

[0121] (5),

[0122] For the input feature set Perform K-means clustering to obtain Each cluster center is a set of granular centers. ,in Indicates the first The center of each particle; let the particle be assigned to the center of the first particle. The feature index set of each cluster is Its elements are binary pairs ,in For sample index, For view indexing; obtain the sphere radius based on the average distance within the cluster:

[0123] (6),

[0124] in, Indicates the first The radius of each sphere, Indicate the cardinality of the set; apply a unit norm constraint to the initial sphere centers:

[0125] (7),

[0126] in, This indicates that the left-hand variable is updated to the result of the expression on the right.

[0127] To maintain the consistency and numerical stability of distance metrics.

[0128] S32. Calculate the matching score between the sample and the pellet. For any sample index... Index of any view Index of any ball Calculate the squared distance from the sample to the center of the grain. :

[0129] (8),

[0130] And based on squared distance and sphere radius Construct matching scores:

[0131] (9),

[0132] in, For the first The sample at the th The view relative to the first The matching score of each ball. The exponent parameter is used for numerical scaling of the radius to ensure training stability. The parameter is used for smoothing the term to prevent the denominator from being too small; the matching fraction can achieve adaptive balance for long-tailed distributions: when the head granules cover a large area, leading to... When the value is large, the magnitude of the corresponding matching score is suppressed, thus reducing its dominance in gradient backpropagation; when the tail ball coverage area is small, it leads to... When the value is smaller, the discrimination of the corresponding matching score is enhanced, thereby improving the visibility of tail granules in allocation and updating and reducing head adsorption.

[0133] S33. Obtain the predicted probability distribution from the matching scores and form the target distribution for training. For any view index... ,based on Calculate the predicted probability distribution:

[0134] (10)

[0135] in, For the first The sample at the th The view belongs to the first The predicted probability of each ball. Temperature parameters for predicting the distribution; based on Calculate intermediate nonnegative quantities And construct the unnormalized target distribution matrix. :

[0136] (11),

[0137] in, The temperature parameters represent the target distribution. This applies to the unnormalized target distribution matrix. Perform Sinkhorn-Knopp iterative normalization to obtain the target distribution matrix. Specifically, this involves introducing a row scaling vector. With column scaling vector And order:

[0138] (12)

[0139] in, express A set of 3D non-negative real vectors express A set of 3D non-negative real vectors This represents the diagonalization operator, within a preset number of iterations. Column normalization and row normalization are performed alternately within the distribution to ensure that the target distribution simultaneously satisfies the probability constraint of the sample dimension and the occupancy balance constraint of the sphere dimension; where the column margin is set to... And set a smoothing constant. To prevent the denominator from being zero, the column is normalized and updated as follows:

[0140] (13)

[0141] Linear normalization update to:

[0142] (14)

[0143] After the iteration ends, by The target distribution matrix is ​​obtained. To satisfy:

[0144] (15)

[0145] (16)

[0146] This ensures that the overall occupancy of each ball in the current batch is more balanced, alleviating the training degradation caused by a few head balls absorbing too many samples under the long-tail distribution.

[0147] S34, Based on Two Views Figure 1 Unsupervised optimization is performed to maintain consistency, and parameters are updated. The target distribution matrix is ​​based on two views. , With the predicted probability distribution matrix , Constructing Cross-Consistency Loss :

[0148] (17)

[0149] The loss is minimized using gradient descent. Update the feature extraction network Parameters and set of granule centers Repeat step 2 to obtain the updated normalized sample features. and Let the updated set of sphere centers be... After each update, apply a unit norm constraint to the center of the sphere:

[0150] (18)

[0151] Steps 3-5 involve updating the particle radius using a logarithmic domain Kalman filter based on the predicted distribution. For any sample index... Index of any view Generate hard-assigned tags :

[0152] (19)

[0153] Calculate the squared distance from the updated sample to the center of the grain. :

[0154] (20)

[0155] For any number Each ball represents a collection of indices hard-assigned to that ball in the current batch.

[0156] (twenty one),

[0157] And order ,exist At that time, in sets Mid-square distance The mean construction of the first Observed square radius of each sphere :

[0158] (twenty two),

[0159] Update the square radius of the sphere before and the updated square radius observations Mapped to logarithmic field state variables respectively and :

[0160] (twenty three),

[0161] (twenty four),

[0162] in, Let be the smoothing constant for the logarithmic field transformation. Introduce... Representing the state of the logarithm field Uncertainty, process noise This represents the uncertainty increment caused by the slow drift of the radius during training iterations; an uncertainty update is then performed.

[0163] (25)

[0164] Estimating observation noise based on the dispersion of hard-assigned samples : Calculate first exist Within the sample variance :

[0165] (26)

[0166] Again Scale normalization is performed to make it consistent with the variation range of the logarithmic domain observations, resulting in... :

[0167] (27)

[0168] Therefore, the Kalman gain is calculated. :

[0169] (28)

[0170] Updating state variables in the logarithmic field With uncertainty :

[0171] (29)

[0172] (30)

[0173] exist Keep Unchanged. Write the updated logarithmic field state back to the sphere radius and apply a lower bound truncation of the radius:

[0174] (31),

[0175] (32),

[0176] in To preset the minimum radius; in Keep Unchanged. Repeat steps 3-2 to 3-5 until the preset number of training rounds is reached. .

[0177] Step S40 includes the following steps:

[0178] S41. The normalized features of the two views are added together with equal weights to obtain the sample representation. :

[0179] (33),

[0180] And on Normalize:

[0181] (34),

[0182] right Perform K-means clustering to obtain the final cluster labels.

[0183] S42. The number of clusters selected is 2, and the final cluster labels are shown in Table 1.

[0184] Table 1

[0185]

[0186] The prediction accuracy was calculated based on the clustering results using the following method. :

[0187] (35),

[0188] in To predict whether it is lung adenocarcinoma, the truth label is also the number of lung adenocarcinoma samples. To predict whether it is lung adenocarcinoma, the truth label is the number of samples of lung squamous cell carcinoma. To predict whether it is squamous cell carcinoma of the lung, the ground truth label is also the number of samples with squamous cell carcinoma of the lung. To predict whether it is squamous cell carcinoma of the lung, the true label is the number of samples with lung adenocarcinoma, which is calculated as follows. The accuracy rate is 91.6%, indicating that the diagnostic results provided by this invention have a high degree of accuracy.

[0189] This invention provides a deep granulomatous unsupervised learning method for identifying long-tailed lung cancer subtypes. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A deep granulospheric unsupervised learning method for identifying long-tailed lung cancer subtypes, characterized in that, Includes the following steps: Step 1: Preprocess the lung lesion image dataset and generate two enhanced views from the original samples to obtain two sample image datasets. Step 2: Input the two sample image datasets obtained in Step 1 into the pre-trained deep neural network for feature extraction, obtain two sample feature sets, and normalize the sample features. Step 3: Train based on deep particle unsupervised learning to construct a particle structure characterized by the particle center and radius; During training, matching scores are calculated based on sample features and particle structure. The matching scores are then used to form the predicted distribution of samples to particles. Unsupervised optimization is performed based on the consistency constraints between the two enhanced views to update network parameters and particle center. Finally, the particle radius is updated using the Kalman filter method. Step 4: Use the trained feature extraction network to represent the samples, and use the k-means method to obtain clustering results from the sample representations.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1-1: Perform denoising and intensity normalization on the lung lesion image to obtain a preprocessed image; Steps 1-2 involve performing random data augmentation twice on each preprocessed image to generate two augmented view image datasets; let the original number of samples be... The two augmented view datasets are denoted as follows: and : （1）， in, and These represent augmented view datasets respectively. The Middle Augmented view images of the original samples and the augmented view dataset The Middle An enhanced view image of the original sample.

3. The method according to claim 2, characterized in that, In steps 1-2, the random data augmentation includes one or more of the following: random flipping, random rotation, random cropping, color transformation, contrast or brightness perturbation, blur perturbation, and adding Gaussian noise.

4. The method according to claim 3, characterized in that, Step 2 includes the following steps: Let the pre-trained feature extraction network be... The output feature dimension is The two augmented view datasets are input into the feature extraction network respectively to obtain... Features and Features : （2）， in, , ; Represent the space of real numbers; L2 normalization is performed on the features to obtain Normalization characteristics and Normalization characteristics : （3）， in, Denotes the Euclidean norm. and This serves as the input feature for step 3.

5. The method according to claim 4, characterized in that, Step 3 includes the following steps: Step 3-1: Construct the particle-sphere structure based on the input features; let the number of target cluster categories and the number of particles be respectively... and And satisfy: （4）， in These are adjustable coefficients; the normalized features of the two views are combined to form the input feature set. : （5）， For the input feature set Perform K-means clustering to obtain Each cluster center is a set of granular centers. ,in Indicates the first The center of each particle; let the particle be assigned to the center of the first particle. The feature index set of each cluster is , The elements are binary pairs ,in For sample index, For view indexing; obtain the sphere radius based on the average distance within the cluster: （6）， in, Indicates the first The radius of each sphere, Indicate the cardinality of the set; apply a unit norm constraint to the initial sphere centers: （7）， in, This indicates that the left-hand variable is updated to the result of the expression on the right-hand side. Step 3-2: Calculate the matching score between the sample and the pellet; index any sample. Index of any view Index of any ball Calculate the squared distance from the sample to the center of the grain. : （8）， And based on squared distance and sphere radius Construct matching scores: （9）， in, For the first The sample at the th The view relative to the first The matching score of each ball. For the power exponent parameter, For smoothing term parameters; Step 3-3: Obtain the predicted probability distribution from the matching scores and form the target distribution for training, for any view index. ,based on Calculate the predicted probability distribution: （10）， in, For the first The sample at the th The view belongs to the first The predicted probability of each particle, where exp is the natural exponential function. Temperature parameters for predicting the distribution; based on Calculate intermediate nonnegative quantities And construct the unnormalized target distribution matrix. : （11）， in, The temperature parameters are for the target distribution; for the unnormalized target distribution matrix Perform Sinkhorn-Knopp iterative normalization to obtain the target distribution matrix. Specifically, this includes: introducing row scaling vectors With column scaling vector And order: （12）， in, express A set of 3D non-negative real vectors express A set of 3D non-negative real vectors This represents the diagonalization operator, setting the column boundaries as... And set a smoothing constant. The normalization update formula is: （13）， The row normalization update formula is: （14）， After the iteration ends, by Obtain the target distribution matrix And satisfy: （15）， （16）； Steps 3-4: Perform unsupervised optimization based on the consistency of the two views and update the parameters; Steps 3-5 involve updating the particle radius using logarithmic domain Kalman filtering based on the predicted distribution.

6. The method according to claim 5, characterized in that, Steps 3-4 include: the target distribution matrix based on the two views. , With the predicted probability distribution matrix , Constructing Cross-Consistency Loss : （17）， Minimize the loss using gradient descent Update the feature extraction network Parameters and set of granule centers Repeat step 2 to obtain the updated normalized sample features of the two augmented view datasets. and Let the updated set of particle centers be... After each update, apply a unit norm constraint to the center of the sphere: （18）。 7. The method according to claim 6, characterized in that, Steps 3-5 include: indexing any sample Index of any view Generate hard-assigned tags : （19）， Calculate the squared distance from the updated sample to the center of the grain. : （20）， For any number Each ball, collected from the current batch, is hard-allocated to the [number]th [unit]. The set of indices of each sphere : （21）， And order ,exist At that time, in sets Mid-square distance The mean construction of the first Observed square radius of each sphere : （22）， Update the square radius of the sphere before and the updated square radius observations Mapped to logarithmic field state variables respectively and : （23）， （24）， in, Let be the smoothing constant for the logarithmic field transformation; introduce Representing the state of the logarithm field Uncertainty, process noise This represents the uncertainty increment caused by the slow drift of the radius during training iterations; an uncertainty update is then performed. （25）， Estimating observation noise based on the dispersion of hard-assigned samples : Calculate first exist Within the sample variance : （26）， Again After scaling, we obtain : （27）， Therefore, the Kalman gain is calculated. : （28）， Updating state variables in the logarithmic field With uncertainty : （29）， （30）， exist Keep constant;; Write the updated logarithmic field state back to the sphere radius and apply a lower bound truncation of the radius: （31）， （32）， in To preset the minimum radius; in Keep Unchanged; repeat steps 3-2 to 3-5 until the preset number of training rounds is reached. .

8. The method according to claim 7, characterized in that, In step 4, the normalized features of the two views are added together with equal weights to obtain the sample representation. : （33）， And on Normalization is performed to obtain the normalized sample representation. : （34）， right Perform K-means clustering to obtain the final cluster labels.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images