A dual-consistency spine image segmentation method based on possibility clustering
By employing a biconsistency-based spinal image segmentation method based on probability clustering, medical image segmentation is performed using unlabeled data. This addresses the issues of dependence on labeled data and sensitivity to noise in semi-supervised learning methods, achieving higher segmentation accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-29
AI Technical Summary
Existing semi-supervised learning methods rely on labeled data for training in medical image segmentation tasks and are sensitive to noise and pseudo-labels, which affects model performance and makes it difficult to effectively utilize the inherent patterns in unlabeled data.
A biconsistency-based spine image segmentation method based on probability clustering is adopted. Through data augmentation, feature extraction, cluster information transformation and consistency training, image segmentation is performed using unlabeled data, which reduces the impact of noise and improves the segmentation accuracy.
By effectively utilizing unlabeled data and reducing the impact of noise on segmentation results, the problem of difficult segmentation of diseased intervertebral discs is solved, thus improving the accuracy of medical image segmentation.
Smart Images
Figure CN122115848A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer technology, specifically to a biconsistent spine image segmentation method based on probability clustering. Background Technology
[0002] Deep learning methods have achieved remarkable results in medical image segmentation tasks, and their success lies in the availability of large amounts of labeled data. However, due to privacy, ethical, and legal concerns, medical image data is difficult to obtain, and the labeling process requires expertise from doctors or radiologists, making labeled medical image data even more challenging. To address the difficulty in obtaining labeled data, semi-supervised learning methods have emerged.
[0003] In existing semi-supervised learning methods, the main idea of consistency regularization models is to assume that perturbations of the same input should produce similar outputs. These perturbations include data augmentation or adversarial perturbations. While this approach helps improve model performance, the strength and type of perturbations need to be controlled; otherwise, the model will learn inaccurate features. The main purpose of pseudo-label models is to improve the quality of pseudo-labels, thereby enhancing model performance. However, inaccurate initial predictions are repeatedly used during training, causing error accumulation and ultimately affecting model performance. Furthermore, pseudo-label models require similar distributions of unlabeled and labeled data; otherwise, the quality of the pseudo-labels will affect the model's learning process. To address the shortcomings of these two models, researchers have combined them to propose a holistic model. Compared to single models, the holistic model can maintain consistent predictions for perturbed input images, adapt to common noise in medical images, and effectively improve the quality of pseudo-labels.
[0004] While the aforementioned methods effectively alleviate the model's dependence on labeled data, they essentially rely on labeled data to train the initial model, enhancing the utilization of unlabeled data through pseudo-labels or consistency regularization, rather than discovering the inherent patterns in unlabeled data and then using labeled data to refine these patterns. Discovering the inherent patterns in unlabeled data first is a form of unsupervised learning, and directly effective unsupervised learning methods include clustering and dimensionality reduction. Clustering can be considered an effective means of discovering sample similarity, and contrastive learning, as a similarity learning method, primarily aims to promote similarity within clusters and dissimilarity between clusters. However, in semi-supervised learning, because contrastive learning methods rely on pseudo-labels to cluster unlabeled samples, these methods are sensitive to noisy pseudo-labels. Therefore, how to change the traditional semi-supervised learning method's direct reliance on labeled data to assist learning and reduce the impact of noisy pseudo-labels on model performance is a direction worth exploring. Summary of the Invention
[0005] To overcome the above shortcomings, this invention provides a biconsistent spine image segmentation method based on probability clustering, which can solve the problems of semi-supervised image segmentation methods relying on labeled data to train the initial model and being sensitive to noise.
[0006] The technical solution of the present invention is as follows:
[0007] A biconsistency-based spine image segmentation method based on probability clustering includes the following steps:
[0008] (1) Data augmentation: The input spine image is augmented using a similar region replacement strategy to increase the diversity of labeled and unlabeled spine images;
[0009] (2) Feature extraction and segmentation:
[0010] The main network extracts high-level semantic features from the spine image using a deep encoder; the main network uses a deep encoder to extract high-level semantic features from the input image, where F represents the encoder. This represents the high-level semantic feature extraction process;
[0011] The subnetwork uses two existing semi-supervised segmentation networks for consistency training to segment the input spine image;
[0012] (3) Cluster information transformation:
[0013] Map the high-level semantic features extracted by the main network to the probability clustering space;
[0014] The segmentation results of the subnetwork are converted into probabilistic clustering information, which includes cluster centers, membership degree, and fuzzy index.
[0015] (4) Clustering information initialization: The clustering information of the segmentation results of the sub-network is used as the initialization information of the possible clustering space in the main network; the features in the clustering space are clustered to obtain the membership information;
[0016] (5) Clustering information learning: The main network learns the clustering information of the sub-networks, adjusts the fuzzy index, and thus adjusts the cluster centers and membership degrees in the clustering space of the main network to obtain the final clustering result;
[0017] (6) Consistency training between main network and sub-network: The results of the main network and the results of the sub-network are trained to reduce the difference between the two segmentation results, thereby improving the accuracy of semi-supervised spine image segmentation. At the same time, the segmentation results of the main network serve as the clustering initialization information of its neighboring slices.
[0018] Preferably, the specific steps of the similar region replacement strategy in step (1) above are as follows:
[0019] Divide the two different original images into 9 distinct image blocks using a 3×3 grid, denoted as follows: and , The Hamming distance between them is calculated using average hashing. The Hamming distance represents the similarity between image patches; if the Hamming distance between two image patches is less than 10, they are considered similar and are swapped. The final augmented data is represented as follows:
[0020]
[0021]
[0022] in This indicates a replacement operation.
[0023] Preferably, the semi-supervised segmentation network used in step (2) above is... and , For CDCL network, For an RCPC network, its consistency loss function is:
[0024]
[0025] in, Represented as The prediction results Represented as The prediction results This represents cross-entropy.
[0026] Preferably, the process of converting the segmentation result into probabilistic clustering information in step (3) above is as follows:
[0027] The conversion process is represented as
[0028]
[0029]
[0030] in, This indicates that the subnetwork is connected to the input image. The segmentation results, This represents the transformation of clustering information. Represents the set of cluster centers. This represents the set of membership degrees.
[0031] Preferably, in step (4) above, the features in the clustering space are clustered to obtain membership information. The membership information is represented as...
[0032]
[0033] in, Representing features in the cluster space, , , Used to recalculate the membership degree of features with a membership degree of 0.5.
[0034] Preferably, in step (5) above, the main network uses clustering similarity loss. By learning the clustering information of the subnetworks and adjusting the fuzzy index, the cluster centers and membership degrees in the main network's clustering space are adjusted to obtain the final clustering result; the clustering similarity loss is expressed as:
[0035]
[0036]
[0037] in, This represents the set of cluster centers for the labeled data. This represents the set of cluster centers representing the results predicted by the main network. This represents the set of cluster centers for the subnetwork prediction results. These represent the parameters that the network can learn;
[0038] Based on the update of cluster centers, the calculation of membership also changes accordingly, specifically expressed as follows:
[0039]
[0040] in, Representing features in the cluster space, , , Used to recalculate the membership degree of features with a membership degree of 0.5;
[0041] Given an image with N pixels and C cluster centers, the membership degree of each pixel to a cluster center is a membership matrix, as follows:
[0042]
[0043] The learning process for membership degrees is the same as that for cluster centers, and is represented as follows:
[0044]
[0045] in, This represents the set of membership degrees of the tag data. Indicates the membership degree after the main network update. This represents the set of membership degrees of the subnetwork segmentation results.
[0046] The clustering result from step (5) is input into the decoder to obtain the segmentation result of the main network. This result is compared with the result from step (2), and the difference between the main network and the sub-network is reduced through consistency training to obtain the final segmentation result. This result serves as the initial clustering information for the slices adjacent to the input slice. The consistency loss function between the main network and the sub-network is:
[0047]
[0048] in, Represents cross-entropy. This represents the prediction result of the main network. This represents the prediction result of the subnetwork.
[0049] The initialization process for adjacent slices is represented as follows:
[0050]
[0051]
[0052]
[0053] Where D represents the decoder and F represents the feature encoder. This indicates that the cluster centers are derived from the labels. This indicates that the cluster centers are predicted from the subnetwork.
[0054] The beneficial effects of this invention are:
[0055] This invention introduces probabilistic clustering into semi-supervised learning, enabling it to fully learn from unlabeled data. Probabilistic clustering effectively reduces the impact of noise on segmentation results, effectively solving the segmentation difficulties faced by diseased intervertebral discs due to their size, shape, and high inter-class similarity with surrounding tissues. Attached Figure Description
[0056] Figure 1 This is a schematic diagram of the process of the biconsistency spine image segmentation method based on probability clustering proposed in this invention;
[0057] Figure 2 This is a flowchart of the biconsistency spine image segmentation method based on probability clustering proposed in this invention;
[0058] Figure 3 This is a diagram of the biconsistency spine image segmentation network framework based on probability clustering proposed in this invention.
[0059] Figure 4 This invention proposes a similar region replacement data enhancement strategy.
[0060] Figure 5 This is the adjacent slice initialization strategy proposed in this invention. Detailed Implementation
[0061] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0062] The biconsistency spine image segmentation method based on probability clustering provided by this invention, such as... Figure 1 As shown, the feature is that it includes the following steps:
[0063] Step 1: Data Augmentation;
[0064] Data augmentation methods were used to increase the diversity of labeled and unlabeled spine images;
[0065] Step 1: Perform data augmentation on the spine image using the similar region replacement data augmentation strategy proposed in this invention. The similar region replacement data augmentation strategy is as follows: Figure 4 As shown;
[0066] Divide the two different original images into 9 distinct image blocks using a 3×3 grid, denoted as follows: and , The Hamming distance between them is calculated using average hashing. The Hamming distance represents the similarity between image patches; if the Hamming distance between two image patches is less than 10, they are considered similar and are swapped. The final augmented data is represented as follows:
[0067]
[0068]
[0069] in Indicates a replacement operation;
[0070] Step 2: Feature extraction;
[0071] The main network extracts high-level semantic features from the input image, and the sub-network performs consistency training on the input image to obtain the segmentation result;
[0072] Step 2a: Train the subnetwork consistency between two existing semi-supervised segmentation networks to obtain the segmentation result of the input image;
[0073] Subnetwork consistency loss is expressed as
[0074]
[0075] Step 2b: The main network uses a deep encoder to extract high-level semantic features from the input image;
[0076] Step 3: Cluster information transformation;
[0077] Map the high-level semantic features extracted by the main network to the probability clustering space;
[0078] The segmentation results of the subnetwork are converted into probabilistic clustering information (cluster center, membership degree, fuzzy index).
[0079] Step 3a: The sub-network prediction results are transformed to obtain the corresponding probability clustering information. The transformation process is represented as follows:
[0080]
[0081] Step 3b:
[0082]
[0083] in, This indicates that the subnetwork is connected to the input image. The segmentation results, This represents the transformation of clustering information. Represents the set of cluster centers. This represents the set of membership degrees.
[0084] Step 4: Initialize clustering information;
[0085] Use step 3a as the initialization information for step 3b to calculate the membership degree;
[0086] Step 5: Cluster Information Learning
[0087] The main network uses cluster similarity loss. By learning the clustering information of the subnetworks and adjusting the fuzzy index, the cluster centers and membership degrees in the main network's clustering space are adjusted to obtain the final clustering result; the clustering similarity loss is expressed as...
[0088]
[0089]
[0090] in, This represents the set of cluster centers for the labeled data. This represents the set of cluster centers representing the results predicted by the main network. This represents the set of cluster centers for the subnetwork prediction results. These represent the parameters that the network can learn;
[0091] Based on the update of cluster centers, the calculation of membership also changes accordingly, specifically expressed as follows:
[0092]
[0093] in, Representing features in the cluster space, , , Used to recalculate the membership degree of features with a membership degree of 0.5;
[0094] Given an image with N pixels and C cluster centers, the membership degree of each pixel to a cluster center is a membership matrix, as follows:
[0095]
[0096] The learning process for membership degrees is the same as that for cluster centers, and is represented as follows:
[0097]
[0098] in, This represents the set of membership degrees of the tag data. Indicates the membership degree after the main network update. This represents the set of membership degrees of the subnetwork segmentation results.
[0099] Step 6: The main network and sub-networks are trained in a consistent manner, and the prediction results of the main network are used as the initialization information for its neighboring slices.
[0100] By training the main network and sub-networks to maintain consistency, the difference between their segmentation results is reduced, thereby improving the accuracy of semi-supervised spine image segmentation. Figure 3 The two dotted dashed lines connecting the arrows indicate that when the input slice of the main network... When slicing labels, we will use the cluster centers of the label slices. as input slice Initialization of cluster centers When the input slice is not a label slice, we will subnetwork the input slice. Prediction results Cluster center This serves as the initial cluster center for the input slice in the main network. Next, as Figure 5 As shown, the main network is sliced. Prediction results As the next slice Initialize the cluster centers and then perform feature clustering.
[0101] The consistency loss between the main network and the sub-network is expressed as:
[0102]
[0103] Adjacent slice initialization is represented as
[0104]
[0105]
[0106]
[0107] It should be noted that the clustering method in the main network and the semi-supervised segmentation network in the sub-network can be freely replaced according to different task requirements. The above is merely the optimal implementation of this invention for addressing the noise impact in spine images after similar region replacement data augmentation, and for improving spine image segmentation accuracy by fully learning from unlabeled data. It should be pointed out that those skilled in the art can make several 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.
Claims
1. A biconsistency-based spine image segmentation method based on probability clustering, characterized in that, Includes the following steps: (1) Data augmentation: The input spine image is augmented using a similar region replacement strategy to increase the diversity of labeled and unlabeled spine images; (2) Feature extraction and segmentation: The main network extracts high-level semantic features from the spine image through a deep encoder; The subnetwork uses two existing semi-supervised segmentation networks for consistency training to segment the input spine image; (3) Cluster information transformation: Map the high-level semantic features extracted by the main network to the probability clustering space; The segmentation results of the subnetwork are converted into probabilistic clustering information, which includes cluster centers, membership degree, and fuzzy index. (4) Clustering information initialization: The clustering information of the segmentation results of the sub-network is used as the initialization information of the possible clustering space in the main network; the features in the clustering space are clustered to obtain the membership information; (5) Clustering information learning: The main network learns the clustering information of the sub-networks, adjusts the fuzzy index, and thus adjusts the cluster centers and membership degrees in the clustering space of the main network to obtain the final clustering result; (6) Consistency training between main network and sub-network: The results of the main network and the results of the sub-network are trained to reduce the difference between the two segmentation results, thereby improving the accuracy of semi-supervised spine image segmentation. At the same time, the segmentation results of the main network serve as the clustering initialization information of its neighboring slices.
2. The biconsistency spine image segmentation method based on probability clustering according to claim 1, characterized in that, The specific steps of the similar region replacement strategy in step (1) are as follows: Divide the two different original images into 9 distinct image blocks using a 3×3 grid, denoted as follows: and , The Hamming distance between them is calculated using average hashing. The Hamming distance represents the similarity between image patches; if the Hamming distance between two image patches is less than 10, they are considered similar and are swapped. The final augmented data is represented as follows: in This indicates a replacement operation.
3. The biconsistency spine image segmentation method based on probability clustering according to claim 1, characterized in that, The semi-supervised segmentation network used in step (2) is... and Its consistency loss function is: in, Represented as The prediction results Represented as The prediction results This represents cross-entropy.
4. The biconsistency spine image segmentation method based on probability clustering according to claim 1, characterized in that, The process of converting the segmentation results into probabilistic clustering information in step (3) is as follows: The conversion process is represented as in, This indicates that the subnetwork is connected to the input image. The segmentation results, This represents the transformation of clustering information. Represents the set of cluster centers. This represents the set of membership degrees.
5. The biconsistency spine image segmentation method based on probability clustering according to claim 1, characterized in that, In step (4), the features in the clustering space are clustered to obtain membership information. Membership information is represented as... in, Representing features in the cluster space, , , Used to recalculate the membership degree of features with a membership degree of 0.
5.
6. The biconsistency spine image segmentation method based on probability clustering according to claim 1, characterized in that, In step (5), the main network uses clustering similarity loss. By learning the clustering information of the subnetworks and adjusting the fuzzy index, the cluster centers and membership degrees in the main network's clustering space are adjusted to obtain the final clustering result; the clustering similarity loss is expressed as: in, This represents the set of cluster centers for the labeled data. This represents the set of cluster centers representing the results predicted by the main network. This represents the set of cluster centers for the subnetwork prediction results. These represent the parameters that the network can learn; Based on the update of cluster centers, the calculation of membership also changes accordingly, specifically expressed as follows: in, Representing features in the cluster space, , , Used to recalculate the membership degree of features with a membership degree of 0.5; Given an image with N pixels and C cluster centers, the membership degree of each pixel to a cluster center is a membership matrix, as follows: The learning process for membership degrees is the same as that for cluster centers, and is represented as follows: in, This represents the set of membership degrees of the tag data. Indicates the membership degree after the main network update. This represents the set of membership degrees of the subnetwork segmentation results.
7. The biconsistency spine image segmentation method based on probability clustering according to claim 1, characterized in that, The consistency loss function between the main network and the sub-network in step (6) is: in, Represents cross-entropy. This represents the prediction result of the main network. This represents the prediction result of the subnetwork.