Lumbar vertebra disease semi-supervised classification method and system based on semantic comparison and fusion of uncertain perception

Abstract

The invention relates to the technical field of image classification and recognition, in particular to a semi-supervised classification method and a semi-supervised classification system for lumbar diseases based on semantic contrast fusion uncertain sensing, and adopts a semi-supervised classification network model as a semi-supervised classification framework for lumbar disease recognition. The cross entropy loss of the labeled samples and the consistency loss of the unlabeled samples are used as the loss of the model; selecting the labels with the prediction confidence greater than a threshold value and the prediction uncertainty less than the threshold value as pseudo labels; taking the labeled data with the same label and the unlabeled data with the false label as a positive sample pair, and training by adopting a supervised comparative learning method to obtain comparative learning loss of the model; semi-supervised model loss, uncertainty loss and comparative learning loss are used as total loss functions for joint training of the model. The problems that semantic information of annotated data is not utilized to constrain unannotated data, and a pseudo tag has confirmation deviation are solved, and the lumbar disease recognition accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image classification and recognition, and particularly to a semi-supervised classification method and system for lumbar diseases based on semantic contrast fusion and uncertainty perception. Background Art

[0002] Lumbar diseases are very common and have a relatively high incidence in clinical practice. They are mainly caused by injuries to the soft tissues around the spine, bone hyperplasia, and organic lesions of the lumbar intervertebral discs. Lumbar disc herniation is one of the most common lumbar diseases. As a chronic disease, it receives relatively little attention, which may lead to missing the best treatment opportunity and causing great trouble in the future treatment process. Therefore, early evaluation and diagnosis of lumbar diseases are of great significance. Magnetic resonance images (MRI) can be used to evaluate spinal diseases such as lumbar intervertebral disc degeneration and spinal nerve compression, and have high value for the diagnosis of lumbar diseases. With the rapid growth of the number of images and the rapid development of artificial intelligence technology, intelligent medical image processing and analysis technology has also made remarkable progress.

[0003] Currently, many studies have applied deep learning technology to the diagnosis of lumbar diseases, such as lumbar disc herniation classification, spinal stenosis grading, foraminal stenosis detection and grading, lumbar spondylolisthesis identification, and various lumbar degenerative diseases. These studies usually collect hundreds or thousands of cases for experiments, and professional radiologists manually annotate these images. The performance of intelligent diagnosis models usually depends on the quality and quantity of the labeled data in the training dataset. The amount of lumbar magnetic resonance image data is large, and it takes a lot of time and effort to annotate these images, and it depends on the professional knowledge of doctors, which increases the workload of radiologists and limits the scale of the training dataset. Therefore, it is of great significance to label only a small number of images in a large-scale dataset to achieve a diagnostic performance similar to that of labeling a large number of images, thereby reducing the amount of labeling.

[0004] Semi-supervised learning (SSL) is used to train models in the presence of limited labeled data and a large amount of unlabeled data, alleviating the need for more manually labeled data. One paradigm of SSL is to train a model on a small labeled dataset and input the unlabeled data into the well-trained model to obtain predicted labels. SSL can generalize the representation ability learned from labeled data to unlabeled data and has received extensive attention. Pseudo-labeling and consistency regularization are two main SSL methods, and the currently dominant method has evolved into a hybrid method that combines pseudo-labeling and consistency regularization. Pseudo-labeling usually uses a simple prediction probability threshold to generate pseudo-labels, but when the model overfits to incorrect pseudo-labels, confirmation bias is likely to occur. In the subsequent training phase, these incorrect labels mislead the model's feature learning, resulting in deteriorated model performance. Consistency regularization trains the model by adding noise perturbations to unlabeled images, performing image enhancement transformations, or using the randomness of model predictions to constrain the consistency of unlabeled data predictions as an unsupervised loss. However, most consistency regularization methods utilize the transformation or prediction constraints of unlabeled data and do not consider semantic feature constraints, thus not fully exploiting the semantic features of labeled data.

[0005] In current semi-supervised classification algorithms, there is no direct semantic interaction between labeled data and unlabeled data. The implicit semantic relationship is that the model trained with labeled data generates pseudo-labels for unlabeled data, and the unlabeled data with pseudo-labels and the labeled data are jointly used to update the model. Through the trained model, the semantic information of labeled data is transferred to unlabeled data. However, in the existing semi-supervised learning process, the semantic features of labeled data are not utilized to guide the learning of unlabeled data. Summary of the Invention

[0006] The present invention aims to solve the problem that a large number of labeled images are crucial for deep learning-based lumbar disease diagnosis, but the labeling is time-consuming and not easily obtainable; to overcome this limitation, semi-supervised learning is applied to the training of a small number of labeled images and a large number of unlabeled images. Semi-supervised learning mainly utilizes pseudo-labeling and consistency regularization; however, consistency regularization only constrains the semantic consistency of unlabeled data and does not use the semantic information of labeled data to constrain unlabeled data, and there is confirmation bias in pseudo-labeling. The present invention proposes a semi-supervised classification method and system for lumbar diseases based on semantic contrast fusion and uncertainty perception, which simultaneously considers the prediction uncertainty and semantic similarity between unlabeled images and labeled images, improves the accuracy of pseudo-labels and the performance of the semi-supervised classification network model, and further enhances the accuracy of intelligent identification of lumbar diseases.

[0007] To achieve the above object, the technical solutions adopted are as follows:

[0008] The present invention provides a semi-supervised classification method for lumbar diseases based on semantic contrast fusion and uncertainty perception, including:

[0009] Step 1: Use a semi-supervised classification network model that combines pseudo-label and consistency regularization methods as the semi-supervised classification framework for lumbar disease recognition, and use the cross-entropy loss of labeled samples and the consistency loss of unlabeled samples as the loss of the semi-supervised classification network model;

[0010] Step 2: Train a Dirichlet prior network to minimize the KL divergence between the model and the target Dirichlet distribution, calculate the prediction uncertainty of unlabeled data, and select the label that simultaneously satisfies the prediction confidence being greater than a threshold and the prediction uncertainty being less than a threshold as the pseudo-label;

[0011] Step 3: Use the labeled data with the same label and the pseudo-labeled unlabeled data as positive sample pairs, and regard other samples as negative sample pairs, and use a supervised contrast learning method for training to obtain the contrast learning loss of the model;

[0012] Step 4: Use the semi-supervised model loss, uncertainty loss, and contrast learning loss as the total loss function to jointly train the model and update the semi-supervised classification network model.

[0013] According to the semi-supervised classification method for lumbar diseases based on semantic contrast fusion and uncertainty perception of the present invention, further, the specific steps of Step 1 include:

[0014] Randomly perform two weak augmentation transformations on the labeled data, and perform weak augmentation and strong augmentation transformations on the unlabeled data respectively, and use the transformed data as the input of the model;

[0015] Use a semi-supervised classification network model that combines pseudo-label and consistency regularization methods as the semi-supervised classification framework for lumbar disease recognition;

[0016] Calculate the cross-entropy loss L for the labeled images according to the corresponding lumbar disease categories xc ; calculate the consistency loss L of the predictions of different augmented images for the unlabeled images uc ; the sum of the two losses is used as the loss of the semi-supervised classification network model.

[0017] According to the semi-supervised classification method for lumbar diseases based on semantic contrast fusion and uncertainty perception of the present invention, further, the semi-supervised classification network model that combines pseudo-label and consistency regularization methods uses FixMatch.

[0018] According to the semi-supervised classification method for lumbar diseases based on semantic contrast fusion and uncertainty perception of the present invention, further, the specific steps of Step 2 include:

[0019] Regard the semi-supervised classification network model with Softmax activation as the class distribution expected under the Dirichlet prior;

[0020] Train the Dirichlet prior network to minimize the KL divergence between the model and the target Dirichlet distribution, and use this KL divergence as the uncertainty loss L d ;

[0021] Based on the trained prior network, calculate the prediction uncertainty of the unlabeled data;

[0022] Take the prediction label with a prediction confidence greater than the confidence threshold and a prediction uncertainty less than the uncertainty threshold for the unlabeled data as the pseudo-label.

[0023] According to the semi-supervised classification method for lumbar diseases based on semantic contrast fusion and uncertainty perception of the present invention, further, the specific steps of step 3 include:

[0024] Add two fully connected layers after the feature encoder of the semi-supervised classification network model as the projection network;

[0025] Calculate the feature embeddings of two kinds of augmented labeled data and unlabeled data passing through the feature encoder and the projection network of the semi-supervised classification network model respectively, and perform normalization to obtain the final semantic embeddings;

[0026] Based on the pseudo-label generation method in step 2, select the feature embedding vectors with pseudo-labels to participate in supervised contrast learning, regard the labeled data with the same label and the pseudo-labeled unlabeled data as positive sample pairs, and regard other samples as negative sample pairs;

[0027] Use the InfoNCE loss of contrast learning to calculate the loss L of supervised contrast learning for the feature embeddings of the labeled data with the same label and the pseudo-labeled unlabeled data s .

[0028] According to the semi-supervised classification method for lumbar diseases based on semantic contrast fusion and uncertainty perception of the present invention, further, use the semi-supervised model loss, uncertainty loss, and contrast learning loss as the total loss function, and the calculation formula of the total loss function is as follows:

[0029] L = L xc + L uc + L d + L s

[0030] Among them, L is the total loss function, L xc represents the cross-entropy loss, L uc represents the consistency loss, L d represents the uncertainty loss, L s represents the contrast learning loss.

[0031] According to the semi-supervised classification method for lumbar diseases based on semantic contrast fusion and uncertain perception of the present invention, further, after step 4, it further includes: by observing the performance of the model on the validation set, selecting the best trained model, testing it on the test set, and calculating the test recognition accuracy of the semi-supervised classification network model.

[0032] Furthermore, the present invention also provides a semi-supervised classification system for lumbar diseases based on semantic contrast fusion and uncertain perception, including:

[0033] A pre-training module, which is used to adopt a semi-supervised classification network model that combines pseudo-labeling and consistency regularization methods as the semi-supervised classification framework for lumbar disease recognition, and uses the cross-entropy loss of labeled samples and the consistency loss of unlabeled samples as the loss of the semi-supervised classification network model;

[0034] A pseudo-label generation module, which is used to train a Dirichlet prior network to minimize the KL divergence between the model and the target Dirichlet distribution, calculate the prediction uncertainty of unlabeled data, and select the label that simultaneously satisfies the prediction confidence being greater than the threshold and the prediction uncertainty being less than the threshold as the pseudo-label;

[0035] A contrastive learning training module, which is used to use the labeled data with the same label and the pseudo-labeled unlabeled data as positive sample pairs, and other samples as negative sample pairs, and adopt a supervised contrastive learning method for training to obtain the contrastive learning loss of the model;

[0036] A joint training module, which is used to use the semi-supervised model loss, uncertainty loss, and contrastive learning loss as the total loss function to jointly train the model and update the semi-supervised classification network model.

[0037] Adopting the above technical solutions, the beneficial effects obtained are:

[0038] The present invention mainly faces the problem that when semi-supervised learning is used for training limited labeled images and a large number of unlabeled images, the semantic information of the labeled data is not effectively utilized and the pseudo-label has confirmation bias. Based on a small amount of labeled lumbar image data and a large amount of unlabeled lumbar image data, a deep model that integrates semi-supervised learning and semantic contrast constraints between labeled and unlabeled images is proposed. In order to better utilize the labeled data to guide the training of unlabeled data, supervised contrast learning is introduced to constrain the semantic feature consistency between labeled data with the same label and unlabeled data with pseudo-labels. Therefore, the correctness of the pseudo-label is crucial. Wrong pseudo-labels will confuse contrast learning and have a negative impact on semi-supervised learning. In order to improve the reliability of pseudo-labels and further improve the accuracy of supervised contrast learning, pseudo-labels are generated using prediction confidence and prediction uncertainty, and the KL loss between the predicted Dirichlet distribution and the target Dirichlet distribution is optimized by the training model to calculate the uncertainty. The present invention uses semantic contrast learning based on uncertainty constraints of Dirichlet distribution to narrow the feature distance between unlabeled images and labeled images, while reducing the confirmation bias caused by false pseudo-labels. Since the prediction uncertainty and semantic similarity between unlabeled and labeled images are taken into consideration at the same time, the accuracy of pseudo-labels and the performance of the semi-supervised classification network model are improved, thereby improving the accuracy of intelligent identification of lumbar diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0040] Figure 1 It is a framework diagram of a semi-supervised classification method for lumbar vertebrae diseases based on semantic contrast fusion and uncertainty perception according to an embodiment of the present invention;

[0041] Figure 2 It is a schematic diagram of a supervised semantic contrastive learning training method for lumbar disease according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings of specific embodiments of the present invention to clearly and completely describe the exemplary scheme of the embodiment of the present invention. Unless otherwise defined, the technical terms or scientific terms used in the present invention should be the common meanings understood by people with ordinary skills in the field.

[0043] This embodiment discloses a semi-supervised classification method for lumbar vertebrae diseases based on semantic contrast fusion and uncertain perception. The method framework diagram is shown in FIG. Figure 1 As shown, the following steps are included:

[0044] Step S1: Implement semi-supervised classification for lumbar disease recognition using a semi-supervised classification network model that combines pseudo-labels and consistency regularization. Calculate the cross-entropy loss and consistency loss for labeled images and unlabeled images respectively, and use the sum of the two losses to update the model training as the loss of the semi-supervised classification network model. Specifically, it includes:

[0045] Step S101: Randomly perform two weak augmentation transformations on the labeled lumbar data. The weak augmentation transformations include random cropping, flipping, etc.; perform weak augmentation and strong augmentation transformations on the unlabeled lumbar data respectively, and use the transformed labeled and unlabeled data as the input of the model.

[0046] Step S102: Use a semi-supervised classification network model that combines pseudo-labels and consistency regularization, such as FixMatch, as the semi-supervised classification baseline model for lumbar disease recognition.

[0047] Step S103: Calculate the cross-entropy loss L xc for the labeled images according to the corresponding lumbar disease categories, calculate the consistency loss L uc between the weakly augmented and strongly augmented image predictions for the unlabeled images, and use the sum of the two losses L xc +L uc as the loss of the semi-supervised classification network model.

[0048] Step S2: Train the Dirichlet prior network to minimize the KL divergence between the model and the target Dirichlet distribution. Based on the trained prior network, calculate the prediction uncertainty of the unlabeled data, and use the predicted labels with a prediction confidence greater than the confidence threshold and a prediction uncertainty less than the uncertainty threshold for the unlabeled data as pseudo-labels. Specifically, it includes:

[0049] Step S201: Consider the semi-supervised classification network model with Softmax activation as the expected class distribution under the Dirichlet prior.

[0050] Step S202: Use the KL divergence between the model and the target Dirichlet distribution as the loss L d for uncertainty calculation. For a definite category, it is a sharp Dirichlet distribution; for an uncertain distribution, it is a flat Dirichlet distribution; train the Dirichlet prior network to minimize this uncertainty loss, as shown in the following formulas (1) and (2):

[0051]

[0052] L d =E[KL[Dir(V|α)||P(V|x;θ)]] (2)

[0053] Among them, V is the probability vector, X is the image feature vector, θ is the model parameter vector, α i is the concentration parameter, equal to the output of the Dirichlet prior network, α 0 is the precision of the Dirichlet distribution, controlling the sharpness of the Dirichlet distribution, Dir(·) is the Dirichlet distribution, Dir(V|α) means that the variable V obeys the Dirichlet distribution with parameter α, and P(·) is the conditional distribution.

[0054] Step S203: Based on the trained prior network, the prediction uncertainty of the unlabeled data is calculated, such as confidence, entropy, mutual information or differential entropy.

[0055] Step S204: only the predicted labels when the prediction confidence of the unlabeled data is greater than the confidence threshold and the prediction uncertainty is less than the uncertainty threshold are used as pseudo labels. The constraint function for generating pseudo labels is shown in formula (3):

[0056]

[0057] in, is the predicted probability, is the uncertainty of the prediction, I is the sign function, τ c is the confidence threshold, τ u is the uncertainty threshold.

[0058] Step S3, feature embedding is obtained through feature encoder and projection network, pseudo labels of unlabeled data are obtained based on the pseudo label generation method of step S2 above, labeled data with the same label and pseudo-labeled unlabeled data are used as positive sample pairs, and other samples are regarded as negative sample pairs. Supervised contrastive learning method is used for training to obtain the contrastive learning loss of the model. The supervised semantic contrastive learning training method is as follows Figure 2 As shown; specifically including:

[0059] Step S301: Add two fully connected layers after the feature encoder of the semi-supervised classification network model as a projection network to map the features.

[0060] Step S302: Calculate the feature embedding of the feature encoder and projection network of the semi-supervised classification network model for the labeled data with two random weak enhancements and the unlabeled data with random weak enhancement and strong enhancement, and normalize the obtained feature embeddings to obtain normalized semantic embeddings.

[0061] Step S303: Based on the uncertainty-aware pseudo-label generation method of step S2, feature embedding vectors with pseudo-labels are selected to be added to supervised contrastive learning, and labeled data with the same label and pseudo-labeled unlabeled data are regarded as positive sample pairs, and other samples are regarded as negative sample pairs.

[0062] Step S304: Calculate the loss L of supervised contrastive learning by using the feature embeddings of the labeled data with the same label and the feature embeddings of the pseudo-labeled unlabeled data with the InfoNCE loss commonly used in contrastive learning. s 。

[0063] Step S4: Use the semi-supervised model loss, uncertainty loss, and contrastive learning loss as the total loss function to jointly train the model and update the semi-supervised classification network model.

[0064] L = L xc + L uc + L d + L s (4)

[0065] where L is the total loss function, L xc represents the cross-entropy loss, L uc represents the consistency loss, L d represents the uncertainty loss, L s represents the contrastive learning loss.

[0066] Step S5: The trained semi-supervised classification network model is the intelligent lumbar disease recognition model. By observing the performance of the model on the validation set, select the best trained model, test it on the test set, and calculate the test recognition accuracy of the semi-supervised classification network model.

[0067] Correspondingly, this embodiment also proposes a semi-supervised classification system for lumbar diseases based on semantic contrast fusion and uncertainty perception, including:

[0068] A pre-training module, which uses a semi-supervised classification network model that combines pseudo-labeling and consistency regularization methods as the semi-supervised classification framework for lumbar disease recognition, and uses the cross-entropy loss of the labeled samples and the consistency loss of the unlabeled samples as the loss of the semi-supervised classification network model.

[0069] A pseudo-label generation module, which is used to train the Dirichlet prior network to minimize the KL divergence between the model and the target Dirichlet distribution, calculate the prediction uncertainty of the unlabeled data, and select the label that simultaneously satisfies the prediction confidence greater than the threshold and the prediction uncertainty less than the threshold as the pseudo-label.

[0070] A contrastive learning training module, which uses the labeled data with the same label and the pseudo-labeled unlabeled data as positive sample pairs, and other samples as negative sample pairs, and uses the supervised contrastive learning method for training to obtain the contrastive learning loss of the model.

[0071] The joint training module is used to jointly train a model using the semi-supervised model loss, uncertainty loss, and contrastive learning loss as the total loss function, and update the semi-supervised classification network model.

[0072] Unless otherwise specifically stated, the components, steps, numerical expressions, and values set forth in these embodiments do not limit the scope of the present invention.

[0073] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0074] The units and method steps of the various examples described in connection with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the composition and steps of the various examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art can use different methods to implement the described functions for each specific application, but such implementation is not considered to exceed the scope of the present invention.

[0075] A person of ordinary skill in the art can understand that all or part of the steps in the above methods can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium, such as a read-only memory, a magnetic disk, or an optical disc, etc. Optionally, all or part of the steps of the above embodiments can also be implemented using one or more integrated circuits. Correspondingly, the various modules / units in the above embodiments can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of the combination of hardware and software.

[0076] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, a person of ordinary skill in the art should understand that: any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A semi-supervised classification method for lumbar diseases based on semantic contrast fusion and uncertainty perception, characterized in that: include: Step 1: A semi-supervised classification network model integrating pseudo-labels and consistency regularization method is used as a semi-supervised classification framework for lumbar disease recognition, and the cross entropy loss of labeled samples and the consistency loss of unlabeled samples are used as the loss of the semi-supervised classification network model; Step 2: Train the Dirichlet prior network to minimize the KL divergence between the model and the target Dirichlet distribution, calculate the prediction uncertainty of the unlabeled data, and select the labels that satisfy both the prediction confidence greater than the threshold and the prediction uncertainty less than the threshold as pseudo labels; Step 3: The labeled data with the same label and the pseudo-labeled unlabeled data are regarded as positive sample pairs, and other samples are regarded as negative sample pairs. The supervised contrastive learning method is used for training to obtain the contrastive learning loss of the model. Step 4: Use the semi-supervised model loss, uncertainty loss, and contrastive learning loss as the total loss function to jointly train the model and update the semi-supervised classification network model.

2. The semi-supervised classification method for lumbar vertebrae diseases based on semantic contrast fusion and uncertainty perception according to claim 1 is characterized in that: The step 1 specifically includes: Randomly perform two weak enhancement transformations on the labeled data, perform weak enhancement and strong enhancement transformations on the unlabeled data, and use the transformed data as the input of the model; A semi-supervised classification network model integrating pseudo-labels and consistency regularization method is used as a semi-supervised classification framework for lumbar disease identification. Calculate the cross entropy loss L for the labeled images according to the corresponding lumbar disease category xc ; Calculate the consistency loss L of different enhanced image predictions for unlabeled images uc ; Both losses and the loss as a semi-supervised classification network model.

3. The semi-supervised classification method for lumbar vertebrae diseases based on semantic contrast fusion and uncertainty perception according to claim 2 is characterized in that: The semi-supervised classification network model integrating pseudo-label and consistency regularization method adopts FixMatch.

4. The semi-supervised classification method for lumbar vertebrae diseases based on semantic contrast fusion and uncertainty perception according to claim 1 is characterized in that: The step 2 specifically includes: The semi-supervised classification network model with Softmax activation is considered as the expected class distribution under the Dirichlet prior; The Dirichlet prior network is trained to minimize the KL divergence between the model and the target Dirichlet distribution, using the KL divergence as the uncertainty loss L d ; Based on the trained prior network, calculate the prediction uncertainty of unlabeled data; The predicted label of the unlabeled data whose prediction confidence is greater than the confidence threshold and whose prediction uncertainty is less than the uncertainty threshold is taken as the pseudo label.

5. The semi-supervised classification method for lumbar vertebrae diseases based on semantic contrast fusion and uncertainty perception according to claim 1 is characterized in that: The step 3 specifically includes: Add two fully connected layers as projection networks after the feature encoder of the semi-supervised classification network model; Calculate the feature embedding of the two enhanced labeled data and unlabeled data through the feature encoder of the semi-supervised classification network model and the projection network respectively, and normalize them to obtain the final semantic embedding; Based on the pseudo-label generation method in step 2, feature embedding vectors with pseudo-labels are selected to join supervised contrastive learning. Labeled data with the same label and pseudo-labeled unlabeled data are regarded as positive sample pairs, and other samples are regarded as negative sample pairs. The feature embedding of the labeled data with the same label and the feature embedding of the pseudo-labeled unlabeled data are used to calculate the loss L of supervised contrastive learning using the InfoNCE loss of contrastive learning. s .

6. The semi-supervised classification method for lumbar vertebrae diseases based on semantic contrast fusion and uncertainty perception according to claim 1 is characterized in that: Using semi-supervised model loss, uncertainty loss and contrastive learning loss as the total loss function, the calculation formula of the total loss function is as follows: L=L xc +L uc +L d +L s Among them, L is the total loss function, L xc represents the cross entropy loss, L uc represents the consistency loss, L d Represents uncertainty loss, L s represents the contrastive learning loss.

7. The semi-supervised classification method for lumbar vertebrae diseases based on semantic contrast fusion and uncertainty perception according to claim 1 is characterized in that: After step 4, the method also includes: selecting the best training model by observing the performance of the model on the validation set, testing it on the test set, and calculating the test recognition accuracy of the semi-supervised classification network model.

8. A semi-supervised classification system for lumbar diseases based on semantic contrast fusion and uncertainty perception, characterized in that: include: A pre-training module is used to adopt a semi-supervised classification network model that integrates pseudo labels and consistency regularization methods as a semi-supervised classification framework for lumbar disease recognition, and use the cross entropy loss of labeled samples and the consistency loss of unlabeled samples as the loss of the semi-supervised classification network model; The pseudo-label generation module is used to train the Dirichlet prior network to minimize the KL divergence between the model and the target Dirichlet distribution, calculate the prediction uncertainty of the unlabeled data, and select the labels that satisfy both the prediction confidence greater than the threshold and the prediction uncertainty less than the threshold as pseudo-labels; The contrastive learning training module is used to treat the labeled data with the same label and the pseudo-labeled unlabeled data as positive sample pairs, and other samples as negative sample pairs, and use the supervised contrastive learning method for training to obtain the contrastive learning loss of the model; The joint training module is used to jointly train the model using the semi-supervised model loss, uncertainty loss and contrastive learning loss as the total loss function to update the semi-supervised classification network model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images