Noise-containing intestinal ultrasound image classification method and system based on pseudo label relaxation contrast loss
By adopting pseudo-label relaxation contrast loss and semi-supervised learning methods in deep neural networks, the model overfitting problem caused by noise labels is solved, the performance and robustness of intestinal ultrasound image classification is improved, and the cost of expert labeling is reduced .
Patent Information
- Application Number
- CN202411019544.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Deep neural networks in ultrasonic intestinal image datasets result in noise labeling due to the lack of labeling experts, resulting in overfitting of the model and reducing discriminant performance and generalization performance.
Using a method based on pseudo-label slack comparison loss, two DNN models are robustly trained through semi-supervised learning and pseudo-label slack comparison learning, and the labeled and labelless sets are divided using the cross loss function and Jensen–Shannon divergence loss value. Calculate the slack comparison loss and update the model parameters.
The test performance and robust performance of the DNN model for the intestinal ultrasound image set with noise-containing labels is improved, saving expert labeling costs.
Smart Images

Figure CN119091184A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of computer vision, and in particular relates to a method and system for classifying noisy intestinal ultrasound images based on pseudo-label relaxation contrast loss. Background Art
[0002] Deep Neural Networks (DNN) are widely used in the field of computer vision, but the training of DNN requires the support of a large number of images with accurate labels, and the labeling of such images requires the experience of domain experts.
[0003] In the process of collecting ultrasound intestinal image datasets, although various hospitals and clinics have accumulated a large amount of intestinal image data, due to the lack of labeling experts, such samples often have inaccurate labels (noise labels), and noisy label samples can easily cause DNN overfitting, resulting in a decrease in DNN discrimination performance and generalization performance. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a noisy intestinal ultrasound image classification method and system based on pseudo-label relaxation contrast loss, which can utilize pseudo-label relaxation contrast loss to improve the test performance and robust performance of the DNN model when facing a set of intestinal ultrasound images with noisy labels, thereby saving expert annotation costs.
[0005] The present invention provides the following technical solutions:
[0006] In a first aspect, a method for classifying noisy intestinal ultrasound images based on pseudo-label relaxation contrast loss is provided, comprising the following steps:
[0007] S1: Two DNN models are gently trained based on the original noisy intestinal ultrasound image dataset;
[0008] S2: Use each DNN model to divide the original noisy intestinal ultrasound image dataset to obtain a labeled set and an unlabeled set corresponding to the current DNN model;
[0009] S3: Use semi-supervised learning and pseudo-label relaxed contrastive learning to perform robust training on the two DNN models respectively to update the two DNN models; when robust training the first DNN model, the labeled set and unlabeled set divided by the second DNN model are used; when robust training the second DNN model, the labeled set and unlabeled set divided by the first DNN model are used;
[0010] S4: Repeat steps S2-S3 until a preset total number of training rounds is reached, and the final DNN model is output to classify noisy intestinal ultrasound images using the final DNN model.
[0011] Optionally, in step S3, the current DNN model is robustly trained using semi-supervised learning and pseudo-label relaxed contrastive learning to update the current DNN model, including:
[0012] Calculate the SSL loss of all samples based on the labeled set and the unlabeled set;
[0013] Use the mapping head to map the extracted high-dimensional features into a low-dimensional space to form low-dimensional features;
[0014] Based on the mapped low-dimensional features, the relaxed contrast loss is calculated for all samples in the labeled set and the unlabeled set;
[0015] The SSL loss and relaxed contrast loss of all samples are combined to calculate the total loss, and backpropagation and model parameter update are performed according to the total loss to update the current DNN model.
[0016] Optionally, based on the mapped low-dimensional features, for all samples in the labeled set and the unlabeled set, when calculating the relaxed contrast loss, it is necessary to construct a negative sample set;
[0017] The constructing of the negative sample set comprises:
[0018] For samples xi , assuming its prediction result vector;
[0019] According to the sample xi The prediction result vector is used to find the index corresponding to the first k components with the highest probability in the prediction result.
[0020] The indexes and samples corresponding to the top k components with the highest probability in all the observed labels that are not equal to the given labels are predicted. xi The set of samples with no intersection in the corresponding indexes is regarded as the negative sample set;
[0021] The negative sample set N i The formula is:
[0022]
[0023] in, is the sample in the first k components with the highest probability in the prediction result xj The corresponding index, For sample xi The corresponding index;
[0024] The relaxation contrast loss L plr for:
[0025]
[0026] in, is the number of original noisy intestinal ultrasound image datasets, It is a sample xi After transformation enhancement, the low-dimensional features mapped by the mapping head, is the low-dimensional feature after another enhancement, τ is the preset sharpening parameter, It is the low-dimensional feature set of the negative sample set.
[0027] Optionally, the SSL loss L ssl for:
[0028]
[0029] in, is a labeled set divided by the DNN model numbered m, is the unlabeled set divided by the DNN model numbered m, L l Based on labeled collection The classification loss, L ul Based on unlabeled collection The classification loss, L reg is the regularization loss, λ u is the preset weight coefficient.
[0030] Optionally, in step S1, a cross loss function is used to perform mild training on two DNN models; the two DNN models have the same structure and different initialization parameters;
[0031] The cross loss function L ce for:
[0032]
[0033] Among them, p m (x i ) is the DNN model numbered m based on the ultrasound image x i The output value, x i is the i-th ultrasound image, m is 1 or 2, n is the number of samples in the original noisy intestinal ultrasound image dataset, y i is the ultrasound image x i The observation label of .
[0034] Optionally, in step S2, the original noisy intestinal ultrasound image dataset is divided to obtain a labeled set and an unlabeled set corresponding to the current DNN model, including:
[0035] Calculate the JSD loss value of all samples in the original noisy intestinal ultrasound image dataset under the current DNN model;
[0036] According to the category of each sample, the JSD loss value is divided into c subsets, and the JSD loss values in each subset are sorted in ascending order;
[0037] Calculate the number of samples R that need to be selected in each subset;
[0038] All subsets are divided into samples to obtain labeled sets and unlabeled sets corresponding to the current DNN model;
[0039] The sample division of the current subset includes: judging whether the number of samples in the current subset is greater than the number of samples R that need to be selected. If so, the first R samples of the current subset are placed in the labeled set, and the rest are placed in the unlabeled set; otherwise, all samples of the current subset are placed in the labeled set.
[0040] Optionally, the formula for calculating the JSD loss value of all samples in the original noisy intestinal ultrasound image dataset under the current DNN model is:
[0041]
[0042] Among them, d i m is the JSD loss value of the i-th sample under the DNN model numbered m, is the intestinal ultrasound image x i The observation label, p m (x i ) The DNN model numbered m is based on the ultrasound image x i The output value of ; JSD(·) is the Jensen–Shannon divergence function;
[0043] The formula for calculating the number of samples R that need to be selected in each subset is:
[0044]
[0045] Among them, R is the number of samples that need to be selected, r is the noise ratio of the original noisy intestinal ultrasound image dataset, c is the number of subsets divided by the JSD loss value, and n is the number of samples in the original noisy intestinal ultrasound image dataset.
[0046] In a second aspect, a noisy intestinal ultrasound image classification system based on pseudo-label relaxation contrast loss is provided, comprising:
[0047] A gentle training module, used to perform gentle training on two DNN models based on the original noisy intestinal ultrasound image dataset;
[0048] A set partitioning module is used to partition the original noisy intestinal ultrasound image dataset using each DNN model to obtain a labeled set and an unlabeled set corresponding to the current DNN model;
[0049] A robust training module, for performing robust training on two DNN models respectively by using semi-supervised learning and pseudo-label relaxed contrastive learning to update the two DNN models; when performing robust training on the first DNN model, the labeled set and the unlabeled set divided by the second DNN model are used; when performing robust training on the second DNN model, the labeled set and the unlabeled set divided by the first DNN model are used;
[0050] The module is repeatedly executed to feed back to the set partitioning module and the robust training module until a preset total number of training rounds is reached, and a final DNN model is output to classify noisy intestinal ultrasound images using the final DNN model.
[0051] According to a third aspect, a computer device is provided, comprising a processor and a memory; wherein, when the processor executes a computer program stored in the memory, the steps of the noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss as described in any one of the first aspects are implemented.
[0052] In a fourth aspect, a computer-readable storage medium is provided for storing a computer program; when the computer program is executed by a processor, the steps of the noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss as described in any one of the first aspects are implemented.
[0053] Compared with the prior art, the present invention has the following beneficial effects:
[0054] (1) The present invention proposes a noisy intestinal ultrasound image classification method and system based on pseudo-label relaxation contrast loss, which can utilize pseudo-label relaxation contrast loss to improve the test performance and robust performance of the DNN model when facing a set of intestinal ultrasound images with noisy labels, thereby saving the cost of experts in classifying noisy intestinal ultrasound images.
[0055] (2) The present invention combines the balanced selection strategy with the pseudo-label relaxation contrast loss, thereby further improving the test performance and robustness of the DNN model when facing a set of intestinal ultrasound images containing noisy labels. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 It is a schematic diagram of the overall structure of the noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss of the present invention;
[0057] Figure 2 It is a flow chart of a balanced screening strategy when the present invention uses two DNN models to divide the data set. DETAILED DESCRIPTION
[0058] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.
[0059] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0060] Embodiment 1
[0061] like Figure 1 As shown, a noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss is provided, comprising the following steps:
[0062] S1: Two DNN models were gently trained based on the original noisy intestinal ultrasound image dataset.
[0063] Specifically, for two DNN models with exactly the same structure but inconsistent initialization parameters (model number is m = {1, 2}), the cross-entropy loss function (CE) is used to calculate the given noisy label data set. Conduct a joint w epochs of warm-up training.
[0064] The original noisy intestinal ultrasound image dataset is Where n is the number of samples in the dataset, is an ultrasound image of the intestine x i The observed label has a certain probability of not being equal to the true label of the image, and c represents the number of categories contained in the dataset.
[0065] The cross loss function L ce for:
[0066]
[0067] Among them, p m (x i ) is the DNN model numbered m based on the ultrasound image x i The output value, x i is the i-th ultrasound image, m is 1 or 2, n is the number of samples in the original noisy intestinal ultrasound image dataset, y i is the ultrasound image x i The observation label of .
[0068] S2: Use each DNN model to divide the original noisy intestinal ultrasound image dataset to obtain a labeled set and an unlabeled set corresponding to the current DNN model.
[0069] S2.1: Using the first DNN model, the original noisy intestinal ultrasound image dataset is divided to obtain a labeled set and an unlabeled set corresponding to the first DNN model.
[0070] S2.2: Using the second DNN model, the original noisy intestinal ultrasound image dataset is divided to obtain a labeled set and an unlabeled set corresponding to the second DNN model.
[0071] When using the first DNN model and the second DNN model to divide the data set, a balanced screening strategy is added to ensure that the number of samples in each category in the filtered labeled set is the same, so as to reduce the impact of the long-tail effect on the model accuracy.
[0072] In this embodiment, if Figure 2 As shown, step S2.1 and step S2.2 further include the following steps (S2-a to S2-d):
[0073] S2-a: Calculate the JSD loss value of all samples in the original noisy intestinal ultrasound image dataset under the current DNN model.
[0074] Specifically, the formula for calculating the JSD loss value of all samples in the original noisy intestinal ultrasound image dataset under the current DNN model is:
[0075]
[0076] Among them, d i m is the JSD loss value of the i-th sample under the DNN model numbered m, is the intestinal ultrasound image x i The observation label, p m (x i ) The DNN model numbered m is based on the ultrasound image x i The output value of ; JSD(·) is the Jensen–Shannon divergence function;
[0077] The specific expression of the Jensen–Shannon divergence function is:
[0078]
[0079] More specifically, p m (x i )=h(f(x i θ m );φ m ), where h(·; φ m ) and f(·;θ m ) is the feature extractor and classifier, φ m and θ m are the parameters of the DNN model numbered m.
[0080] S2-b: According to the category of each sample, the JSD loss value is divided into c subsets, and the JSD loss values in each subset are sorted in ascending order.
[0081] Specifically, within each JSD value subset, the sample indexes are sorted in ascending order according to the size of the JSD value. The smaller the JSD loss value is, the smaller the probability that it is noise is.
[0082] S2-c: Calculate the number of samples R that need to be selected in each subset.
[0083] The formula for calculating the number of samples R that need to be selected in each subset is:
[0084]
[0085] Among them, R is the number of samples that need to be selected, r is the known noise ratio of the original noisy intestinal ultrasound image dataset, c is the number of subsets divided by the JSD loss value, and n is the number of samples of the original noisy intestinal ultrasound image dataset.
[0086] S2-d: All subsets are divided into samples to obtain labeled sets and unlabeled sets corresponding to the current DNN model; the sample division of the current subset includes: judging whether the number of samples in the current subset is greater than the number of samples R that need to be selected, if so, the first R samples of the current subset are placed in the labeled set, and the rest are placed in the unlabeled set; otherwise, all samples of the current subset are placed in the labeled set.
[0087] Of course, in some other embodiments, whether it is the method of dividing the labeled set and the unlabeled set in step S2.1 or step S2.2, other existing technologies can be used for determination, for example, using GMM to estimate the current loss distribution probability of the sample, etc., and the original labels of each sample are retained in the labeled set.
[0088] In this embodiment, for the first DNN model, after the warm-up training is completed, at the beginning of each epoch, for the model numbered m=1, the Jensen–Shannon divergence loss of all training samples on the model is first calculated, and then the Jensen–Shannon divergence loss of all samples is used as the posterior probability of containing noisy labels, and the data set is divided based on the probability, and the samples with higher probabilities retain their original labels as the labeled set Remove the labels from the remaining samples and use them as the unlabeled set
[0089] For the second DNN model, after the warm-up training, at the beginning of each epoch, for the model numbered m=2, the Jensen–Shannon divergence loss of all training samples on the model is first calculated. Then, the Jensen–Shannon divergence loss of all samples is used as the posterior probability of the noisy label, and the data set is divided based on this probability. The samples with higher probabilities retain their original labels as the labeled set. Remove the labels from the remaining samples and use them as the unlabeled set
[0090] S3: Use semi-supervised learning and pseudo-label relaxed contrastive learning to perform robust training on the two DNN models respectively to update the two DNN models; when robust training the first DNN model, the labeled set and unlabeled set divided by the second DNN model are used, and when robust training the second DNN model, the labeled set and unlabeled set divided by the first DNN model are used.
[0091] Specifically, S3.1: Using the existing semi-supervised learning method, combined with pseudo-label relaxed contrastive learning technology, the two subsets divided by the second DNN model and Perform robust training on the DNN model numbered m=1.
[0092] S3.2: Using the existing semi-supervised learning method, combined with pseudo-label relaxed contrastive learning technology, the two subsets divided by the first DNN model are and Robust training is performed on the DNN model numbered m=2.
[0093] Both step S3.1 and step S3.2 include:
[0094] S3-a: Calculate the SSL loss of all samples based on the labeled set and the unlabeled set.
[0095] Specifically, the SSL loss L ssl for:
[0096]
[0097] in, is a labeled set divided by the DNN model numbered m, is the unlabeled set divided by the DNN model numbered m, L l Based on labeled collection The classification loss, L ul Based on unlabeled collection The classification loss of L reg is the regularization loss, which can make the prediction value of the classifier not biased towards a specific category; u is the preset weight coefficient.
[0098] S3-b: Use the mapping head to map the extracted high-dimensional features to the low-dimensional space to form low-dimensional features.
[0099] For the DNN model numbered m, it contains a mapping header in is the parameter of the mapping header, The feature extractor f(·;θ m ) extracts high-dimensional features f(x i ) is mapped to a low-dimensional space, making it easier to distinguish them.
[0100] S3-c: Based on the mapped low-dimensional features, the relaxed contrast loss is calculated for all samples in the labeled set and the unlabeled set.
[0101] Specifically, when calculating the relaxed contrast loss, a set of negative samples needs to be constructed.
[0102] Constructing a negative sample set includes:
[0103] For the sample x i , assuming that its prediction result vector p(x i )=[p i,1 ,p i,2 ,…,p i,c ];
[0104] According to the sample xi The prediction result vector is used to find the index corresponding to the first k components with the highest probability in the prediction result. The indexes corresponding to the first k components with the highest probability in the prediction results for all observation labels that are not equal to the given label and samples xi Corresponding index The set of samples without intersection is regarded as the negative sample set;
[0105] The negative sample set N i The formula is:
[0106]
[0107] in, is the sample in the first k components with the highest probability in the prediction result xj The corresponding index, For sample xi The corresponding index.
[0108] Specifically, the relaxed contrast loss L plr for:
[0109]
[0110] in, is the number of original noisy intestinal ultrasound image datasets, It is a sample xi After transformation enhancement, the low-dimensional features mapped by the mapping head, is a low-dimensional feature after another enhancement, wherein the another enhancement generally refers to performing different augmentation operations on the original data, such as image cropping, rotation, scaling or color adjustment, etc.; τ is a preset sharpening parameter, It is the low-dimensional feature set of the negative sample set.
[0111] For both the first DNN model and the second DNN model, PLReMix relaxed contrastive learning loss is introduced to improve the robust performance of both the first DNN model and the second DNN model.
[0112] S3-d: Combine the SSL loss and relaxed contrast loss of all samples to calculate the total loss, and perform backpropagation and model parameter update based on the total loss to update the current DNN model.
[0113] Specifically, the total loss L tot for:
[0114] L tot =L ssl +L plr
[0115] Among them, L ssl is the SSL loss, L plr is the relaxed contrast loss.
[0116] S4: Repeat steps S2-S3 until the preset total number of training rounds T is reached. tot , and output the final DNN model to classify noisy intestinal ultrasound images using the final DNN model.
[0117] Specifically, the semi-supervised learning method of the first DNN model and the second DNN model can be Co-Teaching, Co-Training or Self-Training; the two trained DNN models can be selected, integrated or fused to output a final DNN model, and the final DNN model can be used to classify noisy intestinal ultrasound images.
[0118] Embodiment 2
[0119] A noisy intestinal ultrasound image classification system based on pseudo-label relaxation contrast loss is provided, comprising:
[0120] A gentle training module, used to perform gentle training on two DNN models based on the original noisy intestinal ultrasound image dataset;
[0121] A set partitioning module is used to partition the original noisy intestinal ultrasound image dataset using each DNN model to obtain a labeled set and an unlabeled set corresponding to the current DNN model;
[0122] A robust training module, for performing robust training on two DNN models respectively by using semi-supervised learning and pseudo-label relaxed contrastive learning to update the two DNN models; when performing robust training on the first DNN model, the labeled set and the unlabeled set divided by the second DNN model are used; when performing robust training on the second DNN model, the labeled set and the unlabeled set divided by the first DNN model are used;
[0123] The module is repeatedly executed to feed back to the set partitioning module and the robust training module until a preset total number of training rounds is reached, and a final DNN model is output to classify noisy intestinal ultrasound images using the final DNN model.
[0124] For more specific processes of the above method, please refer to the corresponding contents disclosed in the aforementioned embodiments, which will not be repeated here.
[0125] Embodiment 3
[0126] A computer device is provided, comprising a processor and a memory; wherein, when the processor executes a computer program stored in the memory, the steps of the above-mentioned method for classifying noisy intestinal ultrasound images based on pseudo-label relaxation contrast loss are implemented.
[0127] For more specific processes of the above method, please refer to the corresponding contents disclosed in the aforementioned embodiments, which will not be repeated here.
[0128] Embodiment 4:
[0129] A computer-readable storage medium is provided for storing a computer program. When the computer program is executed by a processor, the steps of the above-mentioned method for classifying noisy intestinal ultrasound images based on pseudo-label relaxation contrast loss are implemented.
[0130] For more specific processes of the above method, please refer to the corresponding contents disclosed in the aforementioned embodiments, which will not be repeated here.
[0131] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system, device, and storage medium disclosed in the embodiment, since they correspond to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0132] Those skilled in the art can clearly understand that the technology in the embodiments of the present invention can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution in the embodiments of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention or some parts of the embodiments.
[0133] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.
Claims
1. A noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss, characterized in that: The following steps are involved: S1: Two DNN models are gently trained based on the original noisy intestinal ultrasound image dataset; S2: Use each DNN model to divide the original noisy intestinal ultrasound image dataset to obtain a labeled set and an unlabeled set corresponding to the current DNN model; S3: Use semi-supervised learning and pseudo-label relaxed contrastive learning to perform robust training on the two DNN models respectively to update the two DNN models; when robust training the first DNN model, the labeled set and unlabeled set divided by the second DNN model are used; when robust training the second DNN model, the labeled set and unlabeled set divided by the first DNN model are used; S4: Repeat steps S2-S3 until a preset total number of training rounds is reached, and the final DNN model is output to classify noisy intestinal ultrasound images using the final DNN model.
2. The noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss according to claim 1 is characterized in that: In step S3, the current DNN model is robustly trained using semi-supervised learning and pseudo-label relaxed contrastive learning to update the current DNN model, including: Calculate the SSL loss of all samples based on the labeled set and the unlabeled set; Use the mapping head to map the extracted high-dimensional features into a low-dimensional space to form low-dimensional features; Based on the mapped low-dimensional features, the relaxed contrast loss is calculated for all samples in the labeled set and the unlabeled set; The SSL loss and relaxed contrast loss of all samples are combined to calculate the total loss, and backpropagation and model parameter update are performed according to the total loss to update the current DNN model.
3. The noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss according to claim 2 is characterized in that: Based on the mapped low-dimensional features, for all samples in the labeled set and the unlabeled set, when calculating the relaxed contrast loss, a negative sample set needs to be constructed; The constructing of the negative sample set comprises: For the sample x i , assuming its prediction result vector; According to the sample x i The prediction result vector is used to find the index corresponding to the first k components with the highest probability in the prediction result. The indexes and samples corresponding to the top k components with the highest probability in all the observed labels that are not equal to the given labels are predicted. xi The set of samples with no intersection in the corresponding indexes is regarded as the negative sample set; The negative sample set N i The formula is: in, is the sample in the first k components with the highest probability in the prediction result xj The corresponding index, For sample xi The corresponding index; The relaxation contrast loss L plr for: in, is the number of original noisy intestinal ultrasound image datasets, It is a sample xi After transformation enhancement, the low-dimensional features mapped by the mapping head, is the low-dimensional feature after another enhancement, τ is the preset sharpening parameter, It is the low-dimensional feature set of the negative sample set.
4. The noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss according to claim 2 is characterized in that: The SSL loss L ssl for: in, is a labeled set divided by the DNN model numbered m, is the unlabeled set divided by the DNN model numbered m, L l Based on labeled collection The classification loss, L ul Based on unlabeled collection The classification loss, L reg is the regularization loss, λ u is the preset weight coefficient.
5. The noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss according to claim 1, characterized in that: In step S1, a cross loss function is used to gently train two DNN models; the two DNN models have the same structure but different initialization parameters; The cross loss function L ce for: Among them, p m (x i ) is the DNN model numbered m based on the ultrasound image x i The output value, x i is the i-th ultrasound image, m is 1 or 2, n is the number of samples in the original noisy intestinal ultrasound image dataset, y i is the ultrasound image x i The observation label of .
6. The noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss according to claim 1, characterized in that: In step S2, the original noisy intestinal ultrasound image dataset is divided to obtain a labeled set and an unlabeled set corresponding to the current DNN model, including: Calculate the JSD loss value of all samples in the original noisy intestinal ultrasound image dataset under the current DNN model; According to the category of each sample, the JSD loss value is divided into c subsets, and the JSD loss values in each subset are sorted in ascending order; Calculate the number of samples R that need to be selected in each subset; All subsets are divided into samples to obtain labeled sets and unlabeled sets corresponding to the current DNN model; The sample division of the current subset includes: judging whether the number of samples in the current subset is greater than the number of samples R that need to be selected. If so, the first R samples of the current subset are placed in the labeled set, and the rest are placed in the unlabeled set; otherwise, all samples of the current subset are placed in the labeled set.
7. The noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss according to claim 6 is characterized in that: The formula for calculating the JSD loss value of all samples in the original noisy intestinal ultrasound image dataset under the current DNN model is: Among them, d i m is the JSD loss value of the i-th sample under the DNN model numbered m, is the intestinal ultrasound image x i The observation label, p m (x i ) The DNN model numbered m is based on the ultrasound image x i The output value of ; JSD(·) is the Jensen-Shannondivergence function; The formula for calculating the number of samples R that need to be selected in each subset is: Among them, R is the number of samples that need to be selected, r is the noise ratio of the original noisy intestinal ultrasound image dataset, c is the number of subsets divided by the JSD loss value, and n is the number of samples in the original noisy intestinal ultrasound image dataset.
8. A noisy intestinal ultrasound image classification system based on pseudo-label relaxation contrast loss, characterized in that: include: A gentle training module, used to perform gentle training on two DNN models based on the original noisy intestinal ultrasound image dataset; A set partitioning module is used to partition the original noisy intestinal ultrasound image dataset using each DNN model to obtain a labeled set and an unlabeled set corresponding to the current DNN model; A robust training module, for performing robust training on two DNN models respectively by using semi-supervised learning and pseudo-label relaxed contrastive learning to update the two DNN models; when performing robust training on the first DNN model, the labeled set and the unlabeled set divided by the second DNN model are used; when performing robust training on the second DNN model, the labeled set and the unlabeled set divided by the first DNN model are used; The module is repeatedly executed to feed back to the set partitioning module and the robust training module until a preset total number of training rounds is reached, and a final DNN model is output to classify noisy intestinal ultrasound images using the final DNN model.
9. A computer device, characterized in that: It comprises a processor and a memory; wherein, when the processor executes the computer program stored in the memory, the steps of the noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss described in any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium, characterized in that: Used to store computer programs; when the computer programs are executed by the processor, the steps of the noisy intestinal ultrasound image classification method based on pseudo-label relaxation contrast loss described in any one of claims 1 to 7 are implemented.
Citation Information
Patent Citations
Image recognition model training method and device and image recognition method and device
CN112307860A
Noise label sample screening learning method and system based on oversampling strategy
CN117421657A
Noise label-containing image learning method and system based on balanced selection and comparative learning
CN118072101A
Fusion of cluster labeling algorithms by analyzing sub-clusters
US20160217201A1
