Cooperative auxiliary label correction method for medical image analysis

Abstract

Labels in a medical image data set often have noise or wrong labeling, which greatly affects the performance of a deep neural network (DNNs) in medical image analysis. In order to solve the problem, the invention provides a novel framework, namely a Co-label correction assisting network (CNLC), which is used for detecting and correcting the damaged labels at the same time. The framework is composed of two core modules: a noise detection module and a noise cleaning module. The noise detection module predicts a sample label by using a convolutional neural network (CNN), and divides samples into clean, uncertain and damaged classes through an anti-noise loss function. And the noise cleaning module corrects uncertain and damaged sample labels based on a graph convolutional network (GCN) model, and maintains a local topological relation between samples at the same time. In order to optimize the cooperation effect of the two modules, a double-layer optimization algorithm is designed, so that the label detection process and the label correction process are alternately carried out, and finally the robustness of the model and the accuracy of label correction are improved. Experimental results on three widely used medical image data sets show that the CNLC framework is obviously superior to the current advanced method in the aspect of processing the label noise problem.

Description

Technical Field

[0001] The present invention relates to the application field of deep neural networks, and more specifically, to a label correction method for processing a medical image dataset containing incorrect labels. Background Art

[0002] In medical image analysis, the performance of deep neural networks depends heavily on the quality of the dataset. However, the annotation process of medical image datasets is usually complex and expensive, and prone to incorrect labels. Incorrect labels can not only severely degrade the performance of deep learning models, but also lead to incorrect diagnosis results. Therefore, it is of great significance to develop an effective method that can automatically detect and correct incorrect labels.

[0003] Existing label correction methods can be roughly divided into two categories: robustness enhancement methods and label correction methods. Robustness enhancement methods use various techniques (such as data enhancement, loss regularization, etc.) to reduce the sensitivity of the model to incorrect labels and output a more robust model; label correction methods attempt to detect and correct incorrect labels in the dataset. However, existing methods usually have the following problems: 1) they cannot detect and correct all incorrect labels at the same time; 2) they usually ignore the relationship between samples, resulting in poor correction effects; 3) they fail to improve the robustness of the model itself. Summary of the invention

[0004] The present invention provides a collaborative assisted label correction framework, which aims to solve the problems in the above-mentioned prior art. The framework includes two main modules: a noise detector and a noise cleaner. The noise detector calculates the classification probability of samples based on a convolutional neural network (CNN) and uses an anti-overfitting loss function to divide all training samples into three categories: clean, undetermined and wrong. The noise cleaner is based on a graph convolutional network (GCN) and uses the local topological structure between samples to correct the detected wrong labels.

[0005] In addition, the present invention also designs a two-layer optimization algorithm, which alternately optimizes the noise detector and the noise cleaner in each iteration process, thereby enhancing the effect of label correction. Experimental results show that the proposed method exhibits superior performance on multiple medical image datasets. Compared with previous methods, the present invention has the following beneficial effects:

[0006] 1) Improve the accuracy of label correction, dual module design: The CNLC framework combines the Noise Detector and the Noise Cleaner. The former is responsible for identifying incorrect labels in the dataset, and the latter is responsible for label correction based on the class-based graph convolutional network (GCN). The two work together to make the label correction process more accurate and efficient;

[0007] 2) Wide applicability: CNLC can make full use of labeled and unlabeled samples, enhancing the applicability of the model on datasets with incomplete labels. Therefore, CNLC is not only applicable to datasets with complete labels, but also suitable for datasets with a large number of undetermined or unlabeled samples;

[0008] 3) Applicable to high-noise datasets. Experimental results show that CNLC performs particularly well in processing high-noise datasets. Compared with existing label correction methods (such as Co-teaching, Self-Ensemble Label Correction, etc.), CNLC can effectively process data with different noise levels, especially when there is high noise (noise rate reaches 40%), it still maintains high classification accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a workflow diagram of the network framework for assisting label correction of the present invention;

[0010] Figure 2 It is a structural diagram of the module in the present invention; DETAILED DESCRIPTION

[0011] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0012] Example

[0013] Figure 1 : is a workflow diagram of the network framework for assisting label correction of the present invention. Figure 1 As shown, the workflow diagram of the present invention assisting the label correction network framework includes the following specific steps:

[0014] S101: Model initialization:

[0015] A standard convolutional neural network (CNN) model is used to initialize the noise detection module and load the medical image dataset for training.

[0016] S102: Forward propagation and loss calculation:

[0017] For each input sample x i Perform forward propagation and output predicted probability

[0018] Based on the predicted probability Calculate the loss value of the sample with the true label In order to prevent the model from overfitting to the noisy labels, the anti-noise loss function L is used. r To replace the traditional cross entropy loss function:

[0019]

[0020] The first term is the cross entropy loss, which represents the classification loss of the model for each sample, and the second term is a time-related smoothing term used to prevent overfitting.

[0021] S103: Determine L r Is the loss small enough (less than ε? r ).

[0022] S104: Sample division:

[0023] According to the loss values ​​calculated above, all samples are sorted. All training samples are divided into three categories: Clean samples: samples with smaller loss values, the labels are inferred to be correct. Uncertain samples: samples with intermediate loss values, the model cannot accurately determine whether their labels are correct. Corrupted samples: samples with larger loss values, the labels are inferred to be corrupted or noisy.

[0024] Division criteria: The n1 samples with the highest loss values ​​are considered clean samples. A threshold is usually set (for example, the top 5% or 10% of samples) to select these samples. The n1 samples with the lowest loss values ​​are considered corrupted samples (for example, the 5% or 10% of samples with the largest loss values). Samples between clean samples and corrupted samples (the middle 90% or 80%) are considered uncertain samples.

[0025] In practical applications, the specific thresholds for dividing samples (e.g., 5% or 10%) can be adjusted according to the characteristics of different data sets. These thresholds can also be adjusted dynamically to gradually optimize the ratio of clean samples to corrupted samples as model training proceeds, ensuring more accurate label classification.

[0026] S105: Output embedded features: The noise detection module also generates embedded features for each sample, which will be used in the noise cleaning module for further analysis and correction of labels.

[0027] The task of the noise detection module is to identify the labeled noise samples present in the training set and divide the samples into clean samples, uncertain samples and corrupted samples. This process is completed by convolutional neural network (CNN).

[0028] S106: GCN model construction:

[0029] A category-based graph convolutional network (GCN) model is built for each category c. The GCN model adjusts the predicted labels through the local topological relationship (sample similarity) between samples, thereby improving the accuracy of label correction.

[0030] S107: Noise rate estimation:

[0031] Use the Gaussian mixture model (GMM) to model the loss distribution of samples in the training set and estimate the proportion of noise labels (ie, the noise rate) r:

[0032]

[0033] where v i is the judgment result of whether sample i is a noise sample, and n is the total number of samples in the training set.

[0034] S108: Sample selection:

[0035] According to the noise rate r, a portion of the damaged samples and uncertain samples are selected for further correction. The clean samples are used as positive samples to participate in the semi-supervised learning training of GCN.

[0036] S109: Semi-supervised learning:

[0037] The noise cleaning module uses binary cross entropy loss L bce and mean square error loss L mse Jointly optimize the GCN model:

[0038] L ssl =L bce +L mse

[0039] Where L bce is the binary cross entropy loss for labeled samples:

[0040]

[0041] where z i is the true label of sample i, q i is the predicted probability of the sample in the current round of training.

[0042] L mse is the mean squared error loss for unlabeled samples, representing the difference between the current prediction and the previous prediction:

[0043]

[0044] where q i is the prediction result of sample i in the current round, is the predicted value of the sample in the previous round.

[0045] S110: Determine L ssl Is it small enough (less than ε ssl );

[0046] S111: Label correction:

[0047] Use the GCN model to predict the labels of corrupted samples and uncertain samples, and use the predicted probability q ic Correct the label. The specific label correction formula is:

[0048]

[0049] where q ic is the probability that sample i is predicted to be category c, and the category corresponding to the maximum probability is selected as the corrected label.

[0050] The CNLC framework uses a two-layer optimization algorithm to alternately optimize the noise detection module and the noise cleaning module to ensure continuous improvement of model performance and label correction accuracy. The optimization process is as follows:

[0051] Upper layer optimization (noise detection module optimization): First optimize the noise detection module to minimize the anti-noise loss function L r , in order to improve the classification ability of clean samples and noise samples.,The embedded features and preliminary label division results of the noise detection module are passed to the,noise cleaning module.

[0052] Lower layer optimization (noise cleaning module optimization): Using the embedded features generated by the noise detection module, the noise cleaning module optimizes its GCN model through semi-supervised learning to minimize the semi-supervised loss L ssl , correct the noise labels. The corrected labels are fed back to the noise detection module to update the label division of the samples.

[0053] Alternating iteration: The optimization process of the above two modules is continuously alternating to form a closed-loop optimization system. In each round of iteration, the noise detection module produces more accurate label divisions, and the noise cleaning module performs better corrections on the damaged labels. This process continues until the loss function of the noise cleaning module converges, and finally an optimized model and accurate labels are obtained.

[0054] The two-level optimization objective function of the CNLC framework is:

[0055]

[0056]

[0057] Where θ is the parameter of the noise detection module, ω c are the parameters of the graph convolutional network (GCN) in the noise cleaning module; f t(x;θ) represents the output of the noise detection module in round t; A represents the prediction result of the noise cleaning module for category c in round t; c and E c are the adjacency matrix and feature matrix of category c respectively.

[0058] Optimization of the noise detection module: First, optimize the noise detection module to generate embedded features and initial label divisions (clean, corrupted, uncertain) for all samples. The optimization goal of the noise detection module is to minimize the noise-resistant loss function L r , in order to improve the model's ability to recognize noise samples.

[0059] Optimization of the noise cleaning module: The noise cleaning module uses the embedded features generated by the noise detection module to correct the damaged labels. The noise cleaning module minimizes L through semi-supervised learning. ssl , to improve the accuracy of label correction.

[0060] Alternating optimization process: In each round of iteration, the noise detection module is optimized first, followed by the noise cleaning module. The labels generated by the noise detection module are used to guide the correction process of the noise cleaning module, and the labels corrected by the noise cleaning module are fed back to the noise detection module, forming a closed-loop iterative optimization. This process is repeated until the correction results of the noise cleaning module converge, and finally an optimized model and accurate labels are obtained.

[0061] In order to better illustrate the technical effect of the present invention, the present invention is experimentally verified using specific examples. In this experimental verification, the effectiveness of the present invention was evaluated using four node classification benchmark datasets in the field of data analysis. The experiment was conducted on three widely used medical image datasets, including BreakHis: Contains 7,909 breast cancer histopathology images, of which 2,480 are benign samples and 5,429 are malignant samples. 5,537 images were randomly selected for training and 2,372 images were used for testing. ISIC: Contains 12,000 digital images of skin diseases, of which 6,000 are normal samples and 6,000 are melanoma samples. 9,600 images were randomly selected for training and the rest were used for testing. NIHCC: Contains 10,280 chest X-ray images, of which 5,110 are normal samples and 5,170 are lung disease samples. 8,574 images were randomly selected for training and the remaining samples were used for testing.

[0062] In order to evaluate the performance of the model in the presence of noisy labels, the experiments artificially introduced different degrees of noisy labels (noise rate ∈ = {0.2, 0.4}) on these datasets. This forced the model to deal with inaccurate labels, thereby testing its robustness under label noise conditions.

[0063] To verify the effectiveness of the CNLC framework, the experiment compared it with the following six popular label correction or robust training methods:

[0064] CE (Cross-Entropy): Traditional cross-entropy loss as a baseline method. Co-teaching (CT): Co-training of two models to cope with label noise. Nested Co-teaching (NCT): An improved Co-teaching method that uses a nested structure to enhance robustness. Self-Paced Resistance Learning (SPRL): A learning method that resists overfitting. Co-Correcting (CC): A method in which two models correct each other's labels. Self-Ensemble Label Correction (SELC): A label correction method based on self-ensemble.

[0065] The experiment uses the following four evaluation indicators to measure the performance of different methods:

[0066] ACC (Accuracy): The proportion of samples predicted correctly by the model. SEN (Sensitivity): The proportion of positive samples correctly identified. SPE (Specificity): The proportion of negative samples correctly identified. AUC (Area Under the ROC Curve): A comprehensive indicator reflecting the classification performance of the model.

[0067] Table 1 shows the classification results of the three data sets in this embodiment (mean ± standard deviation).

[0068]

[0069] Table 1

[0070] As shown in Table 1, it can be seen from the results in Table 1 that the CNLC of the present invention achieves the best results in both noise conditions of the three data sets. CNLC is significantly better than the existing methods under all noise conditions, especially in high noise environments. Compared with the baseline method CE, the accuracy of the CNLC framework under noisy label conditions is improved by 10%-20%. This verifies the effectiveness of the present invention.

[0071] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A collaborative auxiliary framework for label correction in deep neural networks, particularly for medical image datasets, characterized in that The framework includes: S1: a noise detection module, which processes the training samples using a convolutional neural network (CNN), predicts the class probability of the samples, and divides the training samples into clean samples, uncertain samples, and corrupted samples based on the calculated loss value; S2: Provide an anti-noise loss function for optimizing the noise detection module to prevent the model from overfitting to the noise label. The loss function combines the cross entropy loss with a time-related smoothing term and is defined by the following formula: Where b is the number of samples in the batch, p i [y i ] represents the predicted probability of the i-th sample, is the label of the sample, C is the number of categories, and λ(t) is a time-related hyperparameter. S3: The noise detection module divides the training samples into the following categories based on the loss value: clean samples: samples with smaller loss values, whose labels are considered to be correct; uncertain samples: samples with medium loss values, whose labels the model cannot accurately determine whether are correct; corrupted samples: samples with larger loss values, whose labels are speculated to be corrupted or contain noise. S4: further includes a noise cleaning module, which uses a category-based graph convolutional network (GCN) model to correct corrupted and uncertain sample labels, specifically including the following steps: S4.1: Build a graph convolutional network (GCN) model for each category to ensure that similar samples have similar labels in the same category by preserving the local topological structure between samples; S4.2: The noise cleaning module corrects the labels of uncertain and corrupted samples using a semi-supervised learning method based on the embedded features generated from the noise detection module. S4.2.1: The noise cleaning module estimates the noise ratio in the training sample through the Gaussian mixture model (GMM). The noise rate is calculated by the following formula: where v i is the judgment result of whether sample i is a noise sample, and n is the total number of samples in the training set. S4.2.2: The noise cleaning module is optimized by the following semi-supervised learning loss function: L ssl =L bce +L mse Where L bce is the binary cross entropy loss for labeled samples: where z i is the true label of sample i, q i is the predicted probability of the sample in the current round of training. L mse is the mean squared error loss for unlabeled samples, representing the difference between the current prediction and the previous prediction: where q i is the prediction result of sample i in the current round, is the predicted value of the sample in the previous round.

2. The method according to claim 1, wherein the noise cleaning module uses a graph convolutional network (GCN) model to correct the labels of uncertain samples and corrupted samples, and determines the corrected labels according to the following formula: where q ic is the probability that sample i is predicted to be category c, and the category corresponding to the maximum probability is selected as the corrected label.

3. The method according to claims 1-2 further comprises a two-layer optimization algorithm, wherein the optimization algorithm alternately optimizes the noise detection module and the noise cleaning module to improve the robustness of the model and the accuracy of label correction, specifically comprising: Upper layer optimization: Optimize the noise detection module by minimizing the anti-noise loss function L r Improve the model's ability to recognize noise labels; Lower-level optimization: optimize the noise cleaning module by minimizing the semi-supervised learning loss function L ssl Enables the model to achieve efficient label correction on both labeled and unlabeled samples.

4. The method according to claim 3, wherein the objective function of the two-level optimization algorithm is defined as follows: Where θ is the parameter of the noise detection module, ω c are the parameters of the graph convolutional network (GCN) in the noise cleaning module; f t (x;θ) represents the output of the noise detection module in round t; A represents the prediction result of the noise cleaning module for category c in round t; c and E c are the adjacency matrix and feature matrix of category c respectively.

Images