Medical image recognition model training method and system based on gradient change difference
By adopting a regularization method based on gradient variation differences in medical image recognition models, the impact of noise labels on model training is solved, and higher recognition accuracy and robustness are achieved.
Patent Information
- Application Number
- CN202411575802.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-06
- Publication Date
- 2025-02-25
AI Technical Summary
In traditional medical imaging diagnosis, noise labeling problems make it difficult for the model to effectively process the complexity and diversity of images during training, resulting in a degradation of recognition performance.
The regularization method based on the gradient change difference is adopted. By calculating the gradient change difference between two consecutive rounds during the training process, the regularization loss function of noise label learning is defined to suppress the class gradient offset and sample gradient diffusion caused by noise labels.
It effectively reduces misjudgment caused by noise labels, improves the accuracy of medical image recognition, and improves the robustness of the model to noise labels.
Smart Images
Figure CN119514629A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a medical image recognition model training method and system based on gradient variation difference. Background Art
[0002] Traditional medical imaging diagnosis relies on the doctor's experience and knowledge, but is affected by subjective factors and working conditions, resulting in uncertainty in the diagnosis results. Medical institutions have accumulated a large amount of medical imaging data covering all parts of the body, such as CT, MRI, etc., which provide an important basis for disease diagnosis. However, the diversity of massive data increases the difficulty for doctors to make accurate diagnoses. Artificial intelligence technology, especially deep learning, has achieved initial success in the diagnosis of lung diseases such as lung cancer and pneumonia by automatically extracting image features. However, the problem of noise labels in medical images is still a challenge. Traditional noise processing methods are difficult to cope with the complexity and diversity of images, so more efficient and accurate solutions are needed. In the traditional medical imaging diagnosis process, doctors often need to spend a lot of time and energy to manually annotate or correct these noise labels, which not only increases the time cost of diagnosis, but also may introduce human errors. More efficient and accurate solutions are needed to deal with different types and degrees of noise labels.
[0003] When using noisy labels to train deep neural networks, the model often "remembers" incorrect labels, resulting in reduced recognition performance. In noisy label learning methods based on neural networks, the loss value guides the optimization process of the model. Deep learning usually uses algorithms such as gradient descent to update parameters to gradually reduce the loss function. When designing the loss function, this type of method mainly considers modifying or reducing the loss of samples that may be mislabeled, focusing on designing robust loss functions, aiming to reduce the impact of noisy labels during the back-propagation process and maintain good model performance. For example, Bootstrap loss adds a perceptual loss term on the basis of traditional classification loss to reduce the impact of noisy labels on the model from the perspective of prediction consistency.
[0004] In noisy label learning, regularization plays an important role in preventing neural networks from overfitting noisy labels. From this perspective, regularization methods in noisy label learning generally regard the performance degradation caused by noisy data as an overfitting problem. By introducing regularization methods, these methods effectively suppress the overfitting of the model to noisy data and provide more robust and robust generalization capabilities. For example, the SR method makes the loss robust to noisy labels by limiting the output vector of the network. It is a simple and effective sparse regularized robust learning strategy.
[0005] Robust loss functions and regularization methods are designed to suppress the model's overfitting to noisy label samples by constraining the loss or adding regularization terms, but they do not consider the changes in the noisy label samples themselves and the essential changes in the samples as a whole during the neural network training process. Summary of the invention
[0006] The object of the present invention is to provide a medical image recognition model training method and system based on gradient change difference to solve at least one technical problem existing in the above-mentioned background technology.
[0007] In order to achieve the above object, the present invention adopts the following technical solutions:
[0008] In a first aspect, the present invention provides a medical image recognition model training method based on gradient change difference, comprising:
[0009] Acquire training data; the training data includes multiple medical images and labels of annotated images;
[0010] The neural network is trained using the training data to obtain the medical image recognition model; wherein the gradient change difference between two consecutive rounds in the training process is calculated and used to define a regularized loss function for noise label learning.
[0011] As a further limitation of the first aspect, the noise label samples affect the gradient distribution of the corresponding category samples, including category gradient offset and sample gradient diffusion; wherein the category gradient mean and variance are used to describe the difference in changes in sample gradient distribution, the gradient mean reflects the central tendency of the gradient, and the gradient variance characterizes the degree of dispersion of the gradient distribution; the gradient data is stored in a dictionary, and the dictionary maintains a data structure for each class to store statistical information about the gradient; the dictionary uses the class label as the key value, and the corresponding value is a data structure containing the mean and variance of the gradient of that class.
[0012] As a further limitation of the first aspect, at the end of each training round, the gradient mean and variance of all categories in the current round are calculated, and these statistics are stored in the gradient data dictionary to establish a dynamic gradient statistics database to capture the gradient changes of the model at different periods.
[0013] As a further limitation of the first aspect, the gradient statistic is expressed as:
[0014]
[0015] In the formula, y i is a label in the dataset D, n i is the label y in D i The number of samples, g j For label yi The gradient data corresponding to the jth sample in the sample, For label y i The mean of the gradient data of the samples, For label y i The variance of the sample gradient data.
[0016] As a further limitation of the first aspect, the mean and variance of the gradient of the round with the best recognition effect on the data set during the historical training process are used to construct the regularization loss. The purpose of this regularization loss is to constrain the current sample so that its gradient is closer to the center of the gradient distribution in the gradient data corresponding to the current label.
[0017] As a further qualification of the first aspect, the distance metric chosen to construct the loss is:
[0018]
[0019] In the formula, B is the number of samples in the current batch, d i is the distance metric of the i-th sample, G i is the gradient vector of the i-th sample, and The label y corresponding to sample i i The mean and variance of the gradient statistics of ∈ are very small constants used to avoid division by zero.
[0020] In a second aspect, the present invention provides a medical image recognition model training system based on gradient change difference, comprising:
[0021] An acquisition module, used to acquire training data; the training data includes a plurality of medical images and labels of annotated images;
[0022] A training module is used to train a neural network using training data to obtain the medical image recognition model; wherein the gradient change difference between two consecutive rounds in the training process is calculated and used to define a regularized loss function for noise label learning.
[0023] In a third aspect, the present invention provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the medical image recognition model training method based on gradient change differences as described in the first aspect is implemented.
[0024] In a fourth aspect, the present invention provides a computer device comprising a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the medical image recognition model training method based on gradient change differences as described in the first aspect.
[0025] In a fifth aspect, the present invention provides an electronic device comprising: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes instructions for implementing the medical image recognition model training method based on gradient change differences as described in the first aspect.
[0026] The beneficial effects of the present invention are as follows: the method focuses on how to reduce the influence of noise labels on neural network classification in the problem of disease image recognition. The method does not rely on any specific neural network framework and can be easily added to a common classification network framework. The parameter scale of the model is not increased, so the training cost and inference cost remain unchanged, ensuring that the model performance is improved without sacrificing efficiency. Theoretically, it has good interpretability, can effectively reduce false positives caused by noise labels, reduce misdiagnosis or missed diagnosis in disease auxiliary diagnosis, and thus improve the accuracy of image recognition.
[0027] Additional advantages of the present invention will be more clearly given in the following description or learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0029] Figure 1 This is a gradient calculation flow chart described in an embodiment of the present invention. DETAILED DESCRIPTION
[0030] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below by the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.
[0031] It should be understood by those skilled in the art that unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this invention belongs.
[0032] It should also be understood that terms, such as those defined in commonly used dictionaries, should be understood to have a meaning consistent with that in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless as defined herein.
[0033] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements and / or groups thereof.
[0034] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. Different embodiments or examples described in this specification and features of different embodiments or examples may be combined and combined by those skilled in the art without contradiction.
[0035] To facilitate understanding of the present invention, the present invention is further explained below with reference to specific embodiments in conjunction with the accompanying drawings, and the specific embodiments do not constitute a limitation on the embodiments of the present invention.
[0036] Those skilled in the art should understand that the drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily necessary for implementing the present invention.
[0037] Based on the deep learning model, the present invention proposes a regularization method based on gradient change difference to suppress the class gradient offset and sample gradient diffusion caused by noise labels. This method uses the basic neural network model to extract features from images, but the difference is that the sample gradient is continuously corrected by the class average gradient and gradient variance to constrain the neural network optimization process. The present invention aims to reveal the phenomenon of class gradient drift and sample gradient diffusion in the process of noise label learning, and designs a new noise label learning method for lung disease identification. This method suppresses the class gradient offset and sample gradient diffusion caused by noise labels without increasing the complexity of the model. The additional cost of this method is only the gradient calculation and storage of class gradient statistics in the previous training stage. While maintaining efficient training, it significantly improves the performance of the model in dealing with image noise label problems.
[0038] Example 1
[0039] In this embodiment 1, a medical image recognition model training system based on gradient change difference is first provided, including: an acquisition module for acquiring training data; the training data includes multiple medical images and labels of annotated images. A training module is used to train a neural network using the training data to obtain the medical image recognition model; wherein the gradient change difference between two consecutive rounds in the training process is calculated, and the regularized loss function of noise label learning is defined based on this.
[0040] In this embodiment, the above-mentioned system is used to implement a medical image recognition model training method based on gradient change differences, including: obtaining training data; the training data includes multiple medical images and labels of annotated images; using the training data to train a neural network to obtain the medical image recognition model; wherein, the gradient change difference between two consecutive rounds during the training process is calculated, and this is used to define a regularized loss function for noise label learning.
[0041] Specifically, in the above-mentioned model training of this embodiment, for the noise label learning for lung disease image recognition, a regularization method based on gradient change difference is designed, and a processing mechanism for distinguishing noise labels from correct labels is introduced to effectively distinguish normal labels and noise labels in lung images. Specifically, the gradient change difference between two consecutive rounds in the training process is calculated, and this is used to define the regularization loss function of noise label learning. The specific solution process is as follows: Figure 1 shown.
[0042] The following introduces the gradient statistics and gradient change difference constraint regularization loss function L reg .
[0043] In this example, the goal is to solve the problem of noisy labels in image recognition by using the distribution difference of sample gradients. First, the image is input into the neural network to extract features, and then the gradient statistics are calculated: the gradient class mean and the gradient variance. The gradient is calculated using the cross entropy loss function as an example, focusing on the changes in the last fully connected layer of the neural network. Take the image x i The gradient size For example, the fully connected layer (FC) of the neural network inputs a feature vector in the network, denoted as h = [h1,…,h L ] T , the weight matrix of the FC layer is w=[w1,…,w C ] T ∈R C×f The output of the FC layer after the softmax function is y i =[y1,…,y C] is the true label. Therefore, the gradient It is expressed as:
[0044]
[0045] In the formula, y i For sample x i The real label, is the output of the FC layer after being processed by the softmax function, h is an input feature vector, and C is the number of sample categories. Noise label samples affect the gradient distribution of samples of this category, including category gradient offset and sample gradient diffusion. In this embodiment, the category gradient mean and variance are used to describe the difference in changes in sample gradient distribution. The gradient mean reflects the central tendency of the gradient, while the gradient variance characterizes the degree of dispersion of the gradient distribution. In the specific calculation, the gradient data is stored in a dictionary, and the dictionary maintains a data structure for each class to store statistical information about the gradient. This dictionary uses the class label as the key value, and the corresponding value is a data structure containing the mean and variance of the gradient of this class. Specifically, the gradient statistic is expressed as:
[0046]
[0047] In the formula, y i is a label in the dataset D, n i is the label y in D i The number of samples, g j For label y i The gradient data corresponding to the jth sample in the sample, For label y i The mean of the gradient data of the samples, is its variance. This means that each key in the gradient data dictionary is a label y i , the corresponding value is a tuple containing the mean and variance. At the end of each training round, the mean and variance of the gradients of all categories in the current round are calculated, and these statistics are stored in the gradient data dictionary. In this way, we have established a dynamic gradient statistics database that can capture the gradient changes of the model at different times. The gradient distribution will change during the training process, so in the training, we choose to use the gradient data of the current best round to update the dictionary, so that the gradient information learned by the model can be dynamically tracked, avoiding the influence of instability during the training process.
[0048] During training, the mean and variance of the gradient of the best round of recognition on the dataset during historical training are used to construct the regularization loss. The purpose of this regularization loss is to constrain the current sample so that its gradient is closer to the center of the gradient distribution in the gradient data corresponding to the current label. By constraining the gradient of the current sample to the center of the gradient corresponding to the label, the sensitivity of the model to noisy samples can be reduced. Similarly, by constraining the gradient of the current sample to be close to the center of the gradient distribution of other samples with the same label, some of the effects of label noise can be offset, and the feature information of clean samples can be better learned. We choose to construct the loss using a distance metric, calculated as follows:
[0049]
[0050] In the formula, B is the number of samples in the current batch, d i is the distance metric of the i-th sample, G i is the gradient vector of the i-th sample, and The label y corresponding to sample i i The mean and variance of the gradient statistics of ∈ are very small constants to avoid division by zero. The meaning of the loss function is to use a distance metric to measure the gap between the center of the class gradient data corresponding to the current sample and the label. Finally, the loss function can be used to regularize the average gradient difference in a category-by-category manner during network training to suppress the influence of noise labels in lung disease images.
[0051] Example 2
[0052] This embodiment 2 provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the medical image recognition model training method based on gradient change difference as described above is implemented. The method includes:
[0053] Acquire training data; the training data includes multiple medical images and labels of annotated images;
[0054] The neural network is trained using the training data to obtain the medical image recognition model; wherein the gradient change difference between two consecutive rounds in the training process is calculated and used to define a regularized loss function for noise label learning.
[0055] Example 3
[0056] This embodiment 3 provides a computer device, including a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, the processor calls the program instructions to execute the above-mentioned medical image recognition model training method based on gradient change difference, the method comprising:
[0057] Acquire training data; the training data includes multiple medical images and labels of annotated images;
[0058] The neural network is trained using the training data to obtain the medical image recognition model; wherein the gradient change difference between two consecutive rounds in the training process is calculated and used to define a regularized loss function for noise label learning.
[0059] Example 4
[0060] This embodiment 4 provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the above-mentioned medical image recognition model training method based on gradient change difference, the method comprising:
[0061] Acquire training data; the training data includes multiple medical images and labels of annotated images;
[0062] The neural network is trained using the training data to obtain the medical image recognition model; wherein the gradient change difference between two consecutive rounds in the training process is calculated and used to define a regularized loss function for noise label learning.
[0063] In summary, the noise label learning method for disease identification described in the embodiment of the present invention targets the common problem of noise labels, fully explores the essential changes of the model caused by noise labels, and designs a special processing mechanism for the expression of change differences and regularization methods, thereby enhancing the robustness of the model to noise labels. By suppressing the class gradient drift and sample gradient diffusion caused by noise labels, this method effectively reduces the misjudgment caused by noise labels, improves the accuracy of disease-assisted diagnosis, and provides an innovative idea for solving the problem of image noise labels. It is a strategy that has not been used in previous similar studies.
[0064] Compared with other lung disease image noise label processing technologies, the present invention focuses on how to reduce the impact of noise labels on neural network classification in disease image recognition problems. The invented method does not rely on any specific neural network framework and can be easily added to the common classification network framework. It can be effectively integrated with existing technologies in noise label processing, such as robust loss functions and regularization techniques. The parameter scale of the model is not increased, so the training cost and inference cost remain unchanged, ensuring that the model performance is improved without sacrificing efficiency. In theory, it has good interpretability, can effectively reduce false positives caused by noise labels, reduce misdiagnosis or missed diagnosis in disease-assisted diagnosis, and thus improve the accuracy of lung disease image recognition.
[0065] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0066] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0067] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0068] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0069] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative work on the basis of the technical solution disclosed in the present invention should be included in the scope of protection of the present invention.
Claims
1. A medical image recognition model training method based on gradient change difference, characterized in that: include: Get training data; The training data includes a plurality of medical images and labels of annotated images; The neural network is trained using the training data to obtain the medical image recognition model; wherein the gradient change difference between two consecutive rounds in the training process is calculated and used to define a regularized loss function for noise label learning.
2. The medical image recognition model training method based on gradient change difference according to claim 1, characterized in that: Noise label samples affect the gradient distribution of corresponding category samples, including category gradient offset and sample gradient diffusion; wherein, the category gradient mean and variance are used to describe the difference in sample gradient distribution changes, the gradient mean reflects the central tendency of the gradient, and the gradient variance characterizes the degree of dispersion of the gradient distribution; the gradient data is stored in a dictionary, and the dictionary maintains a data structure for each class to store statistical information about the gradient; the dictionary uses the class label as the key value, and the corresponding value is a data structure containing the mean and variance of the gradient of this class.
3. The medical image recognition model training method based on gradient change difference according to claim 2, characterized in that: At the end of each training round, the gradient mean and variance of all categories in the current round are calculated, and these statistics are stored in the gradient data dictionary to build a dynamic gradient statistics database to capture the gradient changes of the model at different periods.
4. The medical image recognition model training method based on gradient change difference according to claim 3 is characterized in that: The gradient statistic is expressed as: In the formula, y i is a label in the dataset D, n i is the label y in D i The number of samples, g j For label y i The gradient data corresponding to the jth sample in the sample, For label y i The mean of the gradient data of the samples, For label y i The variance of the sample gradient data.
5. The medical image recognition model training method based on gradient change difference according to claim 4 is characterized in that: The regularization loss is constructed using the mean and variance of the gradient of the round with the best recognition effect on the data set during the historical training process. The purpose of this regularization loss is to constrain the current sample so that its gradient is closer to the center of the gradient distribution in the gradient data corresponding to the current label.
6. The medical image recognition model training method based on gradient change difference according to claim 5, characterized in that: The distance metric chosen to construct the loss is: In the formula, B is the number of samples in the current batch, d i is the distance metric of the i-th sample, G i is the gradient vector of the i-th sample, and The label y corresponding to sample i i The mean and variance of the gradient statistics of ∈ are very small constants used to avoid division by zero.
7. A medical image recognition model training system based on gradient change difference, characterized in that: include: An acquisition module is used to acquire training data; The training data includes a plurality of medical images and labels of annotated images; A training module is used to train a neural network using training data to obtain the medical image recognition model; wherein the gradient change difference between two consecutive rounds in the training process is calculated and used to define a regularized loss function for noise label learning.
8. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by the processor, the medical image recognition model training method based on gradient change difference as described in any one of claims 1-6 is implemented.
9. A computer device, characterized in that: It includes a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the medical image recognition model training method based on gradient change difference as described in any one of claims 1-6.
10. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes instructions for implementing the medical image recognition model training method based on gradient change differences as described in any one of claims 1-6.