Schizophrenia classification method and device based on multi-center sMRI and domain self-adaption
By acquiring sMRI data from multiple medical centers, data augmentation and preprocessing are performed, data is migrated to the target domain using a generative adversarial network, and the generated target domain training set is combined to train the schizophrenia classification model, which solves the problem of insufficient data sample size and lack of unified standards in the existing technology, and significantly improves classification accuracy.
Patent Information
- Application Number
- CN202510159953.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-13
- Publication Date
- 2025-05-16
AI Technical Summary
In the prior art, the sample size of the human brain structural magnetic resonance image data is not sufficient to meet deep learning. Direct reference to the classification model trained by other medical centers is not ideal in the current medical center image data. Moreover, due to the differences in the acquisition instruments, the data does not have unified standards and cannot be compared horizontally.
By obtaining sMRI data from multiple medical centers, data augmentation and preprocessing are performed separately, augmenting source domain and target domain image data are generated, and the source domain data is migrated to the target domain using the generative adversarial network, and the generated target domain training set is combined to train the schizophrenia classification model.
The classification accuracy of the classification model for target domain data is significantly improved, and the problem of insufficient data sample size and lack of unified data standards is solved, and effective comparison of cross-domain data and efficient training of models is realized.
Smart Images
Figure CN120014360A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of computer image processing, and in particular to a schizophrenia classification method and device based on multi-center sMRI and domain adaptation. Background Art
[0002] In the past, the diagnosis of mental illness was usually based on the subjective judgment of doctors based on scale evaluation, which was relatively lacking in objectivity. In recent years, schizophrenia has a high incidence and insufficient detection accuracy, which has led to an increasing demand for objective diagnostic methods for schizophrenia in clinical practice. Thanks to the development of computer science, auxiliary diagnostic methods based on neuroimaging and computer science are objective and have become an important auxiliary means for the diagnosis of schizophrenia.
[0003] Magnetic resonance imaging (MRI) technology provides a tool for non-invasive objective measurement of the human brain. Structural magnetic resonance imaging (sMRI) obtains information by measuring the distribution of protons through a magnetic resonance instrument, and has good soft tissue resolution and multi-directional arbitrary slice capabilities. Therefore, the human brain image data obtained by structural magnetic resonance imaging technology can be used as the basis for the diagnosis of schizophrenia.
[0004] However, problems were also found in the application. First, due to the different brands of acquisition instruments used in structural magnetic resonance imaging technology, the collected sMRI data often do not have a unified standard and cannot be compared horizontally; second, the amount of sample data from a single medical center is small, which is difficult to meet the data volume required for deep learning, resulting in unsatisfactory performance and classification accuracy of the classification model; third, directly referencing the classification model that has been trained by other medical centers and performing classification on the image data of the current medical center will result in unsatisfactory classification accuracy. Summary of the invention
[0005] Purpose of the invention: The present invention provides a schizophrenia classification method and device based on multi-center sMRI and domain adaptation, aiming to solve the problem in the prior art that the sample size of human brain structural magnetic resonance image data is insufficient to meet deep learning, and the classification accuracy is not ideal when directly citing the classification models that have been trained by other medical centers; further, due to the differences in acquisition instruments, human brain structural magnetic resonance image data do not have a unified standard and cannot be compared horizontally.
[0006] Technical solution: The present invention provides a schizophrenia classification method based on multi-center sMRI and domain adaptation, including: obtaining a human brain structural magnetic resonance image of a first medical center as original source domain image data; obtaining a human brain structural magnetic resonance image of a second medical center as original target domain image data; enhancing the original source domain image data and the original target domain image data respectively to obtain enhanced source domain image data and enhanced target domain image data; migrating the enhanced source domain image data to a target domain to generate secondary target domain image data, combining the secondary target domain image data and the enhanced target domain image data to obtain a target domain training set, and training a schizophrenia classification model; inputting the human brain structural magnetic resonance image to be detected from the second medical center into the schizophrenia classification model to determine the type of schizophrenia.
[0007] Specifically, the original source domain image data and the original target domain image data both include samples with schizophrenia and healthy samples.
[0008] Specifically, the original image data is biased and corrected, the skull and non-brain tissues are removed, spatial registration is performed, the image is aligned to the standard anatomical structure, cropped to the brain region, and the image size is adjusted to a standard size; image intensity normalization is performed; the image data is interpolated, and the interpolated image data is used as enhanced source domain image data and enhanced target domain image data.
[0009] Specifically, the image data is interpolated using the MultiMix method.
[0010] Specifically, the image intensity is normalized, and then the process also includes: extracting brain region features from the image data.
[0011] Specifically, the schizophrenia classification model has been trained using enhanced source domain image data before being trained using the target domain training set.
[0012] Specifically, a generative adversarial network is used to migrate the enhanced source domain image data to the target domain to obtain secondary target domain image data.
[0013] Specifically, a first proportion of image data is extracted from the secondary target domain image data, a second proportion of image data is extracted from the enhanced target domain image data, and the extracted image data are combined to obtain a target domain training set.
[0014] Specifically, a third proportion of image data is extracted from the enhanced target domain image data and used as a test set to test the performance of the schizophrenia classification model.
[0015] The present invention also provides a schizophrenia classification device based on multi-center sMRI and domain adaptation, comprising: a data acquisition unit, a data preprocessing unit, a data migration unit and an application unit, wherein: the data acquisition unit is used to acquire a human brain structural magnetic resonance image of a first medical center as original source domain image data; acquire a human brain structural magnetic resonance image of a second medical center as original target domain image data; the data preprocessing unit is used to enhance the original source domain image data and the original target domain image data, respectively, to obtain enhanced source domain image data and enhanced target domain image data; the data migration unit is used to migrate the enhanced source domain image data to the target domain, generate secondary target domain image data, combine the secondary target domain image data and the enhanced target domain image data to obtain a target domain training set, and train a schizophrenia classification model; the application unit is used to input the human brain structural magnetic resonance image to be detected from the second medical center into the schizophrenia classification model to determine the type of schizophrenia.
[0016] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: it references sample data from multiple centers to meet the number of samples required for deep learning of the classification model, migrates the source domain data to the target domain, and conducts secondary training on the model to improve the classification accuracy of the classification model for the target domain (current medical center) data; further, it crops and normalizes the structural magnetic resonance image data of the human brain to establish a unified standard. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 A schematic diagram of the process of the schizophrenia classification method based on multi-center sMRI and domain adaptation provided by the present invention;
[0018] Figure 2 A schematic diagram of the steps of preprocessing the structural magnetic resonance image of the human brain provided by the present invention. DETAILED DESCRIPTION
[0019] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0020] See also Figure 1 , which is a flowchart of the schizophrenia classification method based on multi-center sMRI and domain adaptation provided by the present invention.
[0021] In the embodiment of the present invention, a structural magnetic resonance image of a human brain of a first medical center is obtained as original source domain image data; and a structural magnetic resonance image of a human brain of a second medical center is obtained as original target domain image data.
[0022] In the embodiment of the present invention, both the original source domain image data and the original target domain image data include samples with schizophrenia and healthy samples.
[0023] In a specific implementation, the source domain data used in the embodiment of the present invention can be collected and provided by the first medical center (certain brain hospital), and the target domain data can be collected and provided by the second medical center (certain mental health center). The data set composition can be that the original model training set, test set, and validation set are all from the first medical center (certain brain hospital), and the deep domain adaptive network model training set, test set, and validation set are from the new target domain samples generated by migrating from the source domain to the target domain and the second medical center (certain mental health center). Each data set is composed of data sets of schizophrenia and normal people. The source domain data has a total of 286 people's nuclear magnetic resonance images, including 182 people in the first schizophrenia group and 104 people in the normal control group. The target domain data has a total of 361 people's nuclear magnetic resonance images, including 197 people in the first schizophrenia group and 164 people in the normal control group. All first-episode schizophrenia patients did not take medication, and they had no history of other neurological diseases or serious drug diseases.
[0024] See also Figure 2 , which is a schematic diagram of the steps of preprocessing the structural magnetic resonance image of the human brain provided by the present invention.
[0025] In the embodiment of the present invention, the original source domain image data and the original target domain image data are enhanced respectively to obtain enhanced source domain image data and enhanced target domain image data.
[0026] In an embodiment of the present invention, bias field correction is performed on the original image data (original source domain image data and original target domain image data), the skull and non-brain tissues are removed, spatial registration is performed, alignment is made to the standard anatomical structure, cropping is performed to the brain region, and the image size is adjusted to a standard size; image intensity normalization is performed; the image data is interpolated, and the interpolated image data is used as enhanced source domain image data and enhanced target domain image data.
[0027] In the specific implementation, bias field correction is performed on all original MRI images to remove brightness unevenness caused by scanner characteristics or deviations in the scanning process, unify image brightness, and provide standardized data for classification tasks.
[0028] In a specific implementation, the skull and non-brain tissues in the MRI image are removed, the pure brain image is extracted, and the brain tissue image for research is obtained.
[0029] In a specific implementation, the stripped brain tissue images are spatially registered and aligned to the standard anatomical structure in preparation for subsequent voxel analysis.
[0030] In the specific implementation, the processed MRI image is cropped to the brain region and the size can be adjusted to 128×128×100mm 3, generate standardized three-dimensional data for subsequent interpolation and training.
[0031] In the specific implementation, in order to unify the standards of all original image data, the image intensity is normalized. The following formula is used for normalization:
[0032] I'=(II min ) / (I max -I min ),
[0033] Where, I' represents the normalized image intensity. I represents the grayscale value of the original image, and I min Represents the minimum grayscale value of all MRI image data, I max Represents the maximum grayscale value of all MRI image data.
[0034] In the specific implementation, the image intensity is normalized to the range of [0,1]. Image normalization is a key step in data preprocessing. By mapping the image data to a unified range and establishing a unified standard, it is convenient for effective horizontal comparison of image data, overcoming the problem that there is no unified standard for human brain structural magnetic resonance image data and horizontal comparison cannot be performed due to differences in parameters of acquisition instruments. At the same time, it also reduces the computational complexity of model training and reduces the impact of sparse high-intensity noise, thereby improving the convergence effect of the model.
[0035] In the embodiment of the present invention, the MultiMix method is used to interpolate the image data.
[0036] The core formula of the Multi-Mix method is as follows:
[0037]
[0038] For a pair of input labeled samples (x, y) and (x′, y′), MultiMix generates K interpolated samples, where: k is the interpolation coefficient randomly sampled from the Beta distribution, satisfying 0<λ1<λ2<...<λ k <1, ensuring that the interpolation results are orderly and continuous in the input space and output space. h(x) and h(x′) represent the feature representation of samples x and x′ at a certain layer of the network respectively. For input layer interpolation (Input Mixup), h(x) is the original input x; and for hidden layer interpolation (ManifoldMixup), h(x) is the output of a hidden layer of the deep network. Represents new label samples generated by mixed interpolation.
[0039] In specific implementation, due to the scarcity of medical image data, the data set usually has the problem of small sample size and uneven distribution. The data augmentation method can effectively expand the data set, thereby significantly improving the training performance of the model. The preprocessed sMRI is enhanced using the MultiMix method. Multi-Mix is an extension of the standard Mixup data augmentation technology. It generates multiple interpolation samples from a pair of training samples instead of just one interpolation sample. This method improves the generalization ability and robustness of the model by providing more intermediate representations between data points.
[0040] In a specific implementation, when the 286 sMRI data of the source domain are subjected to the above-mentioned data preprocessing process and the MultiMix method is used to perform data enhancement on the 286 sMRI data, K=2 interpolation samples are generated for each pair of samples, and a total of 40,845 enhanced samples are generated. When the 361 sMRI data of the target domain are subjected to the above-mentioned data preprocessing process and the MultiMix method is used to perform data enhancement on the 361 sMRI data, K=2 interpolation samples are generated for each pair of samples, and a total of 130,200 enhanced samples are generated. After obtaining the enhanced samples, the image data originally obtained from the first medical center and the second medical center are no longer used in the enhanced source domain image data and the enhanced target domain image data, that is, the original source domain image data and the original target domain image data are not included in the enhanced source domain image data and the enhanced target domain image data, and the original image data no longer participates in the training process.
[0041] In the embodiment of the present invention, after the image data is normalized, brain region features (such as gray matter, white matter, and cerebrospinal fluid) are extracted from the image data.
[0042] In the specific implementation, after basic correction, registration and normalization of the acquired sMRI image data, the volume features of the brain region are extracted, such as gray matter, white matter and cerebrospinal fluid, to obtain the significant areas of the human brain. The significant areas can be used as image data for subsequent image data interpolation, and enhanced source domain image data and enhanced target domain image data are obtained after interpolation. By extracting brain region volume features, the characteristics of schizophrenia can be better characterized and the classification accuracy can be improved.
[0043] In an embodiment of the present invention, enhanced source domain image data is migrated to a target domain to generate secondary target domain image data, and the secondary target domain image data and the enhanced target domain image data are combined to obtain a target domain training set to train a schizophrenia classification model.
[0044] In the embodiment of the present invention, the schizophrenia classification model has been trained using enhanced source domain image data before being trained using the target domain training set.
[0045] In an embodiment of the present invention, a generative adversarial network is used to migrate enhanced source domain image data to a target domain to obtain secondary target domain image data.
[0046] In the specific implementation, the schizophrenia classification model includes an R3D convolutional neural network, each of which contains 18 layers, including a convolutional layer (Conv), a residual block (Residual Blocks), a mixed attention module (Mixed Attention R3D, MAR3D), a global average pooling layer (Global Average Pooling), a fully connected layer (Fully Connected Layer), and a Softmax layer, and initializes the number of hidden layer neurons, the size of the convolution kernel, the number of iterations, and the learning rate.
[0047] In the specific implementation, the input sMRI data is initially processed to extract basic features and key information of the data. Then, higher-level features are gradually extracted, and the stability of network training is improved through the residual connection mechanism, and the gradient problem in the deep network is effectively alleviated. In the feature learning process, the network adaptively focuses on important areas in the input data through the attention optimization module, further strengthening the capture of key features and improving the recognition ability of schizophrenia-related pathological features. Subsequently, the extracted high-dimensional features are compressed and integrated through the dimensionality reduction module to provide refined feature representation for subsequent classification processing.
[0048] In the specific implementation, the preprocessed enhanced source domain image data is input into the schizophrenia classification model for training, verification, and testing, and finally the optimal convolutional neural network model is obtained and the model hyperparameters are saved for subsequent secondary training.
[0049] In the specific implementation, a target domain deep adaptive network classification model GR3D (Generative Channel-Based R3D, GCBR3D) is established to migrate the source domain image data to the target domain. This model introduces a domain adaptive generative adversarial network (GAN) based on the R3D model of the source domain. The generative adversarial network consists of a feature generator and a feature discriminator, and initializes the number of hidden layer neurons, the size of the convolution kernel, the number of iterations and the learning rate.
[0050] In the specific implementation, the target domain generates an adversarial network model, which consists of a feature generator and a feature discriminator. The feature generator is used to generate features similar to the target domain feature distribution from the source domain data, and the mapping of the source domain and target domain features is achieved through the generation process. The feature discriminator is used to distinguish the generated features from the real target domain features to form an adversarial mechanism. Through the joint training of the generator and the discriminator, the generator is gradually optimized and can generate features that are closer to the target domain distribution, thereby achieving feature alignment between the source domain and the target domain. This domain alignment process effectively reduces the impact of differences in data distribution in different centers on model performance, and significantly improves the classification accuracy and generalization ability of the model on target domain data. Finally, the aligned features are input into the schizophrenia classification model for further classification to ensure the robustness and accuracy of the cross-domain classification task.
[0051] In an embodiment of the present invention, a first proportion of image data is extracted from the secondary target domain image data, a second proportion of image data is extracted from the enhanced target domain image data, and the extracted image data are combined to obtain a target domain training set.
[0052] In the embodiment of the present invention, image data of a third proportion is extracted from the enhanced target domain image data and used as a test set to test the performance of the schizophrenia classification model.
[0053] In a specific implementation, in an embodiment of the present invention, 70% of the generated target domain data (secondary target domain image data) and 30% of the real target domain data (enhanced target domain image data) are combined to form a training set for training a schizophrenia classification model, and the remaining 30% of the generated target domain data and 20% of the real target domain data are combined to form a validation set. Finally, the remaining 50% of the real target domain data is used as a test set.
[0054] In a specific implementation, the four model evaluation indicators of Accuracy, Precision, Recall and F1Score can be used to evaluate the performance of the schizophrenia classification model used.
[0055] In the specific implementation, the experimental results show that compared with the single medical center data method, the multi-medical center-based method provided by the present invention significantly improves the accuracy and generalization ability of cross-domain sMRI image classification. By integrating multi-center data and combining the domain adversarial generative network (GAN) technology, the impact of data distribution differences on model performance is overcome. Machine learning methods are usually limited to an accuracy of less than 90% in the task of schizophrenia classification. For example, the classification accuracy of traditional methods such as the K nearest neighbor algorithm is 85.58%, and the accuracy of DT and RBF-SVM is less than 80%. In contrast, the present invention combines convolutional neural networks (CNN) with domain adaptation methods to significantly improve classification performance. Based on a large number of samples, the R3D convolutional neural network is directly used for classification, and the accuracy can reach 92.59%, and after combining the domain adversarial network (GAN) optimization, the generalization ability can be further enhanced.
[0056] The present invention also provides a schizophrenia classification device based on multi-center sMRI and domain adaptation, comprising: a data acquisition unit, a data preprocessing unit, a data migration unit and an application unit, wherein: the data acquisition unit is used to acquire a human brain structural magnetic resonance image of a first medical center as original source domain image data; acquire a human brain structural magnetic resonance image of a second medical center as original target domain image data; the data preprocessing unit is used to enhance the original source domain image data and the original target domain image data, respectively, to obtain enhanced source domain image data and enhanced target domain image data; the data migration unit is used to migrate the enhanced source domain image data to the target domain, generate secondary target domain image data, combine the secondary target domain image data and the enhanced target domain image data to obtain a target domain training set, and train a schizophrenia classification model; the application unit is used to input the human brain structural magnetic resonance image to be detected from the second medical center into the schizophrenia classification model to determine the type of schizophrenia.
[0057] The schizophrenia classification device based on multi-center sMRI and domain adaptation provided by the present invention, wherein the execution unit for executing methods, steps or functions, the methods, steps or functions executed by it, can refer to the schizophrenia classification method based on multi-center sMRI and domain adaptation provided by the present invention.
Claims
1. A schizophrenia classification method based on multi-center sMRI and domain adaptation, characterized in that: include: Acquire the structural magnetic resonance images of the human brain from the First Medical Center as original source domain image data; acquiring a human brain structural magnetic resonance image from a second medical center as original target domain image data; The original source domain image data and the original target domain image data are enhanced respectively to obtain enhanced source domain image data and enhanced target domain image data; Migrating the enhanced source domain image data to the target domain to generate secondary target domain image data, combining the secondary target domain image data with the enhanced target domain image data to obtain a target domain training set, and training a schizophrenia classification model; The human brain structural magnetic resonance images to be tested from the second medical center are input into the schizophrenia classification model to determine the type of schizophrenia.
2. The schizophrenia classification method based on multi-center sMRI and domain adaptation according to claim 1, characterized in that: The original source domain image data and the original target domain image data both include samples with schizophrenia and healthy samples.
3. The schizophrenia classification method based on multi-center sMRI and domain adaptation according to claim 1, characterized in that: The enhancing the original source domain image data and the original target domain image data respectively includes: The raw image data was corrected for bias fields, the skull and non-brain tissues were removed, spatial registration was performed, the images were aligned to standard anatomical structures, cropped to the brain region, and the image size was adjusted to a standard size; Perform normalization of image intensity; The image data is interpolated, and the interpolated image data is used as enhanced source domain image data and enhanced target domain image data, respectively.
4. The schizophrenia classification method based on multi-center sMRI and domain adaptation according to claim 3, characterized in that: The interpolating the image data comprises: Use the MultiMix method to interpolate the image data.
5. The schizophrenia classification method based on multi-center sMRI and domain adaptation according to claim 3, characterized in that: The method further comprises performing normalization of image intensity, and then: Extract brain region features from image data.
6. The schizophrenia classification method based on multi-center sMRI and domain adaptation according to claim 3, characterized in that: The schizophrenia classification model has been trained using enhanced source domain image data before being trained using the target domain training set.
7. The schizophrenia classification method based on multi-center sMRI and domain adaptation according to claim 3, characterized in that: The step of migrating enhanced source domain image data to a target domain comprises: A generative adversarial network is used to migrate the enhanced source domain image data to the target domain to obtain secondary target domain image data.
8. The schizophrenia classification method based on multi-center sMRI and domain adaptation according to claim 7, characterized in that: The step of combining the secondary target domain image data and the enhanced target domain image data to obtain a target domain training set includes: Image data of a first proportion is extracted from the secondary target domain image data, image data of a second proportion is extracted from the enhanced target domain image data, and the extracted image data are combined to obtain a target domain training set.
9. The schizophrenia classification method based on multi-center sMRI and domain adaptation according to claim 8, characterized in that: The schizophrenia classification model is trained, and then includes: A third scale of image data is extracted from the enhanced target domain image data and used as a test set to test the performance of the schizophrenia classification model.
10. A schizophrenia classification device based on multi-center sMRI and domain adaptation, characterized in that: include: Data acquisition unit, data preprocessing unit, data migration unit and application unit, wherein: The data acquisition unit is used to acquire a structural magnetic resonance image of a human brain from a first medical center as original source domain image data; and acquire a structural magnetic resonance image of a human brain from a second medical center as original target domain image data; The data preprocessing unit is used to enhance the original source domain image data and the original target domain image data respectively to obtain enhanced source domain image data and enhanced target domain image data; The data migration unit is used to migrate the enhanced source domain image data to the target domain, generate secondary target domain image data, combine the secondary target domain image data and the enhanced target domain image data to obtain a target domain training set, and train the schizophrenia classification model; The application unit is used to input the human brain structural magnetic resonance image to be detected from the second medical center into the schizophrenia classification model to determine the type of schizophrenia.