Method for predicting brain glioma subtype through magnetic resonance imaging based on deep learning

Abstract

The invention discloses a method for predicting brain glioma subtypes through magnetic resonance imaging based on deep learning, and relates to the technical field of chemical exchange saturation transfer magnetic resonance imaging. The method comprises the following steps: preprocessing a CEST magnetic resonance image of a patient; voxel Z spectrums and corresponding classification labels in tumor and opposite side normal tissue areas of the CEST magnetic resonance image are obtained; inputting the obtained voxel Z spectrum and the classification label into a convolutional network model for training; and performing classification prediction on test data by using the trained model. According to the method, a voxel Z-spectrum full-spectrum training convolutional network model is utilized to realize classification of tumor tissues and normal tissues and prediction of IDH and MGMT subtypes of glioma. According to the method, the whole voxel Z spectrum serves as network input, Z spectrum information can be better utilized, and the requirement for the training data size is lowered.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical exchange saturation transfer magnetic resonance imaging, and in particular to a method for predicting brain glioma subtypes by magnetic resonance imaging based on deep learning. Background Art

[0002] Classification of glioma subtypes based on molecular profiles is crucial for treatment decisions and survival prediction of central nervous system tumors. Isocitrate dehydrogenase (IDH) mutation status and O6 methylguanine DNA methyltransferase (MGMT) promoter methylation status are important markers for the diagnosis and treatment of brain gliomas. However, IDH mutation status and MGMT promoter methylation status usually need to be obtained through histopathological biopsy, which is invasive and biased. Previous studies have shown that IDH mutation and MGMT promoter methylation can lead to different tumor microenvironments, including acidosis and metabolite accumulation, which can be used to predict brain glioma subtypes.

[0003] Chemical exchange saturation transfer (CEST) magnetic resonance imaging is an emerging molecular imaging technology that can perform non-invasive detection of endogenous metabolites and cellular microenvironments. It has been used clinically to evaluate the metabolism of tumors, cerebral ischemia and other neurological diseases, thereby assisting in the early diagnosis and treatment of stroke and determining tumor subtypes. The principle of CEST imaging technology is as follows: at the resonance frequency of a specific exchangeable proton, a saturated radio frequency pulse is applied to resonate with the exchangeable proton to generate a saturated signal, which is then transferred to water molecules through the exchangeable protons, causing changes in the water signal, thereby indirectly detecting specific proton signals by detecting the water signal. Currently, a variety of different CEST quantitative methods have been proposed for detecting a variety of metabolites in the body, such as Z-spectrum asymmetry analysis, multi-pool Lorentz fitting, and Lorentz difference-based methods.

[0004] Many studies have achieved tumor classification by using CEST quantitative analysis to obtain changes in metabolite signals associated with lesion areas. However, using different CEST quantitative methods may lead to significantly different classification accuracies, or even conflicting results. In addition, most current deep learning tumor classification methods are based on the image domain, requiring a large amount of training data for neural network training, which may be difficult to obtain in clinical settings.

[0005] In summary, in the field of glioma subtype classification, a method that can ensure accuracy, reduce the amount of training data required, and improve the prediction speed of the network model is urgently needed. Summary of the invention

[0006] The present invention aims to provide a method for predicting brain glioma subtypes based on deep learning by magnetic resonance imaging, which can use the full spectrum information of the voxel Z spectrum to quickly and accurately predict the IDH and MGMT status in brain glioma by inputting the patient's voxel Z spectrum. The present invention inputs the full spectrum of the patient's voxel Z spectrum into a convolutional network for learning and training, so that the network learns the Z spectrum features and differences between normal tissue and tumor tissue, and between IDH and MGMT in different states, thereby realizing voxel-level classification and achieving fast and accurate classification results.

[0007] The technical solution of the present invention is:

[0008] A method for predicting brain glioma subtypes using magnetic resonance imaging based on deep learning, comprising the following steps:

[0009] S1. Preprocessing the patient's CEST magnetic resonance image;

[0010] S2, obtaining the voxel Z-spectrum and corresponding diagnostic labels in the tumor and contralateral normal tissue regions of the CEST magnetic resonance image;

[0011] S3, inputting the obtained voxel Z spectrum and diagnostic label into the convolutional network model for training;

[0012] S4. Use the trained network model to perform classification prediction on the test data.

[0013] Furthermore, in step S1, the CEST magnetic resonance image of the patient is preprocessed, specifically:

[0014] Divide the patient data into training and testing sets;

[0015] De-noising the patient's CEST magnetic resonance images;

[0016] The denoised CEST magnetic resonance image is compared with the unsaturated reference image S 0 Perform image registration correction;

[0017] Divide the corrected CEST MRI image by the unsaturated reference image S 0 , get the normalized standard image S sat ;

[0018] Through the standard image S sat Obtain the Z spectrum of each pixel of the patient and perform B 0 Correction.

[0019] Further, in step S2, the voxel Z spectra and corresponding classification labels in the tumor and contralateral normal tissue regions of the CEST magnetic resonance image are obtained, specifically:

[0020] The region of interest in the CEST image was manually segmented using multi-contrast magnetic resonance images to obtain the tumor tissue region and the normal tissue region;

[0021] The tumor tissue area and normal tissue area of ​​the patient under different IDH and MGMT states are marked respectively to obtain the classification labels.

[0022] Further, in step S3, the obtained voxel Z spectrum and classification label are input into the convolutional network model for training, specifically:

[0023] Construct a convolutional network model, which consists of five one-dimensional convolutional layers, one pooling layer, two fully connected layers, and an output classification layer. The convolution kernel size of the one-dimensional convolutional layer is 3, the number of kernels is 32, and the number of neurons in the two fully connected layers is 64 and 32 respectively. Finally, Softmax is used to output the classification results of the network.

[0024] During the network model training process, the cross entropy loss function is used to calculate the loss, and the weights and biases in the model are updated through the error back propagation algorithm.

[0025] Furthermore, in step S4, the trained model is used to perform classification prediction on the test data, specifically:

[0026] After preprocessing the test patient data, the Z spectra of tumor voxels and normal voxels are obtained and input into the trained network model to obtain the classification labels predicted by the model and obtain the accuracy of the test data classification;

[0027] The performance of the network model is evaluated through the confusion matrix and ROC curve of the training set and test set.

[0028] Beneficial technical effects of the present invention:

[0029] The present invention uses the full spectrum of voxel Z spectrum to train the convolutional network model to achieve the classification of tumor tissue and normal tissue, as well as the prediction of glioma IDH and MGMT subtypes. The present invention can quickly and accurately achieve the classification prediction of brain glioma subtypes. Compared with the previous neural network that uses images as training data, the present invention uses the patient's single voxel Z spectrum as training data, which reduces the network's requirements for data volume. In addition, the full spectrum of Z spectrum is input into the network model for training, which can avoid the drawbacks of traditional quantization methods and make full use of all collected frequency offset information. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flowchart of the method for predicting brain glioma subtypes using magnetic resonance imaging based on deep learning of the present invention.

[0031] Figure 2This is a schematic diagram of the overall structure of the classification network model in the method for predicting brain glioma subtypes using magnetic resonance imaging based on deep learning in the present invention.

[0032] Figure 3 The classification result diagram of the brain glioma training set and test set provided by the embodiment of the present invention. Among them, AB is the training confusion matrix and training curve corresponding to the IDH mutation status classification network model, CD is the test confusion matrix and ROC curve corresponding to the IDH mutation status classification network model, EF is the training confusion matrix and training curve corresponding to the MGMT promoter methylation status classification network model, and GH is the test confusion matrix and ROC curve corresponding to the MGMT promoter methylation status classification network model. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Figure 1 This is a flow chart of a method for predicting brain glioma subtypes using magnetic resonance imaging based on deep learning in the present invention. Figure 1 As shown, the method for predicting brain glioma subtypes by magnetic resonance imaging based on deep learning includes the following specific steps:

[0035] S1. Preprocess the patient's CEST magnetic resonance image, including:

[0036] Divide the patient data into training and testing sets;

[0037] Since the medical image acquisition process is accompanied by a certain degree of noise interference, the original patient data needs to be denoised in the data preprocessing stage. The embodiment of the present invention uses the BM4D image denoising algorithm to denoise the acquired patient CEST magnetic resonance images to obtain images with a high signal-to-noise ratio.

[0038] Since the CEST image acquisition time is relatively long, the movement of the patient during the acquisition process may cause the CEST image to shift. Therefore, the patient data needs to be image registered in the data preprocessing stage. In the embodiment of the present invention, the denoised CEST magnetic resonance image is compared with the unsaturated reference image S 0 Perform image registration correction to obtain CEST images that match each other at different saturation frequencies;

[0039] Divide the corrected CEST MRI image by the unsaturated reference image S 0 , get the normalized standard image S sat;

[0040] Due to the presence of B in the CEST magnetic resonance image acquisition process 0 The field is not uniform, so it is necessary to use the standard image S sat The patient voxel Z spectrum is obtained and B 0 Correction, the embodiment of the present invention obtains the B value of the Z spectrum by fitting the water peak to the Z spectrum 0 Deviation, subtract B from the original saturation frequency offset 0 Deviation, and then use the interpolation method to get the corrected Z spectrum.

[0041] S2. Obtain the voxel Z spectra and corresponding classification labels in the tumor and contralateral normal tissue regions of the CEST magnetic resonance image, including:

[0042] In order to obtain the classification label, the embodiment of the present invention uses a magnetic resonance image with high lesion-background contrast, such as T 2 -Flair, T 1 The enhanced MRI images were used as references to manually segment the tumor area and the contralateral normal tissue area in the CEST images;

[0043] According to the segmented area, the label of the voxels in the normal tissue area is Control, and the voxels in the tumor area are marked with different classification labels according to the different IDH and MGMT status of the patients. Specifically, if the patient is IDH wild-type, the label of the voxels in the tumor area is IDH0; if the patient is IDH mutant, the label of the voxels in the tumor area is IDH1; if the patient's MGMT promoter is unmethylated, the label of the voxels in the tumor area is MGMT0; if the patient's MGMT promoter is methylated, the label of the voxels in the tumor area is MGMT1.

[0044] S3. Input the obtained voxel Z spectrum and classification label into the convolutional network model for training, specifically including:

[0045] In an embodiment of the present invention, a training data matrix with a matrix size of m×n and a training label matrix with a matrix size of m×1 are respectively constructed, wherein m is the number of voxel Z spectra and n is the total number of saturation frequency offsets;

[0046] Since the convolutional network can effectively extract the feature information in the Z spectrum, a convolutional network model is constructed. The network model consists of five one-dimensional convolutional layers, one pooling layer, two fully connected layers and an output classification layer, as shown in Figure 2 As shown, the convolution kernel size of the one-dimensional convolution layer is 3, the number is 32, and the number of neurons in the fully connected layer is 64 and 32 respectively; finally, Softmax is used to output the classification results of the network;

[0047] During the network model training process, the cross entropy loss function is used to calculate the loss, and the weights and biases in the model are updated through the error back propagation method.

[0048] S4. Use the trained model to classify and predict the test data, including:

[0049] Input the Z spectrum data of the test set into the trained network model to obtain the classification label predicted by the model and obtain the accuracy of the test data classification;

[0050] The performance of the network model is evaluated through the confusion matrix and ROC curve of the test set.

[0051] Result analysis:

[0052] Figure 3 The results of the training set and test set of brain glioma in the IDH and MGMT status classification network are shown in Figure 2. A total of 86 patient data, of which 31 were IDH wild-type, 32 were IDH mutant, 11 were MGMT unmethylated, and 12 were MGMT methylated. CEST images were collected using a 3T magnetic resonance imaging device with a saturation time of 3s and a saturation power of 0.76μT. 61 frequency offsets from -30ppm to 30ppm were collected. From the training confusion matrix and curve ( Figure 3 It can be clearly seen from Figures A, B, E, and F in the figure that the network model for IDH and MGMT status classification prediction has converged after 5000 iterations. For the test set data, Figure 3 The confusion matrix in Figure C shows that the true positive rates for contralateral normal tissue, IDH wild-type, and IDH mutant were 98.4%, 81.8%, and 87.3%, respectively. Figure 3 The confusion matrix in Figure G shows that the true positive rates for contralateral normal tissue, MGMT promoter unmethylation, and MGMT promoter methylation were 92.4%, 71.8%, and 76.0%, respectively. Figure 3 In Figure D, the AUC values ​​of the contralateral normal tissue, IDH wild-type, and IDH mutant in the IDH status classification model were 99%, 99%, and 96%, respectively, which is better than the 89% AUC reported by Jiang et al. using APTw MRI at 3 T. The present invention uses the full voxel Z spectrum as the network input, which can better utilize the Z spectrum information and reduce the requirement for the amount of training data.

[0053] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting brain glioma subtypes using magnetic resonance imaging based on deep learning, characterized in that: The specific steps include: S1. Preprocessing the patient's CEST magnetic resonance image; S2, obtaining the voxel Z spectra and corresponding classification labels in the tumor and contralateral normal tissue regions of the CEST magnetic resonance image; S3, inputting the obtained voxel Z spectrum and classification label into the convolutional network model for training; S4. Use the trained model to perform classification prediction on the test data.

2. According to claim 1, a method for predicting brain glioma subtypes based on magnetic resonance imaging based on deep learning, characterized in that: The specific method of step S1 is as follows: De-noising the patient's CEST magnetic resonance images; The denoised CEST magnetic resonance image is aligned with the unsaturated reference image S0 for image registration correction; The corrected CEST magnetic resonance image is divided by the unsaturated reference image S0 to obtain the normalized standard image S sat ; Through the standard image S sat The Z spectra of different voxels of the patient were obtained and B0 correction was performed.

3. The method for predicting brain glioma subtypes by magnetic resonance imaging based on deep learning according to claim 1, characterized in that: The specific method of step S2 is as follows: The region of interest in the CEST image was manually segmented using multi-contrast magnetic resonance images to obtain the tumor tissue region and the normal tissue region; The tumor tissue area and normal tissue area of ​​the patient under different IDH and MGMT states are marked respectively to obtain the classification labels.

4. The method for predicting brain glioma subtypes by magnetic resonance imaging based on deep learning according to claim 1, characterized in that: The specific method of step S3 is as follows: The obtained Z spectra of voxels in different states of tumors and Z spectra of voxels in normal tissues and the corresponding classification labels are used as training data; Construct a convolutional network model, which consists of five one-dimensional convolutional layers, one pooling layer, two fully connected layers, and an output classification layer; the convolution kernel size of the one-dimensional convolutional layer is 3, the number of neurons is 32, and the number of neurons in the two fully connected layers is 64 and 32 respectively; finally, use Softmax to output the classification results of the network; During the network model training process, the cross entropy loss function is used to calculate the loss, and the weights and biases in the model are updated through the error back propagation algorithm.

5. The method for predicting brain glioma subtypes by magnetic resonance imaging based on deep learning according to claim 1, characterized in that: The specific method of step S4 is as follows: After preprocessing the test patient data, the Z spectra of the voxels in the tumor area and the normal tissue area are extracted, and the Z spectra are input into the trained network model to obtain the model-predicted labels and prediction accuracy; The performance of the network model is evaluated through the confusion matrix of the training set and the test set and the ROC curve diagram.

Images