Brain glioma recognition device based on 5T multi-nuclear magnetic resonance imaging system

Abstract

The invention discloses a brain glioma recognition device based on a 5T multi-nuclear magnetic resonance imaging system. The brain glioma recognition device comprises an amide proton imaging acquisition module used for receiving a chemical exchange saturation transfer atlas of amide protons and the concentration of the amide protons; the 23Na MRI image acquisition module is used for receiving the 1H T2 weighted MRI image, the 23Na MRI image and the total sodium concentration map; the 31P MRS image acquisition module is used for receiving the 31P MRS image and the phosphorus-containing metabolite concentration of the brain region; and an identification module including a discrimination network and a segmentation network. According to the method, the sample group is input into the discrimination network and the segmentation network of the cross-modal data brain tumor, abnormal signals of different voxels in the brain region can be identified, and the boundary of the brain glioma can be outlined.

Description

Technical Field

[0001] The present invention belongs to the technical field of magnetic resonance imaging, and in particular relates to a brain glioma identification device based on a 5T multi-nuclear magnetic resonance imaging system. Background Art

[0002] Gliomas are tumors that originate from brain glial cells and are the most common primary intracranial tumors. At present, clinical diagnosis mainly relies on imaging diagnoses such as computed tomography (CT) and magnetic resonance imaging (MRI). Diffusion-weighted imaging, diffusion tensor imaging, perfusion-weighted imaging, magnetic resonance spectroscopy, functional MRI, and positron emission tomography are of great significance for the differential diagnosis and treatment effect evaluation of gliomas. MRI is superior to CT in image information. CT mainly shows the density difference between glioma lesions and normal brain tissue, characteristic density manifestations such as calcification, hemorrhage and cystic lesions, the location of lesions, edema and space-occupying effects. MRI has the advantages of high spatial resolution, no tissue penetration depth limitation, and good soft tissue contrast. It is often used for the signal intensity differences and space-occupying effects of hemorrhage, necrosis, edema, etc. of gliomas, and can show the invasion range of lesions. However, the low sensitivity of magnetic resonance imaging limits its more accurate detection of gliomas. In addition, the information provided by single radionuclide detection is limited. For highly invasive brain gliomas, it is a huge challenge to accurately identify them at an early stage and accurately outline their infiltration boundaries.

[0003] Amide proton transfer (APT) imaging is a new type of non-invasive molecular MRI technology that can quantitatively detect endogenous proteins and peptides in tissues, obtain information changes in the microenvironment of tumor cells, and thus reflect the characteristic biological behavior of brain gliomas. The application value of APT imaging of brain tumors has been verified on the basis of biopsy, and APT imaging has shown important prospects in the study of neurodegenerative diseases and psychiatry. The difference in protein content in brain tumors is considered to be the main source of APT imaging signals, but at the same time, the intensity of APT imaging signals will also be affected by other factors. For example, the cystic components in some gliomas with cystic lesions are rich in protein, and their APT imaging signals are higher, but they are not related to the degree of malignancy. Therefore, when analyzing the APT imaging signal of brain tumors, simultaneous multi-nuclear imaging combined with other radionuclides can provide a more accurate quantitative analysis of the amide proton concentration in the brain region.

[0004] Among the endogenous radionuclides that can be used for MRI, in addition to protons ( 1 H), and 23 Na and 31 P.1 H is the most abundant nucleus in the human body and has the highest gyromagnetic ratio. Conventional MRI usually uses protons as imaging observation nuclei. 1 H nucleus (endogenous: 23 Na and 31 P, etc.; exogenous: 129 Xe and 19 Multi-core MRI (F, etc.) is being continuously studied and explored, becoming a new tool for brain disease and brain science research. 1 The chemical shift range of H metabolites detected by traditional MRI technology in brain tissue is mainly 0-5ppm. The chemical shift range is narrow, and the signals of different metabolites are prone to overlap and difficult to distinguish. 1 H has a larger chemical shift range, thus providing more comprehensive functional and metabolic information. In endogenous human multinuclear MRI studies, 23 Na is the second most abundant MRI-observable nuclide that can be detected in living tissues at natural abundance. + It plays a vital role in many cellular functions. 23 Na MRI has also been used to study many brain diseases (brain tumors, cerebral infarction, stroke, etc.). 31 P is a very important endogenous nuclide. 31 P magnetic resonance spectroscopy (MRS) can detect macromolecular metabolites and phospholipid membranes in the body, providing important information for energy metabolism. 31 P MRS has been reported to be used in the study of many brain diseases including Alzheimer's disease, Parkinson's disease, multiple sclerosis, migraine, cerebral ischemia, brain tumors, bipolar disease, major depression and attention deficit / hyperactivity disorder. 31 The unique use of P to directly image brain energy production from the hydrolysis of ATP provides new insights into human brain energetics and its role in supporting neuronal activity and brain function. 31 P MRS imaging has broad application prospects in clinical diagnosis and research. It can provide important information about phosphorus metabolism and its changes in different disease states, which is helpful for early diagnosis, treatment monitoring and drug development of diseases. 31 P MRS imaging research has important scientific significance and application prospects.

[0005] Based on the advantages of high sensitivity and high resolution of the 5T high-field MRI system and the fact that the 1.5T and 3T MRIs currently used in clinical practice cannot meet the needs of accurate disease diagnosis, 1 H / 23 Na / 31Multi-core MRI can obtain 1 H T2-weighted MRI, APT imaging, and sodium concentration quantification. The high-resolution advantage of the 5T high-field MRI system can distinguish overlapping phosphorus-containing metabolites. 31 The P MRS spectrum peak can provide more accurate quantitative analysis of phosphorus-containing metabolites in the brain glioma area (compared to 3T or lower field strength magnetic resonance spectrometers), including adenosine triphosphate, adenosine diphosphate, creatinine phosphate, etc. Based on the advantages of high field and multi-nuclear imaging, the system can accurately identify abnormal signals of different voxels in the brain area, which is of great significance for the early diagnosis and boundary delineation of brain glioma. Summary of the invention

[0006] The purpose of the present invention is to address the above-mentioned problems in the prior art and provide a device for identifying gliomas based on a 5T multi-nuclear magnetic resonance imaging system. Based on the advantages of high sensitivity and high resolution of the 5T high-field magnetic resonance spectrometer, highly sensitive APT imaging of brain regions can be performed to more accurately quantify endogenous proteins and peptides; 23 Na MRI, obtains the total sodium concentration map of different voxels in the brain; 31 P MRS was used to quantitatively analyze phosphorus-containing metabolites in the brain to obtain the concentrations of phosphorus-containing metabolites in different voxels of the brain. 1 H / 23 Na / 31 P multi-core MRI data, construct a sample group, import it into the constructed discriminant network and segmentation network of brain tumor based on cross-modal data, determine whether there is a brain tumor, and outline the boundary.

[0007] The above-mentioned purpose of the present invention is achieved by the following technical means:

[0008] A brain glioma identification device based on a 5T multi-nuclear magnetic resonance imaging system, comprising:

[0009] An amide proton imaging acquisition module is used to receive the chemical exchange saturation transfer spectrum of amide protons, receive the APT images obtained by applying saturation pulses at 3.5ppm and -3.5ppm, and also to obtain the amide proton concentration in the brain area;

[0010] 23 Na MRI image acquisition module, used to receive 1 H T2-weighted MRI images, 23 Na MRI images and maps of total sodium concentration in brain regions;

[0011] 31 P MRS image acquisition module, used to receive 31 P MRS images and concentrations of different phosphorus-containing metabolites in brain regions;

[0012] The sample group generation module is used to generate sample groups. Each sample group includes 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 P MRS images, amide proton concentration in brain regions, 23 Na concentration, concentrations of various phosphorus-containing metabolites in brain regions, and generating training and test sets based on each sample group;

[0013] The recognition model includes a discriminant network and a segmentation network, which are used to train the discriminant network and the segmentation network based on a training set, and are also used to input a sample group to be recognized to obtain a predicted category label and a predicted segmentation label.

[0014] As described above, the discriminant network includes an input layer, a text encoding layer, a first fully connected layer, a first convolutional layer, a second convolutional layer, a first residual block group, a second residual block group, a third residual block group, a fourth residual block group, a global average pooling layer, a second fully connected layer, and an output layer.

[0015] The segmentation network includes an encoder and a decoder.

[0016] In the sample group 1 The H T2-weighted MRI images and the chemical exchange saturation transfer maps of amide protons were input into the input layer of the discriminant network built based on the ResNet architecture to obtain the image data feature map, and then the amide proton concentration, 23 The Na concentration and the concentrations of various phosphorus-containing metabolites in the brain area are input into the text encoding layer of the discriminant network for text encoding to obtain a text vector. The text vector is fused with the image data feature map extracted by the input layer of the discriminant network through the first fully connected layer to obtain a fused feature map. The fused feature map is then passed through the first convolutional layer, the second convolutional layer, the first residual block group, the second residual block group, the third residual block group and the fourth residual block group, the global average pooling layer, and the output layer in sequence to obtain the corresponding predicted category label.

[0017] As described above, the first residual block group contains 3 residual blocks, the second residual block group contains 4 residual blocks, the third residual block group contains 6 residual blocks, and the fourth residual block group contains 3 residual blocks.

[0018] Included in the sample group 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 NaMRI images, 31The P MRS image data is transmitted to the segmentation network based on the U-Net network architecture. In the segmentation network, skip connections are used to connect the layers of the same level in the encoder and decoder parts, and the attention module of multi-scale features is connected during the skip connection process.

[0019] As described above, the attention module of the multi-scale features includes a convolution layer, a maximum pooling layer, and an average pooling layer. The feature image for jump connection output by the downsampling layer of the encoder is input into the attention module as the initial feature image. The initial feature image passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then performs channel splicing. The spliced ​​features pass through a 5×5 convolution layer and are linearly multiplied with the initial feature image to obtain the first step result. The first step result passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then performs channel splicing. After passing through a 3×3 convolution layer, it is linearly multiplied with the first step output result to obtain the second step result. The second step output result passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then performs channel splicing. After passing through a 3×3 convolution layer, it is linearly multiplied with the second step output result and then passes through a sigmoid activation function to generate a feature weight map. The feature weight map is output to the upsampling layer of the decoder of the corresponding level layer.

[0020] The total loss function of the recognition model is the sum of the discriminant network loss function Lc and the segmentation network loss function Ls multiplied by the corresponding weights. Based on minimizing the total loss function, the recognition model is end-to-end trained using the training set, and the parameters of the recognition model are saved, where:

[0021] The discriminant network loss function Lc is based on the following formula:

[0022]

[0023] Among them, N is the number of sample groups input into the discriminant network, y i is the actual category label of the i-th sample group, p i is the predicted category label of the i-th sample group, i represents the sequence number of the sample group in each round of training iteration, i∈{1, 2, …N};

[0024] The segmentation network loss function Ls is based on the following formula:

[0025] L s =-(y t log(y p )+(1-y t )log(1-y p ))

[0026] Among them, y tis the actual segmentation label, y p is the predicted segmentation label.

[0027] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the amide proton imaging acquisition module, 23 Na MRI image acquisition module, 31 p MRS image acquisition module, sample group generation module, and recognition module.

[0028] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the amide proton imaging acquisition module, 23 Na MRI image acquisition module, 31 p MRS image acquisition module, sample group generation module, and recognition module.

[0029] A computer program product, comprising a computer program, which, when executed by a processor, implements the amide proton imaging acquisition module of the above-mentioned device, 23 Na MRI image acquisition module, 31 p MRS image acquisition module, sample group generation module, and recognition module.

[0030] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0031] 1. The field strength used in the present invention is 5T, which has higher magnetic resonance sensitivity than the traditional 1.5T and 3T. It can detect substances that are difficult to detect with traditional magnetic resonance technology, and can more sensitively detect key proteins of brain glioma, sodium concentration in brain areas, and phosphorus-containing metabolites.

[0032] 2. The magnetic resonance spectrometer used in the present invention has a field strength of 5T. Compared with the traditional 1.5T and 3T, it has the characteristics of high resolution, can accurately visualize brain areas with smaller voxels, and can perform more accurate quantitative analysis of the concentrations of various metabolites and proteins in the brain area.

[0033] 3. The present invention can perform high-sensitivity amide proton imaging of brain regions, obtain the concentration of amide protons in brain regions, perform quantitative analysis of protein concentrations in brain regions, and analyze abnormal areas of amide proton concentration by comparing with normal brain tissue.

[0034] 4. The present invention includes a cross-modal data brain tumor discrimination network and segmentation network. During the segmentation process, a multi-scale feature attention module is designed, which uses spatial attention to increase the weight of the key areas of the input features, thereby better restoring the detailed information of these key areas.

[0035] 5. The present invention obtains the amide proton concentration, total sodium concentration map, and different phosphorus-containing metabolite concentration data of different voxels in the brain region through a 5T multi-nuclear magnetic resonance system, and constructs a sample group, which includes 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 P MRS images, amide proton concentration in brain regions, 23 Na concentration, concentration of various phosphorus-containing metabolites in brain regions, and inputting the sample group into the cross-modal data brain tumor discrimination network and segmentation network can identify abnormal signals of different voxels in the brain region and outline the boundaries of gliomas. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the structure of the present invention.

[0037] Figure 2 Schematic diagram of the structure of the attention module for multi-scale features. DETAILED DESCRIPTION

[0038] The exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art. It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should be the common meanings understood by those skilled in the art to which this application belongs. The following describes a method based on 5T proposed in an embodiment of the present application in conjunction with the accompanying drawings. 1 H / 23 Na / 31 P A device for identifying and delineating the boundaries of gliomas in a multi-nuclear magnetic resonance imaging system.

[0039] Following the above background technology, in addition to being able to achieve single-nuclide imaging, the related technology also has a relatively low sensitivity and resolution of MRI due to the low magnetic field strength (1.5T or 3T), which leads to high detection limits and inaccurate quantification in the detection of biomarkers for gliomas. Single nuclide imaging is difficult to accurately quantify the concentration of tumor markers in different voxels in the brain, and it is difficult to accurately outline the boundaries of gliomas. Based on this, the present application embodiment provides a method based on 5T 1 H / 23 Na / 31P multi-nuclear MRI system for glioma identification and boundary delineation. Compared with traditional low-field MRI (1.5T and 3T), the 5T high magnetic field strength can overcome the shortcomings of insufficient sensitivity and resolution of MRI. Compared with single-nuclear MRI, the system can provide 1 H / 23 Na / 31 P multi-nuclear magnetic resonance imaging information is used to obtain the protein concentration, sodium concentration, and phosphorus metabolite concentration of brain glioma. The obtained brain region multi-nuclear magnetic resonance data is imported into the cross-modal data brain tumor discrimination and segmentation network model to determine whether the brain glioma exists and outline its boundaries ( Figure 1 ).

[0040] A brain glioma identification device based on a 5T multi-nuclear magnetic resonance imaging system, comprising: an amide proton imaging acquisition module, 23 Na MRI image acquisition module, 31 P MRS image acquisition module, sample group generation module, and recognition module.

[0041] Amide proton imaging acquisition module:

[0042] The amide proton imaging acquisition module is used to receive the chemical exchange saturation transfer spectrum of amide protons, receive the APT images obtained by applying saturation pulses at 3.5ppm and -3.5ppm, and is also used to obtain the amide proton concentration in the brain area.

[0043] Due to abnormal protein expression in glioma patients, their levels are higher than those in normal brain tissue. The chemical exchange rate of amide protons is usually around 30 Hz, which can produce a magnetization transfer effect, but the relaxation time of mobile proteins and peptides is longer, and the magnetic resonance frequency range of amide groups is narrower, concentrated at ~3.5 ppm. In most tissues, the concentration of amide protons in endogenous mobile peptides and proteins is about 5 to 8 mM. The above conditions make mobile amide protons in brain regions suitable for chemical exchange saturation transfer imaging, and this technology can be used to quantitatively analyze the concentration of amide protons in brain regions.

[0044] Brain amide proton magnetic resonance imaging experiment: First, test the chemical exchange saturation transfer spectrum of amide protons, that is, the Z spectrum. The Y-axis of the chemical exchange saturation transfer spectrum is S / S0, where: S represents the signal intensity of water protons after applying radio frequency pulse excitation at different chemical shift sites, S0 represents the signal intensity of water protons after applying radio frequency pulse excitation, and the X-axis is the saturated chemical shift site. The chemical exchange saturation transfer spectrum is drawn to obtain the chemical exchange saturation transfer spectrum. In the next step of whole-brain APT imaging, saturation pulses are applied to the signal site of amide protons at 3.5ppm and the symmetrical chemical shift signal site at -3.5ppm to perform APT imaging of different voxels in the whole brain. The calculation method of the APT imaging effect of different voxels in the brain region is based on the asymmetric magnetization transfer rate MTR at ±3.5ppm. asym (asymmetrical magnetization transfer rate) is calculated by the difference between asym(3.5ppm) =MTR asym(+3.5ppm) -MTR asym(-3.5ppm) , where MTR asym(+3.5ppm) is the asymmetric magnetization transfer rate at +3.5ppm, MTR asym(-3.5ppm) The asymmetric magnetization transfer rate at -3.5 ppm is the higher protein content in brain tumors, which leads to an increase in exchangeable amide protons and an increase in the corresponding APT imaging signal.

[0045] 23 Na MRI image acquisition module:

[0046] 23 Na MRI image acquisition module, used to receive 1 H T2-weighted MRI images, 23 Na MRI images and maps of total sodium concentration in brain regions.

[0047] In patients with glioma and healthy volunteers 23 Prior to Na MRI imaging, a fast spin echo sequence was performed on a 5T magnetic resonance spectrometer. 1 H T2-weighted MRI imaging, obtained 1 H T2-weighted MRI images using dual-tuning 23 Na- 1 H birdcage head coil for whole brain imaging of glioma patients and healthy volunteers 23 Na MRI imaging, obtained 23 Na MRI images. According to the quantitative analysis software of the 5T magnetic resonance spectrometer, the total sodium concentration (TSC) of the whole brain and the regions of interest delineated in the brain of glioma patients and healthy volunteers was quantitatively analyzed, and the total sodium concentration (TSC) of the whole brain and the regions of interest delineated in the brain were obtained. 23Na concentration.

[0048] 31 P MRS image acquisition module:

[0049] 31 P MRS image acquisition module, used to receive 31 P MRS images and concentrations of different phosphorus-containing metabolites in brain regions.

[0050] Glioma patients and healthy volunteers 31 P MRS image use 31 P / 1 H dual-tuned birdcage coil was used for acquisition. First, T1-weighted fast field echo images in the axial plane were obtained by multiplanar reconstruction for the mid-sagittal plane. 31 Uniform localization of P MRS grids and multivoxel mapping of the whole brain in glioma patients and healthy volunteers 31 P MRS acquisition, obtain 31 P MRS images. Adiabatic RF pulses were delivered with broadband decoupling and an automatic shimming procedure with a maximum acceptable value of 35 Hz, and the size of a single voxel was 25 × 25 × 20 mm. Subsequently, T1-weighted MRI images were acquired to ensure that the position of the patient's head was the same as at the beginning of the examination. Voxel selection: A large number of voxels were obtained using a 3D atlas approach. However, for practical purposes and to avoid multiple comparison errors, a target number of voxels with representative bilateral white matter, deep gray matter, and cortex was determined for analysis. Considering that the differences in metabolite concentrations in different tissue types (e.g., gray matter and white matter) exceeded the differences between the same tissue types in different brain locations (e.g., temporal lobe and frontal lobe), homologous brain regions in patients with glioma and healthy volunteers were selected for quantification of phosphorus-containing metabolite concentrations.

[0051] The linear fitting model processing software (LCmodel) was used to analyze the brain regions 31 P MRS spectra were used to quantify the concentration of phosphorus-containing metabolites in different voxels of the brain. The manual steps for convergence of the spectra included zero-order and first-order phase correction, followed by identification of spectral peaks of phosphorus-containing metabolites such as inorganic phosphate and creatine phosphate, and quantitative analysis of phosphorus-containing metabolites in the brain.

[0052] Sample group generation module:

[0053] The sample group generation module is used to generate sample groups. Each sample group includes 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 P MRS images, amide proton concentration in brain regions, 23Na concentration, concentrations of various phosphorus-containing metabolites in brain regions, and generating training and test sets based on each sample group.

[0054] Identification Model:

[0055] Step 1: Get the training set and the test set, where the training set and the test set contain multiple sample groups, each of which contains 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 PMRS images, amide proton concentration in brain regions, 23 Na concentration and concentrations of various phosphorus-containing metabolites in brain regions. 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 The actual category labels of P MRS image data are tumor and no tumor. 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 The actual segmentation label of the P MRS image data is the tumor segmentation label.

[0056] Step 2: Build a recognition model based on cross-modal data. The recognition model includes a discriminant network and a segmentation network. Set the total loss function.

[0057] Among them, the discriminant network includes an input layer, a text encoding layer, a first fully connected layer, a first convolutional layer, a second convolutional layer, a first residual block group, a second residual block group, a third residual block group, a fourth residual block group, a global average pooling layer, a second fully connected layer and an output layer.

[0058] The segmentation network consists of an encoder and a decoder.

[0059] First, the sample group 1 The H T2-weighted MRI images and the chemical exchange saturation transfer maps of amide protons were input into the input layer of the discriminant network built based on the ResNet architecture to obtain the image data feature map, and then the amide proton concentration, 23 The Na concentration and the concentration of each phosphorus-containing metabolite in the brain region are input into the text encoding layer of the discriminant network for text encoding. 23The Na concentration and the concentration of each phosphorus-containing metabolite in the brain region are encoded into a text vector, and the text vector is fused with the image data feature map extracted by the input layer in the discriminant network built based on the ResNet architecture through the first fully connected layer to obtain a fused feature map. Then, after the fused feature map passes through the first convolution layer and the second convolution layer in sequence, the output result of the second convolution layer is input into the first residual block group, which contains 3 residual blocks, and then the feature image output by the first residual block group is input into the second residual block group, which contains 4 residual blocks, and then the feature image output by the second residual block group passes through the subsequent third residual block group and the fourth residual block group, which contain 6 and 3 residual blocks respectively; finally, the feature image output by the fourth residual block group is input into the output layer after passing through a global average pooling layer and the second fully connected layer, and the output of the output layer obtains the corresponding predicted category label.

[0060] Then the samples that were identified as having tumors were 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 The P MRS image data is transmitted to the segmentation network based on the U-Net network architecture. In the segmentation network, skip connections are used to connect the layers of the same level in the encoder and decoder parts. At the same time, the attention module of multi-scale features is connected in the skip connection process. The attention module includes convolutional layer, maximum pooling layer, and average pooling layer. The feature image for skip connection output by the downsampling layer of the encoder is input into the attention module as the initial feature image. The initial feature image passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then is spliced ​​on the channel. The spliced ​​features pass through a 5×5 convolution layer and are linearly multiplied with the initial feature image to obtain the first step result. The first step result passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then is spliced ​​on the channel. After passing through a 3×3 convolution layer, it is linearly multiplied with the first step output result to obtain the second step result. Then, the second step output result passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then is spliced ​​on the channel. After passing through a 3×3 convolution layer, it is linearly multiplied with the second step output result and then passes through a sigmoid activation function to generate a feature weight map. The feature weight map is output to the upsampling layer of the decoder of the corresponding level layer. By analogy, each skip connection passes through the multi-scale feature attention module. The above-mentioned channel splicing is to splice the output data after the maximum pooling layer and the output data after the average pooling layer in the channel dimension.

[0061] Step 3: Set the total loss function; the total loss function is divided into two parts: the discriminant network loss function L c and segmentation network loss function L s The total loss function is the weighted sum of the discriminant network loss function Lc and the segmentation network loss function Ls, as shown in the formula:

[0062]

[0063] Among them, N is the number of sample groups input into the discriminant network, y i is the actual category label of the i-th sample group. In this embodiment, if it is a positive classification example, then y i The value is 1. If it is a negative example, then y i The value is 0; p i is the predicted category label of the i-th sample group, indicating the value of the classification prediction obtained by the discriminant network as a positive example. i represents the sequence number of the sample group in each round of training iteration, i∈{1, 2, …N}.

[0064] L s =-(y t log(y p )+(1-y t )log(1-y p ))

[0065] Among them, y t is the actual segmentation label, which represents the value of the tumor segmentation label of the voxel in the sample group in the training set, y t ∈{0,1}, the positive example is y t The value is 1, and the negative example is y t Taking the value as 0 makes the segmentation module loss function L s Suitable for evaluation of binary classification tasks; p To predict the segmentation label, it indicates the probability that the voxel in the tumor segmentation result obtained by the segmentation network is a positive example. If the voxel in the tumor segmentation result is a positive example, then y p The value is 1, otherwise, y p The value is 0.

[0066] Step 4: Based on minimizing the total loss function set in step 3, the brain tumor recognition model of cross-modal data constructed in step 2 is end-to-end trained using the training set generated in step 1, and the parameters of the recognition model are saved;

[0067] According to the recognition model built in step 2, the network learning rate is initialized to 0.0001, the batch size is set to 8, and the network training is performed on the PyTorch platform using the Adam optimizer. 1The H T2-weighted MRI images and the chemical exchange saturation transfer maps of amide protons were input into the constructed discriminant network to obtain the predicted category labels, and then the samples in the group with tumors were identified. 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 The P MRS image data was transmitted to the brain tumor segmentation network, and the network training was performed according to the total loss function set in step 4. After the total number of iterations of network training reached 200, the training was stopped and the corresponding network parameters were saved.

[0068] Step 5: Group the samples to be identified 1 The H T2-weighted MRI images and the chemical exchange saturation transfer maps of amide protons are input into the discriminant network of the trained recognition model to obtain the predicted category label; 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 The P MRS image data is input into the segmentation network of the recognition model trained in step 4 to obtain the predicted segmentation label.

[0069] Those skilled in the art can understand that the amide proton imaging acquisition module, 23 Na MRI image acquisition module, 31 The whole or part of the processes of the P MRS image acquisition module, the sample group generation module, and the identification module can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the amide proton imaging acquisition module, 23 Na MRI image acquisition module, 31 The entire or partial process of the P MRS image acquisition module, sample group generation module, and recognition module.

[0070] In one embodiment, a computer device is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the amide proton imaging acquisition module, 23 NaMRI image acquisition module, 31 P MRS image acquisition module, sample group generation module, and recognition module.

[0071] In one embodiment, a computer-readable storage medium is provided on which a computer program is stored. When the computer program is executed by a processor, the amide proton imaging acquisition module,23 Na MRI image acquisition module, 31 P MRS image acquisition module, sample group generation module, and recognition module.

[0072] In one embodiment, a computer program product is provided, including a computer program, which, when executed by a processor, implements the amide proton imaging acquisition module of the above-mentioned device, 23 Na MRI image acquisition module, 31 P MRS image acquisition module, sample group generation module, and recognition module.

[0073] It should be noted that the embodiments described in the present invention are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the described embodiments or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. A brain glioma identification device based on a 5T multi-nuclear magnetic resonance imaging system, characterized in that: include: An amide proton imaging acquisition module is used to receive the chemical exchange saturation transfer spectrum of amide protons, receive the APT images obtained by applying saturation pulses at 3.5ppm and -3.5ppm, and also to obtain the amide proton concentration in the brain area; 23 Na MRI image acquisition module, used to receive 1 H T2-weighted MRI images, 23 Na MRI images and maps of total sodium concentration in brain regions; 31 P MRS image acquisition module, used to receive 31 P MRS images and concentrations of different phosphorus-containing metabolites in brain regions; The sample group generation module is used to generate sample groups. Each sample group includes 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 P MRS images, amide proton concentration in brain regions, 23 Na concentration, concentrations of various phosphorus-containing metabolites in brain regions, and generating training and test sets based on each sample group; The recognition model includes a discriminant network and a segmentation network, which are used to train the discriminant network and the segmentation network based on a training set, and are also used to input a sample group to be recognized to obtain a predicted category label and a predicted segmentation label.

2. According to claim 1, a brain glioma identification device based on a 5T multi-nuclear magnetic resonance imaging system is characterized in that: The discriminant network includes an input layer, a text encoding layer, a first fully connected layer, a first convolutional layer, a second convolutional layer, a first residual block group, a second residual block group, a third residual block group, a fourth residual block group, a global average pooling layer, a second fully connected layer and an output layer. The segmentation network includes an encoder and a decoder.

3. According to claim 2, a brain glioma identification device based on a 5T multi-nuclear magnetic resonance imaging system is characterized in that: In the sample group 1 The H T2-weighted MRI images and the chemical exchange saturation transfer maps of amide protons were input into the input layer of the discriminant network built based on the ResNet architecture to obtain the image data feature map, and then the amide proton concentration, 23 The Na concentration and the concentrations of various phosphorus-containing metabolites in the brain area are input into the text encoding layer of the discriminant network for text encoding to obtain a text vector. The text vector is fused with the image data feature map extracted by the input layer of the discriminant network through the first fully connected layer to obtain a fused feature map. The fused feature map is then passed through the first convolutional layer, the second convolutional layer, the first residual block group, the second residual block group, the third residual block group and the fourth residual block group, the global average pooling layer, and the output layer in sequence to obtain the corresponding predicted category label.

4. The device for identifying glioma based on a 5T multi-nuclear magnetic resonance imaging system according to claim 3, characterized in that: The first residual block group includes 3 residual blocks, the second residual block group includes 4 residual blocks, the third residual block group includes 6 residual blocks, and the fourth residual block group includes 3 residual blocks.

5. According to claim 3, a 5T-based 1 H / 23 Na / 31 A device for identifying and delineating the boundaries of gliomas in a multi-nuclear magnetic resonance imaging system, characterized in that: Included in the sample group 1 H T2-weighted MRI images, chemical exchange saturation transfer maps of amide protons, 23 Na MRI images, 31 The P MRS image data is transmitted to the segmentation network based on the U-Net network architecture. In the segmentation network, skip connections are used to connect the layers of the same level in the encoder and decoder parts, and the attention module of multi-scale features is connected during the skip connection process.

6. The device for identifying glioma based on a 5T multi-nuclear magnetic resonance imaging system according to claim 5, characterized in that: The attention module of the multi-scale feature includes a convolution layer, a maximum pooling layer, and an average pooling layer. The feature image for jump connection output by the downsampling layer of the encoder is input into the attention module as the initial feature image. The initial feature image passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then is spliced ​​on the channel. The spliced ​​features pass through a 5×5 convolution layer and are linearly multiplied with the initial feature image to obtain the first step result. The first step result passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then is spliced ​​on the channel. After passing through a 3×3 convolution layer, it is linearly multiplied with the first step output result to obtain the second step result. The second step output result passes through the maximum pooling layer and the average pooling layer of the attention module respectively, and then is spliced ​​on the channel. After passing through a 3×3 convolution layer, it is linearly multiplied with the second step output result and then passes through a sigmoid activation function to generate a feature weight map. The feature weight map is output to the upsampling layer of the decoder of the corresponding level layer.

7. The device for identifying glioma based on a 5T multi-nuclear magnetic resonance imaging system according to claim 5, characterized in that: The total loss function of the recognition model is the sum of the discriminant network loss function Lc and the segmentation network loss function Ls multiplied by the corresponding weights. Based on minimizing the total loss function, the recognition model is end-to-end trained using the training set, and the parameters of the recognition model are saved, where: The discriminant network loss function Lc is based on the following formula: Among them, N is the number of sample groups input into the discriminant network, y i is the actual category label of the i-th sample group, p i is the predicted category label of the i-th sample group, i represents the sequence number of the sample group in each round of training iteration, i∈{1, 2, …N}; The segmentation network loss function Ls is based on the following formula: L s =-(and t log(y p )+(1-and t )log(1-y p )) Among them, y t is the actual segmentation label, y p is the predicted segmentation label.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, an amide proton imaging acquisition module of the device according to any one of claims 1 to 7 is implemented. 23 Na MRI image acquisition module, 31 P MRS image acquisition module, sample group generation module, and recognition module.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the amide proton imaging acquisition module of the device according to any one of claims 1 to 7 is implemented. 23 Na MRI image acquisition module, 31 P MRS image acquisition module, sample group generation module, and recognition module.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the amide proton imaging acquisition module of the device according to any one of claims 1 to 7 is realized. 23 Na MRI image acquisition module, 31 P MRS image acquisition module, sample group generation module, and recognition module.

Images