Method for extracting features of magnetic resonance medical images based on tensor rank-one decomposition

Abstract

The application designs a magnetic resonance medical image feature extraction method based on tensor rank-one decomposition, aiming at providing a novel and effective magnetic resonance brain image algebraic feature. The main steps are: step one, with the aid of image information extraction technology, the region of interest (ROI) is obtained from the original MRI image; step two, through a mathematical tensor representation method, a novel multi-modal data set is obtained after processing the original tensor containing ROI; step three, a weighted multi-modal optimization problem model is established by using the multi-modal data set and the label set, and a novel magnetic resonance medical image algebraic feature is obtained by solving the optimization problem. The application constructs a novel image feature extraction method for brain tumor MRI image data, which is different from the traditional omics feature, provides a mathematical feature from the perspective of tensor decomposition, and shows effectiveness in the image classification task.

Description

Technical Field

[0001] This invention relates to a general method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition, belonging to the field of medical image processing and feature extraction. Background Technology

[0002] Image features refer to a set of characteristics and properties in an image that distinguish it from other images. After preprocessing, images can be measured and extracted to visually extract features such as brightness, edges, and texture. Data transformation can also yield features such as statistical moments, histograms, and principal components. Medical imaging features primarily refer to radiomics features (also known as radiomics features). This type of feature is a special type of image feature that often targets only the Region of Interest (ROI) in an image. It utilizes various statistical, textural, and morphological methods to extract features, effectively helping doctors obtain key information from patient scans, thereby enabling early diagnosis and treatment.

[0003] The most common radiomics features currently include first-order statistical features, morphological features, and second-order and higher-order texture features. These features are basically used to quantitatively characterize the gray values ​​in the Region of Interest (ROI). For example, first-order features quantify the symmetry, uniformity, and intensity distribution variations of gray values ​​in the region under test, specifically manifested as basic statistical indicators such as the median, mean, maximum, minimum, standard deviation, skewness, and kurtosis of the gray values ​​in that region.

[0004] From a biological perspective, medical imaging omics features provide medical researchers with a quantitative direction for interpreting tumor pathology. However, for specific classification and diagnosis problems, researchers cannot accurately determine which features most directly influence the determination of a patient's pathological type, thus hindering precise diagnostic prediction. With the development and widespread application of machine learning and deep learning methods in medical imaging pathology analysis, researchers are able to continuously expand the size of their research datasets, hoping to use data-driven approaches to uncover patterns in certain pathological attributes from medical images. Therefore, image-feature-based pathology analysis and prediction has become the most popular and accurate research trend. For example, researchers at Duke University School of Medicine, using the ResNet model, took the T2 and T1ce sequences from structural phases as initial inputs and used 14 omics features (after dimensionality reduction) as intermediate inputs, achieving an accuracy of approximately 88.5% in the IDH classification task on a three-center dataset. Experimental testing showed that after incorporating the novel image features provided in this invention, the model's prediction accuracy, sensitivity, and AUC index were further improved. Summary of the Invention

[0005] Purpose of the invention: To address the problems existing in the prior art, this invention provides a general method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition. The main objective is to provide reliable image feature information for medical image diagnosis and treatment, to interpret the relationship between image data and diagnostic targets from a mathematical perspective, and to provide a reliable image analysis and type processing index for medical image research.

[0006] Technical solution: A general method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition, characterized by the following steps:

[0007] S1: Using image information extraction technology, the region of interest (ROI) is obtained from the original MRI medical images. Medical images generally refer to the lesion or tumor area in the image.

[0008] S2: Cropping ROIs from the original image at a fixed size to obtain new tensor data, and then processing the original tensor containing ROIs using a mathematical tensor representation method to obtain a new multimodal dataset.

[0009] S3: Using a multimodal dataset and a label set, a weighted multimodal optimization problem model is established, and the optimization problem is solved to obtain a novel algebraic feature of magnetic resonance medical images.

[0010] As a further improvement to the present invention, the specific steps include:

[0011] Step 1: Using an image information extraction model based on deep learning and statistical learning, extract the region of interest from the preprocessed raw medical images and label it on the raw data to obtain a standard medical image dataset with ROI.

[0012] Step 1.1: Construct a deep convolutional neural network model based on the Unet architecture. Using publicly available 3D medical image datasets such as BraTS (Brain Tumor Segmentation Challenge), UCSF (University of California, San Francisco Glioma Dataset), and UPenn (University of Pennsylvania Glioma Dataset), divide the image training set and validation set into scientifically reasonable parts. Set appropriate loss functions (Focal Dice Loss) and model parameters to complete the training of the image information extraction model. By comparing the performance of existing models, select the model with the best validation accuracy as the information extraction algorithm to achieve the most accurate image ROI segmentation.

[0013] Step 1.2: Considering that the publicly available data used for model learning are all standard data, and all original images are processed into a standard image space, maintaining standard pixel spacing, standard dimensions, and standard spatial coordinates, it is necessary to perform standardized preprocessing on the original medical images before using the trained information extraction algorithm: transform the image data uniformly into a standard three-dimensional pixel space, ensuring that each voxel is 1mm. 3 Then, the image pixel distribution pattern is adjusted to a t-distribution through intensity normalization, and all images corresponding to each case are registered to a fixed target image. After completing the image preprocessing, the processed image data is input into the model using the information extraction algorithm in step 1.1 to predict the ROI corresponding to each group of cases.

[0014] Step 2: Using windows of the same size, crop the processed image data to ensure that the ROI in each image can be included in the cropping window as much as possible, resulting in a set of ROI tensor data with reduced resolution. Through a tensor compression mathematical representation method, the ROI tensor is processed into a rank-one vector form to obtain a new type of multimodal dataset.

[0015] Step 2.1: Mark the ROIs in the standardized image on the original image, and crop the region using a window of the same size based on the marked area, ensuring that the ROI in each image is included within the cropping window as much as possible. This yields a set of image tensor data with reduced resolution, including the cropped tensor of the original image and the cropped tensor data of the information-extracted mask image. Here, a fixed window of dimension 64*64*64 is generally selected for cropping, resulting in a 6.4-dimensional window in the original image. 3 cm 3 The pruning space can be optimized to better match the subsequent experimental model while minimizing space loss when pruning the ROI.

[0016] Step 2.2: Based on the characteristics of medical image tensor data, a rank-one representation method for non-negative tensors is used to calculate the projections of the cropped 3D image tensor onto the three dimensions, obtaining vectors in three modalities as the rank-one representations of the tensors, thus constructing a novel multimodal dataset X = {X1, X2, X3}, where X... i The lengths of i = 1, 2, and 3 are all 64, achieving a kind of tensor information compression.

[0017] Step 3: Using the multimodal dataset and its corresponding label set, establish a weighted multimodal optimization problem model, use a special least squares method to solve the optimization problem, solve for the weight values ​​in the original problem, and calculate the feature matrix corresponding to each group of images based on the weights, as a new type of algebraic feature of magnetic resonance medical images.

[0018] Step 3.1: Using the multimodal dataset X and its corresponding label set Y, construct a block diagonal matrix by performing a centering operation on the dataset and label set. And the data matrix A=X T Y T YX, consider a label-based mathematical approximation representation W T P s X s Given s = 1, 2, 3, a novel weighted multimodal optimization problem model based on image data is established.

[0019] Step 3.2: Solve the optimization problem using a partially orthogonal least squares method to obtain the weight parameters W and eigenvectors P in the original problem. s By substituting the parameter values ​​into the mathematical approximation formula, the feature matrix corresponding to each group of images can be calculated. As a novel algebraic feature of magnetic resonance medical imaging.

[0020] As a further improvement of this invention, it is applied to the fields of medical image classification and diagnosis, three-dimensional image processing, and information extraction. In the field of medical image classification and diagnosis, algebraic features are uniformly extracted from standardized image data of different types and attributes. The image data and feature data are then used as input to a classifier model to accurately predict the type of medical image, providing doctors with a basis for diagnosis. In the field of three-dimensional image processing and information extraction, unlike commonly used brightness, color, texture, edge, and various statistical features, image data is tensorized to process the image into a rank-one representation, achieving image information compression. Simultaneously, based on the rank-one information and category information of the known data, an algebraic feature representing the category attribute of the image itself is provided, serving as an effective guiding parameter for image analysis and processing.

[0021] Beneficial effects:

[0022] 1. This invention uses an efficient algorithm and cleverly applies the fast solution technique of tensor operation to quickly extract algebraic features of medical images from the dataset. Compared with the extraction of traditional omics features, the feature extraction time corresponding to the image data of each case can be reduced by more than 60%.

[0023] 2. By combining traditional omics features and novel algebraic features as input data for medical diagnosis and treatment models, compared with using traditional omics features alone, it can improve the ACC (accuracy) by 1.5%, the AUC (area under curve) by 1.2%, and the sensitivity by 7% on diagnosis and treatment classification tasks on multi-center datasets, effectively improving the prediction performance of traditional diagnosis and treatment classification models.

[0024] 3. The algebraic features obtained by using tensor operations, from rank-one decomposition to solving optimization problems, are all guaranteed by rigorous mathematical theory. The resulting algebraic features can interpret the tendency of the type attributes of medical images from a probabilistic perspective, thereby more effectively improving the model's ability to identify image types. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method steps described in this invention.

[0026] Figure 2 These are raw medical image data provided by common datasets. Their dimensions and pixel spacing vary, and most of them are 2D images from different centers. They need to be fully standardized to process large amounts of image data. From left to right, they are transverse, sagittal and coronal scan images of T2 sequence.

[0027] Figure 3 SE-ResNet is a deep learning-based classification convolutional network. Its basic framework consists of multiple SE (sequence and excitation) layers. Image features are combined with the results of the input image after passing through the SE layers, and the model classification results are obtained through multiple fully connected layers. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention.

[0029] Example 1

[0030] like Figure 1 As shown, this invention provides a method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition, the specific steps of which are as follows:

[0031] Step 1: Using MRI images of gliomas as the experimental target, the region of interest is extracted from the preprocessed raw medical images using an image information extraction model based on deep learning and statistical learning. The region of interest is then labeled on the raw data to obtain a standard medical image dataset with ROI.

[0032] Step 1.1: Referring to the authorized invention patent (application number: CN202110519585.8, patent name: 3D Brain Lesion Segmentation Method Based on Optimal Transmission), a deep convolutional neural network segmentation model based on SegResUnet and C2FNAS was constructed. A total of 1251 MRI images of gliomas were collected using publicly available 3D medical image datasets such as BraTS, UCSF, and UPenn. Each case dataset included labels for three tumor regions: WT (wholetumor), TC (tumorcore), and ET (enhanced tumor). These tumor regions represent the ROIs of interest to the physicians. 1000 images were randomly selected as the training set, and the remaining 251 images were used as the test set. The loss function was set to Dice Focal Loss, and the training iterations were set to 1000 times, with each iteration using four sequence images from the case. The training model for WT was set to C2FNAS, while the training models for TC and ET were set to SegResUnet, which can achieve the most accurate tumor segmentation. During the cross-validation phase, the accuracy of WT, TC, and ET reached 93.71%, 91.57%, and 88.19%, respectively. Finally, the online validation accuracy of WT, TC, and ET on BraTS' official platform reached 92.14%, 88.23%, and 84.11%, respectively, which is among the top ten in the world.

[0033] Step 1.2: Considering that the publicly available data used for model learning are all standard data, all original images are processed into a standard image space, maintaining standard pixel spacing, standard dimensions, and standard spatial coordinates. Therefore, before using the trained information extraction algorithm, it is necessary to perform standardized preprocessing on the original medical images: (1) Use the standardized resampling interpolation method to uniformly transform the two-dimensional and three-dimensional image data into a standard three-dimensional pixel space, ensuring that each voxel is 1mm. 3 (2) By using intensity normalization, the gray values ​​are limited to the range of [0, 1000] and the pixel distribution is matched with the t-distribution to minimize the image differences caused by different scanning methods of different institutions; (3) The T2 sequence in each case is set as the target image for registration, and the remaining sequence images are registered to the fixed target image through a rigid registration algorithm to ensure the uniformity of the images of each group of cases; (4) The craniotomy algorithm based on BET algorithm and neural network model is used to extract the brain region from the original MRI image. The remaining part will have a negative impact on ROI extraction; (5) Each brain is transformed to dimension 128 one-to-one through the optimal transfer algorithm. 3The cubes are used as input data for the segmentation model. After completing the above image preprocessing, the processed glioma MRI data are input into the segmentation model using the information extraction algorithm in step 1.1 to predict the WT, TC, and ET regions in cube form for each case, which are the ROIs corresponding to the brain images. It is not necessary to restore them to the original image dimensions.

[0034] Step 2: Using windows of the same size, crop the processed image data to ensure that the ROI in each image can be included in the cropping window as much as possible, resulting in a set of ROI tensor data with reduced resolution. Through a tensor compression mathematical representation method, the ROI tensor is processed into a rank-one vector form to obtain a new type of multimodal dataset.

[0035] Step 2.1: Mark the ROIs in the standardized cube image on the original image, and crop the region using a 64*64*64 dimension window according to the marked area, ensuring that the ROI in each image is included within the cropping window as much as possible. This results in a set of 64*64*64 image tensor data, containing the cropped tensor of the original image and the cropped tensor data of the information-extracted mask image. For each case, T1ce and T2 sequences are selected for ROI cropping, ultimately obtaining three sets of cropped T1ce, T2, and WT labeled images, which will serve as input data for subsequent medical diagnostic experiments.

[0036] Step 2.2: Using a rank-one representation method for nonnegative tensors, for any given sequence, the tensor data T∈R... 64×64×64 , can be used As an approximation of T, where X1∈R 64 ,X2∈R 64 ,X3∈R 64 They are three one-dimensional vectors. The outer product operation between pointers uses the alternating least squares method to compute the optimal rank-one approximation of a tensor, by randomly selecting the initial vector X. 1,0 ∈R 64 ,X 2,0 ∈R 64 ,X 3,0 ∈R 64 In each iteration, two vectors are fixed, and the third vector is updated by solving a least-squares problem related to the tensor T. This process updates all three vectors until the change in the three vectors between two consecutive iterations is less than a certain tolerance, at which point the iteration stops. Assuming the iteration terminates at step k, the result is calculated iteratively. Let X1 = X 1,k X² = X 2,kX3 = X 3,k Due to tensor Since the rank of the vector is 1, it is used as the rank-1 representation of the tensor T. The above algorithm is used to calculate the projections of the cropped 3D image tensor onto the three dimensions, obtaining vectors in the three modalities as the rank-1 representations of the tensor. This constitutes a novel multimodal dataset X = {X1, X2, X3}, where X... i The length of each tensor is 64 for i = 1, 2, and 3, thus achieving information compression of the tensor.

[0037] Step 3: Using the multimodal dataset and its corresponding label set, establish a weighted multimodal optimization problem model, and use a partially orthogonal least squares method to solve the optimization problem, solve for the weight values ​​in the original problem, and calculate the feature matrix corresponding to each group of images based on the weights, as a new type of algebraic feature of magnetic resonance medical images.

[0038] Step 3.1: Utilize the multimodal dataset X and its corresponding label set Y, and perform a centering operation on X for both the dataset and the label set. s =X s H s Y = YH s s = 1, 2, 3, where Construct a block diagonal matrix And the data matrix A=X T Y T YX, consider a label-based mathematical approximation representation W T P s X s A novel weighted multimodal optimization problem model based on image data is established.

[0039] Step 3.2: Solve the optimization problem using a partially orthogonal least squares method to obtain the weight parameters W in the original problem, and then combine the weight values ​​with the eigenvector P. s Substituting the mathematical approximation formula back into the equation, we can calculate the feature matrix corresponding to each group of images. As a novel algebraic feature of magnetic resonance medical imaging.

[0040] Example 2

[0041] like Figure 1 As shown, this invention provides a method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition, which is mainly used for medical diagnosis and classification of glioma MRI images. The specific steps are as follows:

[0042] Step 1: Use real MRI images of gliomas from hospitals, as well as three sets of publicly available data from UCSF, UPenn, and EGD, as experimental data for Example 2 to test the effectiveness of the segmentation model. The remaining implementation steps are the same as in Example 1.

[0043] Step 1.1: Based on PyTorch and Monai libraries, a segmentation model and training program were written using Python. C2FNAS is a segmentation model that adaptively changes network parameters. To achieve the best results, the convolutional layer parameters were fixed to the model parameters corresponding to the highest experimental accuracy during the program writing stage. Considering the missing image sequences in the experimental data, only T2 and T1ce sequences were used as training data. The remaining implementation steps are the same as in Example 1. The segmentation accuracy is shown in Table 1. C2FNAS has the most outstanding WT segmentation effect, while SegResUnet has the most outstanding segmentation accuracy in TC and ET tasks. Although ResUnet has the shortest network training time, considering that the training time does not affect the efficiency of subsequent image processing, SegResUnet will be used to generate TC and ET tumor regions, and C2FNAS will be used to generate WT segments. Dice Focal Loss was used as the loss function for segmentation training, which can effectively improve the learning accuracy of the model in the "class imbalance" semantic segmentation problem. The learning rate was set to 10. -4 The number of iterations is set to 1000 to ensure the convergence of model learning;

[0044] Table 1. Segmentation accuracy of the three models verified by cross-validation and online validation.

[0045]

[0046] Step 1.2: Considering that the publicly available data used for model learning are all standard data, such as... Figure 2 It can be seen that the experimental data in this example are all in their original, unprocessed form, obtained from image scanning machines of different institutions. Therefore, they all need to be processed to a standard image space to maintain standard pixel spacing, standard dimensions, and standard spatial coordinates. Before using the trained information extraction algorithm, the experimental data are standardized and preprocessed using Python programming: (1) A linear interpolation algorithm is used to reconstruct the three-dimensional image of all images, and the pixel spacing in each direction is resampled to 1 mm to ensure that each unit voxel is 1 mm. 3(2) Using the IntensityWindowing algorithm in the SimpleITK library, the grayscale values ​​are mapped to the interval [0, 1000], and the pixel distribution conforms to the standard t-distribution; the remaining implementation steps are the same as in Example 1; the processed image is substituted into the segmentation model in step 1.1 to obtain the segmentation result, and the tensor shape obtained by the optimal transfer algorithm is maintained, without the need to restore to the original brain image, and the dimension is maintained at 128. 3 After obtaining the ROI generation results, 82 cases were randomly selected for segmentation prediction. The results were then submitted to hospital doctors for evaluation. After the doctors repaired the ROIs, the segmentation results were returned as the gold standard for segmentation. By calculating the DICE segmentation accuracy between the gold standard and the model segmentation results, the final results were evaluated as follows: 93% of the images achieved an accuracy of 90% or higher on WT segmentation, and 88% of the images achieved an accuracy of 90% or higher on TC segmentation, reaching a high level.

[0047] Step 2: Since the processed experimental data has the same dimensions, it is only necessary to use a window of the same size to crop its ROI region. The rest of the implementation steps are the same as in Example 1.

[0048] Step 2.1: Use Python programming to perform edge search on the image to find the widest boundary of the ROI and use it as the initial point for cropping. In the case of multiple disconnected ROI regions, directly select the ROI with the largest region as the target for subsequent research. The rest of the implementation steps are the same as in Example 1.

[0049] Step 2.2: Write a tensor rank-one alternating least squares algorithm using MATLAB. Taking T1ce image data as an example, treat it as a tensor T∈R 64×64×64 Alternating iterations yield the rank-one decomposition X1∈R 64 ,X2∈R 64 ,X3∈R 64 The remaining implementation steps are the same as in Example 1, forming a new multimodal dataset X = {X1, X2, X3}, where X i The dimensions of i = 1, 2, and 3 are all 1*64, thus achieving tensor information compression;

[0050] Step 3: For the multimodal datasets of T1ce and T2 and their corresponding label sets, use MATLAB to write a weighted multimodal optimization problem model. The model architecture of the two is the same, the only difference is that the multimodal datasets input to the model are different. The data labels are divided into two types: 0 and 1. The 0 label represents the wild type of IDH gene and the 1 label represents the mutant type of IDH gene. The rest of the implementation steps are the same as in Example 1.

[0051] Step 3.1: Perform centering operations on the multimodal dataset X and its corresponding label set Y: X = XH and Y = YH, where H = I - eeT / N∈R N×N Let I be the identity matrix, and e = [1,…,1]. T ∈R N And N is the number of data points. Furthermore, a multimodal optimization problem model is established.

[0052] Step 3.2: Construct a data matrix A = X using the centered multimodal dataset X and the label set Y. T Y T YX and the block diagonal matrix Find the eigenvectors corresponding to the two largest eigenvalues ​​in the generalized eigenvalue problem AP = CPΛ, and denote the eigenma matrix formed by arranging them as P; use the formula W = (P T CP) -1 P T XY T The weighting parameter W is calculated. Mathematical analysis shows that W and P... s This is the optimal solution to the optimization problem model in step 3.1, where P s satisfy And s = 1, 2, 3. Therefore, for s = 1, 2, 3, It can be viewed as three optimal approximations of the label set Y, and the feature matrix This is the novel algebraic feature proposed in this invention.

[0053] Using glioma image datasets from three different institutions—UCSF, UPenn, and EGD (Erasmus glioma dataset)—a total of 1482 data points with IDH gene classification labels were collected. Applying this patented algorithm to IDH gene classification experiments yielded good results. Specifically, the UCSF and UPenn datasets were used as the training set, and the EGD dataset as the independent validation set. A neural network model was constructed as the classifier. The T1ce and T2 sequences from the glioma image data were used as experimental objects. A set of standard images and their ROIs were obtained through the segmentation model in step 1. Then, through step 2, a multimodal dataset in a standard data format was obtained. The portion of the multimodal data corresponding to the training set was used as the algorithm objective in step 3 to solve for the weight parameters W and feature vector P required for algebraic features. s Substituting this into the entire multimodal data, we can obtain the algebraic features corresponding to each group of cases.

[0054] Table 2. Performance of three feature combination methods in the IDH classification task.

[0055] Omics characteristics 88.51％ 88.25％ 83.18％ Omics + Algebraic Features 89.72％ 89.39％ 90.32％ Algebraic characteristics 89.51％ 88.76％ 90.32％

[0056] In the classification experiment, the classification accuracy of traditional omics features and novel algebraic features were compared. Table 2 shows the three different feature combinations. Figure 3 It is a common deep learning classifier. The results of ACC, AUC and sensitivity after image data enters the classifier are as follows: (1) When using omics features alone, the classification accuracy ACC can reach 89.51%, the AUC can reach 88.25%, and the sensitivity is only 83.18%; (2) When omics features and algebraic features are combined, the ACC can be increased to 89.72%, the AUC can be increased to 89.39%, and the sensitivity can reach 90.32%; (3) When algebraic features are used alone as model input, the ACC is the same as when omics features are used alone, but the AUC is increased to 88.76%, and the sensitivity is also 90.32%; In summary, in the IDH classification task of multi-center data, algebraic features have played a certain positive role in the experiment. Compared with traditional omics features, algebraic features can also provide a reliable basis for the type diagnosis of image data.

[0057] The MRI medical image feature extraction method based on tensor rank-one decomposition can also be applied to medical image data such as meningiomas, brain metastases, and liver tumors. After standardized preprocessing, the optimal transport mapping or ellipsoidal parameterization method can be remembered to achieve tensor standardization of medical images. Based on the rank-one decomposition of tensors, the algebraic features corresponding to each image can be extracted and used as an important indicator for image type determination. In addition to the classification task of IDH genotyping, this feature can also be used for genotyping tasks such as MGMT and 1p / 19q. According to different image types, similar feature approximation optimization problems can be constructed to obtain the corresponding algebraic features for classification. Current segmentation algorithms and medical image preprocessing algorithms are relatively mature. Inputting medical image data can enable automated image processing and ROI extraction. In the future, algebraic feature extraction and image classification and typing will be added. For medical images of various cases, a complete medical image diagnosis and treatment system can be quickly realized from image input to ROI recognition and then to the output of gene classification and typing results. The classification results can be used to provide doctors with reliable diagnosis and treatment directions and plans, thereby significantly improving the medical efficiency of radiologists, increasing the cure rate, and reducing the mortality rate.

[0058] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition, characterized in that, Includes the following steps: S1: Using image information extraction technology, obtain the region of interest (ROI) from the original MRI medical images; S2: Cropping ROIs from the original image at a fixed size to obtain new tensor data, and then processing the original tensor containing ROIs using a mathematical tensor representation method to obtain a new multimodal dataset. Step 2.1: Mark the ROIs in the standardized image, and crop the region according to the marked area using a window of the same size to ensure that the ROI in each image can be included in the cropping window, and obtain a set of image tensor data with reduced resolution; Step 2.2: Based on the characteristics of medical image tensor data, a rank-one representation method for non-negative tensors is used to calculate the projections of the cropped 3D image tensor in three dimensions, obtaining vectors in three modalities as the rank-one representations of the tensors, thus constructing a novel multimodal dataset. ; S3: Using a multimodal dataset and label set, a weighted multimodal optimization problem model is established, and the optimization problem is solved to obtain a novel algebraic feature of magnetic resonance medical images; Step 3.1: Using the multimodal dataset X and its corresponding label set Y, consider a label-based mathematical approximation representation. , A novel weighted multimodal optimization problem model based on image data is established. ; Step 3.2: Solve the optimization problem using the least squares method to obtain the weight parameters W in the original problem, and then combine the weight values ​​with the eigenvectors. Substituting the mathematical approximation formula back into the equation, we can calculate the feature matrix corresponding to each group of images. As a novel algebraic feature of magnetic resonance medical imaging.

2. The method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition according to claim 1, characterized in that, Specifically, it includes the following steps: Step 1: Using an image information extraction model based on deep learning and statistical learning, extract the region of interest from the preprocessed raw medical images and label it on the raw data to obtain a standard medical image dataset with ROI. Step 2: Crop the image data in the standard dataset into windows of the same size to ensure that the ROI in each image can be included in the cropping window, and obtain a set of ROI tensor data with reduced resolution. Through a tensor compression mathematical representation method, process the ROI tensor into a rank-one vector form to obtain a new type of multimodal dataset. Step 3: Using the multimodal dataset and its corresponding label set, establish a weighted multimodal optimization problem model, and use a special least squares method to solve the optimization problem, solve for the weight values ​​in the original problem, and calculate the feature matrix corresponding to each group of images based on the weights, as a new type of algebraic feature of magnetic resonance medical images.

3. The method for extracting magnetic resonance medical image features based on tensor rank-one decomposition according to claim 2, characterized in that, Step 1 includes the following steps: Step 1.1: Using numerous publicly available medical image datasets, train an image information extraction model based on deep learning and statistical learning. By comparing the performance of existing models, select the model with the best validation accuracy as the information extraction algorithm. Step 1.2: Standardize the original medical images by transforming them into a standard three-dimensional pixel space. Using the information extraction algorithm from Step 1.1, input the standardized image data into the model to obtain the ROI prediction for each group of images.

4. The method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition according to claim 1, characterized in that, It is applied in the fields of medical image classification and diagnosis, three-dimensional image processing, and information extraction.

5. The method for extracting features from magnetic resonance medical images based on tensor rank-one decomposition according to claim 4, characterized in that, In the field of medical image classification and diagnosis, algebraic features are uniformly extracted from standardized image data of different types and attributes. The image data and feature data are used as inputs to the classifier model to accurately predict the type of medical image and provide doctors with a basis for diagnosis.

6. The method for extracting magnetic resonance medical image features based on tensor rank-one decomposition according to claim 4, characterized in that, In the field of 3D image processing and information extraction, unlike commonly used brightness, color, texture, edge and various statistical features, this method tensors the image data to process it into a rank-one representation, thereby compressing the image information. At the same time, based on the rank-one information and category information of the known data, it provides an algebraic feature that represents the category attribute of the image itself, which can serve as an effective guiding parameter for image analysis and processing.

Images