Random forest method, system and terminal for marking magnetic resonance spine image based on multi-feature fusion

Abstract

The application provides a kind of random forest magnetic resonance spine image marking method, system and terminal based on multi-feature fusion, by pre-processing and feature extraction processing to the magnetic resonance spine image to be marked obtain caudal vertebra feature and other vertebral body feature, then based on the spine recognition model obtained by training random forest, obtain the magnetic resonance spine recognition result from the extracted feature, and combine the position relationship of each vertebral body in three-dimensional space, finally obtain the magnetic resonance spine marking result.The application adopts trained random forest to identify vertebral body, not only the model training and prediction time is less, but also the hardware requirement is lower, in the case without GPU, also can obtain faster prediction speed and high-precision prediction result.And through the position relationship of each vertebral body in three-dimensional space, vertebral body marking optimization is carried out, which increases the marking accuracy and improves the algorithm robustness, and is better suitable for spine discontinuity, image blur, image deformation and metal implant or non-autonomous motion caused artifact etc.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image data processing, in particular to a random forest spine image labeling method, system and terminal based on multi-feature fusion. BACKGROUND

[0002] The spine is an important part of the human musculoskeletal system, but due to incorrect posture and other reasons, the number of people of all ages suffering from spine-related diseases is increasing. Magnetic resonance imaging (MRI) technology is one of the effective methods for checking human spine diseases. However, in the process of treatment and diagnosis, manual labeling of MR images of the spine not only consumes a lot of time, but also may have labeling errors due to personnel differences. Fast and accurate positioning and segmentation of vertebrae not only saves time and reduces personnel costs, but also helps doctors diagnose spinal diseases, determine treatment plans and evaluate the effectiveness of the plans.

[0003] Segmentation of the target spine image is one of the important links. Traditional image segmentation methods, such as region growing, threshold segmentation, and pixel value-based segmentation, mainly segment different regions of the image according to image gray scale and boundary gradient information. However, for spine MRI image labeling, the existing technology has the following difficulties:

[0004] (1) These methods are mostly used for computed tomography (CT) images. For spine MRI images, due to the presence of a large amount of soft tissue information, the boundary may be unclear and not distinct;

[0005] (2) For magnetic resonance images deviating from the center layer, the vertebral body features are not obvious, making it difficult to segment the vertebral body.

[0006] (3) Due to motion, metal, and magnetic field inhomogeneity, magnetic resonance spine images may have deformation and artifacts, which is one of the difficulties in spine labeling.

[0007] (4) Due to scoliosis patients and spine boundary layers, the spine display is discontinuous, and there is also a risk of missing labeling and mislabeling. SUMMARY

[0008] In view of the above-mentioned shortcomings of the prior art, the present application aims to provide a random forest spine image labeling method, system and terminal based on multi-feature fusion, which solves the above-mentioned problems of the prior art.

[0009] To achieve the above object and other related objects, the present application provides a random forest magnetic resonance spine image marking method based on multi-feature fusion, which comprises: pre-processing a magnetic resonance spine image to be marked to obtain a pre-processed image; performing feature extraction processing according to the pre-processed image to obtain corresponding caudal vertebra features and other vertebra features; obtaining a magnetic resonance spine recognition result according to the caudal vertebra features and the other vertebra features based on a spine recognition model obtained by training a random forest; performing vertebra marking according to the magnetic resonance spine recognition result based on the positional relationship of each vertebra in a three-dimensional space calculated from the positional information of each vertebra in each layer of the magnetic resonance image, to obtain a magnetic resonance spine marking result.

[0010] In an embodiment of the present application, the training method of the spine recognition model comprises: pre-processing a plurality of magnetic resonance spine images with caudal vertebra and other vertebra labels to obtain pre-processed images corresponding to each magnetic resonance spine image; performing feature extraction processing on each pre-processed image to extract the caudal vertebra features and the other vertebra features corresponding to each magnetic resonance spine image to obtain a feature data training set; training a random forest based on the feature data training set to obtain the spine recognition model.

[0011] In an embodiment of the present application, the pre-processing comprises: sequentially performing image filtering, image enhancement, image sharpening and image mask extraction on the input magnetic resonance spine image to obtain the pre-processed image.

[0012] In an embodiment of the present application, the feature extraction processing comprises: performing HOG feature extraction according to the pre-processed image, and performing 2DPCA dimension reduction on the extracted HOG features to obtain reduced dimension spine HOG features; performing GLCM feature extraction according to the pre-processed image to obtain gridized GLCM features; serially fusing the reduced dimension HOG features and the gridized GLCM features to obtain the caudal vertebra features and taking the reduced dimension HOG features as the other vertebra features.

[0013] In an embodiment of the present application, the HOG feature extraction method comprises: performing unit segmentation according to the pre-processed image, and calculating the gradient and amplitude of each unit; uniformly dividing the gradient and amplitude of each unit into a plurality of gradient directions, and accumulating the amplitudes of the same gradient direction of each unit according to a weight to form a gradient histogram of each unit; normalizing each unit, and connecting the gradient histograms to obtain the corresponding HOG features.

[0014] In an embodiment of the present application, the corresponding gray level co-occurrence matrix is calculated according to the preprocessed image; the feature parameters are calculated based on the gray level co-occurrence matrix; wherein the feature parameters include: the feature parameters for representing the uniformity and roughness of the image gray level distribution, the feature parameters for representing the depth and thickness of the image texture, the feature parameters for representing the image information amount, the feature parameters for representing the image texture brightness information, and the feature parameters for representing the image gray level correlation; based on the window of the set size, the window sliding is performed according to the preprocessed image, and the sliding window features of each window are calculated based on the calculated feature parameters to form the grid GLCM features.

[0015] In an embodiment of the present application, the magnetic resonance spine recognition result obtained according to the caudal vertebra feature and the other vertebra features includes: identifying the caudal vertebra based on the caudal vertebra feature to obtain the spine recognition result corresponding to the identified caudal vertebra; identifying the other vertebra in the spine recognition result corresponding to the identified caudal vertebra based on the other vertebra features to obtain the magnetic resonance spine recognition result corresponding to the identified all vertebrae.

[0016] In an embodiment of the present application, the position relationship of each vertebra in the three-dimensional space calculated based on the position information of each vertebra in each layer of the magnetic resonance image, the vertebra labeling according to the magnetic resonance spine recognition result to obtain the magnetic resonance spine labeling result includes: calculating the IOU value according to the magnetic resonance spine recognition result, and removing the redundant information in the magnetic resonance spine recognition result; calculating the position relationship of each vertebra in the three-dimensional space based on the position information of each vertebra in the magnetic resonance spine recognition result after removing the redundant information and the position information of each vertebra in the other layers of the magnetic resonance spine image; labeling each vertebra in the magnetic resonance spine recognition result after removing the redundant information based on the position relationship of each vertebra in the three-dimensional space to obtain the magnetic resonance spine labeling result.

[0017] To achieve the above object and other related objects, the present application provides a random forest magnetic resonance spine image labeling system based on multi-feature fusion, which comprises: an image preprocessing module for preprocessing the magnetic resonance spine image to be labeled to obtain a preprocessed image; a feature extraction module connected to the image preprocessing module, for performing feature extraction processing according to the preprocessed image to obtain corresponding caudal vertebra features and other vertebra features; a vertebra recognition module connected to the feature extraction module, for obtaining a magnetic resonance spine recognition result according to the caudal vertebra features and the other vertebra features based on a spine recognition model obtained by training a random forest; and a labeling module connected to the vertebra recognition module, for performing vertebra labeling according to the magnetic resonance spine recognition result based on the position relationship of each vertebra in the three-dimensional space calculated based on the position information of each vertebra in each layer of the magnetic resonance image to obtain a magnetic resonance spine labeling result.

[0018] To achieve the above object and other related objects, the present application provides a random forest magnetic resonance spine image labeling terminal based on multi-feature fusion, comprising one or more memories and one or more processors; the one or more memories are used for storing a computer program; the one or more processors are connected to the memories and are used for running the computer program to execute the random forest magnetic resonance spine image labeling method based on multi-feature fusion.

[0019] As described above, the present application is a random forest magnetic resonance spine image labeling method, system and terminal based on multi-feature fusion, which has the following beneficial effects: the present application obtains caudal vertebra features and other vertebra features through preprocessing and feature extraction processing of the magnetic resonance spine image to be labeled, obtains a magnetic resonance spine recognition result from the extracted features based on a spine recognition model obtained by training a random forest, and finally obtains a magnetic resonance spine labeling result by combining the positional relationship of each vertebra in a three-dimensional space. The present application uses a trained random forest to recognize vertebrae, which not only has less model training and prediction time, but also has lower hardware requirements, and can obtain faster prediction speed and high-precision prediction results without GPU assistance. Moreover, the vertebrae are labeled and optimized by the positional relationship of each vertebra in a three-dimensional space, which increases the labeling accuracy and improves the algorithm robustness, and is better suitable for labeling work in the case of discontinuous spine, image blur, image deformation, metal implant or non-autonomous motion artifacts, etc. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of a random forest magnetic resonance spine image labeling method based on multi-feature fusion in an embodiment of the present application is shown.

[0021] Figure 2 A flowchart of GLCM feature extraction in an embodiment of the present application is shown.

[0022] Figure 3 A flowchart of the training mode of a spine recognition model in an embodiment of the present application is shown.

[0023] Figure 4 A model training flowchart in an embodiment of the present application is shown.

[0024] Figure 5 A magnetic resonance spine image spine labeling diagram in an embodiment of the present application is shown.

[0025] Figure 6 A structure diagram of a random forest magnetic resonance spine image labeling system based on multi-feature fusion in an embodiment of the present application is shown.

[0026] Figure 7A structure diagram of a random forest magnetic resonance spine image marking terminal based on multi-feature fusion is shown in an embodiment of the application. DETAILED DESCRIPTION

[0027] Other advantages and benefits of the present application will become apparent to those skilled in the art upon reading the following detailed description of embodiments of the present application. The present application can be implemented or applied in other different embodiments and the details of the present description can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0028] It should be noted that in the following description, reference is made to the accompanying drawings, which form a part of the disclosure. It is understood that other embodiments can be utilized and mechanical, structural, electrical, and operational changes can be made without departing from the spirit and scope of the present application. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the embodiments of the present application are defined only by the appended patent claims. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. Spatially relative terms, such as "upper", "lower", "left", "right", "below", "below", "bottom", "top", and the like, can be used herein for ease of description to describe one element or feature's relationship to another element(s) or feature(s) as illustrated in the figures.

[0029] Throughout this specification, when it is said that a certain part is "connected" to another part, it includes not only the case of "direct connection" but also the case of "indirect connection" in which other elements are interposed therebetween. In addition, when it is said that a certain part "includes" a certain constituent element, unless it is specifically stated to the contrary, other constituent elements are not excluded, but it means that other constituent elements can be further included.

[0030] The first, second, and third terms mentioned therein are used to explain various parts, components, regions, layers, and / or sections, but are not limited thereto. These terms are only used to distinguish a certain part, component, region, layer, or section from other parts, components, regions, layers, or sections. Therefore, the first part, component, region, layer, or section described below can be referred to as the second part, component, region, layer, or section within the scope of the present application.

[0031] Also, as used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context indicates otherwise. It will be further understood that the terms "comprises", "comprising", "includes" and / or "including" when used herein, specify the presence of stated features, operations, elements, components, items, and / or groups but do not preclude the presence or addition of one or more other features, operations, elements, components, items, and / or groups. As used herein, the terms "or" and "and / or" are to be interpreted as inclusive, i.e., as meaning one or any combination of the items. Thus, "A, B or C" or "A, B and / or C" means any of the following: A; B; C; A and B; A and C; B and C; A, B and C. Only when the combination of elements, functions, or operations are inherently mutually exclusive is an exception to this definition presented.

[0032] The addition of machine learning provides a new direction for image segmentation, through the learning of image texture, contour, line feature information, surface feature information, statistical features, etc. level, better achieve the target segmentation extraction of specific area. Dong X et al. proposed a method combining scale invariant feature transform matching algorithm (SIFT) and binary classification model (SVM) to classify vertebrae, realize automatic detection of vertebrae; Wu Yuhao detects the vertebrae contour through the model, and combines the particle filter algorithm to locate the intervertebral disc; and Wu Xiangyuan et al. use random forest algorithm to realize automatic detection of CT image vertebrae. In recent years, deep learning has been increasingly applied to magnetic resonance image recognition. Shumao Pang applied deep learning algorithm to spinal MR image processing, and achieved good application effect. However, deep learning has large amount of calculation, and it often needs high hardware equipment, and training and recognition time. The deep learning used in the prior art for labeling makes it have more resource consumption compared with machine learning while having good recognition accuracy.

[0033] In machine learning algorithms, the random forest algorithm is widely used in face recognition, license plate recognition, medical image segmentation and other fields. Xue Feng proposed a classification method for large-scale vehicle logo images, which combines SVM and random forest to realize the recognition of vehicle logo images; Guo JinXin and Coote applied the random forest algorithm to face recognition and obtained high recognition rate; Ben Glocker applied the random forest algorithm to CT images to realize automatic positioning and recognition of vertebrae. In summary, the random forest algorithm has good application in image segmentation and recognition, and has achieved certain application results. In the process of training and identifying vertebrae, the selection of features is very important. Common feature extraction algorithms such as SIFT, local binary pattern (LBP), histogram of oriented gradient (HOG), gray level co-occurrence matrix (GLCM), Haar feature template (Haar) and the like, among which HOG and GLCM can better describe the information of the spine.

[0034] Therefore, the present application provides a random forest magnetic resonance spine image labeling method based on multi-feature fusion. The caudal vertebra features and other vertebra features are obtained by preprocessing and feature extraction processing of the magnetic resonance spine image to be labeled. Then, the spine recognition model obtained by training the random forest is used to obtain the magnetic resonance spine recognition result from the extracted features, and the position relationship of each vertebra in the three-dimensional space is calculated to finally obtain the magnetic resonance spine labeling result. The trained random forest is used for vertebra recognition in the present application, which not only has less model training and prediction time, but also has lower hardware requirements. Without the help of GPU, the prediction speed and high-precision prediction result can be obtained. Moreover, the vertebra labeling is optimized by the position relationship of each vertebra in the three-dimensional space, which increases the labeling accuracy and improves the robustness of the algorithm, and is better suitable for labeling work in the case of discontinuous spine, image blur, image deformation, metal implant or pseudo image caused by involuntary movement.

[0035] The embodiments of the present application will be described in detail below with reference to the accompanying drawings. The present application can be embodied in various different forms, and is not limited to the embodiments described herein.

[0036] As Figure 1 A flowchart of a random forest magnetic resonance spine image labeling method based on multi-feature fusion in an embodiment of the present application is shown.

[0037] The method comprises:

[0038] Step S1: Preprocessing the magnetic resonance spine image to be labeled to obtain a preprocessed image.

[0039] In an embodiment, the preprocessing comprises: sequentially performing image filtering, image enhancement, image sharpening and image mask extraction on the input magnetic resonance spine image to obtain a preprocessed image.

[0040] The image filtering is mainly for processing noise, and can be Gaussian filtering, median filtering, etc. The image mask extraction is to remove the background region and retain the main information for calculation.

[0041] Step S2: performing feature extraction processing according to the preprocessed image to obtain corresponding caudal vertebra features and other vertebra features.

[0042] In an embodiment, step S2 comprises:

[0043] performing MSSW processing on the preprocessed image; specifically, the MSSW processing is performed by using a multiple-scales sliding window method (MSSW) to scale the preprocessed image at different scales, search and extract features; the scaling can be performed by interpolation or convolution. The MSSW processing can be well applied to images with different fields of view (FOV) and pixel spacings (Pixel Spacing).

[0044] performing feature extraction processing on the preprocessed image after the MSSW processing.

[0045] In an embodiment, since it is necessary to well represent the characteristics of the target image, i.e., to well distinguish the target region from the non-target region. The traditional single feature information often cannot fully represent the characteristics of the target region, and using more features will also cause redundancy, which not only increases the calculation time, but also may cause overfitting phenomenon, resulting in poor recognition effect.

[0046] Therefore, the feature dimension reduction and feature fusion are used. Specifically, the LBP feature of the image can be calculated, the feature information is recalculated to generate a new feature, or the original image is generated to a co-occurrence matrix, and the co-occurrence matrix is further generated to a histogram feature. The feature dimension reduction described above can be one-dimensional dimension reduction or two-dimensional dimension reduction. The two-dimensional dimension reduction can be performed on a single feature or a fused feature.

[0047] The feature extraction processing comprises:

[0048] The pre-processed image after MSSW processing is subjected to HOG feature extraction, and the extracted HOG feature is subjected to 2DPCA dimension reduction to obtain the dimension-reduced spine HOG feature; it should be noted that the feature dimension reduction described above can be one-dimensional dimension reduction or two-dimensional dimension reduction; preferably, two-dimensional dimension reduction is adopted, that is, single feature dimension reduction or fused feature calculation can be performed.

[0049] The pre-processed image after MSSW processing is subjected to GLCM feature extraction to obtain the grid GLCM feature; the specific performance can be to calculate the image LBP feature, recalculate the co-occurrence matrix of the feature information, generate a new feature, or generate a co-occurrence matrix from the original image, and further generate a histogram feature from the co-occurrence matrix.

[0050] The dimension-reduced HOG feature and the grid GLCM feature are serially fused to obtain the tail vertebra feature, and the dimension-reduced HOG feature is used as the feature of other vertebrae. Specifically, the HOG feature after 2DPCA dimension reduction is serially fused with the grid GLCM feature, and is used as the feature of the tail vertebra part for calculation. The HOG feature after 2DPCA dimension reduction is used for calculation of other vertebrae parts. The feature fusion described herein includes fusion of two or more features, which can be stacked in series, or the two features can be fused in terms of calculation method. For example, the input of the nth feature can be one or more outputs of the first n-1 features.

[0051] In a specific embodiment, the pre-processed image is subjected to HOG feature extraction, and the extracted HOG feature is subjected to 2DPCA dimension reduction to obtain the dimension-reduced spine HOG feature, which mainly includes two parts, one is HOG feature extraction, and the other is 2DPCA dimension reduction of the extracted HOG feature;

[0052] Specifically, the HOG feature extraction method includes:

[0053] The pre-processed image after MSSW processing is segmented into a plurality of cells, and the gradient and amplitude of each cell are calculated;

[0054] The gradient and amplitude of each cell are uniformly divided into a plurality of gradient directions, and the amplitudes of the same gradient direction of each cell are accumulated according to the weight to form the gradient histogram of each cell; specifically, the calculated gradient and amplitude of each cell are uniformly divided into f directions, i.e. bins, and the gradient direction adopted in this patent has no positive and negative distinction, and the direction f is set to 9, i.e. 180° is divided into 9 intervals. The amplitudes of all points in the same gradient direction of the same cell are accumulated according to the weight to form the corresponding gradient histogram; let G(x, y) represent the gradient amplitude of the image pixel point, and θ(x, y) represent the gradient direction, then:

[0055]

[0056] θ(x, y) = tan -1 (G y (x, y) / G x (x, y)) (2)

[0057] Each cell is normalized and the gradient histograms are concatenated to obtain the corresponding HOG feature.

[0058] The spine image HOG feature has certain redundancy, if directly used, not only increases the training and prediction time, but also may cause overfitting due to more redundant information, reduces the vertebral body segmentation accuracy, thus needs to be dimensionally reduced. Principal Component Analysis (PCA) retains the trend and pattern while simplifying the complexity of high-dimensional data. And 2DPCA can obtain the covariance matrix of the image by directly acting on the original image, which is more accurate and efficient than PCA.

[0059] Specifically, the way of 2DPCA dimension reduction of the extracted HOG feature includes:

[0060] Based on the HOG feature as a sample A, the projection space is X, and the projection feature vector Y is represented as:

[0061] Y = AX; (3)

[0062] The key point of 2DPCA is to find the best projection space X.

[0063] J(C) = X T E{(X-E(X)) T (X-E(X))}; (4)

[0064] That is

[0065] J(X) = tr(S x ) = X T G t X; (5)

[0066] In the formula, S x is the covariance matrix of Y, tr(S x ) is the trace of S x , G t represents the covariance matrix of the sample, so when J(X) is maximized, the corresponding projection space X is the best projection space. G t The first m largest eigenvalues correspond to eigenvectors X1, X2, …, Xm m , which satisfy the standard orthogonal condition, that is:

[0067]

[0068] Then the sample image A is projected to the best vector X1, X2, …, X m The projection feature vector obtained by projection:

[0069] Y i = A [X1, X2, …, X m ]; (7)

[0070] In a specific embodiment, the Gray level co-occurrence matrix (GLCM) determines the texture information of the image by describing the relationship between the pixel points, and can better describe the texture uniformity, thickness and other information.

[0071] Therefore, according to the pre-processed image, GLCM feature extraction is performed to obtain the grid GLCM feature; the GLCM feature extraction method includes:

[0072] According to the pre-processed image processed by the MSSW, the corresponding gray level co-occurrence matrix is calculated; specifically, in order to reduce the calculation cost, 16 gray levels are adopted, and θ takes values of 0°, 45°, 90° and 135°, and P(i,j) represents the number of i and j, and the corresponding gray level co-occurrence matrix is calculated by the existing gray level co-occurrence matrix calculation method;

[0073] Based on the gray level co-occurrence matrix, the feature parameters are calculated; wherein, the feature parameters include: feature parameters for representing the uniformity and roughness of the image gray distribution, feature parameters for representing the depth and thickness of the image texture, feature parameters for representing the image information amount, feature parameters for representing the image texture brightness information, and feature parameters for representing the image gray correlation;

[0074] Specifically, the calculated feature parameters include:

[0075] 1. Angular Second Moment (ASM), used to represent the uniformity and roughness of the image gray distribution, and the calculation formula is:

[0076] ASM = ∑ i ∑ j P(i,j) 2 ; (8)

[0077] 2. Contrast is the contrast information of the pixel and its field, and is used to represent the depth and thickness of the image texture, and the calculation formula is:

[0078] CON = ∑ i ∑ j (i-j) 2 P(i,j); (9)

[0079] 3. Homogeneity is used to represent the change of local texture, and the calculation formula is:

[0080]

[0081] 4. Entropy is used to represent the information amount of the image, and the calculation formula is:

[0082] ENT = -∑ i ∑ j P(i, j) log P(i, j); (11)

[0083] 5. Average is used to represent the brightness information of the texture, and the calculation formula is:

[0084]

[0085] 6. Correlation is used to represent the gray correlation of the image, and the calculation formula is:

[0086]

[0087] In the formula, u1 = ∑ i i∑ j P(i, j), and u2 = ∑ j j∑ i P(i, j).

[0088] In order to reduce the redundant information of GLCM, the vertebral feature extraction is performed in a gridding manner in the application, and the specific implementation manner is as follows:

[0089] Based on the window with a set size, the window sliding is performed on the image, and the sliding window features of each window are calculated based on the calculated feature parameters, so as to compose the gridding GLCM features. Specifically, as Figure 2 , a sliding window with a size of n*m is set, and the entire image is traversed with d x , d y as the step length, and the window can be overlapped in the sliding process. The mean value of the feature values in the window at each sliding window is calculated, that is, the mean value of the above six feature parameters in the window, the feature parameters in all areas passed by the sliding window are linked, and the gridding GLCM features of the image are composed.

[0090] Step S3: based on the vertebral recognition model obtained by the training random forest, the magnetic resonance vertebral recognition result is obtained according to the tail vertebra feature and other vertebral features.

[0091] In an embodiment, as Figure 3 , the training manner of the vertebral recognition model comprises:

[0092] First step: a plurality of magnetic resonance spine images with caudal vertebrae and other vertebrae labels are respectively preprocessed as sample data, and preprocessed images corresponding to each magnetic resonance spine image are obtained; specifically, the input each magnetic resonance spine image is sequentially subjected to image filtering, image enhancement, image sharpening and image mask extraction to obtain the preprocessed image corresponding to each magnetic resonance spine image.

[0093] Second step: according to each preprocessed image, feature extraction processing is performed to extract the caudal vertebrae features and other vertebrae features corresponding to each magnetic resonance spine image to obtain a feature data training set; specifically, HOG feature extraction is performed on each preprocessed image, and 2DPCA dimension reduction is performed on the extracted HOG features to obtain the dimension-reduced spine HOG features corresponding to each magnetic resonance spine image; GLCM feature extraction is performed on each preprocessed image to obtain the grid-based GLCM features corresponding to each magnetic resonance spine image; the dimension-reduced HOG features and the grid-based GLCM features of each magnetic resonance spine image are serially fused to obtain the caudal vertebrae features corresponding to the magnetic resonance spine image, and the dimension-reduced HOG features are taken as the other vertebrae features, and finally the feature data training set is formed, that is, the caudal vertebrae features and the other vertebrae features of each magnetic resonance spine image. It should be noted that the HOG feature extraction and GLCM feature extraction adopted in this embodiment are similar to the HOG feature extraction and GLCM feature extraction in the manner of marking the magnetic resonance spine image to be marked described in the above embodiment, and therefore will not be described here.

[0094] Wherein, the 2DPCA dimension reduction is performed in the following manner:

[0095] Suppose the sample data set is A = {A1, A2, …, A n}, the projection space is X, and the projection feature vector Y is represented as:

[0096] Y = AX (8)

[0097] The key point of 2DPCA is to find the best projection space X.

[0098] J(X) = X T E{(X-E(X)) T (X-E(X))} (9)

[0099] That is

[0100] J(X) = tr(S x ) = X T G t X (10)

[0101] In the formula, S x is the covariance matrix of Y, tr(S x ) is the trace of S x , and Gt where X represents the covariance matrix of the samples, then the projection space X corresponding to the maximum J(X) is the optimal projection space.

[0102] G t The first m largest eigenvalues correspond to eigenvectors X1, X2, …, X m , and at the same time satisfy the standard orthogonal, that is:

[0103]

[0104] Then the sample image A i to the optimal vector X1, X2, …, X m The projection feature vector obtained by projection is:

[0105] Y i = A i [X1, X2, …, X m ] (14)

[0106] Step 3: training a random forest based on the feature data training set to obtain the spine recognition model.

[0107] Random forest is an algorithm based on decision tree, which integrates multiple trees through the Bagging idea of ensemble learning. As the name implies, it uses a large number of "trees" and forms a "forest" by selecting a training set, generating a decision tree, and voting for the decision tree. Three steps realize classification and have high generalization ability.

[0108] The spine recognition model trained by the application adopts feature fusion random forest (FMRF), and the recognition process is as follows: first, the identification of the remaining vertebrae is completed on the basis of the identification of the tail vertebrae, so the identification of the tail vertebrae is particularly important.

[0109] When extracting features of the tail vertebrae, both image texture features and boundary information should be included, and the redundant part of the image features should be extracted to prevent overfitting and long calculation time. In addition, since the features of the tail vertebrae S1 are similar to those of the lumbar vertebrae and thoracic vertebrae, if all the vertebrae are labeled as the same label, the tail vertebrae recognition label deviation is large, so in the calculation, the tail vertebrae part is identified separately. Therefore, the HOG features of each magnetic resonance spine image in the feature data training set after 2DPCA dimension reduction are adopted, the features are serially fused with the grid GLCM features, and the obtained tail vertebrae features are used to train the tail vertebrae recognition part of the model. The HOG features of each magnetic resonance spine image after tail vertebrae recognition are used as other vertebrae features after 2DPCA dimension reduction, and the other tail vertebrae recognition part of the model is trained.

[0110] In an embodiment, obtaining the magnetic resonance spine recognition result according to the caudal vertebra feature and the other vertebra features comprises: identifying the caudal vertebra based on the caudal vertebra feature to obtain a spine recognition result corresponding to the identified caudal vertebra; and identifying the other vertebrae in the spine recognition result corresponding to the identified caudal vertebra based on the other vertebra features to obtain a magnetic resonance spine recognition result corresponding to all the identified vertebrae.

[0111] Step S4: based on the positional relationship of each vertebra in the three-dimensional space calculated from the positional information of each vertebra at each layer of the magnetic resonance image, marking the vertebrae according to the magnetic resonance spine recognition result to obtain a magnetic resonance spine marking result.

[0112] In an embodiment, step S4 comprises:

[0113] According to the IOU value calculated from the magnetic resonance spine recognition result, the redundant information in the magnetic resonance spine recognition result is removed based on the calculated IOU value; the IOU value is the intersection over union, by calculating the intersection over union value of the identified region, the identified result with larger overlapping information is deleted, and the redundancy is reduced.

[0114] Based on the positional information of each vertebra in the magnetic resonance spine recognition result after removing the redundant information and the positional information of each vertebra in the other magnetic resonance spine images of each layer, the positional relationship of each vertebra in the three-dimensional space is calculated; specifically, since the magnetic resonance spine image to be marked is only a 2D image of a single layer of the magnetic resonance image, the magnetic resonance spine recognition result thereof is also only the identification result of each vertebra of a single layer.

[0115] In actual application, different magnetic resonance images of the spine may have different FOV, PixelSpacing and other parameters, so that the sizes of the vertebrae are different. In addition, since there is a discontinuous display of single-layer vertebrae for non-central layer images and patients with scoliosis, only marking from the current 2D image may easily result in incorrect marking. Therefore, the spatial position information of each vertebra of the magnetic resonance spine image of the other layers collected and the spatial position information of each vertebra identified in the current layer are used to calculate the positional relationship of each vertebra in the three-dimensional space, and based on the positional relationship, the vertebrae can be more accurately marked;

[0116] Specifically, the vertebrae identified in the magnetic resonance spine recognition result after removing the redundant information are used to calculate the spatial position, the slice image of the current sequence is counted, the positional relationship of each vertebra in the three-dimensional space is calculated, and the spine marking work is completed. Here, the counting mainly refers to arranging the identification result after removing the redundant information, combining the 3D spatial information, and further processing.

[0117] Based on the positional relationship of each vertebra in the three-dimensional space, each vertebra in the magnetic resonance spine recognition result after removing the redundant information is marked to obtain a magnetic resonance spine marking result.

[0118] The advantages of marking the positional relationship of each vertebral body in three-dimensional space include: for discontinuous images of the vertebral body, the images can be correctly and completely marked; for images in which the boundary of the vertebral body is blurred, positioning is difficult, and most of the features of the vertebral body are similar and difficult to distinguish, the images can also be relatively accurately marked based on the reference of other images; for deformation and artifacts of magnetic resonance spine images due to motion, metal, and uneven magnetic field, the images can also be marked by referring to other images.

[0119] In order to better describe the random forest magnetic resonance spine image marking method based on multi-feature fusion, the following specific embodiments are provided for description;

[0120] Embodiment 1: A method for automatically marking a vertebral body of a magnetic resonance spine image.

[0121] There are the following difficulties in marking a spine magnetic resonance image: the image contains a lot of soft tissue information, the boundary of the vertebral body is blurred, and positioning is difficult; the image may have deformation due to uneven magnetic field; due to metal, motion, and other problems, the image may have artifacts, and part of the information of the tail vertebra is missing; most of the features of the vertebral body are similar and difficult to distinguish, the vertebral body in the rectangular frame has similar features; the spine of a scoliosis patient and the boundary layer of the spine are not continuous. And the traditional image segmentation has certain limitations, and the segmentation method based on deep learning has higher requirements for hardware, and the training and recognition time is relatively increased.

[0122] In view of the above problems, a method for automatically marking a vertebral body of a magnetic resonance spine image is provided.

[0123] The method includes two parts, one is the model training part, and the other is the image recognition and marking part;

[0124] Among them, as Figure 4 The specific implementation of the model training part includes:

[0125] 1. Image preprocessing: In order to better obtain image information and reduce noise images, first, the sample images with vertebral body labels are preprocessed, including (1) image filtering, such as Gaussian filtering, median filtering, etc., for processing noise points; (2) image enhancement, (3) image sharpening, (4) image mask extraction, that is, removing the background area and retaining the main information for calculation.

[0126] 2. Image feature extraction; mainly including HOG feature extraction, GLCM feature extraction, 2DPCA dimension reduction, and feature fusion to obtain a feature set.

[0127] Among them, the HOG feature extraction step is as follows:

[0128] Step 1: Divide the image region to be extracted into several cells, and calculate the gradient and amplitude of each cell.

[0129] Step 2: Divide the gradient and amplitude calculated in each cell into f directions, i.e. bins. The gradient direction used in this patent does not distinguish between positive and negative directions, and the direction f is set to 9, i.e. 180° is divided into 9 intervals. The amplitudes of all points with the same gradient direction in the same cell are added according to the weight to form the corresponding gradient histogram.

[0130] Step 3: Normalize each cell and connect the histograms to obtain the HOG feature result.

[0131] The GLCM feature extraction steps are as follows:

[0132] Step 1: Use 16 gray levels, and set θ to 0°, 45°, 90°, and 135°. Let P(i,j) represent the number of i and j occurrences. The feature parameters calculated by the gray level co-occurrence matrix include Angular Second Moment (ASM), which represents the uniformity and roughness of the image gray scale distribution; Contrast, which represents the contrast information of the pixels and their fields, represents the depth and thickness of the image texture; Homogenity reflects the local texture change; Entropy represents the information amount of the image; Average reflects the brightness information of the texture; Correlation reflects the image gray scale correlation. The above parameters are obtained using formulas (8)-(14).

[0133] Step 2: Reduce GLCM redundancy information, and use a grid method to extract the vertebral feature, i.e. set a sliding window with size n*m, and traverse the entire image with steps d x 、d y , and the window can overlap during sliding. Calculate the mean value of the feature values in the window at each sliding step, i.e. the mean value of the above six feature parameters in the window, link the feature parameters in all regions passed by the sliding window, and form the grid GLCM feature of the image.

[0134] The HOG feature after 2DPCA dimensionality reduction described above is serially fused with the grid GLCM feature to serve as the feature of the tail vertebra. The HOG feature after 2DPCA dimensionality reduction is used for the features of other vertebral parts to obtain a feature set containing the features of each sample.

[0135] 3. Model training; the HOG feature after 2DPCA dimensionality reduction is used, and the feature is serially fused with the grid GLCM feature to calculate the feature of the coccyx part. The HOG feature after 2DPCA dimensionality reduction is used to calculate the other vertebrae parts. In addition, since the feature of the coccyx S1 is similar to the feature of the lumbar vertebrae and thoracic vertebrae, if all the vertebrae are set as the same label, the recognition label deviation of the coccyx is large, so in the calculation, the coccyx part is marked separately. After the coccyx and other vertebrae features are extracted, the random forest algorithm is used for training to generate the corresponding model.

[0136] As Figure 5 , the specific implementation of the image recognition and marking part includes:

[0137] 1. Image preprocessing: In order to better obtain image information and reduce noise images, the image to be marked is first preprocessed, including (1) image filtering, such as Gaussian filtering, median filtering, etc., for noise processing; (2) image enhancement, (3) image sharpening, (4) image mask extraction, that is, removing the background area and retaining the main information for calculation.

[0138] 2. Multiple-scale sliding window method (MSSW) is used to scale, search and extract features from the preprocessed image. The scaling can be performed by interpolation or convolution.

[0139] 3. Image feature extraction; mainly including HOG feature extraction, GLCM feature extraction, 2DPCA dimensionality reduction, coccyx feature and other vertebrae features obtained by feature fusion. The feature extraction steps are similar to the feature extraction steps of the training model.

[0140] 4. Random forest model is used for prediction to obtain the magnetic resonance spine recognition result.

[0141] 5. Further prediction of the image vertebrae can be performed in various ways, such as (1) combining the information between different layers to calculate the spatial coordinates and the centroid of the aggregation point; (2) collecting all the results after MSSW calculation, and using upsampling or downsampling to make all the data reach the same pixel size, and calculating the overlap degree; (3) removing the redundant information of the recognition result by calculating the IOU value of the current image recognition result.

[0142] Then the single-layer recognized vertebrae are calculated for spatial position, the current sequence slice image is counted, the position relationship of each vertebrae in three-dimensional space is calculated, and the spine labeling work is completed. The counting here mainly refers to sorting the magnetic resonance spine recognition results after removing the redundant information, combining the 3D space information, and further processing.

[0143] 7. vertebral body marking; taking the tail vertebra as the reference point, based on the identification of the tail vertebra, other vertebral bodies are marked. It can well identify and mark the images with missing information of the middle layer, boundary layer and tail vertebra. Combined with the 3D information of the whole sequence image, it can also correctly and completely mark the discontinuous images of the vertebral body.

[0144] In this embodiment, the tail vertebra is identified, and the identification of other spinal vertebrae is completed based on the tail vertebra marking. The 2DPCA dimension-reduced HOG and the gridding GLCM algorithm are used to extract the fusion features of the tail vertebra, the fusion features are combined with the random forest to complete the identification and marking of the tail vertebra. The 2DPCA dimension-reduced HOG features combined with the random forest are used to identify and mark the remaining vertebral bodies. A multi-scale sliding window search method is used for image traversal, the redundant information of the identification result is removed by calculating the IOU, and the matrix coordinate information between the layers of the magnetic resonance image is combined to optimize the marking result of the whole sequence image, thereby increasing the identification accuracy and improving the robustness of the algorithm.

[0145] Similar to the principle of the above embodiment, the application provides a random forest magnetic resonance spinal vertebra image marking system based on multi-feature fusion.

[0146] The following provides specific embodiments in combination with the drawings:

[0147] As Figure 6 A structure schematic diagram of a random forest magnetic resonance spinal vertebra image marking system based on multi-feature fusion in an embodiment of the application is shown.

[0148] The system comprises:

[0149] An image preprocessing module 1, which is used for preprocessing the magnetic resonance spinal vertebra image to be marked to obtain a preprocessed image;

[0150] A feature extraction module 2, which is connected to the image preprocessing module 1 and is used for performing feature extraction processing according to the preprocessed image to obtain corresponding tail vertebra features and other vertebral body features;

[0151] A vertebral body identification module 3, which is connected to the feature extraction module 2 and is used for obtaining a magnetic resonance spinal vertebra identification result according to the tail vertebra features and the other vertebral body features based on a spinal vertebra identification model obtained by training a random forest;

[0152] A marking module 4, which is connected to the vertebral body identification module 3 and is used for performing vertebral body marking according to the magnetic resonance spinal vertebra identification result based on the positional relationship of each vertebral body in the three-dimensional space calculated from the positional information of each vertebral body in each layer of the magnetic resonance image, to obtain a magnetic resonance spinal vertebra marking result.

[0153] It should be noted that Figure 6The division of various modules in the system embodiments is only a logical function division, and all or part of the actual implementation can be integrated into one physical entity, or can be physically separated. And these modules can all be implemented in the form of software called by a processing element; or all be implemented in the form of hardware; or part of the modules are implemented in the form of software called by a processing element, and part of the modules are implemented in the form of hardware;

[0154] For example, each module can be one or more integrated circuits configured to implement the above method, such as one or more application specific integrated circuits (ASICs), or one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs), etc. For another example, when a certain module above is implemented in the form of a processing element scheduling program code, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules can be integrated together to implement in the form of a system on a chip (SOC).

[0155] Since the implementation principle of the random forest magnetic resonance spine image labeling system based on multi-feature fusion has been described in the foregoing embodiments, it is not repeated here.

[0156] In an embodiment, the training manner of the spine recognition model comprises: pre-processing a plurality of magnetic resonance spine images with caudal vertebrae and other vertebrae labels to obtain pre-processed images corresponding to each magnetic resonance spine image; performing feature extraction processing according to each pre-processed image to extract caudal vertebrae features and other vertebrae features corresponding to each magnetic resonance spine image to obtain a feature data training set; and training a random forest based on the feature data training set to obtain the spine recognition model.

[0157] In an embodiment, the pre-processing comprises: sequentially performing image filtering, image enhancement, image sharpening, and image mask extraction on the input magnetic resonance spine image to obtain the pre-processed image.

[0158] In an embodiment, the feature extraction process comprises: performing HOG feature extraction on the preprocessed image, and performing 2DPCA dimension reduction on the extracted HOG features to obtain reduced dimension spine HOG features; performing GLCM feature extraction on the preprocessed image to obtain grid GLCM features; and serially fusing the reduced dimension HOG features and the grid GLCM features to obtain tail vertebra features and taking the reduced dimension HOG features as other vertebra features.

[0159] In an embodiment, the manner of performing HOG feature extraction comprises: performing cell segmentation on the preprocessed image, and calculating the gradient and amplitude of each cell; uniformly dividing the gradient and amplitude of each cell into multiple gradient directions, and accumulating the amplitudes of the same gradient direction of each cell according to weights to form a gradient histogram of each cell; normalizing each cell, and connecting the gradient histograms to obtain the corresponding HOG features.

[0160] In an embodiment, the manner of performing GLCM feature extraction comprises: calculating the corresponding gray level co-occurrence matrix according to the preprocessed image; calculating feature parameters based on the gray level co-occurrence matrix; wherein the feature parameters comprise: feature parameters for representing the uniformity and roughness of the image gray level distribution, feature parameters for representing the depth and thickness of the image texture, feature parameters for representing the image information amount, feature parameters for representing the image texture brightness information, and feature parameters for representing the image gray level correlation; performing window sliding according to the preprocessed image based on a window of a set size, and calculating the sliding window features of each window based on the calculated feature parameters to form the grid GLCM features.

[0161] In an embodiment, obtaining the magnetic resonance spine recognition result according to the tail vertebra features and the other vertebra features comprises: identifying the tail vertebra based on the tail vertebra features to obtain a spine recognition result corresponding to the identified tail vertebra; and identifying the other vertebrae in the spine recognition result corresponding to the identified tail vertebra based on the other vertebra features to obtain a magnetic resonance spine recognition result corresponding to the identified all vertebrae.

[0162] In an embodiment, based on the position relationship of each vertebra in three-dimensional space calculated from the position information of each vertebra in each layer of the magnetic resonance image, performing vertebra labeling according to the magnetic resonance spine recognition result to obtain a magnetic resonance spine labeling result comprises: calculating an IOU value according to the magnetic resonance spine recognition result, and removing redundant information in the magnetic resonance spine recognition result; calculating the position relationship of each vertebra in three-dimensional space based on the position information of each vertebra in the magnetic resonance spine recognition result after removing the redundant information and the position information of each vertebra in other layers of the magnetic resonance spine image; and labeling each vertebra in the magnetic resonance spine recognition result after removing the redundant information based on the position relationship of each vertebra in three-dimensional space to obtain a magnetic resonance spine labeling result.

[0163] As Figure 7 A structure diagram of a multi-feature fusion based random forest magnetic resonance spine image labeling terminal 30 in an embodiment of the present application is shown.

[0164] The multi-feature fusion based random forest magnetic resonance spine image labeling terminal 30 comprises a memory 31 and a processor 32. The memory 31 is configured to store a computer program; and the processor 32 is configured to execute the computer program, so as to implement the multi-feature fusion based random forest magnetic resonance spine image labeling method as described above. Figure 1 The multi-feature fusion based random forest magnetic resonance spine image labeling method as described above.

[0165] Optionally, the number of the memories 31 can be one or more, and the number of the processors 32 can be one or more. Figure 7 In the embodiment, one is taken as an example.

[0166] Optionally, the processor 32 in the multi-feature fusion based random forest magnetic resonance spine image labeling terminal 30 loads one or more instructions corresponding to a process of an application program into the memory 31 according to the steps as described above, and executes the application program stored in the first memory 31 by the processor 32, so as to implement various functions in the multi-feature fusion based random forest magnetic resonance spine image labeling method as described above. Figure 1 The multi-feature fusion based random forest magnetic resonance spine image labeling method as described above. Figure 1 The multi-feature fusion based random forest magnetic resonance spine image labeling method as described above.

[0167] Optionally, the memory 31 can include but is not limited to a high-speed random access memory, a non-volatile memory, such as one or more disk storage devices, flash memory devices or other non-volatile solid-state storage devices; the processor 32 can include but is not limited to a central processing unit (CPU), a network processor (NP) and the like; and can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0168] Optionally, the processor 32 can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0169] The application further provides a computer readable storage medium storing a computer program, the computer program realizing the method as shown in the random forest magnetic resonance spine image marking method based on multi-feature fusion when running. Figure 1 The computer readable storage medium can include, but is not limited to, a floppy disk, an optical disk, a CD-ROM (compact disk read-only memory), a magneto-optical disk, a ROM (read-only memory), a RAM (random access memory), an EPROM (erasable programmable read-only memory), an EEPROM (electrically erasable programmable read-only memory), a magnetic card or an optical card, a flash memory, or other types of media / machine readable media suitable for storing machine executable instructions. The computer readable storage medium can be a product not connected to a computer device, or a component connected to a computer device.

[0170] Compared with the prior art, the application has the advantages including:

[0171] 1. The random forest algorithm is adopted in the application, the training and prediction time of the algorithm is less relative to the conventional deep learning method, the hardware requirement is lower, the fast prediction speed can be obtained without GPU assistance, and the application is easy to implement in engineering.

[0172] 2. The HOG after 2DPCA dimension reduction and the gridding GLCM algorithm are adopted to extract the fusion features of the coccyx, the HOG after 2DPCA dimension reduction is adopted to extract the features of other vertebrae. The risk of overfitting is reduced through dimension reduction, the key feature loss is prevented by combining the gridding GLCM information, and the overall calculation cost is lower. Through simulation, the application can also better identify the image with partial coccyx information loss.

[0173] 3. Since the coccyx features are prominent and easy to identify, the application first identifies and labels the coccyx part, and then completes the labeling work of the remaining vertebrae, so that the labeling work efficiency is greatly improved.

[0174] 4、The application adopts a multi-scale sliding window search method to traverse the image, removes redundant information of the recognition result by calculating IOU, and optimizes the overall sequence image labeling result by combining the matrix coordinate information between each layer of the magnetic resonance image, thereby increasing the recognition accuracy and improving the algorithm robustness. Since the overall sequence position relationship is combined, it is better applicable to the labeling work under the discontinuous condition of the spine.

[0175] To sum up, the random forest magnetic resonance spine image labeling method, system and terminal based on multi-feature fusion of the application obtain the coccyx features and other vertebral features by pre-processing and feature extraction processing of the magnetic resonance spine image to be labeled, obtain the magnetic resonance spine recognition result from the extracted features based on the spine recognition model obtained by training the random forest, and finally obtain the magnetic resonance spine labeling result by combining the position relationship of each vertebra in the three-dimensional space. The application adopts the trained random forest to recognize the vertebra, which not only has less model training and prediction time, but also has lower hardware requirements, and can obtain faster prediction speed and high-precision prediction result without GPU assistance. Moreover, the vertebra labeling is optimized by the position relationship of each vertebra in the three-dimensional space, which increases the labeling accuracy and improves the algorithm robustness, and is better applicable to the labeling work under the conditions of discontinuous spine, image blur, image deformation, metal implant or pseudo-image caused by involuntary movement. Therefore, the application effectively overcomes the various shortcomings in the prior art and has high industrial utilization value.

[0176] The above embodiments only exemplarily illustrate the principles and effects of the application, and are not used to limit the application. Any person skilled in the art can modify or change the above embodiments without departing from the spirit and scope of the application. Therefore, all equivalent modifications or changes completed by those skilled in the art without departing from the spirit and technical thought of the application should be covered by the claims of the application.

Claims

1. A random forest magnetic resonance spine image labeling method based on multi-feature fusion, characterized by, The method comprises: pre-processing a to-be-labeled magnetic resonance spine image to obtain a pre-processed image; According to the pre-processed image, feature extraction processing is performed to obtain corresponding caudal vertebra features and other vertebra features; wherein, according to the pre-processed image, feature extraction processing is performed to obtain corresponding caudal vertebra features and other vertebra features, which comprises: performing HOG feature extraction on the pre-processed image, and performing 2DPCA dimension reduction on the extracted HOG features to obtain dimension-reduced spine HOG features; performing GLCM feature extraction on the pre-processed image to obtain grid-based GLCM features; serially fusing the dimension-reduced HOG features and the grid-based GLCM features to obtain caudal vertebra features and taking the dimension-reduced HOG features as other vertebra features; Based on the spine recognition model obtained by the training random forest, the magnetic resonance spine recognition result is obtained according to the caudal vertebra features and the other vertebra features; Based on the position relationship of each vertebra in the three-dimensional space calculated from the position information of each vertebra in each layer of the magnetic resonance image, the vertebra labeling is performed according to the magnetic resonance spine recognition result to obtain a magnetic resonance spine labeling result. 2.The method of claim 1, wherein the method comprises: The training method of the spine recognition model comprises: Pre-processing a plurality of magnetic resonance spine images with caudal vertebra and other vertebra labels to obtain pre-processed images corresponding to each magnetic resonance spine image; According to each pre-processed image, feature extraction processing is performed to extract caudal vertebra features and other vertebra features corresponding to each magnetic resonance spine image to obtain a feature data training set; Based on the feature data training set, the random forest is trained to obtain the spine recognition model. 3.The method of claim 1 or 2, wherein, The pre-processing comprises: The input magnetic resonance spine image is sequentially subjected to image filtering, image enhancement, image sharpening and image mask extraction to obtain a pre-processed image.

4. The method according to claim 1 or 2, wherein the method is a multi-feature fusion based random forest magnetic resonance spine image labeling method, characterized in that, The feature extraction processing comprises: HOG feature extraction is performed on the pre-processed image, and the extracted HOG features are subjected to 2DPCA dimension reduction to obtain dimension-reduced spine HOG features; GLCM feature extraction is performed on the pre-processed image to obtain grid-based GLCM features; The dimension-reduced HOG features and the grid-based GLCM features are serially fused to obtain caudal vertebra features and the dimension-reduced HOG features are taken as other vertebra features.

5. The method of claim 4, wherein the method is based on multi-feature fusion random forest magnetic resonance spine image labeling. The HOG feature extraction method comprises: According to the pre-processed image, unit segmentation is performed, and the gradient and amplitude of each unit are calculated; The gradient and amplitude of each unit are evenly divided into a plurality of gradient directions, and the amplitudes of the same gradient direction of each unit are accumulated according to a weight to form a gradient histogram of each unit; Each unit is normalized, and the gradient histograms are connected to obtain corresponding HOG features.

6. The method of claim 4, wherein the method is based on multi-feature fusion random forest magnetic resonance spine image labeling. The GLCM feature extraction method comprises: According to the pre-processed image, a corresponding gray level co-occurrence matrix is calculated; Based on the gray level co-occurrence matrix, feature parameters are calculated; wherein, the feature parameters comprise: feature parameters for representing the uniformity and roughness of the image gray level distribution, feature parameters for representing the depth and thickness of the image texture, feature parameters for representing the image information amount, feature parameters for representing the image texture brightness information, and feature parameters for representing the image gray level correlation; Based on the window of the set size, the window is slid according to the pre-processed image, and the sliding window feature of each window is calculated based on the calculated feature parameter to compose the grid GLCM feature. 7.The method of claim 1, wherein The magnetic resonance spine recognition result is obtained according to the caudal vertebra feature and other vertebra features, and the magnetic resonance spine recognition result includes: The caudal vertebra is recognized based on the caudal vertebra feature to obtain the spine recognition result corresponding to the recognized caudal vertebra; The other vertebra is recognized based on the other vertebra feature in the spine recognition result corresponding to the recognized caudal vertebra to obtain the magnetic resonance spine recognition result corresponding to all the recognized vertebrae. 8.The method of claim 1, wherein, The vertebrae are labeled according to the magnetic resonance spine recognition result based on the position relationship of the vertebrae in the three-dimensional space calculated from the position information of the vertebrae in each layer of the magnetic resonance image to obtain the magnetic resonance spine labeling result, and the magnetic resonance spine labeling result includes: The IOU value is calculated according to the magnetic resonance spine recognition result, and the redundant information in the magnetic resonance spine recognition result is removed; The position relationship of the vertebrae in the three-dimensional space is calculated based on the position information of the vertebrae in the magnetic resonance spine recognition result after removing the redundant information and the position information of the vertebrae in other layers of the magnetic resonance spine image; The vertebrae in the magnetic resonance spine recognition result after removing the redundant information are labeled based on the position relationship of the vertebrae in the three-dimensional space to obtain the magnetic resonance spine labeling result.

9. A multi-feature fusion based random forest magnetic resonance spine image labeling system, characterized in that, The system includes: An image preprocessing module for preprocessing the magnetic resonance spine image to be labeled to obtain a pre-processed image; A feature extraction module connected to the image preprocessing module for performing feature extraction processing according to the pre-processed image to obtain corresponding caudal vertebra features and other vertebra features; wherein the feature extraction processing according to the pre-processed image to obtain corresponding caudal vertebra features and other vertebra features includes: performing HOG feature extraction according to the pre-processed image, and performing 2DPCA dimension reduction on the extracted HOG features to obtain dimension-reduced spine HOG features; performing GLCM feature extraction according to the pre-processed image to obtain grid GLCM features; and serially fusing the dimension-reduced HOG features and the grid GLCM features to obtain caudal vertebra features and taking the dimension-reduced HOG features as other vertebra features; A vertebra recognition module connected to the feature extraction module for obtaining a magnetic resonance spine recognition result according to the caudal vertebra features and the other vertebra features based on a spine recognition model obtained by a training random forest; A labeling module connected to the vertebra recognition module for labeling vertebrae according to the magnetic resonance spine recognition result based on the position relationship of the vertebrae in the three-dimensional space calculated from the position information of the vertebrae in each layer of the magnetic resonance image to obtain a magnetic resonance spine labeling result.

10. A multi-feature fusion-based random forest magnetic resonance spine image marking terminal, characterized by, It includes: One or more memories and one or more processors; The one or more memories are used to store a computer program; The one or more processors connected to the memories are used to run the computer program to perform the method of any one of claims 1 to 8.

Images