Multi-scale muscle image segmentation method for paravertebral muscle MRI

Abstract

The invention discloses a multi-scale muscle image segmentation method for paravertebral muscle MRI, and relates to the field of image segmentation, and the method comprises the steps: carrying out the preprocessing of an original paravertebral muscle magnetic resonance image, and obtaining a paravertebral muscle magnetic resonance image to be segmented; inputting the paravertebral muscle magnetic resonance image to be segmented into the image classification network structure for primary segmentation processing to obtain a primary muscle segmentation result; performing secondary segmentation processing on the paravertebral muscle magnetic resonance image to be segmented by using the fusion segmentation network structure, and obtaining a secondary muscle segmentation result based on a processing result; and carrying out splicing and convolution processing on the primary muscle segmentation result and the secondary muscle segmentation result, obtaining a segmentation result of paravertebral muscle information based on a processing result, and verifying the accuracy of the segmentation result by using an identification technology. According to the invention, based on the multi-scale mixed feature fusion network, the detection rate and accuracy of small muscles and multi-class muscles on the same image are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of image segmentation, and in particular to a multi-scale muscle image segmentation method for paraspinal muscle MRI. Background Art

[0002] MRI imaging of paraspinal muscles is mainly used to observe the muscle groups located on both sides of the spine, including the multifidus muscles, erector spinae muscles and rotator muscles. Their health is crucial to the stability and function of the spine. Through MRI imaging, the structure, volume, degree of fat infiltration and presence of other pathological changes of the paraspinal muscles can be segmented in detail.

[0003] Traditional paraspinal muscle MRI segmentation methods rely on shallow features such as shape, texture, and simple linear iterative clustering for initial image segmentation, but the accuracy of these techniques is often low. The emergence and development of deep learning has led to convolutional neural networks (CNNs) with powerful representation capabilities, which have significantly improved the segmentation of paraspinal muscle MRI images, such as Unet, which is a key model for medical segmentation. The architecture of Unet includes an encoding structure for feature extraction and a decoding structure that uses deconvolution for upsampling to ensure that the final segmentation matches the size of the input image. It is often used for accurate pixel-level segmentation. In addition, based on Unet, some variants have been developed, including Unet++, Resunet, and Transunet, which have made many breakthroughs in their respective medical image segmentation fields. They still have inherent limitations: the convolution kernel effectively captures local features, but it is difficult to handle long-term dependencies, which greatly limits their ability to model global context.

[0004] In order to ensure the accurate expression of the target muscle information, the algorithm under study not only requires extensive semantic information for simple classification of pixels in the image, but also requires sufficient local texture and multi-scale feature information as auxiliary supplements. However, the current image segmentation algorithms based on paraspinal muscle MRI generally have problems such as inaccurate target feature extraction and poor tissue boundary prediction capabilities.

[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0006] 1. Technical issues to be solved

[0007] In view of the shortcomings of the prior art, the present invention provides a multi-scale muscle image segmentation method for paraspinal muscle MRI, which has the purpose of effectively improving the detection rate and accuracy of small muscles and multi-category muscles in the same image, thereby solving the problems of inaccurate target feature extraction location and poor tissue boundary prediction ability.

[0008] (II) Technical solution

[0009] In order to achieve the above purpose of effectively improving the detection rate and accuracy of small muscles and multi-category muscles in the same image, the specific technical solutions adopted by the present invention are as follows:

[0010] A multi-scale muscle image segmentation method for paraspinal muscle MRI, the multi-scale muscle image segmentation method comprising:

[0011] S1, preprocessing the original paraspinal muscle magnetic resonance image to obtain the paraspinal muscle magnetic resonance image to be segmented, and inputting the paraspinal muscle magnetic resonance image to be segmented into the graph classification network structure for initial segmentation processing to obtain a first-level muscle segmentation result;

[0012] S2, performing secondary segmentation processing on the paraspinal muscle magnetic resonance image to be segmented using a fusion segmentation network structure, and obtaining a secondary muscle segmentation result based on the processing result;

[0013] S3. The first-level muscle segmentation result and the second-level muscle segmentation result are concatenated and convolved, and the segmentation result of the paraspinal muscle information is obtained based on the processing result, and the accuracy of the segmentation result is verified by using recognition technology.

[0014] Preferably, the original paraspinal muscle magnetic resonance image is preprocessed to obtain the paraspinal muscle magnetic resonance image to be segmented, and the paraspinal muscle magnetic resonance image to be segmented is input into the graph classification network structure for initial segmentation processing, and the first-level muscle segmentation result is obtained, including:

[0015] S11, performing denoising and adaptive gain control processing on the original paraspinal muscle magnetic resonance image based on a non-uniformity correction algorithm, and obtaining a paraspinal muscle magnetic resonance image to be segmented based on the processing result;

[0016] S12, adjusting the size of the paraspinal muscle magnetic resonance image to be segmented to a preset size, and inputting the size into a graph classification network structure, wherein the graph classification network structure uses an encoder to extract input feature information of the paraspinal muscle magnetic resonance image to be segmented;

[0017] S13. Use the feature splicing operator to perform feature fusion on the input feature information to generate a feature map, and perform pooling splicing processing on the feature map based on the enhancement processor and the fusion device, and obtain the first-level muscle segmentation result according to the processing result.

[0018] Preferably, the original paraspinal muscle magnetic resonance image is subjected to denoising and adaptive gain control processing based on a non-uniformity correction algorithm, and the paraspinal muscle magnetic resonance image to be segmented is obtained based on the processing result, including:

[0019] S111, acquiring an original paraspinal muscle magnetic resonance image, and calculating a gain parameter matrix and a bias parameter matrix of the original paraspinal muscle magnetic resonance image using a calibration-based non-uniformity correction algorithm;

[0020] S112, according to the gain parameter matrix and the bias parameter matrix, using the box filter method to perform guided filtering on the original paraspinal muscle magnetic resonance image, and then performing gradient descent update to obtain preliminary image data;

[0021] S113, performing median filtering on the preliminary image data based on the box filtering method to obtain the median filtering result, and performing histogram statistics on the median filtering result to obtain the image muscle distribution data;

[0022] S114, acquiring intermediate image data based on the image muscle distribution data, performing saturation processing on the image data, and obtaining a magnetic resonance image of the paraspinal muscles to be segmented according to the processing result.

[0023] Preferably, the intermediate image data is acquired based on the image muscle distribution data, and saturation processing is performed on the image data, and the magnetic resonance image of the paraspinal muscles to be segmented is obtained according to the processing result, including:

[0024] S1141, performing adaptive gain control processing on the image muscle distribution data, obtaining intermediate image data based on the processing result, and dividing the image muscle distribution data according to the intermediate image data to determine the image muscle distribution range;

[0025] S1142, determining a gamma correction parameter according to the muscle distribution range of the image, performing correction processing on the intermediate image data using the gamma correction parameter, and performing saturation truncation processing on the intermediate image data based on the correction result and a preset saturation truncation threshold;

[0026] S1143 , adaptively grayscale stretch the saturated intermediate image data, and use the stretched intermediate image data as the paraspinal muscle magnetic resonance image to be segmented.

[0027] Preferably, the feature information is inputted for feature fusion to generate a feature map by using a feature splicing operator, and the feature map is subjected to pooling splicing processing based on an enhancement processor and a fusion device, and the first-level muscle segmentation result is obtained according to the processing result, including:

[0028] S131, using a feature splicer to perform a splicing operation on the input feature information in the scale channel dimension, and generating a feature map based on the splicing result;

[0029] S132, performing average pooling processing on the feature map along the image height and width dimensions based on the enhancement processor, and generating a one-dimensional vector according to the pooling result and the vector splicing formula;

[0030] S133, using a weight distribution processor to perform dimensionality reduction and dimensionality increase processing on the one-dimensional vector, and generating a weight vector according to the processing result, and scaling the feature map based on the weight vector to obtain a scaled feature map;

[0031] S134, using a fuser to perform average pooling of the scaled feature map in two dimensions, namely, height and width, and obtaining a comprehensive weight vector based on the pooling result and the weight fusion operation;

[0032] S135. Apply the comprehensive weight vector to the feature map for recalibration, and obtain a segmentation image as a first-level muscle segmentation result according to the recalibration result.

[0033] Preferably, the expression of the vector concatenation formula is:

[0034]

[0035] In the formula, f represents the vector concatenation result, δ represents vector concatenation, F1 represents feature fusion, W represents the feature map width, H represents the feature map height, and x i、j、C Represents the value of channel C at position (i, j) in the feature map.

[0036] Preferably, the weight fusion operation expression is:

[0037]

[0038] The recalibration calculation formula is:

[0039]

[0040] In the formula, S total represents the comprehensive weight vector, S h Represents the weight in the height direction of the image, S w represents the weight in the width direction of the image, A′ represents the recalibrated feature map, and A represents the feature map.

[0041] Preferably, the paraspinal muscle magnetic resonance image to be segmented is subjected to secondary segmentation processing using a fusion segmentation network structure, and the secondary muscle segmentation result obtained based on the processing result includes:

[0042] S21, adjusting the paraspinal muscle magnetic resonance image to be segmented to a preset size and inputting the image into a fusion segmentation network structure, and extracting secondary information features of the paraspinal muscle magnetic resonance image to be segmented using the fusion segmentation network structure;

[0043] S22. Use the jump connection operator to fuse the secondary information features, realize the splicing processing on the channel dimensions of different scales, and obtain the segmentation processing image based on the splicing result to obtain the secondary muscle segmentation result.

[0044] Preferably, the first-level muscle segmentation result and the second-level muscle segmentation result are concatenated and convolved, the segmentation result of the paraspinal muscle information is obtained based on the processing result, and the accuracy of the segmentation result is verified by using recognition technology, including:

[0045] S31, converting the first-level muscle segmentation result and the second-level muscle segmentation result into tensors, splicing the converted tensors, and performing double convolution on the spliced ​​tensors according to a convolution kernel size of a preset size;

[0046] S32, obtaining a segmentation result of the paraspinal muscle information based on the convolution result, obtaining muscle position information according to the segmentation result of the paraspinal muscle information, and randomly generating a sample triplet using the muscle position information;

[0047] S33. Verify the sample triples based on the pre-stored verification information to obtain verification results, and verify the accuracy of the segmentation results based on the verification results and the principal component analysis algorithm.

[0048] Preferably, verifying the sample triples based on the pre-stored verification information to obtain a verification result, and verifying the accuracy of the segmentation result based on the verification result and the principal component analysis algorithm includes:

[0049] S331, analyzing the comparison threshold between the pre-stored verification information and the sample triplet, and extracting feature information of the segmentation result of the paraspinal muscle information based on the comparison threshold and the principal component analysis algorithm;

[0050] S332. The extracted feature information is represented as principal component data, and the Euclidean distance between the principal component data and the pre-stored verification data is calculated, and the accuracy of the segmentation result is verified based on the Euclidean distance.

[0051] (III) Beneficial effects

[0052] Compared with the prior art, the present invention provides a multi-scale muscle image segmentation method for paraspinal muscle MRI, which has the following beneficial effects:

[0053] (1) Based on a multi-scale hybrid feature fusion network, the present invention effectively improves the detection rate and accuracy of small muscles and multi-category muscles in the same image. In particular, for small and irregularly shaped muscles, it can accurately locate small muscle targets and identify the fuzzy boundaries of muscle areas, thereby ensuring accurate segmentation of the muscle body and its edges.

[0054] (2) The present invention pre-processes MRI before multi-scale muscle image segmentation of paraspinal muscles to reduce the computational burden of the subsequent segmentation algorithm. By enhancing the features in the image and retaining the structural information in the image, the purpose of maintaining the continuity of the muscle structure is achieved, providing a more accurate basis for subsequent quantitative analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0056] Figure 1 is a flow chart of a multi-scale muscle image segmentation method for paraspinal muscle MRI according to an embodiment of the present invention;

[0057] Figure 2 is a structural diagram of IMENet in a multi-scale muscle image segmentation method for paraspinal muscle MRI according to an embodiment of the present invention;

[0058] Figure 3 is a structural diagram of MENet in a multi-scale muscle image segmentation method for paraspinal muscle MRI according to an embodiment of the present invention;

[0059] Figure 4 is a structural diagram of MFAM in a multi-scale muscle image segmentation method for paraspinal muscle MRI according to an embodiment of the present invention;

[0060] Figure 5 is a HA structure diagram in a multi-scale muscle image segmentation method for paraspinal muscle MRI according to an embodiment of the present invention;

[0061] Figure 6 is an example diagram of a paraspinal muscle MRI image in a multi-scale muscle image segmentation method for paraspinal muscle MRI according to an embodiment of the present invention;

[0062] Figure 7 is an original input image according to an embodiment of the present invention;

[0063] Figure 8 is a ground-truth graph according to an embodiment of the present invention;

[0064] Fig. 9 is a segmentation visualization diagram of the Unet algorithm according to an embodiment of the present invention;

[0065] Fig.10 is a segmentation visualization diagram of the Unet-tm algorithm according to an embodiment of the present invention;

[0066] Fig.11 is a segmentation visualization diagram of the Unet-tr algorithm according to an embodiment of the present invention;

[0067] Fig.12 is a segmentation visualization diagram of the MHFNet algorithm according to an embodiment of the present invention;

[0068] Fig.13 This is one of the comparison diagrams of the actual segmentation effects of the IMENet algorithm and the Unet algorithm on the paraspinal muscle MRI target according to an embodiment of the present invention;

[0069] Fig.14 This is a second comparison diagram of the actual segmentation effects of the IMENet algorithm and the Unet algorithm on the paraspinal muscle MRI target according to an embodiment of the present invention;

[0070] Fig.15 This is a third comparison diagram of the actual segmentation effects of the IMENet algorithm and the Unet algorithm on the paraspinal muscle MRI target according to an embodiment of the present invention;

[0071] Fig.16 This is the fourth comparison diagram of the actual segmentation effects of the IMENet algorithm and the Unet algorithm for the paraspinal muscle MRI target according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments. They can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and advantages of the present invention.

[0073] According to an embodiment of the present invention, a multi-scale muscle image segmentation method for paraspinal muscle MRI is provided.

[0074] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1-Figure 5 As shown, according to the multi-scale muscle image segmentation method for paraspinal muscle MRI according to an embodiment of the present invention, the multi-scale muscle image segmentation method includes:

[0075] S1. Preprocess the original paraspinal muscle magnetic resonance image to obtain the paraspinal muscle magnetic resonance image to be segmented, and input the paraspinal muscle magnetic resonance image to be segmented into the graph classification network structure for initial segmentation processing to obtain a first-level muscle segmentation result.

[0076] In this embodiment, the original paraspinal muscle magnetic resonance image is preprocessed to obtain the paraspinal muscle magnetic resonance image to be segmented, and the paraspinal muscle magnetic resonance image to be segmented is input into the graph classification network structure for initial segmentation processing, and the first-level muscle segmentation result obtained includes:

[0077] S11. De-noising and adaptive gain control processing are performed on the original paraspinal muscle magnetic resonance image based on a non-uniformity correction algorithm, and the paraspinal muscle magnetic resonance image to be segmented is obtained based on the processing result.

[0078] Specifically, based on the non-uniformity correction algorithm, the original paraspinal muscle magnetic resonance image is subjected to denoising and adaptive gain control processing, and the paraspinal muscle magnetic resonance image to be segmented is obtained based on the processing result, including:

[0079] S111, acquiring an original paraspinal muscle magnetic resonance image, and calculating a gain parameter matrix and a bias parameter matrix of the original paraspinal muscle magnetic resonance image using a calibration-based non-uniformity correction algorithm;

[0080] S112, according to the gain parameter matrix and the bias parameter matrix, using the box filter method to perform guided filtering on the original paraspinal muscle magnetic resonance image, and then performing gradient descent update to obtain preliminary image data;

[0081] S113, performing median filtering on the preliminary image data based on the box filtering method to obtain the median filtering result, and performing histogram statistics on the median filtering result to obtain the image muscle distribution data;

[0082] S114, acquiring intermediate image data based on the image muscle distribution data, performing saturation processing on the image data, and obtaining a magnetic resonance image of the paraspinal muscles to be segmented according to the processing result.

[0083] Wherein, the intermediate image data is obtained based on the image muscle distribution data, and the image data is subjected to saturation processing, and the paraspinal muscle magnetic resonance image to be segmented is obtained according to the processing result, including:

[0084] S1141, performing adaptive gain control processing on the image muscle distribution data, obtaining intermediate image data based on the processing result, and dividing the image muscle distribution data according to the intermediate image data to determine the image muscle distribution range;

[0085] S1142, determining a gamma correction parameter according to the muscle distribution range of the image, performing correction processing on the intermediate image data using the gamma correction parameter, and performing saturation truncation processing on the intermediate image data based on the correction result and a preset saturation truncation threshold;

[0086] S1143 , adaptively grayscale stretch the saturated intermediate image data, and use the stretched intermediate image data as the paraspinal muscle magnetic resonance image to be segmented.

[0087] S12. The size of the paraspinal muscle magnetic resonance image to be segmented is adjusted to a preset size, and the size is input into a graph classification network structure. The graph classification network structure uses an encoder to extract input feature information of the paraspinal muscle magnetic resonance image to be segmented.

[0088] S13. Use the feature splicing operator to perform feature fusion on the input feature information to generate a feature map, and perform pooling splicing processing on the feature map based on the enhancement processor and the fusion device, and obtain the first-level muscle segmentation result according to the processing result.

[0089] Among them, the feature splicing operator is used to perform feature fusion on the input feature information to generate a feature map, and the feature map is subjected to pooling splicing processing based on the enhancement processor and the fusion device. The first-level muscle segmentation results obtained according to the processing results include:

[0090] S131, using a feature splicer to perform a splicing operation on the input feature information in the scale channel dimension, and generating a feature map based on the splicing result;

[0091] S132, performing average pooling processing on the feature map along the image height and width dimensions based on the enhancement processor, and generating a one-dimensional vector according to the pooling result and the vector splicing formula;

[0092] S133, using a weight distribution processor to perform dimensionality reduction and dimensionality increase processing on the one-dimensional vector, and generating a weight vector according to the processing result, and scaling the feature map based on the weight vector to obtain a scaled feature map;

[0093] S134, using a fuser to perform average pooling of the scaled feature map in two dimensions, namely, height and width, and obtaining a comprehensive weight vector based on the pooling result and the weight fusion operation;

[0094] S135. Apply the comprehensive weight vector to the feature map for recalibration, and obtain a segmentation image as a first-level muscle segmentation result according to the recalibration result.

[0095] It should be explained that the graph classification network structure is an IMENet network. During the segmentation process, the size of the paraspinal muscle magnetic resonance image (MRI data set) to be segmented is adjusted to an appropriate size and input into the MHFNet neural network. The MHFNet neural network is used to extract features from the MRI data set and output the segmentation information of the target.

[0096] Specifically, Figure 3 and Figure 5 As shown in the figure, Double Conv represents double convolution, BN represents batch normalization, Relu represents Relu activation function, Copy and Crop represents adaptive cropping, MAXPOOL represents maximum pooling, Upsample represents upsampling, Identity represents identity function, Activation represents activation operation; Input represents input, Avg Pool represents average pooling, BatchNorm represents batch normalization, Sigmoid represents Sigmoid activation function, and Recalibration represents feature recalibration.

[0097] The segmentation information is a three-dimensional tensor (B, N, H, W), where B stands for batch-size, indicating the number of images processed at one time, N stands for num-class, indicating the number of categories predicted by the model, and H and W stand for Height and Width, indicating the height and width of the image respectively.

[0098] The network structure of the MHFNet neural network includes three operation structural units and two main functional parts. The operation structural units include an encoder operation unit, a decoder operation unit and a feature splicing operation unit. The main functional parts include an MFAM module and a HA module. The encoder operation unit is composed of a convolutional layer, a batch normalization layer, a maximum pooling layer and a Relu activation function layer in sequence, which is used to extract the input feature information of the image. The decoder operation unit is composed of a convolutional layer, a batch normalization layer, an upsampling layer and a Relu activation function layer in sequence, which is used for multi-scale feature extraction of the image and more comprehensive capture of important image information.

[0099] The feature splicing operation unit is composed of a cropping layer and a splicing layer in sequence, and is used to perform feature fusion on the extracted input features, that is, to form a new feature map by splicing the feature map cropped by the encoder and the feature map upsampled by the decoder on channel dimensions of different scales.

[0100] Among them, the MFAM module is used to enhance the Unet network's ability to extract multi-scale local feature information and is integrated into the above-mentioned encoder operation unit and decoder operation unit. The HA module is used to improve the feature fusion and perception capabilities of the above-mentioned feature splicing operation unit. The MFAM module consists of four parts in sequence: Squeeze operation, Excitation operation, Split operation, and scaling operation;

[0101] Specifically, the squeeze operation performs average pooling on the feature map with an input size of B×C×H×W along the two dimensions H and W, generating two one-dimensional vectors with sizes of B×C×H×1 and B×C×1×W, respectively, whose lengths are equal to the number of channels C of the input feature map;

[0102] Then the two generated one-dimensional vectors are concatenated to generate a one-dimensional vector of size B×C×(H+W). The one-dimensional vector concatenation formula is:

[0103]

[0104] In the formula, f represents the vector concatenation result, δ represents vector concatenation, F1 represents feature fusion, W represents the feature map width, H represents the feature map height, and x i、j、C Represents the value of channel C at position (i, j) in the feature map.

[0105] Specifically, the Excitation operation structure includes a dimensionality reduction layer, a dimensionality increase layer, and a Sigmoid activation function layer. The dimensionality reduction layer selects to simultaneously introduce a four-branch Dense Layer with a size of B×C×(H+W) and simultaneously uses four fully connected layers to reduce the dimension at a scaling ratio of r, and obtains four dimensionality reduction layers with a size of B×C×(H+W) / r;

[0106] The dimensionality reduction operation formula is:

[0107] S1 = RELU (T1 × Z + b1);

[0108] Where S1 is the result after dimensionality reduction (shape is C / r), T1 is the weight matrix, Z is the input channel descriptor vector, and b1 is the bias term;

[0109] The four dimension reduction layers of size B×C×(H+W) / r are concatenated and then passed to the dimension increase layer, which uses another fully connected layer to restore the feature dimension to the original dimension B×C×(H+W);

[0110] The formula for dimension-raising operation is:

[0111] S2=T2×S1+b2;

[0112] In the formula, T2 represents the weight matrix after dimensionality reduction, and b2 represents the bias term after dimensionality reduction;

[0113] At the same time, the Sigmoid activation function is applied to obtain the final weight vector S:

[0114] S = sigmoid(S2);

[0115] The Split operation is used to obtain one-dimensional vectors of size B×C×H×1 and B×C×1×W respectively and concatenated to restore the image to its original size. Finally, the scaling operation is used to process the original feature map to obtain the final scaled feature map. The formula for the scaling operation is:

[0116] X'=XeS;

[0117] Where ⊙ represents the element-by-element multiplication operation, X represents the original feature map, and X′ represents the scaled feature map.

[0118] Specifically, the HA module is composed of a coordinate information extraction operation, a weight generation operation, and a feature recalibration operation. The coordinate information extraction operation performs average pooling on the feature map of the input size C×H×W along the H and W dimensions, and obtains an intermediate representation through a fully connected layer and an activation function. The weight generation operation generates weights S in the H and W directions through a fully connected layer and a Sigmoid activation function. h and S w, and then a comprehensive weight vector is obtained by fusing the weights in the height and width directions. The weight fusion operation formula is:

[0119]

[0120] In the formula, Represents an element-wise multiplication operation.

[0121] Finally, the feature recalibration operation is used to recalibrate the original feature map by applying the generated weights to the original feature map. The feature recalibration formula is:

[0122]

[0123] In the formula, A′ represents the recalibrated feature map, and A represents the original feature map.

[0124] S2. Perform secondary segmentation processing on the paraspinal muscle magnetic resonance image to be segmented using a fusion segmentation network structure, and obtain a secondary muscle segmentation result based on the processing result.

[0125] In this embodiment, the paraspinal muscle magnetic resonance image to be segmented is subjected to secondary segmentation processing using a fusion segmentation network structure, and the secondary muscle segmentation results obtained based on the processing results include:

[0126] S21, adjusting the paraspinal muscle magnetic resonance image to be segmented to a preset size and inputting the image into a fusion segmentation network structure, and extracting secondary information features of the paraspinal muscle magnetic resonance image to be segmented using the fusion segmentation network structure;

[0127] S22. Use the jump connection operator to fuse the secondary information features, realize the splicing processing on the channel dimensions of different scales, and obtain the segmentation processing image based on the splicing result to obtain the secondary muscle segmentation result.

[0128] It should be explained that the fusion segmentation network structure is a Unet++ neural network based on Resnet50, and the segmentation result is a three-dimensional vector (B, N, H, W). The network structure of the Unet++ neural network of Resnet50 includes 4 operation structural units;

[0129] The operation structure unit includes a Resnet50 pre-training operation unit, an upsampling operation unit, a downsampling operation unit, and a skip connection operation unit. The upsampling operation unit is composed of a convolution layer, a batch normalization layer, a maximum pooling layer, and a Relu activation function layer in sequence, and is used to extract input feature information of the image.

[0130] Specifically, the downsampling operation unit is composed of a convolutional layer, a batch normalization layer, an upsampling layer and a Relu activation function layer in sequence, which is used to extract multi-scale local features of the image and more comprehensively capture the multi-scale texture information of the image. The Resnet50 pre-training operation unit is composed of pre-trained Resnet50 weights. The jump connection operation unit is composed of a splicing layer, a multi-level connection layer and an aggregation layer, which is used to perform feature fusion on the extracted input features, that is, to form a new feature map by splicing the feature map extracted from the down-sampling part with the feature map of the up-sampling part on channel dimensions of different scales.

[0131] S3. The first-level muscle segmentation result and the second-level muscle segmentation result are concatenated and convolved, and the segmentation result of the paraspinal muscle information is obtained based on the processing result, and the accuracy of the segmentation result is verified by using recognition technology.

[0132] In this embodiment, the first-level muscle segmentation result and the second-level muscle segmentation result are spliced ​​and convolved, the segmentation result of the paraspinal muscle information is obtained based on the processing result, and the accuracy of the segmentation result is verified by using recognition technology, including:

[0133] S31, converting the first-level muscle segmentation result and the second-level muscle segmentation result into tensors, splicing the converted tensors, and performing double convolution on the spliced ​​tensors according to a convolution kernel size of a preset size;

[0134] S32, obtaining a segmentation result of the paraspinal muscle information based on the convolution result, obtaining muscle position information according to the segmentation result of the paraspinal muscle information, and randomly generating a sample triplet using the muscle position information;

[0135] S33. Verify the sample triples based on the pre-stored verification information to obtain verification results, and verify the accuracy of the segmentation results based on the verification results and the principal component analysis algorithm.

[0136] Among them, verifying the sample triples based on the pre-stored verification information to obtain the verification results, and verifying the accuracy of the segmentation results based on the verification results and the principal component analysis algorithm include:

[0137] S331, analyzing the comparison threshold between the pre-stored verification information and the sample triplet, and extracting feature information of the segmentation result of the paraspinal muscle information based on the comparison threshold and the principal component analysis algorithm;

[0138] S332. The extracted feature information is represented as principal component data, and the Euclidean distance between the principal component data and the pre-stored verification data is calculated, and the accuracy of the segmentation result is verified based on the Euclidean distance.

[0139] It should be explained that the segmentation results output by the Unet++ neural network based on Resnet50 and MHFNet are converted into tensors, the converted tensors are concatenated, and then double convolution is performed to obtain the final segmentation result, in which the convolution kernel size in the double convolution is 3×3, padding = 1, and stride is 2.

[0140] In order to facilitate understanding of the above technical solutions of the present invention, the working principle or operation mode of the present invention in the actual process is described in detail below.

[0141] This embodiment tests the segmentation of paraspinal muscle MRI images and provides an adaptive test environment and parameter settings. The specific process and steps are as follows:

[0142] Acquire paraspinal muscle images of patients with lumbar disc herniation. The specific configuration of the specific experimental environment is shown in Table 1:

[0143] Table 1 Experimental environment

[0144]

[0145]

[0146] The specific parameter settings during the neural network training process are shown in Table 2:

[0147] Table 2 Neural network training parameter settings

[0148] Name Value Max iterations 5000 Batch size 4 Optimizer type Adam Initial learning rate <![CDATA[10 -3 ]]> Exponential decay rate (0.9，0.999) Minimum learning rate <![CDATA[10 -5 ]]> Power 0.9

[0149] This example conducts an ablation study on MFAM and HA on the Data-spine dataset to evaluate the segmentation performance of IMENet on paraspinal muscle MRI images. According to the type of fusion of MFAM and HA modules with the Unet algorithm, this example proposes three different network structures based on Unet, as shown in Table 3:

[0150] Table 3 Algorithm network structure table

[0151]

[0152]

[0153] According to the parameter settings in Table 1 and Table 2, Unet, Unet-tr, Unet-tm, and MHFNet were trained and tested 10 times, and the indicators of MIOU, mDice, mF1-Score and params (parameters) were counted, where MIOU, mDice, and mF1-Score represent the average values ​​of IOU (intersection over union), Dice, and F1-Score, respectively. F1-Score combines precision and recall. The experimental results are shown in Table 4. In order to intuitively compare the segmentation effects of the MFAM module and the HA module on paraspinal muscle MRI images, Figures 7 to 12 The actual segmentation effect diagram of Unet, Unet-tr, Unet-tm, and MHFNet.

[0154] In order to intuitively demonstrate the segmentation of paraspinal muscle MRI images by IMENet algorithm, this embodiment Figures 7 to 12 The segmentation accuracy of each muscle in the MRI image was statistically analyzed in detail. The specific segmentation of each muscle was statistically analyzed based on the IOU, precision (prediction rate), recall (recall rate), F1-Score, and Dice evaluation indicators. The values ​​of the Unet and MHFNet algorithms near the MIOU average value were used as the evaluation results of the experiment, as shown in Tables 5 and 6:

[0155] Table 4 Algorithm test results

[0156] algorithm MIOU mDice mF1-Score params U-Net 0.8935 0.8918 0.8965 7.5M Unet-tr 0.9212 0.9257 0.9211 7.56M Unet-tm 0.9211 0.9257 0.9206 7.72M MHFNet 0.9335 0.9262 0.9298 7.83M

[0157] Table 5 MIOU = 0.9346 MHFNet algorithm muscle segmentation evaluation table

[0158]

[0159]

[0160] Table 6 MIOU = 0.9046 Unet algorithm target muscle segmentation evaluation table

[0161] Class IOU F1-Score precision recall Dice Background 0.9932 0.9949 0.9945 0.9954 0.9947 1 0.9474 0.9575 0.9717 0.9443 0.9554 2 0.8125 0.8018 0.8452 0.7693 0.7919 3 0.8158 0.8068 0.8306 0.7865 0.7971 4 0.9457 0.9561 0.9471 0.9655 0.9539 Left 0.9054 0.9176 0.9186 0.9167 0.9135 Right 0.9121 0.9245 0.9326 0.9167 0.9207

[0162] Among them, the effectiveness of MFAM and HA modules, as shown in Table 4, MFAM and HA modules significantly enhance the performance indicators of Unet network on the self-made Data-spine dataset. Unet-tr and Unet-tm can improve MIOU, mDice, and mF1-Score evaluation indicators by 2.77%, 3.39%, and 2.46% respectively compared with Unet, and Unet-tm by 2.76%, 3.39%, and 2.41% respectively compared with Unet, while keeping the number of parameters almost unchanged. Figure 6 As shown in the figure, there are misclassifications in the basic Unet model, which leads to inaccurate segmentation edges and inaccurate extraction sites. The added MFAM module and HA enhance the network's ability to process spatial information, thereby enhancing the prediction effect of tissue boundaries and improving the detection rate and accuracy of small muscles and multi-category muscles on the same image. These results indicate that the MFAM and HA modules play a key role in extracting multi-scale local features of MRI, thereby improving the overall segmentation ability of the model.

[0163] In order to verify the performance improvement of the IMENet algorithm for paraspinal muscle MRI image segmentation, under the same configuration environment, this example compares the segmentation indicators of the SOTA algorithm Unet and the IMENet algorithm for paraspinal muscle MRI muscles on the Data-spine dataset. The comparison results are shown in Table 7. Figures 13 to 16 This is a comparison of the actual segmentation effects of Unet and IMENet. These data and visualization results indicate that the model is good at depicting clear and accurate boundaries, especially for small and irregularly shaped polyps. It accurately locates small muscle targets and identifies the fuzzy boundaries of muscle areas, thereby ensuring accurate segmentation of the muscle body and its edges.

[0164] Table 7 Comparison results

[0165] Model MIOU mDice mF1-Score params U-Net 0.8935 0.8918 0.8965 7.5M Ours 0.9402 0.9414 0.9407 5M

[0166] Qualitative results analysis:

[0167] Table 7 shows the evaluation metrics of IMENet segmentation. Figures 13 to 16 The visual segmentation effect of IMENet is shown. From these data and visualizations, it can be seen that IMENet is proficient in paraspinal muscle MRI segmentation. Specifically, it has the following three advantages:

[0168] Versatility in various scenarios: IMENet consistently produces segmentation results with bright foreground and clear background, which are closely consistent with the ground truth labels, which demonstrates that the model can effectively adapt to different segmentation scenarios.

[0169] Adaptability to different acquisition environments: IMENet shows remarkable adaptability to different conditions of image acquisition, such as changes in brightness, reflection, and shadow. This adaptability is crucial for processing large paraspinal muscle MRI targets. The model can effectively capture the characteristics of the muscle and maintain the internal consistency of the segmentation results. These properties emphasize the robustness and effectiveness of MHFNet under different imaging conditions, enhancing its practicality in practical medical applications.

[0170] From the data in Table 7, we can see that IMENet improves MIOU, mDice, and mF1-Score by 4.67%, 4.96%, and 4.42% respectively compared with Unet. Figures 13 to 16 The visualization results show that the IMENet model is good at depicting clear and accurate boundaries, especially for small and irregularly shaped polyps. It accurately locates small muscle targets and identifies the fuzzy boundaries of muscle regions, thus ensuring accurate segmentation of the muscle body and its edges.

[0171] By comparing with classic networks, the segmentation accuracy of IMENet algorithm is greatly improved compared with SOTA algorithms (such as: U-Net, Unet++, ResUnet, TransUnet, Fcn, Deeplabv3plus), among which MIOU, mDice, and mF1-Score are respectively improved by 4.79%, 4.98%, and 5.07% compared with Unet. Compared with the optimal algorithm Deeplabv3plus in evaluation indicators, they are respectively improved by 3.7%, 2.55%, and 2.62%.

[0172] In summary, with the aid of the above technical solution of the present invention, the present invention is based on a multi-scale hybrid feature fusion network, which effectively improves the detection rate and accuracy of small muscles and multi-category muscles on the same image, especially for small and irregularly shaped muscles, and can accurately locate small muscle targets and identify the fuzzy boundaries of muscle areas, thereby ensuring accurate segmentation of the muscle body and its edges. Before the multi-scale muscle image segmentation of paraspinal muscle MRI, the present invention pre-processes the MRI to reduce the computational burden of the subsequent segmentation algorithm, and retains the structural information in the image by enhancing the features in the image, thereby achieving the purpose of maintaining the coherence of the muscle structure, and providing a more accurate basis for subsequent quantitative analysis.

[0173] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-scale muscle image segmentation method for paraspinal muscle MRI, characterized in that: The multi-scale muscle image segmentation method includes: S1, preprocessing the original paraspinal muscle magnetic resonance image to obtain the paraspinal muscle magnetic resonance image to be segmented, and inputting the paraspinal muscle magnetic resonance image to be segmented into the graph classification network structure for initial segmentation processing to obtain a first-level muscle segmentation result; S2, performing secondary segmentation processing on the paraspinal muscle magnetic resonance image to be segmented using a fusion segmentation network structure, and obtaining a secondary muscle segmentation result based on the processing result; S3. The first-level muscle segmentation result and the second-level muscle segmentation result are concatenated and convolved, and the segmentation result of the paraspinal muscle information is obtained based on the processing result, and the accuracy of the segmentation result is verified by using recognition technology.

2. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 1, characterized in that: The original paraspinal muscle magnetic resonance image is preprocessed to obtain the paraspinal muscle magnetic resonance image to be segmented, and the paraspinal muscle magnetic resonance image to be segmented is input into the graph classification network structure for initial segmentation processing to obtain the first-level muscle segmentation result including: S11, performing denoising and adaptive gain control processing on the original paraspinal muscle magnetic resonance image based on a non-uniformity correction algorithm, and obtaining a paraspinal muscle magnetic resonance image to be segmented based on the processing result; S12, adjusting the size of the paraspinal muscle magnetic resonance image to be segmented to a preset size, and inputting the size into a graph classification network structure, wherein the graph classification network structure uses an encoder to extract input feature information of the paraspinal muscle magnetic resonance image to be segmented; S13. Use the feature splicing operator to perform feature fusion on the input feature information to generate a feature map, and perform pooling splicing processing on the feature map based on the enhancement processor and the fusion device, and obtain the first-level muscle segmentation result according to the processing result.

3. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 2, characterized in that: The method of performing denoising and adaptive gain control processing on the original paraspinal muscle magnetic resonance image based on the non-uniformity correction algorithm and obtaining the paraspinal muscle magnetic resonance image to be segmented based on the processing result includes: S111, acquiring an original paraspinal muscle magnetic resonance image, and calculating a gain parameter matrix and a bias parameter matrix of the original paraspinal muscle magnetic resonance image using a calibration-based non-uniformity correction algorithm; S112, according to the gain parameter matrix and the bias parameter matrix, using the box filter method to perform guided filtering on the original paraspinal muscle magnetic resonance image, and then performing gradient descent update to obtain preliminary image data; S113, performing median filtering on the preliminary image data based on the box filtering method to obtain the median filtering result, and performing histogram statistics on the median filtering result to obtain the image muscle distribution data; S114, acquiring intermediate image data based on the image muscle distribution data, performing saturation processing on the image data, and obtaining a magnetic resonance image of the paraspinal muscles to be segmented according to the processing result.

4. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 3, characterized in that: The method of acquiring intermediate image data based on the image muscle distribution data, performing saturation processing on the image data, and obtaining the paraspinal muscle magnetic resonance image to be segmented according to the processing result comprises: S1141, performing adaptive gain control processing on the image muscle distribution data, obtaining intermediate image data based on the processing result, and dividing the image muscle distribution data according to the intermediate image data to determine the image muscle distribution range; S1142, determining a gamma correction parameter according to the muscle distribution range of the image, performing correction processing on the intermediate image data using the gamma correction parameter, and performing saturation truncation processing on the intermediate image data based on the correction result and a preset saturation truncation threshold; S1143 , adaptively grayscale stretch the saturated intermediate image data, and use the stretched intermediate image data as the paraspinal muscle magnetic resonance image to be segmented.

5. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 4, characterized in that: The feature splicing operator is used to perform feature fusion on the input feature information to generate a feature map, and the feature map is subjected to pooling splicing processing based on the enhancement processor and the fusion device. The first-level muscle segmentation result is obtained according to the processing result, including: S131, using a feature splicer to perform a splicing operation on the input feature information in the scale channel dimension, and generating a feature map based on the splicing result; S132, performing average pooling processing on the feature map along the image height and width dimensions based on the enhancement processor, and generating a one-dimensional vector according to the pooling result and the vector splicing formula; S133, using a weight distribution processor to perform dimensionality reduction and dimensionality increase processing on the one-dimensional vector, and generating a weight vector according to the processing result, and scaling the feature map based on the weight vector to obtain a scaled feature map; S134, using a fuser to perform average pooling of the scaled feature map in two dimensions, namely, height and width, and obtaining a comprehensive weight vector based on the pooling result and the weight fusion operation; S135. Apply the comprehensive weight vector to the feature map for recalibration, and obtain a segmentation image as a first-level muscle segmentation result according to the recalibration result.

6. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 5, characterized in that: The expression of the vector splicing formula is: In the formula, f represents the result of vector concatenation, δ represents weight fusion, F1 represents feature concatenation, W represents the width of the feature map, H represents the height of the feature map, and x i、j、C Represents the value of channel C at position (i, j) in the feature map.

7. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 6, characterized in that: The weight fusion operation expression is: The recalibration calculation formula is: In the formula, S total represents the comprehensive weight vector, S h Represents the weight in the height direction of the image, S w represents the weight in the width direction of the image, A′ represents the recalibrated feature map, and A represents the feature map.

8. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 1, characterized in that: The method of performing secondary segmentation processing on the paraspinal muscle magnetic resonance image to be segmented by using the fusion segmentation network structure and obtaining the secondary muscle segmentation result based on the processing result includes: S21, adjusting the paraspinal muscle magnetic resonance image to be segmented to a preset size and inputting the image into a fusion segmentation network structure, and extracting secondary information features of the paraspinal muscle magnetic resonance image to be segmented using the fusion segmentation network structure; S22. Use the jump connection operator to fuse the secondary information features, realize the splicing processing on the channel dimensions of different scales, and obtain the segmentation processing image based on the splicing result to obtain the secondary muscle segmentation result.

9. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 1, characterized in that: The step of concatenating and convolving the first-level muscle segmentation result and the second-level muscle segmentation result, obtaining the segmentation result of the paraspinal muscle information based on the processing result, and verifying the accuracy of the segmentation result by using recognition technology includes: S31, converting the first-level muscle segmentation result and the second-level muscle segmentation result into tensors, splicing the converted tensors, and performing double convolution on the spliced ​​tensors according to a convolution kernel size of a preset size; S32, obtaining a segmentation result of the paraspinal muscle information based on the convolution result, obtaining muscle position information according to the segmentation result of the paraspinal muscle information, and randomly generating a sample triplet using the muscle position information; S33. Verify the sample triples based on the pre-stored verification information to obtain verification results, and verify the accuracy of the segmentation results based on the verification results and the principal component analysis algorithm.

10. The multi-scale muscle image segmentation method for paraspinal muscle MRI according to claim 9, characterized in that: The step of verifying the sample triples based on the pre-stored verification information to obtain the verification results, and verifying the accuracy of the segmentation results based on the verification results and the principal component analysis algorithm includes: S331, analyzing the comparison threshold between the pre-stored verification information and the sample triplet, and extracting feature information of the segmentation result of the paraspinal muscle information based on the comparison threshold and the principal component analysis algorithm; S332. The extracted feature information is represented as principal component data, and the Euclidean distance between the principal component data and the pre-stored verification data is calculated, and the accuracy of the segmentation result is verified based on the Euclidean distance.

Images