A paraspinal muscle abnormality identification method, system, electronic device and storage medium

Abstract

This invention belongs to the field of intelligent diagnostic technology and provides a method, system, electronic device, and storage medium for identifying paraspinal muscle abnormalities. The method includes: collecting MRI images to be identified, segmenting the paraspinal muscle boundaries, and identifying paraspinal muscle abnormalities. This invention reduces segmentation offset caused by single-slice noise and uneven grayscale by introducing an auxiliary box region center formula to associate pixel values ​​with adjacent slices in the paraspinal muscle MRI segmentation model. It also avoids reliance on manual annotation, achieving accurate and reliable localization of the paraspinal muscle region. By combining the paraspinal muscle MRI segmentation model and the paraspinal muscle abnormality identification model, the segmentation map highlights muscle region features and suppresses background interference, allowing the model to focus more on the detailed features of muscle abnormalities, reducing the rate of missed and false detections. Furthermore, the abnormality identification results with defined boundaries facilitate accurate diagnosis by physicians, reducing the risk of misdiagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology, and in particular to a method, system, electronic device, and storage medium for identifying paraspinal muscle abnormalities. Background Technology

[0002] Current paraspinal muscle abnormality analysis relies primarily on imaging, supplemented by quantitative indicators and algorithms. In imaging, MRI is the mainstream method, clearly displaying muscle contours and fat signals for assessing cross-sectional area, volume, and fat infiltration. Chemical shift MRI can accurately quantify fat fraction, and the OTSU method can assess muscle fiber (FI) without additional scanning, showing high consistency with expert opinions. Ultrasound is non-invasive and inexpensive, allowing dynamic monitoring of muscle thickness. CT assesses FI based on density values, but has poor soft tissue resolution. Among quantitative indicators, CSA reflects muscle atrophy, while FI reflects muscle degeneration, eliminating individual differences compared to CSA. In recent years, deep learning has been increasingly applied to the medical field, such as 2DnnU-Net segmentation of paraspinal muscles, improving efficiency and accuracy.

[0003] However, existing methods have limitations: traditional manual annotation is time-consuming and relies on experience; some MRI techniques are affected by uneven fat distribution; ultrasound multifocus superficial muscle has a limited assessment range; existing algorithms lack 2D / 3D fusion models and do not fully cover multiple spinal pathological types; when existing methods use neural networks to analyze MRI, they often only focus on processing the current slice and lack analysis of information from adjacent slices, which can easily lead to blurred boundaries; some anomaly recognition methods often directly use neural networks to predict abnormal areas, which are easily affected by noise and tissue overlap, resulting in blurred boundaries of the abnormal area images output by the model and low accuracy of abnormal area recognition. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method, system, electronic device and storage medium for identifying paraspinal muscle abnormalities, and to solve the problems of existing methods relying on personal experience, losing information of adjacent slices, and being susceptible to noise or tissue overlap.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for identifying paraspinal muscle abnormalities includes: Collect MRI images of the target patient to be identified; The MRI image to be identified is input into a paraspinal muscle MRI segmentation model with a preset embedded auxiliary bounding box region centering formula for boundary segmentation, resulting in a paraspinal muscle boundary segmentation map; the auxiliary bounding box region centering formula is: , , ;in, ; Pixel weights; This is the adjustment coefficient; Indicates the pixel value of the current slice; This represents the pixel value at the same coordinates as the previous slice; The coordinates of all subsequent images are the same pixel values; , These are the pixel x-coordinate and pixel y-coordinate, respectively; The pixel mean of category k; , These are the x and y coordinates of category k, respectively; The set of pixels for category k; The MRI image to be identified and the paraspinal muscle boundary segmentation map are input into a pre-trained paraspinal muscle abnormality recognition model for abnormality detection, and the paraspinal muscle abnormality recognition result is obtained.

[0006] Preferably, the MRI image to be identified is input into the paraspinal muscle MRI segmentation model of the preset embedded auxiliary box region update module for boundary segmentation to obtain a paraspinal muscle boundary segmentation map, including: The MRI image to be identified is subjected to non-uniformity correction, median filtering, and grayscale correction to obtain a preprocessed MRI image; The paraspinal muscle MRI segmentation model is constructed; the paraspinal muscle MRI segmentation model includes: a first segmentation network and a second segmentation network; the first segmentation network includes: an auxiliary box region update module and an MHFNet module connected in sequence; the second segmentation network uses the Unet++ network; The auxiliary bounding box region update module is used to perform region selection and bounding box region update on the preprocessed MRI image to obtain the final set of auxiliary bounding box regions. The preprocessed MRI image and the final set of auxiliary bounding box regions are input into the MHFNet module for initial segmentation processing to obtain the initial segmentation tensor. The preprocessed MRI image and the final set of auxiliary bounding box regions are input into the Unet++ network for secondary segmentation processing to obtain a secondary segmentation tensor. The initial segmentation tensor and the secondary segmentation tensor are fused to obtain a fused tensor. The fused tensor is then convolved to obtain the paraspinal muscle boundary segmentation map.

[0007] Preferably, the pre-training process of the paraspinal muscle abnormality recognition model includes: Collect MRI images of patients with paraspinal muscle abnormalities and normal MRI images of healthy individuals; The paraspinal muscle MRI segmentation model is used to perform boundary segmentation processing on the patient's MRI image and the normal MRI image to obtain an optimized MRI image with a boundary segmentation mask; the optimized MRI image includes: an abnormal MRI image corresponding to the patient's MRI image and a standard MRI image corresponding to the normal MRI image; Define the paraspinal muscle abnormality type and label the optimized MRI images according to the paraspinal muscle abnormality type to obtain the MRI image set to be trained; Morphological dilation is performed on the mask in the set of MRI images to be trained to obtain an enhanced mask; The enhanced mask is used to perform region weighting on the images in the MRI image set to be trained, resulting in a background-suppressed image set; The enhanced mask and the background suppression image set are channel-stitched together to obtain a multi-channel fused input. A ViT-Base model is constructed, and the ViT-Base model is iteratively trained using the multi-channel fusion input to obtain the paraspinal muscle abnormality recognition model.

[0008] Preferably, the MRI image to be identified is subjected to non-uniformity correction, median filtering, and grayscale correction to obtain a preprocessed MRI image, including: The gray-level distribution deviation of the MRI image to be identified is calculated using a calibration algorithm to obtain the gain parameter matrix and the bias parameter matrix; The gain parameter matrix and the bias parameter matrix are smoothed, and the MRI image to be identified is updated using the gradient descent algorithm based on the processed gain parameter matrix and the bias parameter matrix to obtain the initial processed MRI image; A sliding window is set up, and the initial processed MRI image is subjected to sliding median filtering using the sliding window to obtain a secondary processed MRI image. The preprocessed MRI images are obtained by performing gamma correction and saturation truncation on the secondary processed MRI images.

[0009] Preferably, the preprocessed MRI image and the final set of auxiliary bounding box regions are input into the MHFNet module for initial segmentation processing to obtain an initial segmentation tensor, including: The preprocessed MRI images were used to extract features using a graph classification network to obtain magnetic resonance image feature information. The feature image feature information and the final auxiliary box region set are spliced ​​using a feature stitcher to obtain the original feature map. The original feature map is subjected to average pooling using an enhancement processor to obtain enhanced pooled features; The enhanced pooling features are dimension-mapped using a weight allocator to obtain a weight matrix; The original feature map is scaled using the weight matrix to obtain a weighted feature map. The weighted feature map is subjected to average pooling using a fusion processor to obtain an optimized weight vector; The original feature map is calibrated using the optimized weight vector to obtain the calibration result, and the preprocessed MRI image is segmented based on the calibration result to obtain the initial segmentation tensor.

[0010] Preferably, the auxiliary bounding box region update module is used to perform region selection and bounding box region update on the preprocessed MRI image to obtain a final set of auxiliary bounding box regions, including: Each slice image in the preprocessed MRI image is rigidly registered with the previous and next slice images in pixel coordinates; Pre-defined auxiliary bounding box region update parameters; the auxiliary bounding box region update parameters include: initial patch box size, maximum number of iterations, weight adjustment coefficient, and minimum patch box size threshold; Based on the histogram statistics of the preprocessed MRI images and the preset muscle grayscale distribution range, the preprocessed MRI images are thresholded to obtain a preliminary muscle mask. Based on the initial patch box size and translation principle, the initial box of the preliminary muscle mask is generated to obtain the original auxiliary box; The pixel weights are calculated by assigning pixel weights to pixels at the same position based on the pixel values ​​of adjacent slice images. The original auxiliary box is optimized using the auxiliary box region center formula based on the pixel weights to obtain an initial auxiliary box region set. The initial auxiliary box region set is updated to the original auxiliary box, and the process returns to the step "Optimize the original auxiliary box according to the pixel weight using the auxiliary box region center formula to obtain the initial auxiliary box region set" for iteration until the maximum number of iterations is reached, and the final auxiliary box region set is obtained.

[0011] Preferably, a ViT-Base model is constructed, and the ViT-Base model is iteratively trained using the multi-channel fusion input to obtain the paraspinal muscle abnormality recognition model, including: The multi-channel fused input is divided into patch blocks according to a fixed size using an embedding layer; The patch block is feature-mapped using the Embedding layer to obtain the original embedded features; Add positional embeddings to the original embedding features to obtain positional embedding features; The location embedding features are encoded using a Transformer encoder to obtain encoded features; The encoded features are pooled, dimensionally reduced, and subjected to probability statistics using a classification head to obtain the network output. The ViT-Base model is iteratively trained using the cross-entropy loss function based on the network output to obtain the paraspinal muscle abnormality recognition model.

[0012] Preferably, a paraspinal muscle abnormality identification system includes: The data acquisition module is used to collect the MRI images of the target patient. The boundary segmentation module is used to input the MRI image to be identified into a paraspinal muscle MRI segmentation model with a preset embedded auxiliary bounding box region center formula for boundary segmentation, thereby obtaining a paraspinal muscle boundary segmentation map; the auxiliary bounding box region center formula is: , , ;in, ; Pixel weights; This is the adjustment coefficient; Indicates the pixel value of the current slice; This represents the pixel value at the same coordinates as the previous slice; The coordinates of all subsequent images are the same pixel values; , These are the pixel x-coordinate and pixel y-coordinate, respectively; The pixel mean of category k; , These are the x and y coordinates of category k, respectively; The set of pixels for category k; The paraspinal muscle abnormality recognition module is used to input the MRI image to be recognized and the paraspinal muscle boundary segmentation map into a pre-trained paraspinal muscle abnormality recognition model for abnormality detection, and obtain the paraspinal muscle abnormality recognition result.

[0013] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned paraspinal muscle abnormality identification method.

[0014] Preferably, a non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned paraspinal muscle abnormality identification method.

[0015] The present invention discloses the following technical effects: This invention provides a method, system, electronic device, and storage medium for identifying paraspinal muscle abnormalities. By introducing an auxiliary box region center formula to associate adjacent slice pixel values ​​in the paraspinal muscle MRI segmentation model, it solves the problem of losing adjacent slice information in existing methods and achieves accurate and reliable localization of the paraspinal muscle region. By combining the paraspinal muscle MRI segmentation model and the paraspinal muscle abnormality identification model, it solves the problem of being susceptible to noise and tissue overlap caused by directly using neural networks for abnormal region detection, and achieves background interference suppression and highlighting of muscle region features. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the paraspinal muscle abnormality identification process provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the boundary segmentation process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the pre-training process of the paraspinal muscle abnormality recognition model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the MRI image preprocessing workflow provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the iterative training process of the ViT-Base model provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The purpose of this invention is to provide a method, system, electronic device, and storage medium for identifying paraspinal muscle abnormalities, which solves the problems of existing methods relying on personal experience, losing information from adjacent slices, and being susceptible to noise or tissue overlap.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Figure 1 This is a schematic diagram of the paraspinal muscle abnormality identification process provided in an embodiment of the present invention, such as... Figure 1 As shown, the present invention provides a method for identifying paraspinal muscle abnormalities, comprising: Step 100: Collect the target patient's MRI images; Step 200: Input the MRI image to be identified into the paraspinal muscle MRI segmentation model with a preset embedded auxiliary box region center formula for boundary segmentation, and obtain the paraspinal muscle boundary segmentation map; the auxiliary box region center formula is: , , ;in, ; Pixel weights; This is the adjustment coefficient; Indicates the pixel value of the current slice; This represents the pixel value at the same coordinates as the previous slice; The coordinates of all subsequent images are the same pixel values; , These are the pixel x-coordinate and pixel y-coordinate, respectively; The pixel mean of category k; , These are the x and y coordinates of category k, respectively; The set of pixels for category k; Step 300: Input the MRI image to be identified and the paraspinal muscle boundary segmentation map into the pre-trained paraspinal muscle abnormality recognition model for abnormality detection, and obtain the paraspinal muscle abnormality recognition result.

[0022] refer to Figure 2 The MRI image to be identified is input into the paraspinal muscle MRI segmentation model of the preset embedded auxiliary box region update module for boundary segmentation, resulting in a paraspinal muscle boundary segmentation map, including: Step 201: Perform non-uniformity correction, median filtering, and grayscale correction on the MRI image to be identified to obtain a preprocessed MRI image; Step 202: Construct the paraspinal muscle MRI segmentation model; the paraspinal muscle MRI segmentation model includes: a first segmentation network and a second segmentation network; the first segmentation network includes: an auxiliary box region update module and an MHFNet module connected in sequence; the second segmentation network uses a Unet++ network; Step 203: Use the auxiliary box region update module to perform region selection and box region update on the preprocessed MRI image to obtain the final auxiliary box region set; Step 204: Input the preprocessed MRI image and the final auxiliary bounding box region set into the MHFNet module for initial segmentation processing to obtain the initial segmentation tensor; Step 205: Input the preprocessed MRI image and the final auxiliary box region set into the Unet++ network for secondary segmentation processing to obtain the secondary segmentation tensor; Step 206: Fuse the initial segmentation tensor and the secondary segmentation tensor to obtain a fused tensor, and perform convolution on the fused tensor to obtain the paraspinal muscle boundary segmentation map.

[0023] refer to Figure 3 The pre-training process of the paraspinal muscle abnormality recognition model includes: Step 301: Acquire MRI images of patients with paraspinal muscle abnormalities and normal MRI images of healthy individuals; Step 302: Use the paraspinal muscle MRI segmentation model to perform boundary segmentation processing on the patient's MRI image and the normal MRI image to obtain an optimized MRI image with a boundary segmentation mask; the optimized MRI image includes: an abnormal MRI image corresponding to the patient's MRI image and a standard MRI image corresponding to the normal MRI image; Step 303: Define the paraspinal muscle abnormality type and label the optimized MRI images according to the paraspinal muscle abnormality type to obtain the MRI image set to be trained; Step 304: Perform morphological dilation on the mask in the set of MRI images to be trained to obtain an enhanced mask; Step 305: Use the enhancement mask to perform region weighting on the images in the MRI image set to be trained to obtain a background-suppressed image set; Step 306: Perform channel stitching on the enhanced mask and the background suppression image set to obtain a multi-channel fused input; Step 307: Construct the ViT-Base model and iteratively train the ViT-Base model using the multi-channel fusion input to obtain the paraspinal muscle abnormality recognition model.

[0024] refer to Figure 4 The MRI image to be identified is subjected to non-uniformity correction, median filtering, and grayscale correction to obtain a preprocessed MRI image, including: Step 20101: Calculate the grayscale distribution deviation of the MRI image to be identified using a calibration algorithm to obtain the gain parameter matrix and the bias parameter matrix; Step 20102: Smooth the gain parameter matrix and the bias parameter matrix, and update the MRI image to be identified using the gradient descent algorithm based on the processed gain parameter matrix and the bias parameter matrix to obtain the initial processed MRI image; Step 20103: Set a sliding window and use the sliding window to perform sliding median filtering on the initial processed MRI image to obtain a secondary processed MRI image; Step 20104: Perform gamma correction and saturation truncation on the secondary processed MRI image to obtain the preprocessed MRI image.

[0025] Further, the preprocessed MRI image and the final set of auxiliary bounding box regions are input into the MHFNet module for initial segmentation processing to obtain an initial segmentation tensor, including: The preprocessed MRI images were used to extract features using a graph classification network to obtain magnetic resonance image feature information. The feature image feature information and the final auxiliary box region set are spliced ​​using a feature stitcher to obtain the original feature map. The original feature map is subjected to average pooling using an enhancement processor to obtain enhanced pooled features; The enhanced pooling features are dimension-mapped using a weight allocator to obtain a weight matrix; The original feature map is scaled using the weight matrix to obtain a weighted feature map. The weighted feature map is subjected to average pooling using a fusion processor to obtain an optimized weight vector; The original feature map is calibrated using the optimized weight vector to obtain the calibration result, and the preprocessed MRI image is segmented based on the calibration result to obtain the initial segmentation tensor.

[0026] Specifically, the auxiliary bounding box region update module is used to perform region selection and bounding box region update on the preprocessed MRI image to obtain a final set of auxiliary bounding box regions, including: Each slice image in the preprocessed MRI image is rigidly registered with the previous and next slice images in pixel coordinates; Pre-defined auxiliary bounding box region update parameters; the auxiliary bounding box region update parameters include: initial patch box size, maximum number of iterations, weight adjustment coefficient, and minimum patch box size threshold; Based on the histogram statistics of the preprocessed MRI images and the preset muscle grayscale distribution range, the preprocessed MRI images are thresholded to obtain a preliminary muscle mask. Based on the initial patch box size and translation principle, the initial box of the preliminary muscle mask is generated to obtain the original auxiliary box; The pixel weights are calculated by assigning pixel weights to pixels at the same position based on the pixel values ​​of adjacent slice images. The original auxiliary box is optimized using the auxiliary box region center formula based on the pixel weights to obtain an initial auxiliary box region set. The initial auxiliary box region set is updated to the original auxiliary box, and the process returns to the step "Optimize the original auxiliary box according to the pixel weight using the auxiliary box region center formula to obtain the initial auxiliary box region set" for iteration until the maximum number of iterations is reached, and the final auxiliary box region set is obtained.

[0027] refer to Figure 5 A ViT-Base model is constructed, and the ViT-Base model is iteratively trained using the multi-channel fusion input to obtain the paraspinal muscle abnormality recognition model, including: Step 30701: Divide the multi-channel fused input into patches according to a fixed size using an embedding layer; Step 30702: Use the Embedding layer to perform feature mapping on the patch block to obtain the original embedded features; Step 30703: Add positional embeddings to the original embedding features to obtain positional embedding features; Step 30704: Encode the location embedding features using a Transformer encoder to obtain encoded features; Step 30705: Use the classification head to perform pooling, dimensionality reduction mapping, and probability statistics on the encoded features to obtain the network output; Step 30706: Based on the network output, the ViT-Base model is iteratively trained using the cross-entropy loss function to obtain the paraspinal muscle abnormality recognition model.

[0028] Specifically, this embodiment adopts a process of first defining boundaries and then identifying anomalies. The overall process of paraspinal muscle MRI segmentation is as follows: image preprocessing, auxiliary box region updating, and dual-segmentation network fusion segmentation. The target patient's MRI image to be identified is acquired, and image noise and grayscale deviation are eliminated through non-uniformity correction, median filtering, and grayscale correction. An auxiliary box region updating module generates an auxiliary box set for the muscle region by combining the pixel association of adjacent slices. The dual segmentation networks are MHFNet and Unet++. MHFNet captures local muscle detail features based on the auxiliary box set, while Unet++ captures global contextual features based on the auxiliary box set. The two segmentation tensors are fused and convolved to obtain the paraspinal muscle boundary segmentation map.

[0029] Furthermore, preprocessing is required before segmenting MRI images. The preprocessing process in this embodiment includes: 1) Non-uniformity correction: The image grayscale distribution deviation is modeled using a calibration algorithm, and the gain parameter matrix and the offset parameter matrix are output. The gain parameter matrix describes the signal amplification coefficient deviation at different locations, and the offset parameter matrix describes the baseline grayscale shift. Gaussian smoothing is applied to the gain parameter matrix and the bias parameter matrix to avoid artifacts in the corrected image caused by high-frequency noise in the parameter matrix; Based on the smoothed gain parameter matrix and bias parameter matrix, the grayscale difference between the corrected image and the ideal uniform image is minimized using the gradient descent algorithm.

[0030] 2) Median filtering: A sliding window is set up, and for each pixel of each slice, the median value of all pixels within the window is taken as the new value of that pixel to obtain the secondary processed MRI image.

[0031] 3) Grayscale correction: Gamma correction was performed on the MRI images processed in step 2), and the gray values ​​of the corrected images were truncated to the range of [0, 255] to remove extreme gray value anomalies.

[0032] Specifically, existing segmentation models often process single slices individually, resulting in the loss of information from adjacent slices. This leads to significant fluctuations in the generated bounding boxes for adjacent slices. Furthermore, the lack of consideration for adjacent slices reduces the reliability of the generated boxes. Therefore, this embodiment adds an auxiliary bounding box region update mechanism before model segmentation: First, rigidly register the pixel coordinates of three adjacent slices, grouping three pixels from three slices with uniform coordinates into a single group for analysis; set update parameters, including initial patch box size, maximum number of iterations, weight adjustment coefficient, and minimum patch box size threshold; determine the grayscale distribution range of the paraspinal muscles based on the histogram statistics of the preprocessed MRI images, and obtain a preliminary muscle mask through threshold segmentation; generate original auxiliary bounding boxes based on translation principles and the preliminary muscle mask; update the auxiliary bounding boxes using the auxiliary bounding box region center formula, as follows: .

[0033] .

[0034] .

[0035] .

[0036] in, Pixel weights; This is the adjustment coefficient; Indicates the pixel value of the current slice; This represents the pixel value at the same coordinates as the previous slice; The coordinates of all subsequent images are the same pixel values; , These are the pixel x-coordinate and pixel y-coordinate, respectively; The pixel mean of category k; , These are the x and y coordinates of category k, respectively; Let k be the set of pixels of category k.

[0037] The above formula is used to calculate the initial set of auxiliary boxes after optimization. The initial set of auxiliary boxes is used as the new original auxiliary boxes to recalculate the pixel weights and weighted centers to update the box positions and sizes until the maximum number of iterations is reached, and the final set of auxiliary box regions is output.

[0038] Preferably, during segmentation, MHFNet uses a graph classification network to extract features from the input MRI image, outputting magnetic resonance image feature information. The final auxiliary bounding box region set and the graph classification network output are then concatenated by scale channels to obtain a feature map. Average pooling is performed on the feature map to obtain enhanced pooling features; this process compresses the spatial dimension and highlights global features. A weight allocator maps the number of channels in the enhanced pooling features to the number of channels in the feature map. Figure 1The weight matrix is ​​obtained by multiplying the weight matrix element-wise with the feature map to obtain a weighted feature map. In this feature map, the features of the muscle region are amplified, while the features of the background are suppressed. Global average pooling is then performed on the weighted feature map to obtain an optimized weight vector. The feature map is then labeled using the optimized weight vector, and a convolutional layer is used to map the labeled feature map to obtain the initial segmentation tensor. Unet++ is used as a secondary segmentation layer to supplement global context features. Unet++ has higher segmentation accuracy than the traditional Unet. The segmentation results of this model are particularly good in the blurred region where the paraspinal muscle meets fat. Therefore, the segmentation results of MHFNet and Unet++ are combined to further improve the segmentation reliability.

[0039] Specifically, two types of core data were collected: abnormal samples: MRI images of patients with paraspinal muscle abnormalities, and normal samples: MRI images of healthy individuals without paraspinal muscle abnormalities. Abnormal samples typically constitute a low percentage in medical imaging; therefore, the data used for model training includes both abnormal and normal samples to avoid overfitting the model to normal samples. The aforementioned paraspinal muscle MRI segmentation model was used to segment the boundaries of the collected MRI images of patients with abnormalities and those of healthy individuals, outputting optimized MRI images with paraspinal muscle boundary segmentation masks. In paraspinal muscle MRI, the muscle region only occupies 15% to 25% of the total image area, with the remainder consisting of the vertebral body, fat, soft tissue, and other background. Focusing on the muscle region through segmentation masks improves the computational efficiency of the subsequent abnormality recognition model while avoiding interference from background noise in feature extraction. Paraspinal muscle abnormality types were defined based on clinical needs, and radiologists labeled the optimized MRI images to form the MRI image set for training.

[0040] Furthermore, a morphological dilation operation is performed on the paraspinal muscle segmentation mask in the training MRI image set to obtain an enhancement mask. This enhancement mask is then used to perform region-weighted processing on the original images in the training MRI image set to obtain a background-suppressed image set. The enhancement mask and the background-suppressed image set are then concatenated along the channel dimension to obtain a multi-channel fused input. After obtaining the model input, the ViT-Base model is pre-trained. 1) Use an embedding layer to divide the multi-channel fused input into non-overlapping patch blocks of a fixed size.

[0041] 2) The original embedded features are obtained by performing feature mapping on each patch block through the Embedding layer.

[0042] 3) Location embedding: Add location embedding to the original embedding features to obtain location embedding features. Because the paraspinal muscles have irregular shapes, learningable location embedding can better adapt to the features of different muscle locations.

[0043] 4) Input the location embedding features into the Transformer encoder and perform multi-layer feature encoding to obtain encoded features. The multi-head attention mechanism of the Transformer encoder can capture long-distance dependencies between patches and is more suitable for anomaly detection of large-scale tissues such as paraspinal muscles than traditional CNN (which depends on local convolution kernels).

[0044] 5) Process the encoded features using the classification head: Perform global average pooling on the encoded features, reduce the dimensionality of the encoded features using a fully connected layer, and convert the output into the probability distribution of each category using the Softmax function to obtain the network output.

[0045] 6) Loss function optimization: The ViT-Base model is iteratively trained using the cross-entropy loss function until the model converges.

[0046] As an optional implementation, this embodiment also provides a paraspinal muscle abnormality identification system, including: The data acquisition module is used to collect the MRI images of the target patient. The boundary segmentation module is used to input the MRI image to be identified into a paraspinal muscle MRI segmentation model with a preset embedded auxiliary bounding box region center formula for boundary segmentation, thereby obtaining a paraspinal muscle boundary segmentation map; the auxiliary bounding box region center formula is: , , ;in, ; Pixel weights; This is the adjustment coefficient; Indicates the pixel value of the current slice; This represents the pixel value at the same coordinates as the previous slice; The coordinates of all subsequent images are the same pixel values; , These are the pixel x-coordinate and pixel y-coordinate, respectively; The pixel mean of category k; , These are the x and y coordinates of category k, respectively; The set of pixels for category k; The paraspinal muscle abnormality recognition module is used to input the MRI image to be recognized and the paraspinal muscle boundary segmentation map into a pre-trained paraspinal muscle abnormality recognition model for abnormality detection, and obtain the paraspinal muscle abnormality recognition result.

[0047] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned paraspinal muscle abnormality identification method.

[0048] As an optional implementation, this embodiment also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned paraspinal muscle abnormality identification method.

[0049] The beneficial effects of this invention are as follows: This invention reduces segmentation offset caused by single-slice noise and uneven grayscale by introducing an auxiliary box region center formula to associate pixel values ​​with adjacent slices in the paraspinal muscle MRI segmentation model. At the same time, it avoids dependence on manual annotation and achieves accurate and reliable localization of the paraspinal muscle region. By combining the paraspinal muscle MRI segmentation model and the paraspinal muscle abnormality recognition model, the segmentation map is used to highlight muscle region features and suppress background interference, making the model more focused on the detailed features of muscle abnormalities and reducing the false negative and false positive rates. In addition, the abnormality recognition results with defined boundaries facilitate physicians to accurately determine the cause of the disease and reduce the risk of misdiagnosis.

[0050] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0051] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying paraspinal muscle abnormalities, characterized in that, include: Collect MRI images of the target patient to be identified; The MRI image to be identified is input into a paraspinal muscle MRI segmentation model with a preset embedded auxiliary bounding box region centering formula for boundary segmentation, resulting in a paraspinal muscle boundary segmentation map; the auxiliary bounding box region centering formula is: , , ;in, ; Pixel weights; This is the adjustment coefficient; Indicates the pixel value of the current slice; This represents the pixel value at the same coordinates as the previous slice; The coordinates of all subsequent images are the same pixel values; , These are the pixel x-coordinate and pixel y-coordinate, respectively; The pixel mean of category k; , These are the x and y coordinates of category k, respectively; The set of pixels for category k; The MRI image to be identified and the paraspinal muscle boundary segmentation map are input into a pre-trained paraspinal muscle abnormality recognition model for abnormality detection, and the paraspinal muscle abnormality recognition result is obtained.

2. The method for identifying paraspinal muscle abnormalities according to claim 1, characterized in that, The MRI image to be identified is input into the paraspinal muscle MRI segmentation model of the preset embedded auxiliary box region update module for boundary segmentation, resulting in a paraspinal muscle boundary segmentation map, including: The MRI image to be identified is subjected to non-uniformity correction, median filtering, and grayscale correction to obtain a preprocessed MRI image; The paraspinal muscle MRI segmentation model is constructed; the paraspinal muscle MRI segmentation model includes: a first segmentation network and a second segmentation network; the first segmentation network includes: an auxiliary box region update module and an MHFNet module connected in sequence; the second segmentation network uses the Unet++ network; The auxiliary bounding box region update module is used to perform region selection and bounding box region update on the preprocessed MRI image to obtain the final set of auxiliary bounding box regions. The preprocessed MRI image and the final set of auxiliary bounding box regions are input into the MHFNet module for initial segmentation processing to obtain the initial segmentation tensor. The preprocessed MRI image and the final set of auxiliary bounding box regions are input into the Unet++ network for secondary segmentation processing to obtain a secondary segmentation tensor. The initial segmentation tensor and the secondary segmentation tensor are fused to obtain a fused tensor. The fused tensor is then convolved to obtain the paraspinal muscle boundary segmentation map.

3. The method for identifying paraspinal muscle abnormalities according to claim 1, characterized in that, The pre-training process of the paraspinal muscle abnormality recognition model includes: Collect MRI images of patients with paraspinal muscle abnormalities and normal MRI images of healthy individuals; The paraspinal muscle MRI segmentation model is used to perform boundary segmentation processing on the patient's MRI image and the normal MRI image to obtain an optimized MRI image with a boundary segmentation mask; the optimized MRI image includes: an abnormal MRI image corresponding to the patient's MRI image and a standard MRI image corresponding to the normal MRI image; Define the paraspinal muscle abnormality type and label the optimized MRI images according to the paraspinal muscle abnormality type to obtain the MRI image set to be trained; Morphological dilation is performed on the mask in the set of MRI images to be trained to obtain an enhanced mask; The enhanced mask is used to perform region weighting on the images in the MRI image set to be trained, resulting in a background-suppressed image set; The enhanced mask and the background suppression image set are channel-stitched together to obtain a multi-channel fused input. A ViT-Base model is constructed, and the ViT-Base model is iteratively trained using the multi-channel fusion input to obtain the paraspinal muscle abnormality recognition model.

4. The method for identifying paraspinal muscle abnormalities according to claim 2, characterized in that, The MRI image to be identified is subjected to non-uniformity correction, median filtering, and grayscale correction to obtain a preprocessed MRI image, including: The gray-level distribution deviation of the MRI image to be identified is calculated using a calibration algorithm to obtain the gain parameter matrix and the bias parameter matrix; The gain parameter matrix and the bias parameter matrix are smoothed, and the MRI image to be identified is updated using the gradient descent algorithm based on the processed gain parameter matrix and the bias parameter matrix to obtain the initial processed MRI image; A sliding window is set up, and the initial processed MRI image is subjected to sliding median filtering using the sliding window to obtain a secondary processed MRI image. The preprocessed MRI images are obtained by performing gamma correction and saturation truncation on the secondary processed MRI images.

5. The method for identifying paraspinal muscle abnormalities according to claim 2, characterized in that, The preprocessed MRI image and the final set of auxiliary bounding box regions are input into the MHFNet module for initial segmentation processing to obtain the initial segmentation tensor, including: The preprocessed MRI images were used to extract features using a graph classification network to obtain magnetic resonance image feature information. The feature image feature information and the final auxiliary box region set are spliced ​​using a feature stitcher to obtain the original feature map. The original feature map is subjected to average pooling using an enhancement processor to obtain enhanced pooled features; The enhanced pooling features are dimension-mapped using a weight allocator to obtain a weight matrix; The original feature map is scaled using the weight matrix to obtain a weighted feature map. The weighted feature map is subjected to average pooling using a fusion processor to obtain an optimized weight vector; The original feature map is calibrated using the optimized weight vector to obtain the calibration result, and the preprocessed MRI image is segmented based on the calibration result to obtain the initial segmentation tensor.

6. The method for identifying paraspinal muscle abnormalities according to claim 2, characterized in that, The auxiliary bounding box region update module is used to perform region selection and bounding box region update on the preprocessed MRI image to obtain a final set of auxiliary bounding box regions, including: Each slice image in the preprocessed MRI image is rigidly registered with the previous and next slice images in pixel coordinates; Pre-defined auxiliary bounding box region update parameters; the auxiliary bounding box region update parameters include: initial patch box size, maximum number of iterations, weight adjustment coefficient, and minimum patch box size threshold; Based on the histogram statistics of the preprocessed MRI images and the preset muscle grayscale distribution range, the preprocessed MRI images are thresholded to obtain a preliminary muscle mask. Based on the initial patch box size and translation principle, the initial box of the preliminary muscle mask is generated to obtain the original auxiliary box; The pixel weights are calculated by assigning pixel weights to pixels at the same position based on the pixel values ​​of adjacent slice images. The original auxiliary box is optimized using the auxiliary box region center formula based on the pixel weights to obtain an initial auxiliary box region set. The initial auxiliary box region set is updated to the original auxiliary box, and the process returns to the step "Optimize the original auxiliary box according to the pixel weight using the auxiliary box region center formula to obtain the initial auxiliary box region set" for iteration until the maximum number of iterations is reached, and the final auxiliary box region set is obtained.

7. The method for identifying paraspinal muscle abnormalities according to claim 3, characterized in that, A ViT-Base model is constructed, and the ViT-Base model is iteratively trained using the multi-channel fusion input to obtain the paraspinal muscle abnormality recognition model, including: The multi-channel fused input is divided into patch blocks according to a fixed size using an embedding layer; The patch block is feature-mapped using the Embedding layer to obtain the original embedded features; Add positional embeddings to the original embedding features to obtain positional embedding features; The location embedding features are encoded using a Transformer encoder to obtain encoded features; The encoded features are pooled, dimensionally reduced, and subjected to probability statistics using a classification head to obtain the network output. The ViT-Base model is iteratively trained using the cross-entropy loss function based on the network output to obtain the paraspinal muscle abnormality recognition model.

8. A paraspinal muscle abnormality identification system, characterized in that, For implementing the paraspinal muscle abnormality identification method according to claim 1, the system comprises: The data acquisition module is used to collect the MRI images of the target patient. The boundary segmentation module is used to input the MRI image to be identified into a paraspinal muscle MRI segmentation model with a preset embedded auxiliary bounding box region center formula for boundary segmentation, thereby obtaining a paraspinal muscle boundary segmentation map; the auxiliary bounding box region center formula is: , , ;in, ; Pixel weights; This is the adjustment coefficient; Indicates the pixel value of the current slice; This represents the pixel value at the same coordinates as the previous slice; The coordinates of all subsequent images are the same pixel values; , These are the pixel x-coordinate and pixel y-coordinate, respectively; The pixel mean of category k; , These are the x and y coordinates of category k, respectively; The set of pixels for category k; The paraspinal muscle abnormality recognition module is used to input the MRI image to be recognized and the paraspinal muscle boundary segmentation map into a pre-trained paraspinal muscle abnormality recognition model for abnormality detection, and obtain the paraspinal muscle abnormality recognition result.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor to enable the processor to perform a paraspinal muscle abnormality identification method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute a paraspinal muscle abnormality identification method according to any one of claims 1 to 7.

Images