Spine axial plane magnetic resonance image paravertebral muscle fat segmentation and quantitative analysis method and product
By applying multi-task learning segmentation model and PSO-Kapur threshold segmentation algorithm on spinal axial magnetic resonance imaging, we automatically measure the degree of fat infiltration of the lumbar paravertebrae muscles, solving the measurement error problem in the existing methods, and improving the diagnostic efficiency and accuracy.
Patent Information
- Application Number
- CN202411158419.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-11-26
AI Technical Summary
The existing clinical methods have subjective and empirical measurement errors when measuring the degree of paravertebral muscle fat infiltration of lumbar spine, which limits the auxiliary diagnosis efficiency of lumbar spine diseases and the development of related research.
Using a pre-trained multi-task learning segmentation model, combined with CLAHE image enhancement and PSO-Kapur threshold segmentation algorithm, paravertebral muscle fat segmentation and quantitative analysis were performed on spinal axial magnetic resonance images, and segmentation and calculation of related parameters were automated.
It effectively reduces subjective errors caused by manual labeling, improves diagnostic efficiency, realizes fully automated paravertebral muscle fat segmentation and quantitative analysis, and improves the accuracy and efficiency of diagnosis.
Smart Images

Figure CN119027402A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present invention relate to the field of medical image processing technology, and in particular to a method and product for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging. Background Art
[0002] The main cause of low back pain is the degenerative changes of the lumbar paraspinal muscles, which are usually caused by fatty infiltration of the paraspinal muscles. Therefore, measuring the degree of fatty infiltration of the paraspinal muscles can assist in the diagnosis of low back pain and effectively prevent a series of lumbar specific and non-specific diseases and other symptoms.
[0003] Magnetic resonance imaging has excellent soft tissue resolution characteristics and can reveal the details of changes in muscle tissue structure in a more delicate manner. Related studies have shown that the results of MRI fat measurement are significantly consistent with those of biopsy measurement. This method can replace biopsy to complete fat measurement and is currently the best means to evaluate muscle fat infiltration. However, it is common for clinicians to use visual and semi-quantitative measurements to measure paraspinal muscle fat infiltration, but there are often measurement errors caused by subjectivity and experience, which limits the efficiency of auxiliary diagnosis of lumbar diseases and hinders the development of research related to paraspinal muscle fat infiltration. Summary of the invention
[0004] A first aspect of an embodiment of the present invention provides a method for segmenting and quantitatively analyzing paraspinal muscle fat in spinal axial magnetic resonance imaging, the method comprising:
[0005] Acquire an original magnetic resonance image of the spine axial plane of the target object;
[0006] Based on a pre-trained paraspinal muscle segmentation model, each paraspinal muscle image is segmented; the paraspinal muscle segmentation model is obtained by training a multi-task learning segmentation model based on a pre-acquired sample spinal axial plane magnetic resonance original image with paraspinal muscles and intervertebral disc areas annotated;
[0007] Determine the rectangular frame corresponding to the paraspinal muscle image, and use the CLAHE image enhancement algorithm to enhance the image in the rectangular frame; process the enhanced paraspinal muscle image using the PSO-Kapur threshold segmentation algorithm to obtain the paraspinal muscle fat segmentation area;
[0008] Calculate at least one of the following parameters based on the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image: tCSA of each muscle, fCSA of each muscle, Kader grade, tCSA ratio of both sides of multifidus muscle, tCSA ratio of both sides of erector spinae muscle, and fat infiltration rate of each muscle;
[0009] The paraspinal muscle fat segmentation area and the calculated parameters corresponding to each paraspinal muscle image are output.
[0010] Optionally, the multi-task learning segmentation model includes: an encoder, a coarse segmentation decoder, a fine segmentation decoder, a region detection module, and a fusion module. The paraspinal muscle segmentation model is trained by the following steps:
[0011] Collecting the original axial magnetic resonance images of the spine as sample axial magnetic resonance images of the spine, and obtaining the annotation information of the paravertebral muscles and intervertebral disc areas corresponding to each sample axial magnetic resonance image of the spine as sample annotations;
[0012] Input the sample spinal axial plane magnetic resonance original image into the encoder to obtain the sample image code, input the sample image code into the coarse segmentation decoder to obtain the sample coarse segmentation result, input the sample coarse segmentation result into the region detection module to obtain the sample region detection result, input the sample region detection result and the sample coarse segmentation result into the fusion module to obtain the sample fusion result, input the sample image code and the sample fusion result into the fine segmentation decoder to obtain the sample fine segmentation result;
[0013] Based on the sample fine segmentation result and the sample annotation, the parameters of each module are updated until the model converges to obtain a first multi-task learning segmentation model;
[0014] Freeze the model parameters of the encoder and the fusion module in the first multi-task learning segmentation model, and alternately update the parameters of the coarse segmentation module, the region detection module, and the fine segmentation module based on the sample fine segmentation results and annotations until the model converges to obtain a second multi-task learning segmentation model;
[0015] Freeze the model parameters of the encoder, coarse segmentation module, region detection module and fusion module in the second multi-task learning segmentation model, and update the parameters of the fine segmentation module based on the sample fine segmentation results and annotations until the model converges to obtain the paraspinal muscle segmentation model.
[0016] Optionally, the sample region detection result and the sample rough segmentation result are input into a fusion module to obtain a sample fusion result, including:
[0017] The sample area detection results and the sample coarse segmentation results are input into the fusion module. The fusion module maps each muscle recognition frame included in the sample area detection result into a one-hot code image with the same size as the input image to obtain the one-hot code image of the recognition frame of each category. The one-hot code image of the recognition frame of each category and the sample coarse segmentation result are superimposed into an attention map. The sample coarse segmentation result is then jump-connected to the attention map to obtain the sample fusion result as the output of the multi-task fusion module.
[0018] Optionally, the image in the rectangular frame is enhanced by using the CLAHE image enhancement algorithm, and the enhanced paraspinal muscle images are processed by the PSO-Kapur threshold segmentation algorithm to obtain each paraspinal muscle fat segmentation area, including:
[0019] The muscle images of each paraspinal muscle are obtained by segmentation, and the segmentation area rectangular frame of each paraspinal muscle area is calculated;
[0020] Based on the rectangular frames of each segmented area, the images in each segmented area are retained, the grayscale values of pixels outside the segmented area are set to 0, and the CLAHE image enhancement algorithm is used to highlight the local details and image contrast of the image in the segmented area, and the enhanced images of each paraspinal muscle area are obtained respectively;
[0021] The PSO-Kapur algorithm was used to segment the enhanced images of each paraspinal muscle region. The Kapur entropy was used as the objective function to maximize the objective function, and the particle swarm optimization algorithm was used to optimize the running speed of Kapur multi-threshold segmentation to obtain a value close to the optimal solution.
[0022] Obtaining muscle fat segmentation results based on the PSO-Kapur segmentation results, combining the muscle fat segmentation results with the coordinates of the respective segmentation area rectangular frames into corresponding areas of the same image, and then superimposing them onto the original spinal axial magnetic resonance image, and obtaining an image as a threshold segmentation visualization image;
[0023] Among them, the threshold number of Kapur multi-threshold segmentation is 2.
[0024] Optionally, the Kapur entropy is used as the objective function to maximize the objective function, including:
[0025] Use multiple thresholds to divide the image into different segments. The following is the Kapur entropy for each threshold interval:
[0026] ......
[0028] ......
[0030]
[0031] Among them, p i is the probability of event i occurring, H represents Kapur entropy, a, b, c…j…m represent thresholds, p a , p b ,…,p m The grayscale probability of different threshold intervals is represented by a multi-threshold method;
[0032] In order to select the optimal threshold, the following objective function should be maximized:
[0033]
[0034] H i (th) represents the Kapur entropy in the i-th threshold interval when the threshold is th;
[0035] To extend the multi-level threshold expression, Kapur’s objective function is defined as:
[0036] f Kapur (TH) = f Kapur (th i ),i=1,2,3,…,k;
[0037] Different thresholds are represented by TH[th 1, th2, th3…, thk-1], and i corresponds to a specific threshold interval.
[0038] Optionally, a particle swarm optimization algorithm is used to optimize the running speed of Kapur multi-threshold segmentation to obtain a value close to the optimal solution, including:
[0039] Generate initial position and initial velocity;
[0040] Evaluate the objective function for each particle;
[0041] Update the single optimal solution and the global optimal solution, the single optimal solution is the single optimal particle, and the global optimal solution is the global optimal particle
[0042] Update the velocity v and position x of each particle;
[0043] Repeat the particle swarm optimization process until the termination condition is reached;
[0044] In the particle swarm optimization algorithm, the particle velocity update formula is as follows:
[0045] v j (k+1)=wv j (k)+c1r1(p j (k)-x j (k))+c2r2(g(k)-x j (k)),j=1,2,…,np.
[0046] In the particle swarm optimization algorithm, the particle position update formula is as follows:
[0047] x j (k+1)=x j (k)+v j (k);
[0048] Among them, vj (k) and x j (k) are the velocity and position of the jth particle at the kth iteration, respectively, and p j g(k) and g(k) are the single best particle and the global best particle at the kth iteration, respectively; w is the initial weight parameter, c1 is the individual cognitive parameter, and c2 is the social cognitive parameter; r1 and r2 are random numbers between 0 and 1 with uniform distribution.
[0049] A second aspect of an embodiment of the present invention provides a device for segmenting and quantitatively analyzing paraspinal muscle fat in a spinal axial magnetic resonance image, the device comprising:
[0050] A target image acquisition module, used for acquiring an original magnetic resonance image of the spine axial plane of a target object;
[0051] A muscle image segmentation module is used to segment and obtain each paraspinal muscle image based on a pre-trained paraspinal muscle segmentation model; the paraspinal muscle segmentation model is obtained by training a multi-task learning segmentation model based on a pre-acquired sample spinal axial plane magnetic resonance original image labeled with paraspinal muscles and intervertebral disc areas;
[0052] The fat segmentation module is used to determine the rectangular boxes corresponding to the muscle images of each paraspinal muscle, and to enhance the image in the rectangular box using the CLAHE image enhancement algorithm; the enhanced muscle images of each paraspinal muscle are processed using the PSO-Kapur threshold segmentation algorithm to obtain the fat segmentation area of each paraspinal muscle;
[0053] A parameter analysis module, for calculating at least one of the following parameters based on the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image: tCSA of each muscle, fCSA of each muscle, Kader grade, tCSA ratio of both sides of multifidus muscle, tCSA ratio of both sides of erector spinae muscle, and fat infiltration rate of each muscle;
[0054] The output module is used to output the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image and the calculated parameters.
[0055] Optionally, the multi-task learning segmentation model includes: an encoder, a coarse segmentation decoder, a fine segmentation decoder, a region detection module, and a fusion module. The paraspinal muscle segmentation model is trained by the following steps:
[0056] Collecting the original axial magnetic resonance images of the spine as sample axial magnetic resonance images of the spine, and obtaining the annotation information of the paravertebral muscles and intervertebral disc areas corresponding to each sample axial magnetic resonance image of the spine as sample annotations;
[0057] Input the sample spinal axial plane magnetic resonance original image into the encoder to obtain the sample image code, input the sample image code into the coarse segmentation decoder to obtain the sample coarse segmentation result, input the sample coarse segmentation result into the region detection module to obtain the sample region detection result, input the sample region detection result and the sample coarse segmentation result into the fusion module to obtain the sample fusion result, input the sample image code and the sample fusion result into the fine segmentation decoder to obtain the sample fine segmentation result;
[0058] Based on the sample fine segmentation result and the sample annotation, the parameters of each module are updated until the model converges to obtain a first multi-task learning segmentation model;
[0059] Freeze the model parameters of the encoder and the fusion module in the first multi-task learning segmentation model, and alternately update the parameters of the coarse segmentation module, the region detection module, and the fine segmentation module based on the sample fine segmentation results and annotations until the model converges to obtain a second multi-task learning segmentation model;
[0060] Freeze the model parameters of the encoder, coarse segmentation module, region detection module and fusion module in the second multi-task learning segmentation model, and update the parameters of the fine segmentation module based on the sample fine segmentation results and annotations until the model converges to obtain the paraspinal muscle segmentation model.
[0061] Optionally, the sample region detection result and the sample rough segmentation result are input into a fusion module to obtain a sample fusion result, including:
[0062] The sample area detection results and the sample coarse segmentation results are input into the fusion module. The fusion module maps each muscle recognition frame included in the sample area detection result into a one-hot code image with the same size as the input image to obtain the one-hot code image of the recognition frame of each category. The one-hot code image of the recognition frame of each category and the sample coarse segmentation result are superimposed into an attention map. The sample coarse segmentation result is then jump-connected to the attention map to obtain the sample fusion result as the output of the multi-task fusion module.
[0063] Optionally, the fat segmentation module is used to:
[0064] Using the segmented paraspinal muscle images, the segmented rectangular frame of each paraspinal muscle region is calculated;
[0065] Based on the rectangular frames of each segmented area, the images in each segmented area are retained, the grayscale values of pixels outside the segmented area are set to 0, and the CLAHE image enhancement algorithm is used to highlight the local details and image contrast of the image in the segmented area, and the enhanced images of each paraspinal muscle area are obtained respectively;
[0066] The PSO-Kapur algorithm was used to segment the enhanced images of each paraspinal muscle region. The Kapur entropy was used as the objective function to maximize the objective function, and the particle swarm optimization algorithm was used to optimize the running speed of Kapur multi-threshold segmentation to obtain a value close to the optimal solution.
[0067] Obtaining muscle fat segmentation results based on the PSO-Kapur segmentation results, combining the muscle fat segmentation results with the coordinates of the respective segmentation area rectangular frames into corresponding areas of the same image, and then superimposing them onto the original spinal axial magnetic resonance image, and obtaining an image as a threshold segmentation visualization image;
[0068] Among them, the threshold number of Kapur multi-threshold segmentation is 2.
[0069] Optionally, the fat segmentation module is further used for:
[0070] Use multiple thresholds to divide the image into different segments. The following is the Kapur entropy for each threshold interval:
[0071] ......
[0073] ......
[0075]
[0076] Among them, p i is the probability of event i occurring, H represents Kapur entropy, a, b, c…j…m represent thresholds, p a , p b ,…,p m The grayscale probability of different threshold intervals is represented by a multi-threshold method;
[0077] In order to select the optimal threshold, the following objective function should be maximized:
[0078]
[0079] H i (th) represents the Kapur entropy in the i-th threshold interval when the threshold is th;
[0080] To extend the multi-level threshold expression, Kapur’s objective function is defined as:
[0081] f Kapur (TH) = f Kapur (th i ),i=1,2,3,…,k;
[0082] Different thresholds are represented by TH[th1, th2, th3, ..., thk], and i corresponds to a specific threshold interval.
[0083] Optionally, the fat segmentation module is further used for:
[0084] Generate initial position and initial velocity;
[0085] Evaluate the objective function for each particle;
[0086] Update the single optimal solution and the global optimal solution, the single optimal solution is the single optimal particle, and the global optimal solution is the global optimal particle
[0087] Update the velocity v and position x of each particle;
[0088] Repeat the particle swarm optimization process until the termination condition is reached;
[0089] In the particle swarm optimization algorithm, the particle velocity update formula is as follows:
[0090] v j (k+1)=wv j (k)+c1r1(p j (k)-x j (k))+c2r2(g(k)-x j (k)),j=1,2,…,np.
[0091] In the particle swarm optimization algorithm, the particle position update formula is as follows:
[0092] x j (k+1)=x j (k)+v j (k);
[0093] Among them, v j (k) and x j (k) are the velocity and position of the jth particle at the kth iteration, respectively, and p j g(k) and g(k) are the single best particle and the global best particle at the kth iteration, respectively; w is the initial weight parameter, c1 is the individual cognitive parameter, and c2 is the social cognitive parameter; r1 and r2 are random numbers between 0 and 1 with uniform distribution.
[0094] A third aspect of an embodiment of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements the steps in the method for segmentation and quantitative analysis of paraspinal muscle fat in axial magnetic resonance imaging of the spine as described in the first aspect of the present invention.
[0095] A fourth aspect of an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for segmenting and quantifying paraspinal muscle fat in spinal axial plane magnetic resonance imaging as described in the first aspect of the present invention.
[0096] A fifth aspect of an embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the method for segmenting and quantifying paraspinal muscle fat in spinal axial plane magnetic resonance imaging described in the first aspect of the present invention.
[0097] In an embodiment of the present invention, based on the pre-acquired sample spinal axial plane magnetic resonance original image marked with paraspinal muscles and intervertebral disc areas, a multi-task learning segmentation model is trained to obtain a paraspinal muscle segmentation model, and based on the paraspinal muscle segmentation model, the spinal axial plane magnetic resonance original image can be automatically processed to segment and obtain individual paraspinal muscle images; then, image enhancement is performed on the segmented paraspinal muscle images; then, based on the PSO-Kapur threshold segmentation algorithm, the enhanced paraspinal muscle images are processed to obtain paraspinal muscle fat segmentation areas, and then, based on the paraspinal muscle fat segmentation areas corresponding to the individual paraspinal muscle images, relevant parameters are analyzed, thereby realizing a fully automatic segmentation of paraspinal muscle fat and quantitative analysis method, and realizing the function of automatic fat quantification and statistics.
[0098] The fully automatic segmentation and quantitative analysis method of paraspinal muscle fat proposed in the present invention can effectively reduce the subjective error and experience error caused by manual labeling, while greatly improving the diagnostic efficiency of clinicians, and has wide application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0099] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0100] Figure 1 A flowchart showing the steps of the spinal axial plane magnetic resonance imaging paraspinal muscle fat segmentation and quantitative analysis method provided by an embodiment of the present invention is shown;
[0101] Figure 2 A schematic diagram of the data processing flow of the paraspinal muscle segmentation model in the spinal axial plane magnetic resonance imaging paraspinal muscle fat segmentation and quantitative analysis method provided by an embodiment of the present invention is shown;
[0102] Figure 3A schematic diagram showing the operating results obtained by the spinal axial plane magnetic resonance imaging paraspinal muscle fat segmentation and quantitative analysis method provided by an embodiment of the present invention;
[0103] Figure 4 The structure block diagram of the apparatus for segmenting and quantitatively analyzing paraspinal muscle fat in spinal axial magnetic resonance imaging provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0104] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0105] In an embodiment of the present invention, a method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging is provided. Figure 1 As shown, it shows a flowchart of the steps of a method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging according to an embodiment of the present invention, the method comprising:
[0106] S101, obtaining an original magnetic resonance image of the spine axial plane of a target object.
[0107] In an embodiment of the present invention, the target object's spinal axial plane magnetic resonance original image can be acquired with the help of a magnetic resonance imaging device commonly used in the relevant technology (for example: Philips Achieva 1.5 Tesla magnetic resonance scanner). In an embodiment of the present invention, the target object's T2 axial sequence image data can be acquired specifically. During the acquisition process, a fast spin echo sequence is set with a repetition time of 2360 milliseconds, an echo time of 94 milliseconds, an echo chain length of 33, a matrix pixel size of 400*400 pixels, a field of view FOV size of 200*200 mm, and a slice thickness of 4 mm. The acquired spinal axial plane magnetic resonance original image can be saved in DICOM format.
[0108] In the embodiment of the present invention, the original axial magnetic resonance image of the spine of the target object is specifically: an L4-L5 regional image, and the L4-L5 regional image is obtained based on the T2 axial sequence lumbar image of the target object combined with the sagittal image.
[0109] Specifically, the target object's T2 axial sequence lumbar spine image can also be shown to professional technicians, so that the professional technicians can determine the image located in the L4-L5 region based on experience and sagittal image, and then determine the image located in the L4-L5 region as the target object's spinal axial plane magnetic resonance original image.
[0110] S102, based on a pre-trained paraspinal muscle segmentation model, segment and obtain individual paraspinal muscle images; the paraspinal muscle segmentation model is obtained by training a multi-task learning segmentation model based on pre-acquired sample spinal axial plane magnetic resonance original images annotated with paraspinal muscles and intervertebral disc areas.
[0111] In the embodiment of the present invention, T2 axial sequence image data of multiple patients can be collected based on a magnetic resonance imaging device as sample spine axial plane magnetic resonance original images.
[0112] In the embodiment of the present invention, the collected original axial magnetic resonance image set of the spine can also be screened to eliminate T2 axial sequence image data of patients with obesity, systemic diseases, lumbar surgery and scoliosis.
[0113] In an embodiment of the present invention, a professional technician (e.g., a doctor) can annotate the paraspinal muscle area and the intervertebral disc area based on the sample spinal axial plane magnetic resonance original image. Specifically, the paraspinal muscle area includes: left erector spinae, left multifidus, left psoas major, right erector spinae, right multifidus, and right psoas major.
[0114] In the embodiment of the present invention, the labeling information of the original magnetic resonance image of the sample spine axial plane may be obtained by labeling and correction by multiple professional technicians.
[0115] S103, determining the rectangular frames corresponding to the respective paraspinal muscle images, and performing image enhancement on the images within the rectangular frames using the CLAHE image enhancement algorithm; and processing the enhanced paraspinal muscle images using the PSO-Kapur threshold segmentation algorithm to obtain the respective paraspinal muscle fat segmentation regions.
[0116] In an embodiment of the present invention, each paraspinal muscle image segmented by the paraspinal muscle segmentation model can be processed separately. For a certain paraspinal muscle image, its corresponding rectangular frame can be determined first, and the image within the rectangular frame can be enhanced using the CLAHE image enhancement algorithm; the enhanced paraspinal muscle image is processed using the PSO-Kapur threshold segmentation algorithm to obtain the paraspinal muscle fat segmentation area corresponding to the muscle.
[0117] S104, calculating at least one of the following parameters based on the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image: tCSA of each muscle, fCSA of each muscle, Kader grade, tCSA ratio of both sides of multifidus muscle, tCSA ratio of both sides of erector spinae muscle, and fat infiltration rate of each muscle.
[0118] Specifically, for each muscle tCSA, the number of pixels of each paraspinal muscle image can be counted, and then the actual physical area tCSA of each paraspinal muscle muscle region can be obtained according to the physical area calculation formula (actual physical area = (pixel area / scaling ratio^2)*number of pixels).
[0119] For each muscle fCSA, fCSA pixel number = muscle pixel number - fat pixel number, and then according to the physical area calculation formula (actual physical area = (pixel area / scaling ratio^2)*number of pixels), the actual physical area fCSA of the functional muscles in each muscle area of the paraspinal muscles is obtained.
[0120] For the fat infiltration rate of each muscle, fat infiltration rate of each muscle = (tCSA of each muscle - fCSA of each muscle) / tCSA of each muscle.
[0121] According to the Kader classification, the fat infiltration rate can be calculated first, and then the fat infiltration rate can be used for grading: grade 3 is more than 50% fat infiltration, grade 2 is 10%-50% fat infiltration, grade 1 is less than 10% fat infiltration, and grade 0 is no visible fat infiltration.
[0122] For the tCSA ratio of the two sides of the multifidus muscle, the muscle proportion of the two sides of the multifidus muscle = the number of pixels of the left multifidus muscle / the number of pixels of the right multifidus muscle.
[0123] Regarding the muscle ratio on both sides of the erector spinae, the muscle ratio on both sides of the erector spinae = the number of pixels of the left erector spinae / the number of pixels of the right erector spinae.
[0124] Multifidus + erector spinae muscle ratio = (left multifidus muscle pixel number + left erector spinae muscle pixel number) / (right multifidus muscle pixel number + right erector spinae muscle pixel number)
[0125] S105, outputting the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image and the calculated parameters.
[0126] Specifically, in the embodiment of the present invention, a parameter visualization graph may be generated and output based on the tCSA of each muscle, the fCSA of each muscle, and the fat infiltration rate of each muscle.
[0127] In an optional embodiment, a training process of the paraspinal muscle segmentation model described in the above embodiment is provided, such as Figure 2 As shown, it shows a schematic diagram of the data processing flow of the paraspinal muscle segmentation model in an embodiment of the present invention; specifically, the multi-task learning segmentation model includes: an encoder, a coarse segmentation decoder, a fine segmentation decoder, a region detection module, and a fusion module. The paraspinal muscle segmentation model is trained by the following steps:
[0128] S1, collecting the original axial magnetic resonance images of the spine as sample axial magnetic resonance images of the spine, and obtaining the annotation information of the paravertebral muscles and intervertebral disc areas corresponding to each sample axial magnetic resonance image of the spine as sample annotation;
[0129] S2, inputting the sample spinal axial plane magnetic resonance original image into the encoder to obtain the sample image code, inputting the sample image code into the coarse segmentation decoder to obtain the sample coarse segmentation result, inputting the sample coarse segmentation result into the region detection module to obtain the sample region detection result, inputting the sample region detection result and the sample coarse segmentation result into the fusion module to obtain the sample fusion result, inputting the sample image code and the sample fusion result into the fine segmentation decoder to obtain the sample fine segmentation result;
[0130] S3, updating the parameters of each module based on the sample fine segmentation result and the sample annotation until the model converges to obtain a first multi-task learning segmentation model;
[0131] S4, freezing the model parameters of the encoder and the fusion module in the first multi-task learning segmentation model, and alternately updating the parameters of the coarse segmentation module, the region detection module, and the fine segmentation module based on the sample fine segmentation results and annotations until the model converges, thereby obtaining a second multi-task learning segmentation model;
[0132] S5, freezing the model parameters of the encoder, coarse segmentation module, region detection module and fusion module in the second multi-task learning segmentation model, and updating the parameters of the fine segmentation module based on the sample fine segmentation results and annotations until the model converges to obtain a paraspinal muscle segmentation model.
[0133] In an embodiment of the present invention, the encoder is a ResNet-34 encoder with the fully connected layer and the global average pooling layer removed at the end. Each decoder (including: a coarse segmentation decoder and a fine segmentation decoder) has a 5-layer network, each layer is connected to the upsampling channel of the previous layer through a jump connection and decoded by ResNet-D, and finally a three-layer feature map with the same input size is obtained. The region detection module is composed of a 5-layer network, namely, the maximum pooling max_pool, the convolution layer conv1, the convolution layer conv2, the fully connected layer fc1, and the fully connected layer fc2. The step size of each convolution layer is 3, the input and output variables of the fully connected fc1 are 675 and 128 respectively, and the input and output variables of the fully connected fc2 are 128 and 4 respectively. The region detection module outputs the detection box parameters, including: the upper left corner point (x, y) and the width and height (w, h) of the detection box. The fusion module maps each muscle recognition box included in the sample area detection result into a one-hot code image with the same size as the input image, obtains the one-hot code image of the recognition box of each category, superimposes the one-hot code image of the recognition box of each category and the sample coarse segmentation result into an attention map, and then jumps the sample coarse segmentation result to the attention map to obtain the sample fusion result as the output of the multi-task fusion module.
[0134] In an embodiment of the present invention, the training of the multi-task learning segmentation model is divided into three stages. In the first stage, all modules are trained. In the second stage, the coarse segmentation module + region detection module and the fine segmentation module are alternately trained, and the parameters of the coarse segmentation module + region detection module are updated alternately or the parameters of the fine segmentation module are updated. In the third stage, the fine segmentation module is fine-tuned.
[0135] In an embodiment of the present invention, before training, the sample data set can also be preprocessed, the sample spinal axial plane MRI original image can be adjusted to 512*512 pixels, and each sample spinal axial plane MRI original image can be horizontally flipped and rotated at a random angle within plus or minus 30 degrees.
[0136] In the embodiment of the present invention, the Adam optimizer is selected to iteratively optimize the parameters and perform stochastic gradient descent on the loss function. The loss function selects the multi-classification cross entropy loss function L CE and boundary-aware loss function L BCE The hybrid loss function consists of three stages: L1, L2, and L3:
[0137] L1=L3=λ1L CE +λ2L BCE
[0138] L2=L MSE .
[0139] In the embodiment of the present invention, a paraspinal muscle segmentation model is obtained based on multi-task learning U-Net, and the segmentation accuracy of the paraspinal muscle segmentation model can be improved by fusing multi-task learning and attention mechanism.
[0140] In a verification embodiment of the present invention, the model training environment is the pytorch framework under the windows operating system, pytorch version 1.13.1, graphics card model NVIDIA RTX 3070, processor model 11th Gen Intel (R) Core (TM) i7-11800H @ 2.30GHz, memory 32GB. The training set and validation set adopt 5-fold cross validation, and the 169 samples are evenly divided into 5 groups. Each model training uses one group as the validation set and the remaining 4 groups as the training set. The Dice coefficient of the 20 pictures in the test set is 96.3 for the left multifidus muscle, 96.5 for the right multifidus muscle, 96.7 for the left erector spinae muscle, 96.8 for the right erector spinae muscle, 95.3 for the left psoas major muscle, and 95.5 for the right psoas major muscle. The segmentation accuracy on this data set is better than that of the classic models of UNet, DeepLabV3, and MaskRCNN.
[0141] In an embodiment of the present invention, during the application of the paraspinal muscle segmentation model, the input image can be first adjusted to 512*512 pixels and then input into the paraspinal muscle segmentation model. After obtaining the refined segmentation result, the closed connected areas with small areas are removed as the final result of the refined segmentation of each paraspinal muscle area.
[0142] In an optional implementation, step S103 includes:
[0143] S1031, using the segmented paraspinal muscle images, calculate the segmented region rectangular frame of each paraspinal muscle region.
[0144] S1032, based on the rectangular frames of each segmented area, retain the image in each segmented area, set the grayscale value of pixels outside the segmented area to 0, use the CLAHE image enhancement algorithm to highlight the local details and image contrast of the image in the segmented area, and obtain enhanced images of each paraspinal muscle area.
[0145] In the embodiment of the present invention, the CLAHE image enhancement algorithm can perform image enhancement in a local area, which can effectively reduce the illumination non-uniformity effect that may exist in the image in the segmented area, making the overall image brightness more uniform. The CLAHE image enhancement algorithm can perform adaptive processing based on the local features of the image, and can better adapt to the characteristics of the image in different scenes, thereby improving the adaptability and flexibility of the image enhancement effect.
[0146] Specifically, in S1032, the images in the segmented area corresponding to each paraspinal muscle area are processed respectively. Specifically, for the images in one segmented area, step S1032 includes:
[0147] S10321, dividing the image into a plurality of overlapping or non-overlapping blocks.
[0148] In the embodiment of the present invention, a preset segmentation rule may be used to divide the image into a plurality of overlapping or non-overlapping tiles, for example, with a size of 8x8 or 16x16 pixels.
[0149] S10322, calculating a contrast function of each image block to adjust the pixel value of each image block.
[0150] Specifically, a histogram of the gray level of each image block may be calculated, and histogram equalization may be performed on two or more pixels in each image block to increase contrast and clarity.
[0151] In the embodiment of the present invention, the CLAHE image enhancement algorithm can enhance the contrast of the image by performing adaptive histogram equalization on the image blocks in the segmented area, so that the details in the image in the segmented area are clearer and more prominent.
[0152] In the embodiment of the present invention, the CLAHE image enhancement algorithm can effectively enhance the local contrast of the image in the segmented area, thereby improving the detail information of the image in the segmented area, making the fine structure in the image in the segmented area more prominent.
[0153] S10323: merge adjacent adjusted tiles to avoid generating boundaries.
[0154] S10324, cropping the histogram of the merged image block; when the value of the cropped histogram segment is higher than the preset contrast value, evenly distributing the pixels corresponding to the segment to other segments to limit the contrast and balance the image.
[0155] S10325, applying bilinear interpolation to the cropped and merged image to eliminate artifacts in the tile boundary.
[0156] S1033, the PSO-Kapur algorithm is used to segment the enhanced images of each paraspinal muscle region; Kapur entropy is used as the objective function to maximize the objective function, and the particle swarm optimization algorithm is used to optimize the running speed of Kapur multi-threshold segmentation to obtain a value close to the optimal solution.
[0157] In the embodiment of the present invention, the number of thresholds of Kapur multi-threshold segmentation is 2. The problem of how to separate different categories can be solved based on Kapur multi-threshold segmentation.
[0158] In the embodiment of the present invention, the optimal threshold of the image can be automatically determined to achieve adaptive segmentation of the image.
[0159] In an embodiment of the present invention, the PSO-Kapur algorithm (an image segmentation algorithm that combines particle swarm optimization (PSO) and Kapur entropy) uses the PSO algorithm to automatically search for the optimal threshold to maximize Kapur entropy (an indicator for measuring the differences between image classes), thereby determining the optimal segmentation threshold and achieving adaptive segmentation of the image.
[0160] Specifically, the grayscale range of multiple images is defined based on different thresholds, and then the grayscale frequency of each grayscale pixel in the entire image is calculated. Kapur entropy is used to measure the inter-class differences between image grayscale levels under different thresholds. Kapur entropy is a metric that can evaluate image segmentation results. The larger its value, the better the image segmentation effect. The particle swarm optimization algorithm (PSO algorithm) is used for search. The PSO algorithm updates the position and speed of particles in each iteration to find the optimal solution by simulating the process of birds searching for the best food location. In the PSO-Kapur algorithm, each particle represents a potential optimal threshold. The position of the particle is continuously updated by optimizing the Kapur entropy until the optimal solution, that is, the optimal segmentation threshold, is reached. The particles gradually approach the optimal solution in continuous iterations by updating their speed and position in the search space. In this process, the optimal threshold found is also updated. Finally, the PSO-Kapur algorithm will find the optimal segmentation threshold that can maximize the Kapur entropy, thereby achieving adaptive image segmentation.
[0161] Specifically, in the embodiment of the present invention, step S1033 includes:
[0162] S10331, use multiple thresholds to divide the image into different segments, the following is the Kapur entropy of each threshold interval:
[0163] ......
[0165] ......
[0167]
[0168] Among them, p i is the probability of event i occurring, H represents Kapur entropy, a, b, c…j…m represent thresholds, p a , p b ,…,p mThe grayscale probability of different threshold intervals is represented by a multi-threshold method;
[0169] In order to select the optimal threshold, the following objective function should be maximized:
[0170]
[0171] H i (th) represents the Kapur entropy in the i-th threshold interval when the threshold is th;
[0172] To extend the multi-level threshold expression, Kapur’s objective function is defined as:
[0173] f Kapur (TH) = f Kapur (th i ),i=1,2,3,…,k;
[0174] Different thresholds are represented by TH[th1, th2, th3, ..., thk-1], and i corresponds to a specific threshold interval.
[0175] S10332, generate initial position and initial velocity.
[0176] S10333, evaluate the objective function for each particle.
[0177] S10334, updating a single optimal solution and a global optimal solution, wherein the single optimal solution is a single optimal particle, and the global optimal solution is a global optimal particle.
[0178] S10335, update the velocity v and position x of each particle.
[0179] S10336, repeat the particle swarm optimization process until the termination condition is reached.
[0180] In the present embodiment, the termination condition may be: maximizing Kapur entropy.
[0181] In the particle swarm optimization algorithm, the particle velocity update formula is as follows:
[0182] v j (k+1)=wv j (k)+c1r1(p j (k)-x j (k))+c2r2(g(k)-x j (k)),j=1,2,…,np.
[0183] In the particle swarm optimization algorithm, the particle position update formula is as follows:
[0184] x j (k+1)=xj (k)+v j (k);
[0185] Among them, v j (k) and x j (k) are the velocity and position of the jth particle at the kth iteration, respectively, and p j g(k) and g(k) are the single best particle and the global best particle at the kth iteration, respectively; w is the initial weight parameter, c1 is the individual cognitive parameter, and c2 is the social cognitive parameter; r1 and r2 are random numbers between 0 and 1 with uniform distribution.
[0186] In the embodiment of the present invention, the PSO algorithm uses the wisdom of the group to search for the optimal solution by simulating the behavior of biological groups such as bird flocks or fish schools. It can find the global optimal solution or the approximate optimal solution in the search space, and has a strong global search capability for complex multi-dimensional optimization problems. The PSO algorithm uses an iterative method to perform optimization search, and gradually optimizes the objective function value by continuously updating the position and speed of the particles until the stop condition is reached. This iterative optimization process enables the PSO algorithm to quickly find a better solution in the search space. The PSO algorithm is adaptive and can adjust the movement strategy of the particles according to the results of the search process to explore the search space more effectively. This adaptability makes the PSO algorithm have good adaptability and robustness for different types of optimization problems. The implementation of the PSO algorithm is relatively simple, does not require complex mathematical models and a large number of parameter adjustments, and has good parallelization characteristics. It can be easily applied to a parallel computing environment to accelerate the optimization process and improve search efficiency.
[0187] In the embodiment of the present invention, the PSO optimization algorithm can be used to optimize the running speed of Kapur multi-threshold segmentation, and a value close to the optimal solution can be obtained with a number of iterations far less than that of the enumeration method.
[0188] S1034, obtaining muscle fat segmentation results based on the PSO-Kapur segmentation results, combining the muscle fat segmentation results with the coordinates of the respective segmentation area rectangular frames into corresponding areas of the same image, and then superimposing them on the original spinal axial plane magnetic resonance image, and the obtained image is used as a threshold segmentation visualization image.
[0189] In an exemplary embodiment of the present invention, the threshold segmentation visualization image obtained is as follows: Figure 3 As shown, Figure 3A schematic diagram of the operating results obtained by the method for segmentation and quantitative analysis of paraspinal muscle fat in spinal axial plane magnetic resonance imaging provided by an embodiment of the present invention is shown; wherein, the first picture in the first row from left to right is the original spinal axial plane magnetic resonance image of the target object, the second picture is the coarse segmentation visualization result, and the third picture is the fine segmentation visualization result; the first picture in the second row is the image enhancement visualization result, and the second picture is the threshold segmentation visualization result. Figure 3 The right side shows the parameter statistics results, and the third picture in the second row shows the parameter visualization results. The threshold segmentation visualization obtained by segmentation and quantitative analysis of paraspinal muscle fat in the spinal axial plane magnetic resonance imaging provided by the embodiment of the present invention is compared with 20 pictures of manual segmentation results. The JS coefficient is 0.88 for the left multifidus muscle, 0.82 for the right multifidus muscle, 0.97 for the left erector spinae muscle, 0.91 for the right erector spinae muscle, 0.56 for the left psoas muscle, and 0.32 for the right psoas muscle. Except for the psoas major muscles on both sides, the SA parameter is 0.98, which is better than unsupervised methods such as K-Means and Otsu.
[0190] In this exemplary embodiment, the parameters calculated based on the paraspinal muscle fat segmentation region corresponding to each paraspinal muscle image are as follows:
[0191] The tCSA of the right erector spinae muscle was 1614.53 mm^2, the tCSA of the left erector spinae muscle was 1873.93 mm^2, the tCSA of the right multifidus muscle was 353.70 mm^2, the tCSA of the left multifidus muscle was 464.32 mm^2, the tCSA of the right psoas muscle was 277.10 mm^2, and the tCSA of the left psoas muscle was 176.24 mm^2. The fat infiltration rate of the right erector spinae muscle was 0.33, the fat infiltration rate of the left erector spinae muscle was 0.29, the fat infiltration rate of the right multifidus muscle was 0.48, the fat infiltration rate of the left multifidus muscle was 0.53, the fat infiltration rate of the right psoas muscle was 0.31, and the fat infiltration rate of the left psoas muscle was 0.38. The Kader grading is: grade 3 is more than 50% fat infiltration, grade 2 is 10%-50%, grade 1 is less than 10%, and grade 0 is no visible fat infiltration. The calculation results show that except for the left psoas major muscle, which has grade 3 fat infiltration, the other muscles have grade 2 fat infiltration. The proportion of muscles on both sides of the multifidus muscle is 0.86, the proportion of muscles on both sides of the erector spinae muscle is 0.76, and the proportion of muscles on both sides of the multifidus + erector spinae muscle is 1.57.
[0192] In an embodiment of the present invention, the above-mentioned parameters can be displayed in a parameter statistics area. In an embodiment of the present invention, a parameter visualization graph can also be generated based on the tCSA, fCSA, and fat infiltration rate of each muscle, and the parameter visualization result can be displayed in a corresponding position in the parameter statistics area.
[0193] In an embodiment of the present invention, in step S105, the paraspinal muscle fat segmentation area and the calculated parameters corresponding to each paraspinal muscle muscle image can be displayed on the interface. The specific interface design is mainly divided into: address bar, operation button area, image display area, parameter statistics area. The interface can realize full-process interaction. Click to select path, original image, coarse segmentation, fine segmentation, image enhancement, threshold segmentation, calculate parameters, and save image in sequence. It can realize reading of original data, model inference and output of visualization results, image enhancement visualization results, threshold segmentation visualization results, parameter numerical display and parameter visualization results, and image saving to a local folder.
[0194] The present invention provides a high-precision, easy-to-operate, fully automated paraspinal muscle fat segmentation and quantitative analysis method, which is based on spinal axial plane magnetic resonance imaging, uses a pre-trained paraspinal muscle segmentation model for segmentation, uses image enhancement for the segmented area, uses the Kapur threshold segmentation algorithm combined with the PSO optimization algorithm to further segment the segmented area into fat areas, and finally counts and analyzes a series of parameters based on area and signal intensity such as fat infiltration rate in the segmented area. The fully automated paraspinal muscle fat segmentation and quantitative analysis method proposed by the present invention can effectively reduce the subjective error and experience error caused by manual labeling, and greatly improve the diagnostic efficiency of clinicians, and has a wide range of application value.
[0195] Based on the same inventive concept, the embodiment of the present invention also provides a spinal axial plane magnetic resonance imaging paraspinal muscle fat segmentation and quantitative analysis device, such as Figure 4 As shown, it shows a structural block diagram of a spinal axial plane magnetic resonance imaging paraspinal muscle fat segmentation and quantitative analysis device according to an embodiment of the present invention, the device comprising:
[0196] The target image acquisition module 401 is used to acquire the target object's spinal axial plane magnetic resonance original image;
[0197] The muscle image segmentation module 402 is used to segment and obtain each paraspinal muscle image based on a pre-trained paraspinal muscle segmentation model; the paraspinal muscle segmentation model is obtained by training a multi-task learning segmentation model based on a pre-acquired sample spinal axial plane magnetic resonance original image with paraspinal muscles and intervertebral disc areas annotated;
[0198] The fat segmentation module 403 is used to determine the rectangular frames corresponding to the muscle images of the paraspinal muscles, and to enhance the images in the rectangular frames by using the CLAHE image enhancement algorithm; and to process the enhanced muscle images of the paraspinal muscles by using the PSO-Kapur threshold segmentation algorithm to obtain the fat segmentation regions of the paraspinal muscles;
[0199] The parameter analysis module 404 is used to calculate at least one of the following parameters based on the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image: tCSA of each muscle, fCSA of each muscle, Kader grade, tCSA ratio of both sides of multifidus muscle, tCSA ratio of both sides of erector spinae muscle, and fat infiltration rate of each muscle;
[0200] The output module 405 is used to output the paraspinal muscle fat segmentation area and the calculated parameters corresponding to each paraspinal muscle image.
[0201] Based on the same inventive concept, an embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the method for segmenting and quantitatively analyzing paraspinal muscle fat in spinal axial plane magnetic resonance imaging described in any of the above embodiments are implemented.
[0202] Based on the same inventive concept, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps in the method for segmenting and quantitatively analyzing paraspinal muscle fat in spinal axial plane magnetic resonance imaging described in any of the above embodiments are implemented.
[0203] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0204] It will be appreciated by those skilled in the art that the embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the embodiments of the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments of the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0205] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (apparatus), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.
[0206] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0207] These computer program instructions can also be loaded onto a computer or other programmable terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0208] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.
[0209] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.
[0210] The above is a detailed introduction to the paraspinal muscle fat segmentation and quantitative analysis method of spinal axial magnetic resonance imaging provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the idea of the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.
Claims
1. A method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging, characterized in that: The method comprises: Acquire an original magnetic resonance image of the spine axial plane of the target object; Based on a pre-trained paraspinal muscle segmentation model, each paraspinal muscle image is segmented; the paraspinal muscle segmentation model is obtained by training a multi-task learning segmentation model based on a pre-acquired sample spinal axial plane magnetic resonance original image with paraspinal muscles and intervertebral disc areas annotated; Determine the rectangular frames corresponding to the muscle images of each paraspinal muscle, and use the CLAHE image enhancement algorithm to enhance the image in the rectangular frame; process the enhanced muscle images of each paraspinal muscle by the PSO-Kapur threshold segmentation algorithm to obtain the fat segmentation area of each paraspinal muscle; Calculate at least one of the following parameters based on the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image: tCSA of each muscle, fCSA of each muscle, Kader grade, tCSA ratio of both sides of multifidus muscle, tCSA ratio of both sides of erector spinae muscle, and fat infiltration rate of each muscle; The paraspinal muscle fat segmentation area and the calculated parameters corresponding to each paraspinal muscle image are output.
2. The method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging according to claim 1, characterized in that: The multi-task learning segmentation model includes: an encoder, a coarse segmentation decoder, a fine segmentation decoder, a region detection module, and a fusion module. The paraspinal muscle segmentation model is trained by the following steps: Collecting the original axial magnetic resonance images of the spine as sample axial magnetic resonance images of the spine, and obtaining the annotation information of the paravertebral muscles and intervertebral disc areas corresponding to each sample axial magnetic resonance image of the spine as sample annotations; Input the sample spinal axial plane magnetic resonance original image into the encoder to obtain the sample image code, input the sample image code into the coarse segmentation decoder to obtain the sample coarse segmentation result, input the sample coarse segmentation result into the region detection module to obtain the sample region detection result, input the sample region detection result and the sample coarse segmentation result into the fusion module to obtain the sample fusion result, input the sample image code and the sample fusion result into the fine segmentation decoder to obtain the sample fine segmentation result; Based on the sample fine segmentation result and the sample annotation, the parameters of each module are updated until the model converges to obtain a first multi-task learning segmentation model; Freeze the model parameters of the encoder and the fusion module in the first multi-task learning segmentation model, and alternately update the parameters of the coarse segmentation module, the region detection module, and the fine segmentation module based on the sample fine segmentation results and annotations until the model converges to obtain a second multi-task learning segmentation model; Freeze the model parameters of the encoder, coarse segmentation module, region detection module and fusion module in the second multi-task learning segmentation model, and update the parameters of the fine segmentation module based on the sample fine segmentation results and annotations until the model converges to obtain the paraspinal muscle segmentation model.
3. The method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging according to claim 2, characterized in that: The sample area detection results and the sample rough segmentation results are input into the fusion module to obtain the sample fusion results, including: The sample area detection results and the sample coarse segmentation results are input into the fusion module. The fusion module maps each muscle recognition frame included in the sample area detection result into a one-hot code image with the same size as the input image to obtain the one-hot code image of the recognition frame of each category. The one-hot code image of the recognition frame of each category and the sample coarse segmentation result are superimposed into an attention map. The sample coarse segmentation result is then jump-connected to the attention map to obtain the sample fusion result as the output of the multi-task fusion module.
4. The method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging according to claim 1, characterized in that: The image in the rectangular frame is enhanced using the CLAHE image enhancement algorithm, and the enhanced paraspinal muscle images are processed using the PSO-Kapur threshold segmentation algorithm to obtain the paraspinal muscle fat segmentation areas, including: Using the segmented paraspinal muscle images, the segmented rectangular frame of each paraspinal muscle region is calculated; Based on the rectangular frames of each segmented area, the images in each segmented area are retained, the grayscale values of pixels outside the segmented area are set to 0, and the CLAHE image enhancement algorithm is used to highlight the local details and image contrast of the image in the segmented area, and the enhanced images of each paraspinal muscle area are obtained respectively; The PSO-Kapur algorithm was used to segment the enhanced images of each paraspinal muscle region. The Kapur entropy was used as the objective function to maximize the objective function, and the particle swarm optimization algorithm was used to optimize the running speed of Kapur multi-threshold segmentation to obtain a value close to the optimal solution. Obtaining muscle fat segmentation results based on the PSO-Kapur segmentation results, combining the muscle fat segmentation results with the coordinates of the respective segmentation area rectangular frames into corresponding areas of the same image, and then superimposing them onto the original spinal axial magnetic resonance image, and obtaining an image as a threshold segmentation visualization image; Among them, the threshold number of Kapur multi-threshold segmentation is 2.
5. The method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging according to claim 4, characterized in that: Taking Kapur entropy as the objective function to maximize the objective function includes: Use multiple thresholds to divide the image into different segments. The following is the Kapur entropy for each threshold interval: Among them, p i is the probability of event i occurring, H represents Kapur entropy, a, b, c…j…m represent thresholds, p a , p b ,…,p m The grayscale probability of different threshold intervals is represented by a multi-threshold method; In order to select the optimal threshold, the following objective function should be maximized: H i (th) represents the Kapur entropy in the i-th threshold interval when the threshold is th; To extend the multi-level threshold expression, Kapur’s objective function is defined as: f Kapur (TH)=f Kapur (th i ),i=1,2,3,…,k; Different thresholds are represented by TH[th1, th2, th3, ..., thk-1], and i corresponds to a specific threshold interval.
6. The method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging according to claim 5, characterized in that: The particle swarm optimization algorithm is used to optimize the running speed of Kapur multi-threshold segmentation and obtain values close to the optimal solution, including: Generate initial position and initial velocity; Evaluate the objective function for each particle; Update the single optimal solution and the global optimal solution, the single optimal solution is the single optimal particle, and the global optimal solution is the global optimal particle Update the velocity v and position x of each particle; Repeat the particle swarm optimization process until the termination condition is reached; In the particle swarm optimization algorithm, the particle velocity update formula is as follows: v j (k+1)=wv j (k)+c1r1(p j (k)-x j (k))+c2r2(g(k)-x j (k)),j=1,2,…,np. In the particle swarm optimization algorithm, the particle position update formula is as follows: x j (k+1)=x j (k)+v j (k); Among them, v j (k) and x j (k) are the velocity and position of the jth particle at the kth iteration, respectively, and p j g(k) and g(k) are the single best particle and the global best particle at the kth iteration, respectively; w is the initial weight parameter, c1 is the individual cognitive parameter, and c2 is the social cognitive parameter; r1 and r2 are random numbers between 0 and 1 with uniform distribution.
7. A device for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging, characterized in that: The device comprises: A target image acquisition module, used for acquiring an original magnetic resonance image of the spine axial plane of a target object; A muscle image segmentation module is used to segment and obtain each paraspinal muscle image based on a pre-trained paraspinal muscle segmentation model; the paraspinal muscle segmentation model is obtained by training a multi-task learning segmentation model based on a pre-acquired sample spinal axial plane magnetic resonance original image labeled with paraspinal muscles and intervertebral disc areas; The fat segmentation module is used to determine the rectangular boxes corresponding to the muscle images of each paraspinal muscle, and to enhance the image in the rectangular box using the CLAHE image enhancement algorithm; the enhanced muscle images of each paraspinal muscle are processed using the PSO-Kapur threshold segmentation algorithm to obtain the fat segmentation area of each paraspinal muscle; A parameter analysis module, for calculating at least one of the following parameters based on the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image: tCSA of each muscle, fCSA of each muscle, Kader grade, tCSA ratio of both sides of multifidus muscle, tCSA ratio of both sides of erector spinae muscle, and fat infiltration rate of each muscle; The output module is used to output the paraspinal muscle fat segmentation area corresponding to each paraspinal muscle image and the calculated parameters.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for segmenting and quantifying paraspinal muscle fat in spinal axial magnetic resonance imaging according to any one of claims 1 to 7 are implemented.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for segmenting and quantifying the paraspinal muscle fat in spinal axial magnetic resonance imaging according to any one of claims 1 to 7 are implemented.
10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the method for segmenting and quantifying the paraspinal muscle fat in spinal axial magnetic resonance imaging described in any one of claims 1 to 7 are implemented.
Citation Information
Patent Citations
Body composition automatic measurement system based on abdomen CT image and deep learning
CN114305473A
Method, system and equipment for calculating paravertebral muscle fat infiltration rate
CN118298003A
