Data processing method based on automatic segmentation of fat and muscle areas in MRI (Magnetic Resonance Imaging)
Through CT image model migration and FCM algorithm processing, the problem of difficult to ensure the efficiency and accuracy of fat and muscle region segmentation in MRI imaging is solved, and efficient and accurate fat and muscle region segmentation is achieved, reducing the workload of manual labeling and improving the image quality.
Patent Information
- Application Number
- CN202410771463.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-14
- Publication Date
- 2024-11-19
AI Technical Summary
In MRI imaging, it is difficult for the prior art to achieve efficient and high-precision fat and muscle region segmentation during abdominal imaging of inflammatory bowel disease, especially in the DIXON sequence of image quality and partial results due to intestinal movement and tissue relaxation due to intestinal motility and tissue relaxation in DIXON sequences Low signal area problems caused by volume effect.
By acquiring the automatic segmentation model in the CT image, using the single-part segmentation scanning method and FCM algorithm, data processing of the MRI image is carried out, including repeated layer removal, grayscale normalization and translation transformation, and an automatic segmentation model of the MRI image is constructed. Achieve automatic segmentation of fat and muscle areas.
The segmentation efficiency and segmentation accuracy of fat and muscle areas in MRI imaging are improved, the workload of manual labeling is reduced, and unsupervised classification is performed without the use of display thresholds, improving image quality and quantitative results.
Smart Images

Figure CN118982547A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a data processing method based on automatic segmentation of fat and muscle areas in MRI imaging. Background Art
[0002] The high resolution of MRI clearly shows the distribution of intramuscular fat. For sequences that highlight fat structures, such as chemical shift coded imaging, the muscles in the fat phase image appear as low intensity values, and the intermuscular fat appears as high intensity values, showing high contrast. Currently, Lancaster et al. proposed to separate three areas based on the inner and outer contours of subcutaneous fat and the outer contours of visceral fat, and use Gaussian curves to fit the peak of the histogram distribution in the visceral fat contour area to further divide the fat and non-fat areas. This method is an accurate method for semi-automatic fat segmentation in MRI.
[0003] Although existing deep learning methods can achieve automatic segmentation of fat and muscle areas in MRI imaging, the intake of intestinal contrast agents and the disorder of intestinal function of patients during DIXON sequence for abdominal MRI imaging of inflammatory bowel disease make it impossible to achieve long-term single sequence scanning, and thus a single whole-body DIXON scan cannot be achieved. During data acquisition, the pulse sequence needs to be changed regularly to cause the required water / fat relative phase change in the acquired image. In the presence of intestinal movement and limited tissue relaxation time, the increase in scanning time and the change in pulse sequence timing will lead to a significant decrease in the quality of the acquired image. Accordingly, the processed images containing only water and only fat will also suffer from image quality degradation such as increased motion artifacts or image blur. In addition, the existing mask segmentation does not consider the low-signal areas of fat caused by partial volume effects, resulting in large differences in the final quantification of visceral fat, making it difficult to ensure the segmentation efficiency and accuracy of fat and muscle areas in MRI imaging. Summary of the invention
[0004] The purpose of the present invention is to provide a data processing method based on automatic segmentation of fat and muscle areas in MRI imaging, so as to solve the technical problem in the prior art that the segmentation efficiency and segmentation accuracy of fat and muscle areas in MRI imaging are difficult to ensure.
[0005] In order to solve the above technical problems, the present invention specifically provides the following technical solutions:
[0006] The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging is characterized by comprising the following steps:
[0007] Acquire a CT image automatic segmentation model for automatically segmenting fat and muscle regions in CT imaging;
[0008] Acquire multiple MRI images containing fat and muscle areas by using a single-site segmented scanning method;
[0009] The FCM algorithm is used to obtain the true annotation of fat and muscle areas in each MRI image;
[0010] The dataset is composed of MRI images and real annotations of fat and muscle areas;
[0011] Based on the data set, transfer learning is performed on the CT image automatic segmentation model to obtain an MRI image automatic segmentation model to achieve automatic segmentation of fat and muscle areas in MRI images.
[0012] As a preferred embodiment of the present invention, a method for acquiring multiple MRI images containing fat and muscle regions by a single-site segmented scanning method comprises:
[0013] Step S1: Obtain N segmented images by using N-segment scanning, where the value of N includes 2 and 3;
[0014] Step S2: searching for repeated layers in each of the N segmented images;
[0015] Step S3: normalizing the grayscale range of the repeated layers in the N segmented images to 0-255;
[0016] Step S4: Calculating the optimal translation transformation coordinates based on the repeated layers;
[0017] Step S5: translating non-repetitive layers in the segmented image based on the optimal translation transformation coordinates, and reconnecting the translated segmented images to obtain an MRI image.
[0018] As a preferred embodiment of the present invention, in step S1,
[0019] When N=2, segmented images A and B are obtained respectively, and each of the segmented images A and B contains k frames of cross-sectional two-dimensional images, wherein A=(x 1 ,x 2 ,..,x k ), B=(y 1 ,y 2 ,..,y k ), where x 1 ,x 2 ,x k are the 1st, 2nd, and kth cross-sectional 2D images in A, respectively. 1 ,y 2 ,y k are the 1st, 2nd, and kth cross-sectional 2D images in B, respectively. 1 ,x 2 ,..,x kand 1 ,y 2 ,..,y k The ImagePosition[2] information contained in the represented two-dimensional image DICOM indicates the z-axis position of the slice;
[0020] When N=3, segmented images A, B and C are obtained respectively, and each of the segmented images A, B and C contains k frames of cross-sectional two-dimensional images, wherein A=(x 1 ,x 2 ,..,x k ), B=(y 1 ,y 2 ,..,y k ), C=(z 1 ,z 2 ,..,z k ), where x 1 ,x 2 ,x k are the 1st, 2nd, and kth cross-sectional 2D images in A, respectively. 1 ,y 2 ,y k are the 1st, 2nd, and kth cross-sectional 2D images in B, respectively. 1 ,z 2 ,z k are the 1st, 2nd, and kth cross-sectional 2D images in C, respectively. (x 1 ,x 2 ,..,x k )、(y 1 ,y 2 ,..,y k ) and (z 1 ,z 2 ,..,z k ) contains the ImagePosition[2] information that indicates the z-axis position of the slice.
[0021] As a preferred embodiment of the present invention, in step S2,
[0022] When N = 2, the ImagePosition[2] information contained in the two-dimensional image DICOM is used to find the two-dimensional layers in A and B that are in the same z-axis position value range, and obtain the repeated layers in A and B, among which the non-repeated layers in A are A. 1 =(x 1 ,x 2 ,..,x k-m ), the repeat level is A 2 =(x k-m+1 ,x k-m+2 ,..,xk );
[0023] The repeating level in B is B 1 =(y 1 ,y 2 ,..,y m ), the non-repeated level is B 2 =(y m+1 ,y m+2 ,..,y k );
[0024] In the formula, x 1 ,x 2 ,x k-m are the 2D images of the cross section of the 1st, 2nd, and km frames in A, respectively. k-m+1 ,x k-m+2 ,x k are the k-m+1, k-m+2, and k-th cross-sectional 2D images in A, respectively. 1 ,y 2 ,y m are the 1st, 2nd, and mth cross-sectional 2D images in B, respectively. m+1 ,y m+2 ,y k They are the m+1, m+2, and k-th cross-sectional two-dimensional images in B, respectively;
[0025] When N=3, in the ImagePosition[2] information contained in the two-dimensional image DICOM, find the two-dimensional layers in A and B that are in the same z-axis position value range, and obtain the repeated layers in A and B. Find the two-dimensional layers in B and C that are in the same z-axis position value range, and obtain the repeated layers in B and C.
[0026] The non-repeating level in A is A 1 =(x 1 ,x 2 ,..,x k-m ), the repeat level is A 2 =(x k-m+1 ,x k-m+2 ,..,x k );
[0027] The header repetition level in B is B 1 =(y 1 ,y 2 ,..,y m ), the non-repeated level is B 2 =(y m+1 ,y m+2 ,..,y k-n ), the last segment repeats level B 3 =(y k-n+1 ,yk-n+2 ,..,y k );
[0028] The repetition level in C is C 1 =(z 1 ,z 2 ,..,z n ), the non-repeating level is C 2 =(z n+1 ,z n+2 ,..,z k ),
[0029] In the formula, x 1 ,x 2 ,x k-m are the 2D images of the cross section of the 1st, 2nd, and km frames in A, respectively. k-m+1 ,x k-m+2 ,x k are the k-m+1, k-m+2, and k-th cross-sectional 2D images in A, respectively. 1 ,y 2 ,y m are the 1st, 2nd, and mth cross-sectional 2D images in B, respectively. m+1 ,y m+2 ,y k-n are the m+1, m+2, and kn cross-sectional 2D images in B, respectively. k-n+1 ,y k-n+2 ,y k are the k-n+1, k-n+2, and k-th cross-sectional 2D images in B, respectively. 1 ,z 2 ,z n are the 1st, 2nd, and nth cross-sectional 2D images in C, respectively. n+1 ,z n+2 ,z k They are respectively the n+1th, n+2nd, and kth cross-sectional two-dimensional images in C, where kn>m+1.
[0030] As a preferred solution of the present invention, in step S3,
[0031] When N=2, the repeated level in A is A 2 =(x k-m+1 ,x k-m+2 ,..,x k ) is normalized to the grayscale range of 0-255, A 2 The grayscale normalization formula is:
[0032]
[0033] Set the repeated layer in B to B 1=(y 1 ,y 2 ,..,y m ) is normalized to the grayscale range of 0-255, B 1 The grayscale normalization formula is:
[0034]
[0035] In the formula, A 2norm A 2 The gray level normalization result, A 2 is the repeated level in A, B 1norm For B 1 The gray level normalization result, B 1 is the repeated level in B;
[0036] When N=3, the repetition level in A is A 2 =(x k-m+1 ,x k-m+2 ,..,x k ) is normalized to the grayscale range of 0-255, A 2 The grayscale normalization formula is:
[0037]
[0038] Set the repeated layer in B to B 1 =(y 1 ,y 2 ,..,y m ) and B 3 =(y k-n+1 ,y k-n+2 ,..,y k ) are normalized to the range of 0-255, B 1 The grayscale normalization formula is:
[0039]
[0040] B 3 The grayscale normalization formula is:
[0041]
[0042] Change the repetition level in C to C 1 =(z 1 ,z 2 ,..,z n ) is normalized to the grayscale range of 0-255, C 1 The grayscale normalization formula is:
[0043]
[0044] In the formula, A 2norm A 2 The gray level normalization result, A 2 is the repeated level in A, B 1norm For B 1 The gray level normalization result, B 1 Repeat the level for the header in B, B 3norm For B 3 The gray level normalization result, B 3 The end segment of B is repeated, C 1norm C 1 Grayscale normalization result, C 1 is the repetition level in C.
[0045] As a preferred solution of the present invention, in step S4, when N=2, a translation range of [-256, 256] is given;
[0046] B 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated B is iteratively calculated. 1norm and A 1norm The mean square error between
[0047] The coordinates when the mean square error reaches the minimum value are taken as the optimal translation transformation coordinates;
[0048] When N=3, the translation range is given as [-256, 256];
[0049] B 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated B is iteratively calculated. 1norm and A 2norm The mean square error between
[0050] The coordinate when the mean square error reaches the minimum value is taken as B 2 Segment optimal translation transformation coordinates;
[0051] C 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated C is iteratively calculated. 1norm and B 3norm The mean square error between
[0052] The coordinate when the mean square error reaches the minimum value is taken as C 2 The optimal translation transformation coordinates of the segment.
[0053] As a preferred solution of the present invention, in step S5,
[0054] When N = 2, based on the optimal translation transformation coordinates, 2Each non-repeating layer in the A is translated and connected in sequence 1 , A 2 and B after translation 2 , obtaining the MRI image;
[0055] When N = 3, based on B 2 The optimal translation transformation coordinates of segment B 2 Each non-repeating level in the translation is based on B 2 Duan and C 2 The optimal translation transformation coordinates of the segment are 2 Shift the non-repeating levels within;
[0056] Connect A in sequence 1 , A 2 and B after translation 2 , C 2 , and obtain the MRI image.
[0057] As a preferred embodiment of the present invention, a method for obtaining true annotations of fat and muscle areas in each MRI image by using an FCM algorithm includes:
[0058] The FCM algorithm is used to perform rough segmentation of fat and muscle areas based on the water phase map and the fat phase map to obtain a rough segmentation mask;
[0059] Obtaining true annotations of fat and muscle areas in the MRI image based on manual annotation correction in the coarse segmentation mask;
[0060] in,
[0061] The manual annotation correction method comprises:
[0062] The third and fourth category fuzzy classification masks or label maps obtained by operation in the water phase map are used for further manual correction to obtain the true labels of subcutaneous fat and visceral fat, respectively;
[0063] The third and fourth category fuzzy classification masks or label maps obtained by operation in the fat phase map are used for further manual correction to obtain the true labeling of the muscle;
[0064] Among them, the manual annotation correction process was carried out using the open source software ITK-SNAP combined with MRI images without any preprocessing.
[0065] As a preferred solution of the present invention, the rough segmentation method of the FCM algorithm includes:
[0066] The number of pixel categories of the fat phase map and the water phase map is predefined as 4, among which the pixel categories of the fat phase map include cortical bone / background, fat-rich area, bone marrow fat and partial volume effect voxel area; the pixel categories of the water phase map include background, cortical bone / fat, soft tissue and blood vessels and digestive tract fluid;
[0067] The FCM algorithm is used to evaluate the membership value u of category k at pixel j in the fat phase map and the water phase map by minimizing the loss function. j,k , assign each pixel category to each pixel in the fat phase map and the water phase map;
[0068] The loss function is:
[0069]
[0070]
[0071]
[0072] In the formula, y j is pixel j,u j,k is the membership value of category k at pixel j, N is the total number of pixels, c is the number of pixel categories, v k is the data center of the kth category, U is the membership function matrix, and V vector represents the cluster center v k , q is the weight index used to determine the number of fuzzy classifications;
[0073] The membership degree u of the water phase diagram and the fat phase diagram j,k The smallest values correspond to the background, and the membership u jk The values of the largest and second largest two types of voxels are set to 1, and the four types of fat rough segmentation mask images are obtained. The membership degree u in the water phase map jk The largest and next two types of voxel values are set to 1, and four types of muscle rough segmentation mask images including soft tissues such as muscles and abdominal organs of the liver are obtained, and the rough segmentation mask images except the background are removed by OR operation through the water phase map and the fat phase map to obtain the rough mask of the body contour.
[0074] As a preferred solution of the present invention, the true annotations of the fat and muscle areas in the MRI image include masks of subcutaneous fat SAT, visceral fat VAT and muscle SM.
[0075] Compared with the prior art, the present invention has the following beneficial effects:
[0076] In the process of MRI intestinal imaging, the present invention performs segmented scanning of the abdominal area in cross-sectional imaging to avoid layer omission, that is, to avoid information loss, and removes repeated layers between segmented scans. A displacement transformation method is used to enable complete matching of layer connections. The segmented 3D abdominal sequence images are aligned and merged into a complete abdominal 3D image before segmentation. On the one hand, more accurate quantitative results can be provided after subsequent segmentation, and on the other hand, the manual workload required for labeling can be reduced. The FCM algorithm is used to perform initial coarse segmentation of backgrounds such as muscle, fat, and non-muscle fat. Considering the low-signal area in fat caused by the partial volume effect, the image pixels in the cluster are unsupervisedly classified without using a display threshold, thereby improving the segmentation efficiency and accuracy of fat and muscle areas in MRI imaging. BRIEF DESCRIPTION OF THE DRAWINGS
[0077] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.
[0078] Figure 1 A flow chart of a data processing method based on automatic segmentation of fat and muscle areas in MRI imaging provided by an embodiment of the present invention;
[0079] Figure 2 A flowchart of constructing an automatic segmentation model for MRI images provided by an embodiment of the present invention;
[0080] Figure 3 A flow chart of an MRI imaging method based on segmented scanning fusion provided by an embodiment of the present invention;
[0081] Figure 4 The segmented scanning of the MRI image provided in the embodiment of the present invention does not remove the repeated layer fusion map;
[0082] Figure 5 A fusion diagram for removing repeated layers from a segmented scan of an MRI image provided by an embodiment of the present invention;
[0083] Figure 6 A rough segmentation mask flow chart of the FCM algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0084] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0085] like Figure 1 As shown, the present invention provides a data processing method based on automatic segmentation of fat and muscle areas in MRI imaging, characterized in that it includes the following steps:
[0086] Acquire a CT image automatic segmentation model for automatically segmenting fat and muscle regions in CT imaging;
[0087] Acquire multiple MRI images containing fat and muscle areas by using a single-site segmented scanning method;
[0088] The FCM algorithm is used to obtain the true annotation of fat and muscle areas in each MRI image;
[0089] The dataset is composed of MRI images and real annotations of fat and muscle areas;
[0090] Based on the data set, transfer learning is performed on the CT image automatic segmentation model to obtain the MRI image automatic segmentation model, so as to realize automatic segmentation of fat and muscle areas in MRI images.
[0091] The automatic CT image segmentation model in the present invention establishes the segmentation of fat and muscle areas on CT imaging, and trains a deep learning model (TransUNet model) by constructing a data set to achieve automatic segmentation of fat and muscle areas in CT imaging.
[0092] The construction of the CT image segmentation model includes:
[0093] The dataset consisting of the preprocessed CT images and the real annotations of the fat and muscle areas in the CT images is divided into a training set and a test set;
[0094] Based on the training set, the TransUNet model is trained to obtain a CT image segmentation model that inputs the preprocessed CT image and outputs the annotation results of the fat and muscle areas in the CT image;
[0095] Based on the test set, the CT image segmentation model is evaluated for its performance in segmenting fat and muscle regions in CT images.
[0096] like Figure 2As shown, the present invention utilizes transfer learning, so that the source domain is CT images, the source task is to obtain regional fat and muscle annotation data based on CT image segmentation, and the target domain is migrated to MRI images, and the target task is to obtain muscle and different regional fat annotation data based on MRI images.
[0097] The present invention utilizes the characteristic of transfer learning that can save time for manual sample annotation. By migrating the abdominal region fat and muscle segmentation model of CT images to the less-annotated data of MRI images, a high-accuracy segmentation model based on MRI images (i.e., an MRI image automatic segmentation model) is trained, thereby realizing 3D automatic segmentation of abdominal region fat and muscle in MRI images.
[0098] The method for constructing the MRI image automatic segmentation model includes:
[0099] The in-phase image, inverse phase image, water phase image and fat phase image of the MRI image (the in-phase image, inverse phase image, water phase image and fat phase image with co-registration characteristics are collected by DIXON technology) and the data set consisting of the real annotations of the fat and muscle areas in the MRI image are divided into a training set and a test set;
[0100] Based on the training set, transfer learning is performed on the CT image segmentation model to obtain an MRI image automatic segmentation model that inputs the in-phase image, inverted phase image, water phase image, and fat phase image and outputs the annotation results of the fat and muscle areas in the MRI image;
[0101] Based on the test set, the MRI image automatic segmentation model is evaluated for its performance in segmenting fat and muscle regions in MRI images.
[0102] Based on the CT image automatic segmentation model as a pre-trained model, transfer learning is used to obtain the MRI image automatic segmentation model, which can realize the full automatic segmentation of fat and muscle areas in MRI imaging by utilizing the in-phase image, anti-phase image, water phase image and fat phase image of the MRI image.
[0103] In order to avoid information loss and improve the segmentation accuracy and efficiency of the MRI image automatic segmentation model, the present invention divides the cross-sectional imaging area into 2 or 3 segmented areas for scanning during the MRI imaging process to reduce the total scanning time of a single sequence. The segmented scanning is to avoid missing layers.
[0104] The displacement transformation method is used to make the layer connection fully matched, and the segmented 3D sequence images are aligned and merged into a complete 3D image before segmentation. On the one hand, it can provide more accurate quantitative results after subsequent segmentation, and on the other hand, it can reduce the manual workload required for annotation.
[0105] like Figure 3As shown, specifically, the method of acquiring multiple MRI images containing fat and muscle areas by a single-site segmented scanning method includes:
[0106] Step S1: Obtain N segmented images using N-segment scanning, where the value of N includes 2 and 3;
[0107] Step S2: searching for repeated layers in each of the N segmented images;
[0108] Step S3: normalizing the grayscale range of the repeated layers in the N segmented images to 0-255;
[0109] Step S4: Calculate the optimal translation transformation coordinates based on the repeated layers;
[0110] Step S5: translating non-repetitive layers in the segmented image based on the optimal translation transformation coordinates, and reconnecting the translated segmented images to obtain an MRI image.
[0111] In step S1,
[0112] When N=2, segmented images A (such as the upper abdomen segment) and B (such as the lower abdomen and pelvic segment) are obtained respectively. The segmented images A and B both contain k frames of cross-sectional two-dimensional images, where A=(x 1 ,x 2 ,..,x k ), B=(y 1 ,y 2 ,..,y k ), where x 1 ,x 2 ,x k are the 1st, 2nd, and kth cross-sectional 2D images in A, respectively. 1 ,y 2 ,y k are the 1st, 2nd, and kth cross-sectional 2D images in B, respectively. 1 ,x 2 ,..,x k and 1 ,y 2 ,..,y k The ImagePosition[2] information contained in the represented two-dimensional image DICOM indicates the z-axis position of the slice;
[0113] When N=3, segmented images A (such as the upper abdomen), B (such as the middle abdomen) and C (such as the lower abdomen and pelvic cavity) are obtained respectively. The segmented images A, B and C all contain k frames of cross-sectional two-dimensional images, where A=(x 1 ,x 2 ,..,x k ), B=(y 1 ,y2 ,..,y k ), C=(z 1 ,z 2 ,..,z k ), where x 1 ,x 2 ,x k are the 1st, 2nd, and kth cross-sectional 2D images in A, respectively. 1 ,y 2 ,y k are the 1st, 2nd, and kth cross-sectional 2D images in B, respectively. 1 ,z 2 ,z k are the 1st, 2nd, and kth cross-sectional 2D images in C, respectively. (x 1 ,x 2 ,..,x k )、(y 1 ,y 2 ,..,y k ) and (z 1 ,z 2 ,..,z k ) contains the ImagePosition[2] information that indicates the z-axis position of the slice.
[0114] In addition to being used for abdominal segment scanning, the segment scanning in the present invention can also be applied to other parts such as thighs for splicing after segment scanning, which will not be elaborated here.
[0115] In step S2,
[0116] When N = 2, the ImagePosition[2] information contained in the two-dimensional image DICOM is used to find the two-dimensional layers in A and B that are in the same z-axis position value range, and obtain the repeated layers in A and B, among which the non-repeated layers in A are A. 1 =(x 1 ,x 2 ,..,x k-m ), the repeat level is A 2 =(x k-m+1 ,x k-m+2 ,..,x k );
[0117] The repeating level in B is B 1 =(y 1 ,y 2 ,..,y m ), the non-repeated level is B 2 =(y m+1 ,y m+2 ,..,y k );
[0118] In the formula, x 1 ,x 2 ,x k-m are the 2D images of the cross section of the 1st, 2nd, and km frames in A, respectively. k-m+1 ,x k-m+2 ,x k are the k-m+1, k-m+2, and k-th cross-sectional 2D images in A, respectively. 1 ,y 2 ,y m are the 1st, 2nd, and mth cross-sectional 2D images in B, respectively. m+1 ,y m+2 ,y k They are the m+1, m+2, and k-th cross-sectional two-dimensional images in B, respectively;
[0119] When N=3, in the ImagePosition[2] information contained in the two-dimensional image DICOM, find the two-dimensional layers in A and B that are in the same z-axis position value range, and obtain the repeated layers in A and B. Find the two-dimensional layers in B and C that are in the same z-axis position value range, and obtain the repeated layers in B and C.
[0120] The non-repeating level in A is A 1 =(x 1 ,x 2 ,..,x k-m ), the repeat level is A 2 =(x k-m+1 ,x k-m+2 ,..,x k );
[0121] The header repetition level in B is B 1 =(y 1 ,y 2 ,..,y m ), the non-repeated level is B 2 =(y m+1 ,y m+2 ,..,y k-n ), the last segment repeats level B 3 =(y k-n+1 ,y k-n+2 ,..,y k );
[0122] The repetition level in C is C 1 =(z 1 ,z 2 ,..,z n ), the non-repeating level is C 2 =(z n+1 ,z n+2,..,z k );
[0123] In the formula, x 1 ,x 2 ,x k-m are the 2D images of the cross section of the 1st, 2nd, and km frames in A, respectively. k-m+1 ,x k-m+2 ,x k are the k-m+1, k-m+2, and k-th cross-sectional 2D images in A, respectively. 1 ,y 2 ,y m are the 1st, 2nd, and mth cross-sectional 2D images in B, respectively. m+1 ,y m+2 ,y k-n are the m+1, m+2, and kn cross-sectional 2D images in B, respectively. k-n+1 ,y k-n+2 ,y k are the k-n+1, k-n+2, and k-th cross-sectional 2D images in B, respectively. 1 ,z 2 ,z n are the 1st, 2nd, and nth cross-sectional 2D images in C, respectively. n+1 ,z n+2 ,z k They are respectively the n+1, n+2, and k-th cross-sectional two-dimensional images in C; kn>m+1.
[0124] In step S3,
[0125] When N=2, the repeated level in A is A 2 =(x k-m+1 ,x k-m+2 ,..,x k ) is normalized to the grayscale range of 0-255, A 2 The grayscale normalization formula is:
[0126]
[0127] Set the repeated layer in B to B 1 =(y 1 ,y 2 ,..,y m ) is normalized to the grayscale range of 0-255, B 1 The grayscale normalization formula is:
[0128]
[0129] In the formula, A 2norm A 2The gray level normalization result, A 2 is the repeated level in A, B 1norm For B 1 The gray level normalization result, B 1 is the repeated level in B;
[0130] When N=3, the repetition level in A is A 2 =(x k-m+1 ,x k-m+2 ,..,x k ) is normalized to the grayscale range of 0-255, A 2 The grayscale normalization formula is:
[0131]
[0132] Set the repeated layer in B to B 1 =(y 1 ,y 2 ,..,y m ) and B 3 =(y k-n+1 ,y k-n+2 ,..,y k ) are normalized to the range of 0-255, B 1 The grayscale normalization formula is:
[0133]
[0134] B 3 The grayscale normalization formula is:
[0135]
[0136] Change the repetition level in C to C 1 =(z 1 ,z 2 ,..,z n ) is normalized to the grayscale range of 0-255, C 1 The grayscale normalization formula is:
[0137]
[0138] In the formula, A 2norm A 2 The gray level normalization result, A 2 is the repeated level in A, B 1norm For B 1 The gray level normalization result, B 1 Repeat the level for the header in B, B 3norm For B 3 The gray level normalization result, B3 The end segment of B is repeated, C 1norm C 1 Grayscale normalization result, C 1 is the repetition level in C.
[0139] In step S4,
[0140] When N=2, the translation range is given as [-256, 256];
[0141] B 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated B is iteratively calculated. 1norm and A 1norm The mean square error between
[0142] The coordinates when the mean square error reaches the minimum value are taken as the optimal translation transformation coordinates;
[0143] When N=3, the translation range is given as [-256, 256];
[0144] B 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated B is iteratively calculated. 1norm and A 2norm The mean square error between
[0145] The coordinate when the mean square error reaches the minimum value is taken as B 2 Segment optimal translation transformation coordinates;
[0146] C 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated C is iteratively calculated. 1norm and B 3norm The mean square error between
[0147] The coordinate when the mean square error reaches the minimum value is taken as C 2 Segment optimal translation transformation coordinates;
[0148] In step S5,
[0149] When N = 2, based on the optimal translation transformation coordinates, 2 Each non-repeating layer in the A is translated and connected in sequence 1 , A 2 and B after translation 2 , obtain MRI images;
[0150] When N = 3, based on B 2 The optimal translation transformation coordinates of segment B 2 Each non-repeating level in the translation is based on B 2 Duan and C 2 The optimal translation transformation coordinates of the segment are2 Shift the non-repeating levels within;
[0151] Connect A in sequence 1 , A 2 and B after translation 2 , C 2 , and obtain MRI images.
[0152] During the DIXON sequence for abdominal imaging of inflammatory bowel disease, the ingestion of contrast agents in the intestine and the disorder of the patient's intestinal function make it impossible to perform a long-term single-sequence scan, and thus a single whole-body DIXON scan cannot be performed. During the data acquisition process, the pulse sequence needs to be changed regularly to cause the required water / fat relative phase changes in the acquired image. In the presence of intestinal movement and limited tissue relaxation time, the increase in scanning time and the change in pulse sequence timing will lead to a significant reduction in the quality of the acquired image. Accordingly, the processed images containing only water and only fat will also experience image quality degradation such as increased motion artifacts or image blurring. During MRI intestinal imaging, the cross-sectional abdominal area is divided into 2 to 3 segmented areas for scanning to reduce the total scanning time of a single sequence. To avoid missing layers, overlapping layers will be set between the two segments when setting the segmented FOV. The time difference between the cross-sectional layers connected by the segmented scan will cause a certain displacement between the overlapping layers of the human body corresponding to the segmented scan, such as Figure 4 shown.
[0153] The process of segmented scanning for fat and muscle segmentation and quantification in the abdominal area needs to consider the removal of duplicate layers. The present invention adopts a displacement transformation method to enable complete matching of layer connections. Before segmentation, the segmented 3D abdominal sequence images are aligned and merged into a complete abdominal 3D image. On the one hand, it can provide more accurate quantitative results after subsequent segmentation, and on the other hand, it can reduce the manual workload required for labeling. First, the duplicate layers are determined based on the size of the image header information ImagePosition[2]. All layers of the upper abdominal segment scan are retained, and the duplicate layers of the lower abdominal / pelvic segment and the upper abdominal segment are removed. When the scan is divided into three segments, the duplicate layers of the middle segment, the upper abdominal segment, and the pelvic segment are removed. In the algorithm process of realizing the overall segmented scan merging, the duplicate layers are used to calculate the number of movements required for the displacement transformation to achieve complete matching. Figure 5 As shown, taking a two-segment scanning same-phase image as an example, the segmented scanning fusion of the present invention can obtain a complete 3D image.
[0154] In order to improve the segmentation accuracy of the MRI image automatic segmentation model, the present invention uses the FCM algorithm to perform initial rough segmentation of muscle, fat, and non-muscle fat background, considers the low signal area of fat caused by the partial volume effect, and performs unsupervised classification of image pixels in the cluster without using a display threshold, as follows:
[0155] Methods for obtaining true annotations of fat and muscle regions in each MRI image using the FCM algorithm include:
[0156] The FCM algorithm is used to perform rough segmentation of fat and muscle areas based on the water phase map and the fat phase map to obtain a rough segmentation mask;
[0157] The true annotation of fat and muscle areas in MRI images is obtained based on manual annotation correction in the coarse segmentation mask;
[0158] in,
[0159] Manual annotation correction methods include:
[0160] The third and fourth category fuzzy classification masks or label maps obtained by operation in the water phase map are used for further manual correction to obtain the true labels of subcutaneous fat and visceral fat, respectively;
[0161] The third and fourth category fuzzy classification masks or label maps obtained by operation in the fat phase map are used for further manual correction to obtain the true labeling of the muscle;
[0162] Among them, the manual annotation correction process was carried out using the open source software ITK-SNAP combined with MRI images without any preprocessing.
[0163] like Figure 6 As shown, the coarse segmentation method of the FCM algorithm includes:
[0164] The number of pixel categories of the fat phase map and the water phase map is predefined as 4, among which the pixel categories of the fat phase map include cortical bone / background, fat-rich area, bone marrow fat and partial volume effect voxel area; the pixel categories of the water phase map include background, cortical bone / fat, soft tissue and blood vessels and digestive tract fluid;
[0165] The FCM algorithm is used to evaluate the membership value u of category k at pixel j in the fat phase map and the water phase map by minimizing the loss function. j,k , assign each pixel category to each pixel in the fat phase map and the water phase map;
[0166] The loss function is:
[0167]
[0168]
[0169]
[0170] In the formula, y j is pixel j,u j,kis the membership value of category k at pixel j, N is the total number of pixels, c is the number of pixel categories, v k is the data center of the kth category, U is the membership function matrix, and V vector represents the cluster center v k , q is the weight index used to determine the number of fuzzy classifications;
[0171] The membership degree u of the water phase diagram and the fat phase diagram j,k The smallest values correspond to the background, and the membership u jk The values of the largest and second largest two types of voxels are set to 1, and the four types of fat rough segmentation mask images are obtained. The membership degree u in the water phase map jk The largest and next two types of voxel values are set to 1, and four types of muscle rough segmentation mask images including soft tissues such as muscles and abdominal organs of the liver are obtained, and the rough segmentation mask images except the background are removed by OR operation through the water phase map and the fat phase map to obtain the rough mask of the body contour.
[0172] The true annotations of fat and muscle regions in MRI images include masks of subcutaneous fat SAT, visceral fat VAT, and muscle SM.
[0173] In the process of MRI intestinal imaging, the present invention performs segmented scanning of the abdominal area in cross-sectional imaging to avoid layer omission, that is, to avoid information loss, and removes repeated layers between segmented scans. A displacement transformation method is used to enable complete matching of layer connections. The segmented 3D abdominal sequence images are aligned and merged into a complete abdominal 3D image before segmentation. On the one hand, more accurate quantitative results can be provided after subsequent segmentation, and on the other hand, the manual workload required for labeling can be reduced. The FCM algorithm is used to perform initial coarse segmentation of backgrounds such as muscle, fat, and non-muscle fat. Considering the low-signal area in fat caused by the partial volume effect, the image pixels in the cluster are unsupervisedly classified without using a display threshold, thereby improving the segmentation efficiency and accuracy of fat and muscle areas in MRI imaging.
[0174] The above embodiments are only exemplary embodiments of the present application and are not intended to limit the present application. The protection scope of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and protection scope of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present application.
Claims
1. A data processing method based on automatic segmentation of fat and muscle regions in MRI imaging, characterized in that: The following steps are involved: Acquire a CT image automatic segmentation model for automatically segmenting fat and muscle regions in CT imaging; Acquire multiple MRI images containing fat and muscle areas by using a single-site segmented scanning method; The FCM algorithm is used to obtain the true annotation of fat and muscle areas in each MRI image; The dataset is composed of MRI images and real annotations of fat and muscle areas; Based on the data set, transfer learning is performed on the CT image automatic segmentation model to obtain an MRI image automatic segmentation model to achieve automatic segmentation of fat and muscle areas in MRI images.
2. The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging according to claim 1, characterized in that: Methods for acquiring multiple MRI images containing fat and muscle regions by a single-site segmented scanning method include: Step S1: Obtain N segmented images by using N-segment scanning, where the value of N includes 2 and 3; Step S2: searching for repeated layers in each of the N segmented images; Step S3: normalizing the grayscale range of the repeated layers in the N segmented images to 0-255; Step S4: Calculating the optimal translation transformation coordinates based on the repeated layers; Step S5: translating non-repetitive layers in the segmented image based on the optimal translation transformation coordinates, and reconnecting the translated segmented images to obtain an MRI image.
3. The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging according to claim 2, characterized in that: In step S1, When N=2, segmented images A and B are obtained respectively, and each of the segmented images A and B contains k frames of cross-sectional two-dimensional images, where A=(x1, x2, .., x k ), B=(y1,y2,..,y k ), where x1, x2, x k are the 1st, 2nd, and kth cross-sectional 2D images in A, y1, y2, y k are the 1st, 2nd, and kth cross-sectional 2D images in B, x1, x2, .., x k and y1,y2,..,y k The ImagePosition[2] information contained in the represented two-dimensional image DICOM indicates the z-axis position of the slice; When N=3, segmented images A, B and C are obtained respectively, and each of the segmented images A, B and C contains k frames of cross-sectional two-dimensional images, where A=(x1, x2, .., x k ), B=(y1,y2,..,y k ), C=(z1,z2,..,z k ), where x1, x2, x k are the 1st, 2nd, and kth cross-sectional 2D images in A, y1, y2, y k are the 1st, 2nd, and kth cross-sectional 2D images in B, z1, z2, z k are the 1st, 2nd, and kth cross-sectional 2D images in C, (x1, x2, .., x k )、(y1,y2,..,y k ) and (z1,z2,..,z k ) contains the ImagePosition[2] information that indicates the z-axis position of the slice.
4. The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging according to claim 3, characterized in that: In step S2, When N=2, the ImagePosition[2] information contained in the two-dimensional image DICOM is used to find the two-dimensional slices in A and B that are in the same z-axis position value range, and obtain the repeated slices in A and B, where: The non-repeating layers in A are A1=(x1,x2,..,x k-m ), the repetition level is A2=(x k-m+1 ,x k-m+2 ,..,x k ); The repetition level in B is B1=(y1,y2,..,y m ), the non-repeated level is B2=(y m+1 ,y m+2 ,..,y k ); In the formula, x1, x2, x k-m are the 2D images of the cross section of the 1st, 2nd, and km frames in A, respectively. k-m+1 ,x k-m+2 ,x k are the k-m+1, k-m+2, and k-th cross-sectional 2D images in A, y1, y2, y m are the 1st, 2nd, and mth cross-sectional 2D images in B, respectively. m+1 ,y m+2 ,y k They are the m+1, m+2, and k-th cross-sectional two-dimensional images in B, respectively; When N=3, in the ImagePosition[2] information contained in the two-dimensional image DICOM, find the two-dimensional layers in A and B that are in the same z-axis position value range, and obtain the repeated layers in A and B. Find the two-dimensional layers in B and C that are in the same z-axis position value range, and obtain the repeated layers in B and C. The non-repeating layers in A are A1=(x1,x2,..,x k-m ), the repetition level is A2=(x k-m+1 ,x k-m+2 ,..,x k ); The header repetition level in B is B1=(y1,y2,..,y m ), the non-repeated level is B2=(y m+1 ,y m+2 ,..,y k-n ), the last segment repeats level B3=(y k-n+1 ,y k-n+2 ,..,y k ); The repetition level in C is C1=(z1,z2,..,z n ), the non-repeated level is C2=(z n+1 ,z n+2 ,..,z k ); In the formula, x1, x2, x k-m are the 2D images of the cross section of the 1st, 2nd, and km frames in A, respectively. k-m+1 ,x k-m+2 ,x k are the k-m+1, k-m+2, and k-th cross-sectional 2D images in A, y1, y2, y m are the 1st, 2nd, and mth cross-sectional 2D images in B, respectively. m+1 ,y m+2 ,y k-n are the m+1, m+2, and kn cross-sectional 2D images in B, respectively. k-n+1 ,y k-n+2 ,y k are the k-n+1, k-n+2, and k-th cross-sectional 2D images in B, z1, z2, and z n are the 1st, 2nd, and nth cross-sectional 2D images in C, respectively. n+1 ,z n+2 ,z k They are respectively the n+1th, n+2nd, and kth cross-sectional two-dimensional images in C, where kn>m+1.
5. The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging according to claim 4, characterized in that: In step S3, When N=2, the repetition level in A is A2=(x k-m+1 ,x k-m+2 ,..,x k ) is normalized to the range of 0-255, and the A2 grayscale normalization formula is: Set the repetition level in B to B1 = (y1, y2, .., y m ) is normalized to the range of 0-255, and the grayscale normalization formula of B1 is: When N=3, the repetition level in A is A2=(x k-m+1 ,x k-m+2 ,..,x k ) is normalized to the range of 0-255, and the A2 grayscale normalization formula is: Set the repetition level in B to B1 = (y1, y2, .., y m ) and B3=(y k-n+1 ,y k-n+2 ,..,y k ) are normalized to 0-255, and the grayscale normalization formula of B1 is: The B3 grayscale normalization formula is: The repetition level in C is C1=(z1,z2,..,z n ) is normalized to the range of 0-255, and the grayscale normalization formula of C1 is: In the formula, A 2norm is the grayscale normalization result of A2, A2 is the repeated layer in A, B 1norm is the grayscale normalization result of B1, B1 is the header repetition level in B, B 3norm is the grayscale normalization result of B3, B3 is the tail segment repetition level in B, C 1norm is the grayscale normalization result of C1, where C1 is the repeated level in C.
6. The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging according to claim 5, characterized in that: In step S4, When N=2, the translation range is given as [-256, 256]; B 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated B is iteratively calculated. 1norm and A 1norm The mean square error between The coordinates when the mean square error reaches the minimum value are taken as the optimal translation transformation coordinates; When N=3, the translation range is given as [-256, 256]; B 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated B is iteratively calculated. 1norm and A 2norm The mean square error between The coordinates when the mean square error reaches the minimum value are taken as the optimal translation transformation coordinates of segment B2; C 1norm The inner pixel is translated pixel by pixel along the x and y axes of the cross section, and the translated C is iteratively calculated. 1norm and B 3norm The mean square error between The coordinates when the mean square error reaches the minimum value are taken as the optimal translation transformation coordinates of segment C2.
7. The data processing method based on automatic segmentation of fat and muscle regions in MRI imaging according to claim 6, characterized in that: In step S5, When N=2, based on the optimal translation transformation coordinates, each non-repetitive layer in B2 is translated and sequentially connected to A1, A2 and B2 after translation to obtain the MRI image; When N=3, based on the optimal translation transformation coordinates of segment B2, each non-repetitive layer in B2 is translated, and based on the optimal translation transformation coordinates of segments B2 and C2, each non-repetitive layer in C2 is translated; A1, A2 and the translated B2, C2 are connected in sequence to obtain the MRI image.
8. The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging according to claim 1, characterized in that: Methods for obtaining true annotations of fat and muscle regions in each MRI image using the FCM algorithm include: The FCM algorithm is used to perform rough segmentation of fat and muscle areas based on the water phase map and the fat phase map to obtain a rough segmentation mask; Obtaining true annotations of fat and muscle areas in the MRI image based on manual annotation correction in the coarse segmentation mask; in, The manual annotation correction method comprises: The third and fourth category fuzzy classification masks or label maps obtained by operation in the water phase map are used for further manual correction to obtain the true labels of subcutaneous fat and visceral fat, respectively; The third and fourth category fuzzy classification masks or label maps obtained by operation in the fat phase map are used for further manual correction to obtain the true labeling of the muscle; Among them, the manual annotation correction process was carried out using the open source software ITK-SNAP combined with MRI images without any preprocessing.
9. The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging according to claim 8, characterized in that: The coarse segmentation methods of the FCM algorithm include: The number of pixel categories of the fat phase map and water phase map of the MRI image is predefined as 4, among which the pixel categories of the fat phase map include cortical bone / background, fat-rich area, bone marrow fat and partial volume effect voxel area; the pixel categories of the water phase map include background, cortical bone / fat, soft tissue and blood vessels and digestive tract fluid; The FCM algorithm is used to evaluate the membership value u of category k at pixel j in the fat phase map and the water phase map by minimizing the loss function. j,k , assign each pixel category to each pixel in the fat phase map and the water phase map; The loss function is: In the formula, y j is pixel j,u j,k is the membership value of category k at pixel j, N is the total number of pixels, c is the number of pixel categories, v k is the data center of the kth category, U is the membership function matrix, and V vector represents the cluster center v k , q is the weight index used to determine the number of fuzzy classifications; The membership degree u of the water phase diagram and the fat phase diagram is j,k The smallest values correspond to the background, and the membership u jk The values of the largest and second largest two types of voxels are set to 1, and the four types of fat rough segmentation mask images are obtained. The membership degree u in the water phase map jk The largest and next two types of voxel values are set to 1, and four types of muscle rough segmentation mask images including soft tissues such as muscles and abdominal organs of the liver are obtained, and the rough segmentation mask images except the background are removed by OR operation through the water phase map and the fat phase map to obtain the rough mask of the body contour.
10. The data processing method based on automatic segmentation of fat and muscle areas in MRI imaging according to claim 1, characterized in that: The true annotations of the fat and muscle areas in the MRI image include masks of subcutaneous fat SAT, visceral fat VAT and muscle SM.
