A method for automatic positioning of the four-chamber heart based on deep learning
Through the automatic positioning method based on deep learning, the four-cavity heart plane is calculated using the region perception network, which solves the problems of complex, time-consuming and poor accuracy of manual positioning, and realizes the rapid and accurate automation of the four-cavity heart scan and improves the heart positioning efficiency and consistency.
Patent Information
- Application Number
- CN202210075798.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-23
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-01-23
AI Technical Summary
The existing manual cardiac positioning methods are complex and time-consuming, have poor accuracy, and the results of different operators or the same operator are inconsistent, making it difficult to ensure the efficiency and consistency of cardiac scans.
Using the deep learning-based automatic positioning method of the four-cavity heart, the multi-layer chest MR coronal image is scanned, and the area-perceptual positioning network is used to perform cardiac segmentation and planar heat map prediction, the four-cavity heart plane equation is calculated, and the dicom scanning parameters are generated. Realize automatic positioning of the four-chamber core scanning surface.
The cardiac positioning process is simplified, the calculation speed and accuracy is significantly improved, the work burden of doctors is reduced, and the risk of overfitting is reduced, and the four-chamber cardiac scanning parameters can be predicted quickly and accurately.
Smart Images

Figure CN114419157B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for automatic positioning of a four-chamber heart based on deep learning, and belongs to the field of computer image processing. Background Art
[0002] There are three basic scanning planes for MRI heart scans: two-chamber heart, four-chamber heart and short-axis. The two-chamber heart plane reflects the left ventricle and left atrium, the four-chamber heart plane reflects the left and right ventricles and the left and right atria, and the short-axis plane reflects the cross-section of the left and right ventricles. The line connecting the midpoint of the mitral valve and the apex of the left ventricle is defined as the long axis of the left ventricle. With this axis as the axis, the sagittal, coronal and cross-sectional scans of the heart are performed, and the scanning planes obtained are the two-chamber heart, four-chamber heart and short-axis positions.
[0003] After cardiac localization is completed, the patient can undergo subsequent examinations, such as: measuring ejection fraction, ventricular volume, and heart mass; observing the gross morphology of the heart; performing movie sequence scans to evaluate myocardial function; and performing myocardial perfusion and delayed enhancement to diagnose myocardial disease.
[0004] The current method of manual heart positioning is to first scan a human cross-section to locate the line between the left ventricular apex and the midpoint of the mitral valve (left ventricular long axis). Then scan this line to obtain a two-chamber heart plane (heart sagittal plane); then in the two-chamber heart plane, scan the line between the midpoint of the mitral valve and the left ventricular apex to obtain a four-chamber heart plane (heart coronary plane); finally, use the four-chamber heart plane as a reference plane, and continuously scan vertically in the direction of the long axis of the left ventricle to obtain the short axis plane (heart transverse plane). This method is complicated and time-consuming. Even for professionally trained doctors, it takes about 3 minutes to complete the entire scan. In addition, this method has high variability. When performing heart positioning on the same patient, the results of different operators may be inconsistent, and even the results of multiple positioning by the same operator may be different. Summary of the invention
[0005] In view of the problems existing in the manual positioning method, the present invention proposes a method for automatic positioning of the four-chamber heart based on deep learning. After locating the four-chamber heart plane, the doctor can conveniently manually locate the two-chamber heart and the short axis position (both of which are perpendicular to the four-chamber heart plane), which can reduce the doctor's workload and improve the accuracy of heart positioning.
[0006] In order to achieve the above-mentioned purpose, the technical solution of the present invention is as follows: The present invention provides a method for automatic positioning of a four-chamber heart based on deep learning, and the specific steps are as follows:
[0007] Step 1: Scan multi-slice chest MR coronal images.
[0008] Step 2: Overlay the images in step 1 into a 3D image and scale it to voxel spacing = (1,1,1). If the original length, width and height of the 3D image is (a,b,c), and the original voxel spacing = (x,y,z), then the length, width and height of the scaled 3D image is (ax,by,cz). The image is scaled using the 3D linear interpolation algorithm in ITK.
[0009] Step 3: Input the 3D image in step 2 into the region-aware localization network and output the heart segmentation map and the planar heat map. The heart segmentation map includes the left and right ventricles and the left and right atria of the heart. The planar heat map consists of a series of points with values 2 ranging from 0 to 1, and the value of the point represents its probability on the plane. Figure 2 As shown in the figure, the network consists of two parallel Unets, one for segmenting the heart and the other for predicting the plane heat map. The decoder parts of the two Unets are connected using a region-aware module. The region-aware module converts the structural information of the heart into an attention score, which is multiplied with the feature map of the encoder part in the plane prediction Unet to increase the attention to the heart area.
[0010] Step 4: Calculate the four-chamber heart plane equation from the heat map obtained in step 3. The calculation method is as follows: The value of the point with coordinates (xi, yi, zi) is known to be wi, and the fitted plane equation is: z = ax + by + c, then The specific solution process is as follows:
[0011]
[0012]
[0013]
[0014]
[0015] Step 5: Based on the heart segmentation map obtained in step 3 and the plane equation obtained in step 4, calculate the DICOM scanning parameters, including the upper left corner coordinate FOV, the horizontal axis direction vector Vertical axis direction vector The direction vector of the known plane is The side length of the dicom plane is L, and the center of gravity of the intersection of the heart segmentation map and the four-chamber heart plane is G. In order to make the direction of the four-chamber heart scan image conform to the doctor's habit of diagnosis, It should be parallel to the xz plane of the dicom world coordinate system. The parameter calculation process is as follows:
[0016]
[0017]
[0018]
[0019] Step 6: Use the scanning parameters obtained in step 5 to scan and obtain a four-chamber heart scanning plane.
[0020] Compared with the prior art, the advantages of the present invention are as follows: (1) The process is simple. The method only needs to scan the transverse axial chest image. Through the calculation of the above steps, the four-chamber heart scan plane can be obtained; (2) The calculation speed is fast. The method can be deployed on a GPU. The computing speed of the GPU is many times faster than that of the CPU; (3) The accuracy is high. The method adopts a deep learning method and uses many groups of transverse axial chest images for training to improve the accuracy and reduce the risk of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a schematic diagram of the process of the present invention.
[0022] Figure 2 Schematic diagram of the area-aware positioning network model.
[0023] Figure 3 This is a test result diagram of the present invention. DETAILED DESCRIPTION
[0024] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.
[0025] Example: See Figure 1 , Figure 2 The present invention provides a method for automatic positioning of a four-chamber heart based on deep learning, and the specific steps are as follows:
[0026] Step 1: Scan multi-slice chest MR coronal images.
[0027] Step 2: Overlay the image in step 1 into a 3D image and scale it to voxel spacing = (1,1,1). If the original length, width and height of the 3D image is (a,b,c), and the original voxel spacing = (x,y,z), then the length, width and height of the scaled 3D image is (ax,by,cz). The image is scaled using the 3D linear interpolation algorithm in ITK.
[0028] Step 3: Input the 3D image in step 2 into the region-aware localization network and output the heart segmentation map and the planar heat map. The heart segmentation map includes the left and right ventricles and the left and right atria of the heart. The planar heat map consists of a series of points with values 2 ranging from 0 to 1, and the value of the point represents its probability on the plane. Figure 2 As shown in the figure, the network consists of two parallel Unets, one for segmenting the heart and the other for predicting the plane heat map. The decoder parts of the two Unets are connected using a region-aware module. The region-aware module converts the structural information of the heart into an attention score, which is multiplied with the feature map of the encoder part of the plane prediction Unet to increase the attention to the heart area.
[0029] Step 4: Calculate the four-chamber heart plane equation from the heat map obtained in step 3. The calculation method is as follows: The value of the point with coordinates (xi, yi, zi) is known to be wi, and the fitted plane equation is: z = ax + by + c, then The specific solution process is as follows:
[0030]
[0031]
[0032]
[0033]
[0034] Step 5: Based on the heart segmentation map obtained in step 3 and the plane equation obtained in step 4, calculate the DICOM scanning parameters, including the upper left corner coordinate FOV, the horizontal axis direction vector Vertical axis direction vector The direction vector of the known plane is The side length of the dicom plane is L, and the center of gravity of the intersection of the heart segmentation map and the four-chamber heart plane is G. In order to make the direction of the four-chamber heart scan image conform to the doctor's habit of diagnosis, It should be parallel to the xz plane of the dicom world coordinate system. The parameter calculation process is as follows:
[0035]
[0036]
[0037]
[0038] Step 6: Use the scanning parameters obtained in step 5 to scan and obtain a four-chamber heart scanning plane.
[0039] Effect evaluation:
[0040] The present invention proposes a method for automatic positioning of the four-chamber heart based on deep learning. The test results of this method are as follows: Figure 3 As shown in the figure, four groups of test images are shown. The process of each group of tests is: predicting the four-chamber heart scan parameters of the heart from the three-dimensional chest image, and then using this parameter to simulate scanning (cutting the three-dimensional chest image). The test results show that this method can quickly and accurately predict the four-chamber heart scan parameters, providing doctors with an efficient solution for heart positioning.
[0041] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technology can understand and think of any changes or substitutions within the technical scope disclosed by the present invention, which should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A method for automatic positioning of the four-chamber heart based on deep learning. It is characterized in that The specific steps are as follows: Step 1, scanning multiple layers of continuous chest MR coronal images; Step 2, stacking the multiple layers of continuous images in step 1 into a 3D image and scaling it; Step 3, input the 3D image in step 2 into the region-aware localization network, and output the heart segmentation map and planar heat map; Step 4, calculating the four-chamber heart plane equation using a plane fitting algorithm from the heat map obtained in step 3; Step 5: Based on the heart segmentation map obtained in step 3 and the plane equation obtained in step 4, calculate the DICOM scanning parameters, including the upper left corner coordinate FOV, the horizontal axis direction vector Vertical axis direction vector Step 6, using the scanning parameters obtained in step 5, scanning to obtain a four-chamber heart scanning plane; Among them, the region-aware positioning network model mentioned in step 3 can simultaneously predict the heart segmentation map and the plane heat map. The heart segmentation map includes the left and right ventricles and the left and right atria of the heart. The plane heat map consists of a series of points with values of 0-1, and the value of the point represents its probability on the plane. The region-aware positioning network consists of two parallel Unets, one for segmenting the heart and the other for predicting the plane heat map. The decoder part of the two Unets is connected using a region-aware module. The region-aware module converts the structural information of the heart into an attention score, which is multiplied by the feature map of the encoder part in the plane prediction Unet to increase the attention to the heart area; The calculation method of the plane fitting algorithm mentioned in step 4 is: the value of the point with known coordinates (xi, yi, zi) is wi, and the equation of the fitted plane is: z = ax + by + c, then The specific solution process is as follows: The calculated DICOM parameters mentioned in step 5, the upper left corner coordinate FOV, the horizontal axis direction vector Vertical axis direction vector The calculation method is: the direction vector of the known plane is The side length of the dicom plane is L, and the center of gravity of the intersection of the heart segmentation map and the four-chamber heart plane is G. In order to make the direction of the four-chamber heart scan image conform to the doctor's diagnosis habits, Should be parallel to the xz plane of the DICOM world coordinate system, then:
Citation Information
Patent Citations
Deep learning and MRI image-based ventricular function index prediction method
CN106096632A
A heart left ventricle segmentation method based on a deep full convolutional neural network
CN109584254A
Method and device for calculating ultrasonic cardiogram cardiac parameters and measuring myocardial strain
CN111012377A
