A method and system for PET image repositioning

Abstract

The present application relates to a kind of PET image repositioning method and system.The PET image repositioning method is preprocessed to the PET transaxial image of the acquired PET to be repositioned after obtaining preprocessed image, and then the anchor point coordinates are obtained by inputting preprocessed image to detection network, then, the rigid transformation parameters are determined based on anchor point coordinates, then, the PET transaxial image to be repositioned is transformed into PET short axis image based on rigid transformation parameters, to realize the repositioning of PET image, and then the reproducibility problem such as poor existing in manual repositioning of heart PET image in prior art is solved by automatic processing mode.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for repositioning PET images. Background Technology

[0002] Positron emission tomography (PET), as a functional imaging technique, can reflect the functional metabolism of a patient's heart, which is beneficial for a deeper understanding of various physiological mechanisms of the heart and provides a basis for early detection and prognosis assessment of clinical diseases.

[0003] Conventional reconstructed PET images are transverse images perpendicular to the long axis of the patient's chest cavity, rather than perpendicular to the long axis of the left ventricle of the heart (e.g.) Figure 1 (As shown). Direct image analysis and diagnosis from the horizontal axis image can lead to errors in the identification of cardiac regions. Therefore, to meet the clinical needs for cardiac PET image diagnosis, PET reconstructed images usually need to be reoriented to a specific standard plane, namely the short axis (SA). Figure 2 The images shown are the imaging plane, 4-chamber heart slices, and 2-chamber heart slices. In PET cardiac analysis, it is necessary to generate a bullseye image from the image. Since the bullseye image is defined based on the short axis, generating short axis cross-sectional images from the transverse axis slice images is the first step in cardiac PET analysis; this process is called PET cardiac image repositioning. The cross-sectional images in the three transverse axis directions are shown below. Figures 3-5 As shown, the cross-sectional views in three directions along the short axis are as follows: Figures 6-8 As shown.

[0004] Traditional methods for repositioning cardiac PET images typically require manual intervention by the user. This process usually involves selecting a mid-axis slice of the left ventricle (LV) and drawing a line to delineate the long axis of the LV (e.g., ...). Figure 9 (As shown). This process is repeated in the sagittal plane (e.g.) Figure 10 As shown). The major axes of these two LVs define the coordinate transformation from the horizontal axis image to the SA image on different planes (e.g. Figure 11 (As shown). The result of this manual process depends on the operator's actions, and the reproducibility of different people's actions is poor. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a PET image repositioning method and system.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A PET image repositioning method, comprising:

[0008] Obtain the lateral image of the PET image to be repositioned;

[0009] The PET horizontal axis image to be repositioned is preprocessed to obtain a preprocessed image;

[0010] The preprocessed image is input into a detection network to obtain anchor point coordinates; the detection network is a trained deep learning network.

[0011] The rigid transformation parameters are determined based on the anchor point coordinates;

[0012] Based on the rigid transformation parameters, the PET horizontal axis image to be repositioned is transformed into a PET short axis image.

[0013] Preferably, the preprocessing of the PET horizontal axis image to be repositioned to obtain a preprocessed image specifically includes:

[0014] When the PET horizontal axis image to be repositioned is a dynamic image, the image frames within a specific time period in the dynamic image are averaged over time to obtain a static frame image.

[0015] The preprocessed image is obtained by performing interpolation, filtering, and normalization on the static frame image.

[0016] When the PET horizontal axis image to be repositioned is a static image, the preprocessed image is obtained by interpolating, filtering and normalizing each frame of the static image.

[0017] Preferably, the training process of the deep learning network includes:

[0018] Obtain input image samples and preprocess the input image samples to obtain the output image;

[0019] The output image is manually anchored to obtain anchor point position information;

[0020] The anchor point location information is used as the gold standard for network output.

[0021] The position error term is determined based on the gold standard output by the network.

[0022] In the output image, a neighborhood is selected based on the anatomical structure characteristics, with the anchor point in the anchor point location information as the center.

[0023] Construct image fidelity terms based on the neighborhood;

[0024] An energy function is generated based on the position error term and the image fidelity term;

[0025] The energy function is used as the loss function of the deep learning network, and the output image is used as the input, while the anchor point position information is used as the output to train the deep learning network until the value of the loss function is minimized, thus obtaining the trained deep learning network.

[0026] Preferably, the loss function is Loss:

[0027] Loss=PL1+λMSE

[0028] In the formula, PL1 is the position error term, λMSE is the image fidelity term, and λ is the weighting coefficient.

[0029] Preferably, the position error term is PL1:

[0030]

[0031] In the formula, n represents the number of parameters in the network output gold standard, and P m R is the gold standard for predicting the m-th parameter. m Let m be the gold standard corresponding to the m-th parameter, where m = 1, 2, ..., n.

[0032] Preferably, the image fidelity item is MSE:

[0033]

[0034] IP ni S represents the image neighborhood corresponding to the i-th predicted anchor point in the graph. ni For each anchor point, x, y, and z represent the neighborhood of the corresponding gold marker in the image, and x, y, and z are the relative coordinates within that neighborhood.

[0035] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0036] The PET image repositioning method provided by this invention preprocesses the acquired PET horizontal axis image to be repositioned to obtain a preprocessed image, then inputs the preprocessed image into a detection network to obtain anchor point coordinates. Next, rigid transformation parameters are determined based on the anchor point coordinates. Then, based on the rigid transformation parameters, the PET horizontal axis image to be repositioned is transformed into a PET short axis image to achieve PET image repositioning. This automated processing method solves the problems of poor reproducibility in the existing technology when manually repositioning cardiac PET images.

[0037] Corresponding to the PET image repositioning method provided above, the present invention also provides a PET image repositioning system, the system comprising:

[0038] The image acquisition module is used to acquire the horizontal axis image of the PET to be repositioned;

[0039] The image processing module is used to preprocess the PET horizontal axis image to be repositioned to obtain a preprocessed image;

[0040] An anchor point detection module is used to input the preprocessed image into a detection network to obtain anchor point coordinates; the detection network is a trained deep learning network.

[0041] The parameter determination module is used to determine the rigid transformation parameters based on the anchor point coordinates.

[0042] The image positioning module is used to transform the PET horizontal axis image to be repositioned into a PET short axis image based on the rigid transformation parameters.

[0043] Preferably, the image processing module includes:

[0044] The image conversion unit is used to perform time averaging on image frames within a specific time period in the dynamic image when the PET horizontal axis image to be repositioned is a dynamic image, to obtain a static frame image.

[0045] The first processing unit is used to perform interpolation, filtering and normalization processing on the static frame image to obtain the preprocessed image;

[0046] The second processing unit is used to perform interpolation, filtering and normalization processing on each frame of the static image when the PET horizontal axis image to be repositioned is a static image, to obtain the preprocessed image.

[0047] Preferably, it further includes:

[0048] The sample acquisition module is used to acquire input image samples and preprocess the input image samples to obtain an output image;

[0049] Anchor point positioning module is used to manually position anchor points in the output image to obtain anchor point position information;

[0050] The standard generation module is used to generate the network output gold standard based on the anchor point location information;

[0051] An error term determination module is used to determine the position error term based on the gold standard output by the network.

[0052] The neighborhood selection module is used to select a neighborhood in the output image based on the anatomical structure characteristics, with the anchor point in the anchor point location information as the center.

[0053] A fidelity item construction module is used to construct an image fidelity item based on the neighborhood.

[0054] The function generation module is used to generate an energy function based on the position error term and the image fidelity term;

[0055] The network training module is used to train the deep learning network by taking the energy function as the loss function, the output image as the input, and the anchor point position information as the output, until the value of the loss function is minimized, thus obtaining the trained deep learning network.

[0056] Since the technical effects achieved by the PET image repositioning system provided by this invention are the same as those achieved by the PET image repositioning method provided above, they will not be described again here. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of the transverse axis imaging.

[0059] Figure 2 This is a schematic diagram of short-axis imaging;

[0060] Figure 3 This is a cross-sectional view along the horizontal axis.

[0061] Figure 4 This is a sagittal section view;

[0062] Figure 5 This is a coronal section view;

[0063] Figure 6 This is a cross-sectional view of the short axis.

[0064] Figure 7 A cross-sectional view of the 4-chamber core;

[0065] Figure 8 This is a cross-sectional view of the two-chamber heart.

[0066] Figure 9 A schematic diagram showing the long axis positioning of the left ventricle;

[0067] Figure 10 This is a schematic diagram of sagittal plane localization;

[0068] Figure 11 This is a schematic diagram illustrating the coordinate transformation from the horizontal axis image to the SA image;

[0069] Figure 12A flowchart of the PET image repositioning method provided by the present invention;

[0070] Figure 13 This is a block diagram illustrating an embodiment of the PET image repositioning method provided in this invention.

[0071] Figure 14 This is a manually annotated stereoscopic display result diagram provided in an embodiment of the present invention;

[0072] Figure 15 This is a schematic diagram showing the relative coordinate positions of the apex of the heart and the bases of the left and right ventricles, provided in an embodiment of the present invention.

[0073] Figure 16 A schematic diagram of the apex neighborhood;

[0074] Figure 17 A schematic diagram of the central region of the left ventricular basal region;

[0075] Figure 18 A schematic diagram of the central region of the right ventricle;

[0076] Figure 19 A schematic diagram of a deep learning network architecture provided in an embodiment of the present invention;

[0077] Figure 20 This is a schematic diagram of the PET image repositioning system provided by the present invention. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] The purpose of this invention is to provide a PET image repositioning method and system that can solve the problems of poor reproducibility when manually repositioning cardiac PET images in the prior art.

[0080] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0081] like Figure 12 As shown, the PET image repositioning method provided by the present invention includes:

[0082] Step 100: Obtain the PET horizontal axis image to be repositioned. Specifically:

[0083] When the PET axial image to be repositioned is a dynamic image, the image frames within a specific time period of the dynamic image are averaged over time to obtain a static frame image. For example, image frames within a stable tracer uptake period in the dynamic image are selected for time averaging to form a single frame image I(x,y,z)=1 / m∑Frame(x,y,z,ts:te), where ts:te represents the selected time period and m represents the number of time frames corresponding to that time period.

[0084] The preprocessed image is obtained by interpolating, filtering, and normalizing the static frame image. For example, the resolution of a single frame image I(x,y,z) is made consistent in the X, Y, and Z directions by interpolation, then edge-preserving filtering is performed to eliminate noise, and the image is normalized to eliminate the influence of tracer uptake.

[0085] When the PET horizontal axis image to be repositioned is a static image, the preprocessed image is obtained by directly performing interpolation, filtering and normalization on each frame of the static image.

[0086] Step 101: Preprocess the PET horizontal axis image to be repositioned to obtain a preprocessed image.

[0087] Step 102: Input the preprocessed image into the detection network to obtain the anchor point coordinates. The detection network is a trained deep learning network. The training process of the deep learning network includes:

[0088] Obtain input image samples and preprocess them to obtain the output image.

[0089] Manually locate anchor points in the output image to obtain anchor point position information. For example, ... Figure 14 As shown, manual anchor point positioning is performed on the output image. The apex of the left ventricle is drawn, and the line connecting the basal center of line 1 is defined. Line 2 is formed by drawing a perpendicular line from the midpoint of line 1 to the center of the right ventricle. This yields three anchor points: the apex, the basal center, and the right ventricular center. Figure 14 The circular point is the apex anchor point, the rectangular point is the base anchor point, and the triangular point is the right center anchor point.

[0090] The network output gold standard is generated based on the anchor point location information. Nine parameters are formed from the three coordinates of the points corresponding to the three manually calibrated anchor points (apex, base, and right ventricular center), serving as the network output gold standard. The relative coordinate positions of the apex and the base of the left and right ventricles are shown below. Figure 15 As shown.

[0091] The position error term is determined based on the gold standard output from the network. The position error term is PL1:

[0092]

[0093] In the formula, n represents the number of parameters in the network output gold standard, and P m R is the gold standard for predicting the m-th parameter. m Let m be the gold standard corresponding to the m-th parameter, where m = 1, 2, ..., n, for example, n = 9. This is only one method for constructing the error term; the 2-norm can also be used.

[0094] In the output image, the neighborhood is selected based on the anatomical structure characteristics, with the anchor point in the anchor point location information as the center.

[0095] Image fidelity terms are constructed based on neighborhood. For example, using three anchor points as centers, neighborhoods of different sizes and shapes are selected based on their anatomical characteristics (e.g., ...). Figure 16-18 As shown), the constructed image fidelity term is MSE:

[0096]

[0097] IP ni S represents the image neighborhood corresponding to the i-th predicted anchor point in the graph. ni For each anchor point in the image, x, y, and z represent the relative coordinates within that neighborhood. This section only presents one method for constructing the image fidelity term; other methods, such as the 1-norm and structural similarity, can also be used.

[0098] An energy function is generated based on the position error term and the image fidelity term.

[0099] Building deep learning networks (such as Figure 19 As shown, the energy function is used as the loss function of the deep learning network. The network is trained using the output image as input and the anchor point location information as output until the loss function is minimized, resulting in a well-trained deep learning network. The loss function is Loss:

[0100] Loss=PL1+λMSE

[0101] In the formula, PL1 is the position error term, λMSE is the image fidelity term, and λ is the weighting coefficient.

[0102] Step 103: Determine the rigid body transformation matrix based on the anchor point coordinates. The rigid body transformation matrix is ​​calculated using the coordinates of three anchor points, including the translation parameter (T). x ,T y ,T z The calculation process for the rotation matrix R is as follows:

[0103] (x1,y1,z1) are the coordinates of the apex of the heart, (x2,y2,z2) are the coordinates of the base of the left ventricle, and (x3,y3,z3) are the coordinates of the center of the right heart.

[0104] First, calculate the vector z from the apex of the heart to the base of the left ventricle.

[0105] z = (x2 - x1, y2 - y1, z2 - z1)

[0106] Secondly, draw a perpendicular line from the base of the right ventricle to the vector z at the base of the left ventricle, intersecting z to obtain the foot of the perpendicular O. Calculate the coordinates of point O (T). x ,T y ,T z ):

[0107] T x = (x2-x1)*t+x1

[0108] T y = (y2-y1)*t+y1

[0109] T z = (z2-z1)*t+z1

[0110] In the formula,

[0111] Finally, with point O as the origin, z as the z-axis of the minor axis coordinate system, the vector from the center of the right ventricle to point O as the x-axis of the minor axis coordinate system, and the cross product of the z-axis and x-axis as the y-axis of the minor axis coordinate system, the rotation matrix from the minor axis coordinate system to the abscissa coordinate system is obtained:

[0112]

[0113]

[0114] yold = zold × xold

[0115] R = [xold′ yold′ zold′]

[0116] In the formula, R is the rotation matrix, × is the vector cross product, ′ is the transpose symbol, and xold, yold, and zold are the unit vectors corresponding to the unit vectors in the minor axis coordinate system in the horizontal axis coordinate system.

[0117] Step 104: Transform the PET horizontal axis image to be repositioned into a PET short axis image based on rigid transformation parameters. Specifically, using rotation and translation parameters, the range of the heart in the short axis coordinate system is obtained. After sampling, the coordinates of the sample points are mapped to the horizontal axis coordinate system, and the pixel values ​​are obtained using interpolation, thus completing the generation of the short axis image.

[0118] Based on the above description, the implementation framework of the PET image repositioning method provided by the present invention is as follows: Figure 13 As shown.

[0119] This invention is based on a deep learning network, which automatically learns anchor point features and obtains positioning parameters through the anchor point positions, thereby realizing automatic relocalization of PET images.

[0120] In addition, corresponding to the PET image repositioning method provided above, the present invention also provides a PET image repositioning system, such as... Figure 20 As shown, the system includes:

[0121] Image acquisition module 200 is used to acquire the transverse axis image of the PET to be repositioned.

[0122] The image processing module 201 is used to preprocess the PET horizontal axis image to be repositioned to obtain a preprocessed image.

[0123] Anchor point detection module 202 is used to input the preprocessed image into the detection network to obtain anchor point coordinates. The detection network is a trained deep learning network.

[0124] The parameter determination module 203 is used to determine the rigid transformation parameters based on the anchor point coordinates.

[0125] Image positioning module 204 is used to transform the PET horizontal axis image to be repositioned into a PET short axis image based on rigid transformation parameters.

[0126] The image processing module 201 includes:

[0127] The image conversion unit is used to perform time averaging of image frames within a specific time period in the dynamic image when the PET horizontal axis image to be repositioned is a dynamic image, so as to obtain a static frame image.

[0128] The first processing unit is used to perform interpolation, filtering, and normalization processing on the static frame image to obtain a preprocessed image.

[0129] The second processing unit is used to perform interpolation, filtering and normalization processing on each frame of the static image when the PET horizontal axis image to be repositioned is a static image, to obtain a preprocessed image.

[0130] Furthermore, to further improve the accuracy of PET image repositioning, the PET image repositioning system provided by this invention also includes:

[0131] The sample acquisition module is used to acquire input image samples and preprocess the input image samples to obtain the output image.

[0132] The anchor point positioning module is used to manually position anchor points in the output image to obtain anchor point location information.

[0133] The standard generation module is used to generate network output gold standards based on anchor point location information.

[0134] The error term determination module is used to determine the position error term based on the gold standard output by the network.

[0135] The neighborhood selection module is used to select a neighborhood in the output image based on the anchor point position in the anchor point location information and the characteristics of the anatomical structure.

[0136] The fidelity term construction module is used to construct image fidelity terms based on neighborhood.

[0137] The function generation module is used to generate an energy function based on the position error term and the image fidelity term.

[0138] The network training module is used to train the deep learning network by taking the energy function as the loss function, the output image as the input, and the anchor point position information as the output, until the value of the loss function is minimized, thus obtaining the trained deep learning network.

[0139] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0140] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for repositioning PET images, characterized in that, include: Obtain the lateral image of the PET image to be repositioned; The PET horizontal axis image to be repositioned is preprocessed to obtain a preprocessed image; The preprocessed image is input into a detection network to obtain anchor point coordinates; the detection network is a trained deep learning network. The rigid transformation parameters are determined based on the anchor point coordinates; Based on the rigid transformation parameters, the PET horizontal axis image to be repositioned is transformed into a PET short axis image; The training process of the deep learning network includes: Obtain input image samples and preprocess the input image samples to obtain the output image; The output image is manually anchored to obtain anchor point position information; Generate a network output gold standard based on the anchor point location information; The position error term is determined based on the gold standard output by the network. In the output image, a neighborhood is selected based on the anatomical structure characteristics, with the anchor point in the anchor point location information as the center. Construct image fidelity terms based on the neighborhood; An energy function is generated based on the position error term and the image fidelity term; The energy function is used as the loss function of the deep learning network, and the output image is used as the input, while the anchor point position information is used as the output to train the deep learning network until the value of the loss function is minimized, thus obtaining the trained deep learning network.

2. The PET image repositioning method according to claim 1, characterized in that, The preprocessing of the PET horizontal axis image to be repositioned to obtain a preprocessed image specifically includes: When the PET horizontal axis image to be repositioned is a dynamic image, the image frames within a specific time period in the dynamic image are averaged over time to obtain a static frame image. The preprocessed image is obtained by performing interpolation, filtering, and normalization on the static frame image. When the PET horizontal axis image to be repositioned is a static image, the preprocessed image is obtained by interpolating, filtering and normalizing each frame of the static image.

3. The PET image repositioning method according to claim 1, characterized in that, The loss function is: : In the formula, For the position error term, For image fidelity items, These are the weighting coefficients.

4. The PET image repositioning method according to claim 1, characterized in that, The position error term is : In the formula, n To output the number of parameters in the gold standard for the network. P m For the first m The gold standard for predicting parameters. R m For the first m The gold standard corresponding to each parameter m =1,2,... n .

5. The PET image repositioning method according to claim 1, characterized in that, The image fidelity item is : IP ni For the first i Predict the image neighborhood corresponding to the anchor point in the graph, S ni For each anchor point, x, y, and z represent the neighborhood of the corresponding gold marker in the image, and x, y, and z are the relative coordinates within that neighborhood.

6. A PET image repositioning system, characterized in that, include: The image acquisition module is used to acquire the horizontal axis image of the PET to be repositioned; The image processing module is used to preprocess the PET horizontal axis image to be repositioned to obtain a preprocessed image; An anchor point detection module is used to input the preprocessed image into a detection network to obtain anchor point coordinates; the detection network is a trained deep learning network. The parameter determination module is used to determine the rigid transformation parameters based on the anchor point coordinates. The image positioning module is used to transform the PET horizontal axis image to be repositioned into a PET short axis image based on the rigid transformation parameters. The sample acquisition module is used to acquire input image samples and preprocess the input image samples to obtain an output image; Anchor point positioning module is used to manually position anchor points in the output image to obtain anchor point position information; The standard generation module is used to generate the network output gold standard based on the anchor point location information; An error term determination module is used to determine the position error term based on the gold standard output by the network. The neighborhood selection module is used to select a neighborhood in the output image based on the anatomical structure characteristics, with the anchor point in the anchor point location information as the center. A fidelity item construction module is used to construct an image fidelity item based on the neighborhood. The function generation module is used to generate an energy function based on the position error term and the image fidelity term; The network training module is used to train the deep learning network by taking the energy function as the loss function, the output image as the input, and the anchor point position information as the output, until the value of the loss function is minimized, thus obtaining the trained deep learning network.

7. The PET image repositioning system according to claim 6, characterized in that, The image processing module includes: The image conversion unit is used to perform time averaging on image frames within a specific time period in the dynamic image when the PET horizontal axis image to be repositioned is a dynamic image, to obtain a static frame image. The first processing unit is used to perform interpolation, filtering and normalization processing on the static frame image to obtain the preprocessed image; The second processing unit is used to perform interpolation, filtering and normalization processing on each frame of the static image when the PET horizontal axis image to be repositioned is a static image, to obtain the preprocessed image.

Images