Target heart diagram generation model training and generation method, system, device and storage medium

Abstract

The application provides a target heart map generation model training and generation method, system, device and storage medium, and belongs to the technical field of medical image recognition. The target heart map generation model training method comprises the following steps: acquiring a three-dimensional heart function image and a three-dimensional heart anatomical image; constructing a heart target heart map generation model, wherein the heart target heart map generation model comprises a heart short-axis position positioning network and a left ventricular myocardium segmentation network; comparing the three-dimensional heart anatomical image, and jointly training the heart short-axis position positioning network and the left ventricular myocardium segmentation network by using the three-dimensional heart function image to obtain a trained heart target heart map generation model. In the case that the left ventricular myocardium contour in the three-dimensional heart function image is incomplete, the three-dimensional heart function image and the three-dimensional heart anatomical image are compared, the training of the three-dimensional heart function image by the target heart map generation model is completed, and the automatic generation of the heart target heart map is realized.

Description

Technical Field

[0001] This invention relates to the field of medical image recognition technology, specifically to a method, system, device, and storage medium for training and describing a bullseye image generation model. Background Technology

[0002] Following acute myocardial infarction, myocardial reparative fibrosis occurs to stabilize the heart structure and prevent cardiac rupture. Activated myocardial fibroblasts can form reparative scar tissue at the infarct site. However, excessive myocardial fibrosis can lead to poor left ventricular remodeling, heart failure, and even heart failure in the long term, which is closely related to poor patient prognosis. Early detection of poor ventricular remodeling caused by myocardial fibrosis is crucial for preventing the progression of heart failure.

[0003] To facilitate timely diagnosis of cardiac conditions, scanning imaging is often used to identify lesions in the heart. Fibroblast activating protein (FAP) is a membrane-bound serine protease that is highly expressed in activated fibroblasts of damaged myocardium. 68 Ga-FAPI is 68 Ga-labeled fibroblast activating protein (FAP) inhibitors specifically bind to FAP, which is highly expressed on the surface of activated cardiomyocytes. As a state-of-the-art molecular imaging technique, gallium-68-labeled FAP inhibitors (… 68 Ga-FAPI functional imaging can visualize and quantitatively monitor activated cardiomyocytes.

[0004] Cardiac target mapping is an important tool for quantitative analysis of PET myocardial perfusion imaging. It provides a method for combining or comparing different myocardial images in a two-dimensional coordinate system. Specifically, based on the American Heart Association (AHA) method for left ventricular myocardial segmentation and nomenclature, namely the 17-segment cardiac segmentation model, the functional image of the left myocardium in the short-axis plane is projected onto the corresponding position in the two-dimensional target map. In this process, rotating the cardiac functional image from the transverse plane to the short-axis plane and completing the myocardial segmentation are crucial steps in generating the cardiac target map. Currently, some methods and commercial software can automatically generate target maps from myocardial perfusion images; however, these methods are mostly applicable to relatively complete PET / SPECT images of the myocardium, such as FDG PET imaging. However, due to… 68 Ga-FAPI functional imaging is a focal functional imaging technique that cannot capture the complete anatomical morphology of the heart. The outline of the left ventricular myocardium is incomplete. Current methods cannot achieve accurate short-axis localization of the heart and segmentation of the left ventricular myocardium. A large amount of manual correction is still required, which is very time-consuming and the accuracy cannot be guaranteed.

[0005] Therefore, it is necessary to provide training and generation methods, systems, devices, and storage media for bullseye image generation models to solve the above problems. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the prior art, the purpose of the present invention is to provide a method, system, device and storage medium for training and generating a target image generation model, so as to solve the technical problem that the target image model obtained by training cannot accurately locate the short axis of the heart and segment the left ventricular myocardium in the three-dimensional heart function image because the input three-dimensional heart function image cannot contain the complete heart anatomy.

[0007] To achieve the above and other related objectives, this invention provides a method for training a multimodal cardiac target map generation model, comprising the following steps:

[0008] Acquire three-dimensional images of cardiac function and three-dimensional images of cardiac anatomy;

[0009] A cardiac target map generation model is constructed, which includes a cardiac short-axis localization network and a left ventricular myocardial segmentation network.

[0010] By referring to the three-dimensional cardiac anatomical images, the three-dimensional cardiac function images are used to jointly train the cardiac short-axis localization network and the left ventricular myocardial segmentation network to obtain a trained cardiac target map generation model.

[0011] The cardiac target image generation model transforms the three-dimensional cardiac function image into a short-axis functional image, segments the left ventricular myocardial contour from the short-axis functional image, and projects the short-axis functional image into a cardiac target image based on the left ventricular myocardial contour.

[0012] In one embodiment of the present invention, the present invention also provides a multimodal cardiac target map generation method, wherein the multimodal cardiac target map generation method generates a cardiac target map using a multimodal cardiac target map generation model trained by the training method described in any of the above embodiments, and the multimodal cardiac target map generation method includes:

[0013] A three-dimensional cardiac function image is obtained, and the three-dimensional cardiac function image is input into the multimodal cardiac target image generation model;

[0014] The cardiac bullseye generation model transforms the three-dimensional cardiac function image into a short-axis functional image, segments the left ventricular myocardial contour from the short-axis functional image, and projects the short-axis functional image into a cardiac bullseye based on the left ventricular myocardial contour.

[0015] In one embodiment of the present invention, the present invention also provides a training system for a multimodal cardiac target map generation model, the system comprising:

[0016] The data acquisition unit is used to acquire three-dimensional cardiac function images and three-dimensional cardiac anatomy images;

[0017] The model building unit constructs a cardiac target map generation model, which includes a cardiac short-axis localization network and a left ventricular myocardial segmentation network.

[0018] Using the training unit and the three-dimensional cardiac anatomical images, the cardiac short-axis localization network and the left ventricular myocardial segmentation network are jointly trained using the three-dimensional cardiac function images to obtain a trained cardiac target map generation model.

[0019] In one embodiment of the present invention, a computer device is also provided, including a processor coupled to a memory, the memory storing program instructions, which, when executed by the processor, implement the multimodal cardiac target image generation model training method and the multimodal cardiac target image generation method described in any of the above embodiments.

[0020] In one embodiment of the present invention, a computer-readable storage medium is also provided, including a program that, when the program is run on a computer, causes the computer to execute the multimodal cardiac target map generation model training method and the multimodal cardiac target map generation method described in any of the above embodiments.

[0021] In this invention, the target image generation model training and generation method, system, device, and storage medium perform deep learning training on the multimodal cardiac target image generation model, enabling the multimodal cardiac target image generation model to be able to handle situations where the input three-dimensional cardiac function image does not contain the complete cardiac anatomy and the left ventricular myocardial contour is incomplete (e.g., 68 Ga-FAPI-labeled 3D cardiac PET images can be used to affinely transform 3D cardiac functional images into short-axis functional images by referring to the spatial relationship of 3D cardiac anatomy images with complete cardiac contours. At the same time, based on the left ventricular myocardial contour of the 3D anatomy image in the short axis, the left ventricular myocardial contour of the short-axis functional image is delineated, thereby projecting the short-axis functional image into a cardiac target image based on the accurate left ventricular myocardial contour.

[0022] In summary, the target image generation model training and generation method, system, equipment, and storage medium can train the target image generation model on three-dimensional cardiac function images even when the left ventricular myocardial contour is incomplete in the three-dimensional cardiac function images. This is achieved by comparing the three-dimensional cardiac function images with three-dimensional cardiac anatomy images. Compared with existing technologies, this training method reduces the involvement of manual correction, ensures the accuracy of cardiac short-axis positioning and segmentation precision, and realizes the automatic generation of cardiac target images for lesion marking functional imaging. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0024] Figure 1 This is a schematic diagram illustrating the process of generating a cardiac target map using a multimodal cardiac target map generation model in one embodiment of the present invention.

[0025] Figure 2 This is a schematic flowchart of a multimodal cardiac target map generation model training method in one embodiment of the present invention;

[0026] Figure 3 This is a flowchart illustrating step S1 in the multimodal cardiac target map generation model training method according to an embodiment of the present invention.

[0027] Figure 4 This is a flowchart illustrating step S3 in the multimodal cardiac target map generation model training method according to an embodiment of the present invention.

[0028] Figure 5 This is a flowchart illustrating step S31 in the multimodal cardiac target map generation model training method according to an embodiment of the present invention.

[0029] Figure 6 This is a flowchart illustrating step S311 in the multimodal cardiac target map generation model training method according to an embodiment of the present invention.

[0030] Figure 7 This is a flowchart illustrating step S313 in the multimodal cardiac target map generation model training method according to an embodiment of the present invention.

[0031] Figure 8 This is a flowchart illustrating step S32 in the multimodal cardiac target map generation model training method according to an embodiment of the present invention.

[0032] Figure 9 The diagram shown is a structural block diagram of a multimodal cardiac target map generation model training system according to an embodiment of the present invention.

[0033] Figure 10 The diagram shown is a structural schematic of a computer device according to an embodiment of the present invention.

[0034] Component designation explanation:

[0035] 10. Multimodal cardiac target map generation model training system; 11. Data acquisition unit; 12. Model construction unit; 13. Control training unit. Detailed Implementation

[0036] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other. It should also be understood that the terminology used in the embodiments of the present invention is for describing specific implementation schemes and not for limiting the scope of protection of the present invention. Test methods in the following embodiments that do not specify specific conditions are generally performed under conventional conditions or according to the conditions recommended by the respective manufacturers.

[0037] Please see Figures 1 to 10 It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are merely for illustrative purposes to aid those skilled in the art and to facilitate understanding. They are not intended to limit the scope of the invention and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effectiveness and purpose of the invention, should still fall within the scope of the technical content disclosed herein. Furthermore, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity and not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0038] When numerical ranges are given in the embodiments, it should be understood that, unless otherwise stated in the present invention, both endpoints of each numerical range and any value between the two endpoints may be selected. Unless otherwise defined, all technical and scientific terms used in this invention, as well as the prior art known to those skilled in the art and the description of this invention, may be implemented using any prior art methods, devices, and materials similar to or equivalent to those described, used, or made of materials in the embodiments of this invention.

[0039] In this specification, the three-dimensional cardiac anatomical image refers to an electromagnetic scanning imaging technique. This technique allows for scanning the human heart from different angles to obtain a complete anatomical image of the heart in three-dimensional space. The three-dimensional cardiac anatomical image can be a computed tomography (CT) scan or a magnetic resonance imaging (MRI) image. This invention uses MRI images as an example. MRI images utilize the principle of nuclear magnetic resonance, based on the different attenuations of released energy in different structural environments within a substance. By detecting the emitted electromagnetic waves through an applied gradient magnetic field, the location and type of atomic nuclei constituting the object can be determined, allowing for the creation of an internal structural image. MRI imaging technology is widely used in the clinical diagnosis of cardiovascular diseases. It is a non-invasive medical imaging technique for assessing the function and structure of the cardiovascular system. MRI imaging can scan the heart from multiple spatial angles, such as transverse, short-axis, and long-axis views, clearly and completely displaying the ventricular and myocardial contours of the heart in multimodal three-dimensional images.

[0040] Three-dimensional cardiac functional imaging is a molecular imaging technique that creates tomographic images of the heart using radiolabeled radionuclides or radiotracers to highlight localized lesions. These three-dimensional cardiac functional images can be positron emission tomography (PET) or single-photon emission computed tomography (SPECT), among other functional images. This invention uses PET images as an example. PET is a nuclear imaging technique (also known as molecular imaging) that can display metabolic processes in the body. The basis of PET imaging is the detection of gamma-ray pairs indirectly emitted by positron-emitting radionuclides (also known as radiopharmaceuticals, radionuclides, or radiotracers). The tracer is injected intravenously into a bioactive molecule, typically a sugar used for cellular energy. The sensitive detectors of the PET system capture gamma-ray radiation within the body, and software is used to triangulate the emission source, creating a three-dimensional computed tomographic image of the tracer concentration in the body.

[0041] Currently, functional imaging, when applied to cardiac imaging, can only scan transverse images of the heart. Furthermore, because functional imaging requires the injection of radioactive isotopes into the heart to label specific cardiac cardiomyocytes, functional images using specific radioactive isotopes for focal functional imaging of the heart cannot capture the complete cardiac anatomy, hindering further processing and analysis of three-dimensional cardiac functional images by current medical imaging software. For example, when using PET imaging to visualize lesions in myocardial repair fibrosis following acute myocardial infarction, the following methods are often employed... 68 Ga–FAPI labels activated myocardial fiber cells in the heart, but the three-dimensional cardiac PET images scanned under this radionuclide labeling cannot contain the complete outline of the left ventricular myocardium. Pre-medical imaging software has difficulty performing short-axis localization and left ventricular myocardium segmentation in three-dimensional cardiac PET images.

[0042] Therefore, the present invention provides a target image generation model training and generation method, system, device and storage medium. By utilizing the complete ventricular myocardial contour and multimodal images in three-dimensional cardiac anatomical images, the target image generation model is trained to locate and recognize three-dimensional cardiac functional images. This enables the trained target image generation model to complete accurate short-axis localization of the heart and segmentation of the left ventricular myocardium based on incomplete cardiac anatomical morphological information.

[0043] Please see Figure 1 and Figure 2 The target image generation model training and generation method provided by this invention uses the spatial relationship between each transverse section and short axis of the heart in a three-dimensional cardiac anatomical image (such as an MRI image or a CT image) to affine transform a three-dimensional cardiac functional image (such as a PET image or a SPECT image) into a short axis functional image. After the affine transformation, the left ventricular myocardial contour in the short axis functional image is identified and segmented by using the clear ventricular myocardial contour in the three-dimensional cardiac anatomical image. Finally, the identified and segmented left ventricular myocardial region in the short axis functional image is extracted, and the left ventricular myocardial region in the short axis functional image is projected onto the corresponding position of the two-dimensional plane target image according to the AHA-17 segmental cardiac segmentation model establishment method to generate a cardiac target image.

[0044] Please see Figure 2 In one embodiment of the present invention, the multimodal cardiac target map generation model training method includes the following process:

[0045] Step S1: Acquire three-dimensional cardiac function images and three-dimensional cardiac anatomy images; wherein, the three-dimensional cardiac function images can be three-dimensional cardiac PET images or three-dimensional cardiac SPECT images, and the three-dimensional cardiac anatomy images can be three-dimensional cardiac MRI images or three-dimensional cardiac CT images.

[0046] In some embodiments, the three-dimensional cardiac function image acquired in step S1 is a three-dimensional PET image obtained from a cardiac scan labeled with a radionuclide or radiotracer. In one example, the radiotracer is... 68 Ga-FAPI; the acquired three-dimensional cardiac anatomical images are MRI images of a three-dimensional cardiac cine sequence, which should be stored in chronological order within the cardiac cycle.

[0047] Specifically, such as Figure 3 As shown, step S1 includes the following steps:

[0048] S11. Obtain the corresponding three-dimensional cardiac function images and three-dimensional cardiac anatomy images;

[0049] In one example, the three-dimensional cardiac function image is a three-dimensional cardiac PET image, and the three-dimensional cardiac anatomy image is a three-dimensional cardiac MRI image; step S11 specifically involves reading the DICOM format three-dimensional cardiac PET image and the three-dimensional cardiac MRI image, and converting the DICOM format three-dimensional cardiac PET image and the three-dimensional cardiac MRI image into NII format files; wherein, the three-dimensional cardiac PET image is a processed... 68 Ga-FAPI-tagged transverse PET images of the heart, and the three-dimensional cardiac MRI images are three-dimensional cardiac cine sequence MRI images stored in chronological order within the cardiac cycle;

[0050] S12. Preprocess the acquired three-dimensional cardiac function images and three-dimensional cardiac anatomy images;

[0051] In the above example, step S12 specifically includes, for the three-dimensional cardiac PET image, cropping the... 68 The cardiac region in Ga-FAPI-tagged transverse PET images of the heart is used as a 3D cardiac PET image. The cropping area is not limited, but the heart position is kept as centered as possible in the image. The initial position information of the image is saved in the image header file, and the shearing and translation parameters of the affine transformation of the image are updated. For 3D cardiac MRI images, the end-diastolic MRI image of the heart is extracted from the 3D cardiac cine sequence MRI images and saved separately. Finally, the obtained 3D cardiac PET images and 3D cardiac MRI images are normalized respectively.

[0052] Step S2: Construct a cardiac target map generation model, which includes a cardiac short-axis localization network and a left ventricular myocardial segmentation network;

[0053] like Figure 1 As shown, the cardiac target map generation model consists of two parts: a cardiac short-axis localization network and a left ventricular myocardial segmentation network.

[0054] The cardiac short-axis localization network comprises cascaded convolutional layers, fully connected layers, and spatial transformation layers. Each convolutional layer includes a convolutional module consisting of a 3D convolutional layer and a downsampling layer, as well as multiple convolutional modules cascaded thereafter, each consisting of a 3D convolutional layer, a 3D normalization layer, a 3D max pooling layer, and a ReLU activation function layer. The cardiac short-axis localization network estimates the rotation angles and offsets in three directions through cascaded convolutional and fully connected layers. The output of the short-axis localization network predicts the rotation and offset parameters of the affine transformation, which have a size of 1×6 and are stored as a matrix parameter PT = [Prx, Pry, Prz, Ptx, Pty, Ptz]. The first three columns of parameters in the affine transformation feature represent the rotation angles (Prx, Pry, Prz) in the x, y, and z directions when the 3D cardiac function image is rotated towards the short axis, as estimated by the cardiac short-axis localization network. The last three columns represent the required offsets (Ptx, Pty, Ptz) in the x, y, and z directions after the 3D cardiac function image has been pre-rotated, as estimated by the cardiac short-axis localization network. The next step for the cardiac short-axis localization network is to read other pre-stored or preset affine transformation parameters (such as shearing and scaling parameters) to combine with the rotation and offset parameters to form the matrix parameters of the affine transformation feature. Finally, the cardiac short-axis localization network inputs the matrix parameters of the affine transformation features and the three-dimensional cardiac function image into the spatial transformation network to transform the three-dimensional cardiac function image and output a pre-transformed functional image predicting the short axis.

[0055] The left ventricular myocardial segmentation network is a dual-input channel 3D fully convolutional neural network. The left ventricular myocardial segmentation network consists of several convolutional modules with the same structure, including convolutional downsampling modules and convolutional upsampling modules. Each module contains 3D convolutional layers, normalization layers, and pooling layers. The number of convolutional downsampling modules is the same as that of convolutional upsampling modules, and the minimum number of convolutional downsampling modules and convolutional upsampling modules is 4. The maximum number is not limited and is determined by the actual training and testing results of the network, taking the number of modules that perform best on the corresponding dataset.

[0056] Step S3: By comparing the three-dimensional cardiac anatomical images, the cardiac short-axis localization network and the left ventricular myocardial segmentation network are jointly trained using the three-dimensional cardiac function images to obtain a trained cardiac target map generation model.

[0057] In step S3, the backpropagation algorithm is used to perform multi-stage iterative training on the cardiac short-axis localization network and the left ventricular myocardial segmentation network of the cardiac target map generation model to obtain the trained cardiac target map generation model.

[0058] The multi-stage iterative training includes: first, a first-stage training of the cardiac short-axis localization network; after the cardiac short-axis localization network is trained, a second-stage iterative training of the cardiac short-axis localization network and the left ventricular myocardial segmentation network is performed. In the second-stage iterative training, the training of the left ventricular myocardial segmentation network is accelerated and the training of the cardiac short-axis localization network is slowed down by setting the weights of different network loss functions. This stage of training mainly improves the left ventricular myocardial segmentation capability of the left ventricular myocardial segmentation network, while fine-tuning the parameters of the cardiac short-axis localization network.

[0059] Specifically, such as Figure 4 As shown, in some embodiments, the multi-stage iterative training of the cardiac short-axis localization network and the left ventricular myocardial segmentation network of the cardiac target map generation model in step S3 includes the following steps:

[0060] S31. By referring to the spatial transformation process from transverse to short axis in the three-dimensional cardiac anatomical image, the cardiac short axis localization network is trained to transform the three-dimensional cardiac function image to the short axis spatial transformation process.

[0061] S32. By comparing the left ventricular myocardial contours in the short axis of the three-dimensional cardiac anatomy image and the three-dimensional cardiac function image, the left ventricular myocardial segmentation network is trained to segment the left ventricular myocardial region in the short axis of the three-dimensional cardiac function image.

[0062] Furthermore, such as Figure 5 As shown, step S31 includes the following steps:

[0063] S311. Based on the spatial relationship between the transverse and short-axis views in the three-dimensional cardiac anatomical image, transform the corresponding three-dimensional cardiac function image to obtain the short-axis functional image of the three-dimensional cardiac function image.

[0064] Specifically, such as Figure 6 As shown, in some embodiments, step S311 includes the following steps:

[0065] S3111. Based on the spatial relationship between the transverse and short axis views in the three-dimensional cardiac anatomical image, obtain the affine transformation matrix from the transverse to the short axis views in the three-dimensional cardiac anatomical image.

[0066] The affine transformation matrix represents the rotation, scaling, shearing, and translation transformations of a 3D image from the transverse plane to the minor axis. The formula for the affine transformation matrix is ​​as follows:

[0067]

[0068] Specifically, the rotation, scaling, and shearing transformation parameters in the affine transformation matrix can be obtained from Equation 2, and the translation transformation parameters can be controlled by the d, h, and l parameters;

[0069]

[0070] Among them, the rotation transformation parameter Rotation in the affine transformation matrix matrix The rotation angles of the heart's minor axis relative to the transverse plane in the x, y, and z directions are extracted and denoted as rx, ry, and rz. The translation parameters d, h, and l in the affine transformation matrix are used to extract the translation amounts of the heart's minor axis relative to the transverse plane in the x, y, and z directions, denoted as tx, ty, and tz. These rotation and translation parameters are stored as parameters GT = [rx, ry, rz, tx, ty, tz], and then combined with the pre-stored shear transformation parameters Shear. matrix and scaling transformation parameter Zoom matrix Then the affine transformation matrix Affine can be constructed. matrix Since the short-axis rotation of the heart does not involve shearing transformation, the shearing transformation parameter is set to 1, while the scaling transformation parameter retains the scaling parameters of the original three-dimensional cardiac function image during the preprocessing stage.

[0071] S3112. Based on the affine transformation matrix, transform the three-dimensional cardiac function image to obtain the short-axis functional image of the three-dimensional cardiac function image.

[0072] S312. Extract the affine transformation features of the three-dimensional cardiac function image relative to the short axis through the cardiac short axis localization network, and transform the three-dimensional cardiac function image based on the affine transformation features to obtain a pre-transformed functional image.

[0073] S313. By comparing the transformation differences between the pre-transformed functional image and the short-axis functional image, the localization loss function of the cardiac short-axis localization network is obtained.

[0074] Specifically, such as Figure 7 As shown, in some embodiments, step S313 includes the following steps:

[0075] S3131. By comparing the parameter differences between the affine transformation features and the affine transformation matrix, a first loss function is obtained; wherein, the first loss function is the L1 distance between the parameter PT of the affine transformation features extracted by the cardiac short axis localization network and the parameter GT of the affine transformation matrix obtained based on the three-dimensional cardiac anatomical image, and is expressed by the formula L1(GT,PT).

[0076] S3132. By comparing the image differences between the pre-transformed functional image and the short-axis functional image, a second loss function is obtained; wherein, the second loss function is the L2 distance between the pre-transformed functional image and the short-axis functional image, expressed by the formula L2(SaxPET,PSaxPET), where SaxPET is the short-axis functional image and PSaxPET is the pre-transformed functional image.

[0077] S3133. Obtain the localization loss function based on the first loss function and the second loss function; the localization loss function is {L1(GT,PT),L2(SaxPET,PSaxPET)}.

[0078] S314. The cardiac short-axis localization network is trained in multiple stages based on the backpropagation algorithm to minimize the localization loss function and update and optimize the parameters of the cardiac short-axis localization network.

[0079] In training the cardiac short-axis localization network, the network is first trained using a first loss function, and then simultaneously trained using both the first and second loss functions. For example, in the first 50 training epochs, only the first loss function is used; from epochs 50 to 100, both the first and second loss functions are used. The weights of the two loss functions are 10 and 0.1, respectively, and the network's learning rate is 0.001.

[0080] Furthermore, such as Figure 8 As shown, step S32 includes the following steps:

[0081] S321. By comparing the three-dimensional cardiac anatomy image and the three-dimensional cardiac function image, the outline of the left ventricular myocardium in the short-axis functional image is drawn to generate a segmentation label for the left ventricular myocardium in the short-axis functional image.

[0082] Specifically, in some embodiments, step S321 includes the following steps: based on the short-axis functional image, interpolating the three-dimensional cardiac anatomy image in the long axis direction of the heart so that the interpolated anatomy image has the same resolution as the short-axis functional image, and registering the two three-dimensional images; based on the registered three-dimensional cardiac anatomy image in the short-axis ventricular myocardial image, outlining the left ventricular myocardial contour in the short-axis functional image, and generating a left ventricular myocardial segmentation label for the short-axis functional image, SegLVM.

[0083] S322. Input the left ventricular myocardial segmentation label SegLVM of the pre-transformed functional image and the short-axis functional image into the left ventricular myocardial segmentation network, and predict the left ventricular myocardial contour PSegLVM in the pre-transformed functional image through the left ventricular myocardial segmentation network.

[0084] S323. By comparing the left ventricular myocardial contour PSegLVM and the left ventricular myocardial segmentation label SegLVM, the segmentation loss function of the left ventricular myocardial segmentation network is obtained; wherein, the segmentation loss function includes the Dice coefficient loss function and the binary cross entropy loss function (BCE) between the left ventricular myocardial contour PSegLVM and the left ventricular myocardial segmentation label SegLVM, and the segmentation loss function is expressed as {Dice(SegLVM,PSegLVM),BCE(SegLVM,PSegLVM)}.

[0085] S324. The left ventricular myocardial segmentation network is iteratively trained based on the backpropagation algorithm to minimize the segmentation loss function and update and optimize the parameters of the left ventricular myocardial segmentation network.

[0086] In some embodiments, the multi-stage iterative training includes the following: In the first stage of training, the cardiac short-axis localization network is trained using only the first loss function of the localization loss function in the first 50 training rounds; in the 50th to 100th training rounds, the cardiac short-axis localization network is trained using the first and second loss functions of the localization loss function, with weights of 10 and 0.1 respectively, and a learning rate of 0.001 for the cardiac short-axis localization network; In the second stage of training, the cardiac short-axis localization network and the left ventricular myocardial segmentation network are trained using the localization loss function and the segmentation loss function, respectively, in the 100th to 200th training rounds. During training, the loss weights of the first and second loss functions in the localization loss function are both 0.1, and the loss weights of the Dice coefficient loss function and the cross-entropy loss function in the segmentation loss function are 10 and 5 respectively, with a learning rate of 0.001 for the cardiac short-axis localization network and a learning rate of 1e-5 for the left ventricular myocardial segmentation network. In the second stage of training, the left ventricular myocardial segmentation network is mainly trained, and only the parameters of the cardiac short-axis localization network are fine-tuned.

[0087] Please see Figure 1 In one embodiment of the present invention, a multimodal cardiac target map generation method is also provided. The multimodal cardiac target map generation method uses a multimodal cardiac target map generation model trained by the training method described in any of the above embodiments to process three-dimensional cardiac function images and generate cardiac target maps. The multimodal cardiac target map generation method includes the following steps:

[0088] A three-dimensional cardiac function image is obtained and input into the multimodal cardiac target map generation model. The three-dimensional cardiac function image is a transverse PET image of the heart labeled with a radionuclide or radioactive tracer. An exemplary radioactive tracer is...68 Ga–FAPI.

[0089] The cardiac target image generation model, based on the spatial information of a three-dimensional cardiac function image, predicts the matrix parameters of the affine transformation features of the three-dimensional cardiac function image relative to the short axis, specifically rotation and translation transformation parameters. It then uses these affine transformation features to spatially transform the three-dimensional cardiac function image to obtain the short axis functional image. After obtaining the short axis functional image, the corresponding three-dimensional cardiac anatomy image is retrieved and registered with it. Based on the registered three-dimensional cardiac anatomy image, the left ventricular myocardial contour in the short axis functional image is delineated to obtain a left ventricular myocardial segmentation label. Then, based on the left ventricular myocardial segmentation label, the left ventricular myocardial region in the short axis functional image is extracted. Following the AHA-17 segmental cardiac segmentation model establishment method, the three-dimensional functional left myocardial short axis image is projected onto the corresponding position of the two-dimensional planar target image to generate the cardiac target image.

[0090] In one embodiment of the present invention, a multimodal cardiac target map generation model training system 10 is provided, which corresponds one-to-one with the multimodal cardiac target map generation model training method in the above embodiments. For example... Figure 9 As shown, the multimodal cardiac target map generation model training system 10 includes a data acquisition unit 11, a model construction unit 12, and a control training unit 13. Detailed descriptions of each functional module are as follows:

[0091] Data acquisition unit 11 is used to acquire three-dimensional cardiac function images and three-dimensional cardiac anatomy images;

[0092] Model building unit 12 constructs a cardiac target map generation model, which includes a cardiac short-axis localization network and a left ventricular myocardial segmentation network.

[0093] Using the training unit 13 and the three-dimensional cardiac anatomical image, the cardiac short-axis localization network and the left ventricular myocardial segmentation network are jointly trained using the three-dimensional cardiac function image to obtain a trained cardiac target map generation model.

[0094] In one embodiment, the control training unit 13 is specifically used for:

[0095] Using the backpropagation algorithm, the cardiac short-axis localization network and the left ventricular myocardial segmentation network are trained in multiple stages to obtain a trained cardiac target map generation model.

[0096] The multi-stage iterative training includes: training the cardiac short-axis localization network to transform the three-dimensional cardiac function image to the short-axis position by referring to the spatial transformation process from the transverse to the short-axis position in the three-dimensional cardiac anatomy image; and training the left ventricular myocardial segmentation network to segment the left ventricular myocardial region of the three-dimensional cardiac function image in the short-axis position by referring to the left ventricular myocardial contour of the three-dimensional cardiac anatomy image and the three-dimensional cardiac function image in the short-axis position.

[0097] In one embodiment, the control training unit 13 is specifically used for:

[0098] Based on the spatial relationship between the transverse and short-axis views in the three-dimensional cardiac anatomical image, the corresponding three-dimensional cardiac function image is transformed to obtain the short-axis functional image of the three-dimensional cardiac function image.

[0099] The affine transformation features of the three-dimensional cardiac function image relative to the short axis are extracted by the cardiac short axis localization network, and the three-dimensional cardiac function image is transformed based on the affine transformation features to obtain a pre-transformed functional image;

[0100] The localization loss function of the cardiac short-axis localization network is obtained by comparing the transformation differences between the pre-transformed functional image and the short-axis functional image.

[0101] The cardiac short-axis localization network is trained in multiple stages using the backpropagation algorithm to minimize the localization loss function and update and optimize the parameters of the cardiac short-axis localization network.

[0102] In one embodiment, the control training unit 13 is specifically used for:

[0103] Based on the spatial relationship between the transverse and short axis views in the three-dimensional cardiac anatomical image, the affine transformation matrix from the transverse to the short axis views in the three-dimensional cardiac anatomical image is obtained.

[0104] Based on the affine transformation matrix, the three-dimensional cardiac function image is transformed to obtain the short-axis functional image of the three-dimensional cardiac function image.

[0105] In one embodiment, the control training unit 13 is specifically used for:

[0106] The first loss function is obtained by comparing the parameter differences between the affine transformation features and the affine transformation matrix;

[0107] The second loss function is obtained by comparing the image differences between the pre-transformed functional image and the short-axis functional image;

[0108] The localization loss function is obtained based on the first loss function and the second loss function.

[0109] In one embodiment, the control training unit 13 is specifically used for:

[0110] By comparing the three-dimensional cardiac anatomy image and the three-dimensional cardiac function image, the outline of the left ventricular myocardium in the short-axis functional image is delineated, and a segmentation label for the left ventricular myocardium in the short-axis functional image is generated.

[0111] The left ventricular myocardial segmentation labels of the pre-transformed functional image and the short-axis functional image are input into the left ventricular myocardial segmentation network, and the left ventricular myocardial contour in the pre-transformed functional image is predicted by the left ventricular myocardial segmentation network.

[0112] The segmentation loss function of the left ventricular myocardial segmentation network is obtained by comparing the left ventricular myocardial contour and the left ventricular myocardial segmentation label.

[0113] The left ventricular myocardial segmentation network is iteratively trained using the backpropagation algorithm to minimize the segmentation loss function and update and optimize the parameters of the left ventricular myocardial segmentation network.

[0114] In one embodiment, a computer device is provided, the internal structure of which can be shown in the following diagram. Figure 10 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the functions or steps of a dialog mode optimization method.

[0115] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0116] Acquire three-dimensional images of cardiac function and three-dimensional images of cardiac anatomy;

[0117] A cardiac target map generation model is constructed, which includes a cardiac short-axis localization network and a left ventricular myocardial segmentation network.

[0118] By referring to the three-dimensional cardiac anatomical images, the three-dimensional cardiac function images are used to jointly train the cardiac short-axis localization network and the left ventricular myocardial segmentation network to obtain a trained cardiac target map generation model.

[0119] The cardiac target image generation model transforms the three-dimensional cardiac function image into a short-axis functional image, segments the left ventricular myocardial contour from the short-axis functional image, and projects the short-axis functional image into a cardiac target image based on the left ventricular myocardial contour.

[0120] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0121] Acquire three-dimensional images of cardiac function and three-dimensional images of cardiac anatomy;

[0122] A cardiac target map generation model is constructed, which includes a cardiac short-axis localization network and a left ventricular myocardial segmentation network.

[0123] By referring to the three-dimensional cardiac anatomical images, the three-dimensional cardiac function images are used to jointly train the cardiac short-axis localization network and the left ventricular myocardial segmentation network to obtain a trained cardiac target map generation model.

[0124] The cardiac target image generation model transforms the three-dimensional cardiac function image into a short-axis functional image, segments the left ventricular myocardial contour from the short-axis functional image, and projects the short-axis functional image into a cardiac target image based on the left ventricular myocardial contour.

[0125] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0126] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0127] In this invention, the target image generation model training and generation method, system, device, and storage medium perform deep learning training on the multimodal cardiac target image generation model, enabling the multimodal cardiac target image generation model to be able to handle situations where the input three-dimensional cardiac function image does not contain the complete cardiac anatomy and the left ventricular myocardial contour is incomplete (e.g., 68 Ga-FAPI-labeled 3D cardiac PET images can be used to affinely transform 3D cardiac functional images into short-axis functional images by referring to the spatial relationship of 3D cardiac anatomy images with complete cardiac contours. At the same time, based on the left ventricular myocardial contour of the 3D anatomy image in the short axis, the left ventricular myocardial contour of the short-axis functional image is delineated, thereby projecting the short-axis functional image into a cardiac target image based on the accurate left ventricular myocardial contour.

[0128] In summary, the target image generation model training and generation method, system, equipment, and storage medium can train the target image generation model on three-dimensional cardiac function images even when the left ventricular myocardial contour is incomplete in the three-dimensional cardiac function images. This is achieved by comparing the three-dimensional cardiac function images with three-dimensional cardiac anatomy images. Compared with existing technologies, this training method reduces the involvement of manual correction, ensures the accuracy of cardiac short-axis positioning and segmentation precision, and realizes the automatic generation of cardiac target images for lesion marking functional imaging.

[0129] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A training method for a multimodal cardiac target map generation model, characterized in that, The method comprises the following steps: obtaining a three-dimensional cardiac functional image and a three-dimensional cardiac anatomical image; constructing a cardiac target heart map generation model, the cardiac target heart map generation model comprising a cardiac short-axis positioning network and a left ventricular myocardial segmentation network; training the cardiac short-axis positioning network and the left ventricular myocardial segmentation network jointly by using the three-dimensional cardiac functional image in the three-dimensional cardiac anatomical image to obtain a trained cardiac target heart map generation model; the joint training comprises: using a back propagation algorithm to perform multi-stage iterative training on the cardiac short-axis positioning network and the left ventricular myocardial segmentation network to obtain a trained cardiac target heart map generation model; wherein the multi-stage iterative training comprises: training the process of transforming the three-dimensional cardiac functional image to short-axis position by the cardiac short-axis positioning network in the spatial transformation process from transverse position to short-axis position in the three-dimensional cardiac anatomical image; training the process of segmenting the left ventricular myocardial region of the three-dimensional cardiac functional image in short-axis position by the left ventricular myocardial segmentation network in the left ventricular myocardial contour of the three-dimensional cardiac functional image and the three-dimensional cardiac anatomical image in short-axis position; wherein the training the process of transforming the three-dimensional cardiac functional image to short-axis position by the cardiac short-axis positioning network in the spatial transformation process from transverse position to short-axis position in the three-dimensional cardiac anatomical image comprises: transforming the corresponding three-dimensional cardiac functional image based on the spatial relationship between the transverse position and the short-axis position in the three-dimensional cardiac anatomical image to obtain a short-axis functional image of the three-dimensional cardiac functional image; extracting the affine transformation features of the three-dimensional cardiac functional image relative to the short-axis position by the cardiac short-axis positioning network, and transforming the three-dimensional cardiac functional image based on the affine transformation features to obtain a pre-transformation functional image; obtaining a positioning loss function of the cardiac short-axis positioning network by comparing the transformation differences between the pre-transformation functional image and the short-axis functional image; updating and optimizing the parameters of the cardiac short-axis positioning network by minimizing the positioning loss function based on the multi-stage iterative training of the cardiac short-axis positioning network by the back propagation algorithm. 2.The method of claim 1, wherein, the transforming the corresponding three-dimensional cardiac functional image based on the spatial relationship between the transverse position and the short-axis position in the three-dimensional cardiac anatomical image to obtain a short-axis functional image of the three-dimensional cardiac functional image comprises: obtaining an affine transformation matrix from the transverse position to the short-axis position in the three-dimensional cardiac anatomical image based on the spatial relationship between the transverse position and the short-axis position in the three-dimensional cardiac anatomical image; transforming the three-dimensional cardiac functional image based on the affine transformation matrix to obtain a short-axis functional image of the three-dimensional cardiac functional image. 3.The method of claim 2, wherein, the obtaining a positioning loss function of the cardiac short-axis positioning network by comparing the transformation differences between the pre-transformation functional image and the short-axis functional image comprises: obtaining a first loss function by comparing the parameter differences between the affine transformation features and the affine transformation matrix; obtaining a second loss function by comparing the image differences between the pre-transformation functional image and the short-axis functional image; obtaining the positioning loss function based on the first loss function and the second loss function. 4.The method of claim 1, wherein, The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including:

5. A method of multi-modality cardiac target cardiogram generation, the method comprising: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: 6.A training system of a multi-modal cardiac target cardiogram generation model, characterized in that, The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of the three-dimensional cardiac anatomical image and the three-dimensional cardiac functional image, including: The process of segmenting the left ventricular myocardium region of the three-dimensional cardiac functional image in the short axis position by the left ventricular myocardium segmentation network is trained by comparing the left ventricular myocardium contours of Based on the spatial relationship of the cross-sectional view and the short-axis view in the three-dimensional heart anatomical image, a corresponding three-dimensional heart functional image is transformed to obtain a short-axis view functional image of the three-dimensional heart functional image; an affine transformation feature of the three-dimensional heart functional image relative to the short-axis view is extracted through the heart short-axis positioning network, and a pre-transformation functional image is obtained by transforming the three-dimensional heart functional image based on the affine transformation feature; a positioning loss function of the heart short-axis positioning network is obtained by comparing the transformation difference between the pre-transformation functional image and the short-axis view functional image; and the heart short-axis positioning network is iteratively trained in multiple stages based on a back propagation algorithm, so as to minimize the positioning loss function and update and optimize the parameters of the heart short-axis positioning network.

7. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 5.

8. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 5.

Images