An artificial intelligence automatic registration method for nuclear medicine cardiac images

Abstract

The application relates to the technical field of image processing, in particular to an artificial intelligence automatic registration method for nuclear medicine heart images, which comprises the following steps: obtaining nuclear medicine heart SPECT images and chest CT images to be registered, and inputting the images into corresponding heart positioning artificial intelligence models to obtain position parameters of a heart region of SPECT / CT images and generate SPECT / CT mask images containing only the heart region; inputting the SPECT mask images and the CT mask images into a pre-trained SPECT-CT heart registration artificial intelligence model to obtain registration translation parameters between the SPECT images and the CT images, which are used for registering the SPECT images and the CT images. Dynamic SPECT and CT, gated SPECT and CT and static SPECT and CT are automatically and accurately registered in the heart region.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and more specifically, to an artificial intelligence-based automatic registration method for nuclear medicine cardiac images. Background Technology

[0002] SPECT quantitative analysis of myocardial blood flow is increasingly being used in clinical practice. To ensure the accuracy of this quantitative analysis, the dynamic SPECT raw data must be quantitatively reconstructed so that the reconstructed SPECT image format is consistent with the PET image format, allowing for precise measurement of the dynamic changes in radioactivity concentration in the myocardium and blood pool over time, as well as the true radiation distribution of the myocardium under stable conditions. However, a necessary prerequisite for achieving quantitative SPECT reconstruction is accurate physical correction, including tissue attenuation correction using CT data. Before attenuation correction, precise registration of SPECT and CT images in the cardiac region must be ensured; this is a crucial step in guaranteeing the accuracy of dynamic SPECT images.

[0003] In current nuclear medicine SPECT tissue attenuation correction procedures, SPECT-CT registration primarily relies on manual registration in three-dimensional (3D) space by the operator. However, this manual method is highly subjective, and differences in operator experience can lead to varying degrees of registration deviation, severely impacting the accuracy of SPECT tissue attenuation correction. Although existing automated cardiac registration methods are mainly designed for PET and CT image features, the significant differences between SPECT and PET images, such as lower resolution, higher noise, and less information, mean that previous analytical algorithm-based PET-CT cardiac registration methods cannot be directly applied to SPECT-CT cardiac registration.

[0004] Furthermore, dynamic SPECT imaging is subject to various interference factors during image acquisition, and the extent of interference is related to the SPECT and SPECT / CT equipment used. For example, when scanning with the same SPECT / CT machine, the patient's body may move, or when using different SPECT and CT machines, the patient's heart may be located very far from the SPECT and CT locations. When performing dynamic SPECT imaging, the liver has very high uptake and is close to the heart, making it difficult to delineate the cardiac region. In addition, factors such as the different pixel sizes and FOVs used during SPECT data acquisition all pose a significant challenge to the cardiac registration of dynamic SPECT and CT.

[0005] Although gated SPECT myocardial perfusion imaging acquires images when myocardial uptake is in a stable state, and liver interference can be reduced by liver elution after drug injection, similar technical challenges as dynamic SPECT imaging still exist for cardiac registration between gated SPECT and CT, and between static SPECT and CT. Summary of the Invention

[0006] To address the aforementioned technical issues, this application provides an artificial intelligence-based automatic registration method for nuclear medicine cardiac images, enabling automatic and precise registration of cardiac regions between dynamic SPECT and CT, gated SPECT and CT, and static SPECT and CT.

[0007] An AI-based automated registration method for nuclear medicine cardiac images includes: Acquire nuclear medicine cardiac SPECT images and chest CT images to be registered; The SPECT image is input into a pre-trained SPECT cardiac localization artificial intelligence model to obtain the location parameters of the cardiac region in the SPECT image and generate a SPECT mask image containing only the cardiac region. The CT images are input into a pre-trained CT cardiac localization artificial intelligence model to obtain the location parameters of the heart region in the CT images and generate a CT mask image containing only the heart region. The SPECT mask image and CT mask image are input into a pre-trained SPECT-CT cardiac registration artificial intelligence model to obtain the registration translation parameters between the SPECT image and the CT image, which are used to register the SPECT image and the CT image.

[0008] In one possible implementation, after acquiring the nuclear medicine cardiac SPECT image and chest CT image to be registered, the method further includes preprocessing the SPECT image and CT image, including image size preprocessing and image contrast preprocessing, to crop or expand the SPECT image and CT image to a preset size, and to linearly map the voxel values ​​in the image to a fixed range.

[0009] In one possible implementation, the image size preprocessing includes: Read the DICOM image data and voxel size of SPECT and CT images, and crop or expand the SPECT and CT images to a preset size based on the image center; The cropped or expanded images are sampled into 3D images of a preset size using trilinear interpolation or cubic interpolation, resulting in SPECT and CT images that can be used as input to the corresponding artificial intelligence model.

[0010] In one possible implementation, the image contrast preprocessing includes: The maximum gray value is obtained from the gray-level histogram of the SPECT image, which has the largest number of voxels and is greater than a preset number. Replace voxels in the SPECT image with the maximum gray value if their voxel values ​​are greater than the maximum gray value; normalize the SPECT image by using the truncated maximum gray value as the normalized maximum value and the minimum voxel value in the SPECT image as the normalized minimum value.

[0011] In one possible implementation, the SPECT cardiac localization AI model and the CT cardiac localization AI model are three-dimensional convolutional neural network models, and the SPECT-CT cardiac registration AI model is a three-dimensional convolutional neural network model containing cross-attention blocks.

[0012] In one possible implementation, the method for training the SPECT / CT cardiac localization AI model and the SPECT-CT cardiac registration AI model includes: Acquire several SPECT and CT images and label them; wherein, the label content includes the cardiac region in the SPECT image, the cardiac region in the CT image, and the translational displacement from CT to SPECT during SPECT-CT cardiac registration; Acquire mask images of the cardiac region from the SPECT and CT images respectively; Data augmentation is performed on the SPECT images, CT images, and corresponding mask images to construct a training dataset for training the SPECT / CT cardiac localization AI model and the SPECT-CT cardiac registration AI model.

[0013] One possible implementation involves data augmentation of the SPECT image, CT image, and corresponding mask image, including: The original SPECT and CT images were translated randomly in three directions to obtain augmented data for training the SPECT cardiac localization AI model and the CT cardiac localization AI model. The SPECT mask image and the CT mask image are randomly translated a certain distance in three directions to obtain augmented data for training the SPECT-CT cardiac registration artificial intelligence model; The SPECT mask image and CT mask image were randomly scaled in three directions to obtain augmented data for training the SPECT-CT cardiac registration artificial intelligence model.

[0014] In one possible implementation, after training the SPECT / CT cardiac localization AI model and the SPECT-CT cardiac registration AI model, the method further includes: Calculate the consistency parameters and error parameters between the output parameters of the artificial intelligence model and the labeling parameters, including the coefficient of determination, mean absolute error, and Euclidean distance error; Artificial intelligence models that meet the evaluation criteria are selected as candidate models based on preset threshold standards; The registration results of the candidate models are visually evaluated to assess the overlap between the edges of the myocardium and lungs, as well as the overlap between the SPECT edge and the fat on the surface of the heart on CT. The best artificial intelligence model is selected as the final artificial intelligence model for application.

[0015] One possible implementation involves annotating SPECT and CT images to obtain the location parameters of the cardiac region in the SPECT / CT images, including the coordinates of the center of the cuboid and the length, width, and height of the cardiac region in the SPECT / CT images, or the coordinates of the center of the ellipsoid and the length and short longitudinal diameter.

[0016] One possible implementation involves annotating SPECT and CT images, including: The location of the heart in SPECT and CT images can be annotated in any order from the coronal, sagittal, and transverse planes using image annotation software, as well as the registration of the heart in SPECT and CT images can be performed in any order from the coronal, sagittal, and transverse planes.

[0017] Compared with the prior art, the technical solution provided in this application has the following beneficial effects: This application simultaneously and accurately outputs registration displacement parameters as well as the position and size parameters of the cardiac region in SPECT and CT scans. It achieves fast registration and overcomes the challenges of image size diversity caused by machine differences in different clinical environments, varying field of view (FOV), and different image matrix sizes. This allows the trained AI model to process images with arbitrary FOV and matrix sizes. Furthermore, this solution addresses the differences in SPECT image contrast and cardiac signal intensity caused by different acquisition protocols, ensuring stable output from the AI ​​model under varying contrast conditions. This effectively guarantees the accuracy of tissue attenuation correction in SPECT quantitative analysis. Attached Figure Description

[0018] Figure 1 This is a flowchart of an artificial intelligence-based automatic registration method for nuclear medicine cardiac images provided in Embodiment 1 of this application.

[0019] Figure 2This is a schematic diagram illustrating the annotation of the cardiac region on SPECT and CT images by a senior nuclear medicine imaging expert, as provided in Embodiment 1 of this application.

[0020] Figure 3 This is a schematic diagram of selecting a cuboid region or an ellipsoid region in a 3D image, as provided in Embodiment 1 of this application.

[0021] Figure 4 This is a schematic diagram illustrating the registration and annotation of SPECT and CT images by a senior nuclear medicine imaging expert, as provided in Embodiment 1 of this application.

[0022] Figure 5 This is a schematic diagram illustrating the generation of SPECT and CT mask images provided in Embodiment 1 of this application.

[0023] Figure 6 This is a schematic diagram illustrating data enhancement of the original SPECT and CT images and mask images provided in Embodiment 1 of this application.

[0024] Figure 7 This is a schematic diagram of the 3D network architecture of the SPECT cardiac localization artificial intelligence model provided in Embodiment 1 of this application.

[0025] Figure 8 This is a schematic diagram of the 3D network architecture of the SPECT-CT cardiac registration artificial intelligence model provided in Embodiment 1 of this application.

[0026] Figure 9 This is a schematic diagram of the cross-attention block provided in Embodiment 1 of this application.

[0027] Figure 10 This is a schematic diagram of specific quantitative data for verifying the performance of an artificial intelligence model, provided in Embodiment 1 of this application.

[0028] Figure 11 This is a schematic diagram showing the automatic registration effect of SPECT and CT images provided in Embodiment 2 of this application using the artificial intelligence model of this application.

[0029] Figure 12 This is a flowchart of an artificial intelligence-based automatic registration method for nuclear medicine cardiac images provided in Embodiment 3 of this application.

[0030] Figure 13 This is a schematic diagram illustrating image size preprocessing using images acquired by a SPECT device as data, as provided in Embodiment 3 of this application.

[0031] Figure 14 This is a schematic diagram of image contrast preprocessing for normalizing SPECT images, provided in Embodiment 3 of this application.

[0032] Figure 15This diagram illustrates the predicted registration translation parameters obtained by adding a different CT image to the SPECT image provided in Embodiment 3 of this application, based on the SPECT-CT registration artificial intelligence model.

[0033] Figure 16 This diagram illustrates a comparison between the segmentation results of the AI ​​model for SPECT heart region segmentation provided in Embodiment 3 of this application and the segmentation results of a senior nuclear medicine imaging expert, as well as a diagram illustrating the AI ​​model's identification of the heart region in SPECT images.

[0034] Figure 17 This is a schematic diagram comparing the segmentation results of the artificial intelligence model for segmenting the heart region in CT images with the segmentation results of a senior nuclear medicine imaging expert, as well as a schematic diagram of the artificial intelligence model identifying the heart region in CT images.

[0035] Figure 18 This is a schematic diagram comparing the registration results of the artificial intelligence model for SPECT-CT cardiac registration with the registration results of a senior nuclear medicine imaging expert, as provided in Embodiment 3 of this application, and a schematic diagram of the artificial intelligence model identifying the cardiac region in SPECT and CT images. Detailed Implementation

[0036] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0037] Example 1 See Figure 1 This is a flowchart of an artificial intelligence-based automatic registration method for nuclear medicine cardiac images provided in Embodiment 1 of this application. Figure 1 As shown, the specific implementation steps of the above method include: Step 101: Obtain several SPECT images and CT images, and label the SPECT images and CT images.

[0038] The labels for the SPECT and CT images include the cardiac region in the SPECT image, the cardiac region in the CT image, and the translational displacement from CT to SPECT during SPECT-CT cardiac registration.

[0039] In some embodiments, software capable of selecting rectangular or ellipsoidal regions in 3D images is used to annotate the cardiac region in SPECT and CT images. First, the cardiac region in the SPECT and CT images is outlined in the transverse plane, using rectangles to define the length and width of the rectangular region, or ellipses to define the major and minor axes of the ellipsoidal region in the transverse plane. Then, the cardiac region in the SPECT and CT images is outlined in the coronal or sagittal planes, using rectangles to define the height of the rectangular region, or ellipses to define the longitudinal axis of the ellipsoidal region. Of course, in actual operation, the order of outlining in the transverse, coronal, and sagittal planes during the selection process can be arbitrarily changed. Figure 2 As shown, the center coordinates and length, width, height or length and short longitudinal diameter of the cuboid or ellipsoidal region are finally obtained as the location parameters of the cardiac region in SPECT and CT images.

[0040] Among them, such as Figure 3 As shown, the aforementioned cuboid region contains 6 parameters. x , y , z , a , b , c . ( x , y , z ) is used to represent the coordinates of the center of the cuboid in the image. a , b , c () is used to represent the length, width, and height of a cuboid. It should be noted that... x , y , z The origin of the coordinate system can be the center of the image or any corner of the image. The coordinates of any vertex of the cuboid can also be used. Half the length, width, and height of the cuboid can also be used.

[0041] The aforementioned ellipsoidal region also contains 6 parameters. x , y , z , a , b , c . ( x , y , z ) is used to represent the coordinates of the ellipsoid center in the image, a , b , c () is used to represent the diameter of an ellipsoid along three directions. Among them, x , y , z The origin of the coordinate system can be the center of the image or any corner of the image. The coordinates of any vertex of the ellipsoid can also be used. The radius of the ellipsoid can also be used. Example Figure 3The origin of the coordinate axis is the upper left corner of the image, and a, b, c represent the length, width, and height of the cuboid or the major axis of the ellipsoid.

[0042] Using software capable of fusing SPECT and CT scans and parallel displacement for SPECT-to-CT or CT-to-SPECT registration, the registration translation displacement is calculated according to the annotations used during SPECT-CT cardiac registration. First, CT is transferred to SPECT in the transverse plane, then in the coronal plane, and finally in the sagittal plane. Finally, the registration is checked and fine-tuned on all three planes. Figure 4 As shown, the displacement in the three directions is the SPECT-CT cardiac registration displacement. Optionally, the order of registration on the three planes can be interchanged during actual operation. It is also possible to choose to register the SPECT image to the CT image.

[0043] Step 102: Obtain the mask images of the cardiac region from the SPECT and CT images respectively.

[0044] In the embodiments of this application, such as Figure 5 As shown, SPECT and CT images containing only the heart region are generated based on the labeled cuboid or ellipsoidal region of the heart, and used as mask images.

[0045] Step 103: Perform data augmentation on the above SPECT images, CT images, and corresponding mask images to construct a training dataset.

[0046] The aforementioned data augmentation includes increasing the amount of data used to train artificial intelligence models by translating the original image to obtain a new image. Alternatively, data augmentation can be performed through rotation, mirroring, or changing the shape and size of a mask. Any method that manipulates the original image to increase the amount of training data is considered data augmentation.

[0047] In this embodiment, data augmentation includes three methods. First, translation enhancement of the original image increases the amount of data used to train the SPECT and CT cardiac localization AI models. Specifically, the original SPECT and CT images are translated randomly in three directions, ensuring the integrity of the heart region in the image during translation. Simultaneously, the image labels are changed accordingly to correctly identify the heart's position in the translated image. Optionally, the random distance is approximately half the size of the entire image matrix. For example, in a 64×64×64 image matrix, the random translation range is ±32 pixels in each direction. Translation stops when the heart region is about to exceed the image range, satisfying the principle that the heart region is fully displayed in the image.

[0048] Second, the mask image is enhanced through translation, increasing the amount of data used to train the SPECT-CT registration AI model. Specifically, the mask image is randomly translated a certain distance in three directions, ensuring the integrity of the heart region in the image and preventing it from exceeding the image boundaries during the translation process. The image labels are changed accordingly, so that the labels represent the displacement of the mask image for SPECT-CT registration after translation.

[0049] Third, the mask size of the mask image is enhanced to increase the amount of data used to train the SPECT-CT registration AI model. Specifically, the mask is randomly scaled in three directions, randomly increasing or decreasing the area of ​​the image exposed in the mask image, while maintaining the integrity of the heart in the image during the scaling process. The labels do not need to be changed. For example... Figure 6 As shown in the image.

[0050] Step 104: Construct and train the SPECT cardiac localization AI model, the CT cardiac localization AI model, and the SPECT-CT registration AI model.

[0051] In this embodiment, original SPECT images, CT images, and original image translation enhancement images are used to train the SPECT cardiac localization AI model and the CT cardiac localization AI model. A registration AI model is trained using masked images, masked image translation enhancement images, and masked image size enhancement images.

[0052] When training the artificial intelligence model based on the aforementioned training dataset, the SPECT and CT images are first normalized, linearly mapping the voxel values ​​to a fixed interval, such as 0 to 1. Then, a high-performance computing processor is used as the training hardware, and an artificial intelligence model is built based on a deep learning framework. During training, Adam, SDGM, or RMSProp is selected as the optimizer, and weighted mean squared error, MSE, or MAE is chosen as the loss function to optimize the artificial intelligence model. Optionally, the aforementioned high-performance computing processors include Nvidia GPUs, AMD GPUs, or TPUs, and the aforementioned deep learning frameworks include PyTorch, Tensorflow, or MONAI.

[0053] The aforementioned SPECT cardiac localization AI model consists of three deep neural network AI models: 3D convolutional neural network, 3D ResNet (residual neural network), and 3D DenseNet (densely connected network). The 3D convolutional neural network is a neural network composed of an unlimited number of convolutional layers and an unlimited number of fully connected layers. 3D ResNet is a 3D residual neural network with an unlimited number of layers. 3D DenseNet is a 3D densely connected network with an unlimited number of layers. Figure 7As shown, there are 3D convolutional neural networks with 5 layers, 3D ResNet with 18 layers, and 3D DenseNet with 121 layers.

[0054] When building the above-mentioned SPECT cardiac localization artificial intelligence model, the input of the artificial intelligence model is set to a SPECT image with a certain matrix and pixel size, and the output is set to the coordinates of the center of the cuboid and the length, width and height of the cardiac region in the SPECT image, or the center coordinates and the long and short longitudinal axes of the ellipsoid.

[0055] In practical applications, for the SPECT image to be registered, the SPECT image is input into the trained SPECT cardiac localization AI model, which outputs SPECT cardiac region parameters. Then, based on these parameters, a SPECT mask image is generated.

[0056] The aforementioned CT cardiac localization AI model uses the same AI model architecture as the SPECT localization AI model. When building the AI ​​model, the input requirements are set to CT images with a certain matrix and pixel size, and the output is set to the coordinates of the center of a cuboid and its length, width, and height in the CT image, or the coordinates of the center of an ellipsoid and its major and minor longitudinal axes.

[0057] In practical applications, for the CT image to be registered, the CT image is input into the trained CT cardiac localization AI model, which outputs CT cardiac region parameters. Then, based on these CT cardiac region parameters, a CT mask image is generated.

[0058] Building the aforementioned SPECT-CT registration AI model involves constructing a 3D cross-attention convolutional neural network. This network consists of three 3D convolutional neural networks of unlimited layers, one cross-attention block, and one fully connected neural network of unlimited layers. Figure 8 As shown, this is a 3D cross-attention convolutional neural network composed of three three-layer convolutional neural networks and one two-layer fully connected neural network.

[0059] like Figure 9As shown, the aforementioned cross-attention block employs a cross-attention mechanism. The main input passes through two 1×1×N convolutional layers to obtain a feature map K and a feature map V (N can be 1 or more, representing single-head attention or multi-head attention, respectively). The sub-output passes through a convolutional layer of the same size to obtain a feature map Q. Matrix operations are performed on the transposes of feature maps Q and K, and then the resulting matrix is ​​multiplied by feature map V. Finally, the output matrix and the sub-input are concatenated together.

[0060] When building the AI ​​model, the input to the SPECT-CT registration AI model is set to a SPECT mask image with a certain matrix and pixel size, and a CT mask image with a certain matrix and pixel size. The output is set to the displacement in 3D after CT to SPECT registration. In some embodiments, the displacement in 3D after SPECT to CT registration can also be output based on the labeled information.

[0061] In practical applications, SPECT mask images and CT mask images are input into the SPECT-CT registration artificial intelligence model to obtain the predicted registration translation parameters.

[0062] Step 105: For the SPECT and CT images to be registered, based on the above-mentioned SPECT cardiac localization artificial intelligence model, CT cardiac localization artificial intelligence model and SPECT-CT registration artificial intelligence model, determine the registration translation parameters between the SPECT and CT images to be registered, so as to complete the registration of SPECT and CT images.

[0063] Example 2 Based on the above embodiment one, embodiment two of this application evaluates the performance of each cardiac localization artificial intelligence model and registration artificial intelligence model to obtain the best cardiac localization artificial intelligence model and registration artificial intelligence model.

[0064] Specifically, the model performance evaluation mentioned above includes parameter evaluation and visual evaluation. Parameter evaluation is used to assess the degree of difference between the registration AI model output and the label annotation, including consistency parameters and error parameters between the AI ​​model output values ​​and the label annotation values. For example, calculating the coefficient of determination between the AI ​​model output parameters and the label annotation parameters. R 2 , represented as:

[0065] Mean absolute error (MAE) is expressed as:

[0066] The Euclidean distance error is expressed as:

[0067] In the formula, This indicates the registration parameters for label annotation. This represents the registration parameters output by the artificial intelligence model. This represents the average value of the registration parameters used in the label annotation. N This indicates the total number of test data. x represents the difference between the translation parameters in the x-direction of the label annotation and the translation parameters in the x-direction output by the artificial intelligence model. y represents the difference between the translation parameters in the y-direction of the label annotation and the translation parameters in the y-direction output by the artificial intelligence model. z represents the difference between the translation parameters in the z-direction of the label annotation and the translation parameters in the z-direction output by the artificial intelligence model.

[0068] In a specific application scenario, the specific quantitative data and charts for verifying the performance of the artificial intelligence model are shown in Table 1 below. Figure 10 As shown in the image.

[0069] Table 1: Quantitative Data on the Performance of Artificial Intelligence Models

[0070] according to Figure 10 The average Euclidean distance error is 0.87 pixels.

[0071] In this embodiment of the application, a preset threshold is set as the evaluation standard, for example, the average R is selected. 2 AI models with an accuracy >0.8, mean absolute mean error <0.5, and mean Euclidean distance error <1 pixel constitute the group of excellent AI models.

[0072] For the SPECT-CT cardiac registration images registered using the aforementioned excellent model groups, senior nuclear medicine imaging experts conducted a visual evaluation of the registration effect, including: assessing the overlap of the edges of the myocardium and lungs, and the overlap of the SPECT edges with the CT cardiac surface fat, such as... Figure 11 As shown in the figure. A certain number of samples (e.g., >30) are evaluated to select the AI ​​model with the best visual performance as the optimal model, including the optimal SPECT cardiac localization AI model, the optimal CT cardiac localization AI model, and the optimal SPECT and CT cardiac registration AI model.

[0073] Example 3 For SPECT and CT images to be registered, Embodiment 3 of this application provides an artificial intelligence-based automatic registration method for nuclear medicine cardiac images. For example... Figure 12 As shown, the specific implementation steps of the above method include: Step 201: Obtain the nuclear medicine cardiac SPECT image and chest CT image to be registered.

[0074] In some embodiments, after acquiring the SPECT and CT images to be registered, the method further includes image size preprocessing. Specifically, the first step involves reading the DICOM image data and voxel sizes of the SPECT and CT images, and cropping or expanding them to a certain FOV (Field of View) based on the center of the SPECT and CT images, consistent with the FOV used when training the artificial intelligence model. When the original image FOV is greater than this FOV, the image is cropped; when the original image FOV is less than this FOV, the image is expanded with a 0 value. The second step involves sampling the cropped or expanded image into a three-dimensional image of a certain matrix size using trilinear interpolation (cubic interpolation or other methods can also achieve the same effect), consistent with the matrix size used when training the artificial intelligence model, ultimately obtaining the SPECT and CT images used as input to the artificial intelligence model. Figure 13 The example uses images acquired by the SPECT device as training data; however, the FOV and matrix size of other machines can also be used.

[0075] Furthermore, the grayscale value with the largest number of voxels in the grayscale histogram and a certain number of voxels is selected as the maximum grayscale value. Using this maximum grayscale value as the cutoff value, voxels in the SPECT image with voxel values ​​greater than the maximum grayscale value are replaced with the maximum grayscale value. The maximum grayscale value is then used as the maximum value for normalization, and the minimum voxel value in the SPECT image is used as the minimum value for normalization to normalize the SPECT image. It should be noted that other methods can also be used to adjust SPECT contrast to highlight myocardial signals, such as... Figure 14 As shown in the image.

[0076] Step 202: Input the above SPECT images into the pre-trained SPECT cardiac localization artificial intelligence model to obtain the location parameters of the cardiac region in the SPECT images and generate a SPECT mask image containing only the cardiac region.

[0077] Step 203: Input the above CT images into the pre-trained CT cardiac localization artificial intelligence model to obtain the positional parameters of the heart region in the CT images and generate a CT mask image containing only the heart region.

[0078] Step 204: Input the SPECT mask image and CT mask image into the pre-trained SPECT-CT cardiac registration artificial intelligence model to obtain the registration translation parameters between the SPECT image and the CT image, which are used to register the SPECT image and the CT image. For example... Figure 15 As shown in the image.

[0079] The above-mentioned scheme provided in this application is applicable to cardiac registration of SPECT myocardial imaging and chest CT images. It can be used for image fusion, quantitative SPECT tissue attenuation correction, and all other problems requiring SPECT and CT cardiac registration. The artificial intelligence model and application for SPECT cardiac localization and the artificial intelligence model and application for CT cardiac localization provided by this method can also be used for problems requiring automated segmentation of SPECT and CT cardiac regions. SPECT myocardial imaging includes early SPECT dynamic imaging, gated myocardial perfusion imaging, and myocardial perfusion imaging. CT images include CT images acquired simultaneously with SPECT-CT, CT images scanned from different machines, and other chest CT images whose initial acquisition intent was unrelated to SPECT. Furthermore, this application is applicable to cardiac registration of SPECT and CT images with different FOVs acquired by various devices, including general-purpose SPECT machines and dedicated cardiac machines.

[0080] Compared with the prior art, the technical solution provided in this application has the following beneficial effects: This method allows for the training of an artificial intelligence (AI) model for SPECT-CT cardiac registration, which can be applied to image processing workflows for SPECT-CT cardiac registration in various situations, as well as for SPECT and CT cardiac localization. The AI ​​model trained in this process not only accurately provides registration parameters but also the location and size of the cardiac region in both SPECT and CT images. The AI ​​registration model is accurate and fast, outputting registration parameters and the region parameters of the heart in the image within 2 seconds. Furthermore, the difference between its registration and that of senior nuclear medicine imaging experts is only within 5mm. Figure 16 , Figure 17 , Figure 18 As shown in the image.

[0081] This application overcomes the contradiction between the diverse image sizes caused by differences in machines, FOV, and image matrix sizes in different clinical environments and the fixed size input of the artificial intelligence model. A well-trained artificial intelligence model can register SPECT-CT cardiac images with arbitrary FOV and matrix size, such as... Figure 17 As shown in the image.

[0082] Furthermore, this application also addresses the problem that differences in tissue and organ contrast, particularly in cardiac signal intensity, caused by varying acquisition protocols, thus affecting the stable output of artificial intelligence (AI) models. A well-trained AI model can register SPECT images with varying cardiac contrast with CT scans, such as... Figure 18 As shown in the image.

[0083] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based automatic registration method for nuclear medicine cardiac images, characterized in that, include: Acquire nuclear medicine cardiac SPECT images and chest CT images to be registered; The SPECT and CT images are preprocessed, including image size preprocessing and image contrast preprocessing, to crop or expand the SPECT and CT images to a preset size and linearly map the voxel values ​​in the images to a fixed range. The image contrast preprocessing includes: obtaining the gray value with the largest number of voxels in the gray-level histogram of the SPECT image and a number greater than a preset number, as the maximum gray value; Replace the voxels in the SPECT image with the maximum gray value; use the maximum gray value as the normalized maximum value and the minimum voxel value in the SPECT image as the normalized minimum value to normalize the SPECT image. The SPECT image is input into a pre-trained SPECT cardiac localization artificial intelligence model to obtain the location parameters of the cardiac region in the SPECT image and generate a SPECT mask image containing only the cardiac region. The CT images are input into a pre-trained CT cardiac localization artificial intelligence model to obtain the positional parameters of the cardiac region in the CT images and generate a CT mask image containing only the cardiac region; the positional parameters include the coordinates of the center of the cuboid and the length, width and height of the cardiac region in the SPECT and CT images, or the coordinates of the center of the ellipsoid and the major and minor longitudinal axes. The SPECT mask image and CT mask image are input into a pre-trained SPECT-CT cardiac registration artificial intelligence model to obtain the registration translation displacement between the SPECT image and the CT image, which is used to register the SPECT image and the CT image. The method for training the SPECT cardiac localization AI model, the CT cardiac localization AI model, and the SPECT-CT cardiac registration AI model includes: Acquire several SPECT and CT images and label them; wherein, the label content includes the cardiac region in the SPECT image, the cardiac region in the CT image, and the translational displacement from CT to SPECT during SPECT-CT cardiac registration; Acquire mask images of the cardiac region from the SPECT and CT images respectively; Data augmentation is performed on the SPECT images, CT images, and corresponding mask images to construct a training dataset for training the SPECT cardiac localization AI model, the CT cardiac localization AI model, and the SPECT-CT cardiac registration AI model. Data augmentation is performed on the SPECT images, CT images, and corresponding mask images, including: The original SPECT and CT images were translated randomly in three directions to obtain augmented data for training the SPECT cardiac localization AI model and the CT cardiac localization AI model. The SPECT mask image and the CT mask image are randomly translated a certain distance in three directions to obtain augmented data for training the SPECT-CT cardiac registration artificial intelligence model; The SPECT mask image and CT mask image were randomly scaled in three directions to obtain augmented data for training the SPECT-CT cardiac registration artificial intelligence model.

2. The artificial intelligence automatic registration method for nuclear medicine cardiac images according to claim 1, characterized in that, The image size preprocessing includes: Read the DICOM image data and voxel size of SPECT and CT images, and crop or expand the SPECT and CT images to a preset size based on the image center; The cropped or expanded images are sampled into 3D images of a preset size using trilinear interpolation or cubic interpolation, resulting in SPECT and CT images that can be used as input to the corresponding artificial intelligence model.

3. The artificial intelligence automatic registration method for nuclear medicine cardiac images according to claim 1, characterized in that, The SPECT cardiac localization AI model and the CT cardiac localization AI model are three-dimensional convolutional neural network models, and the SPECT-CT cardiac registration AI model is a three-dimensional convolutional neural network model containing cross-attention blocks.

4. The artificial intelligence automatic registration method for nuclear medicine cardiac images according to claim 1, characterized in that, After training the SPECT cardiac localization AI model, the CT cardiac localization AI model, and the SPECT-CT cardiac registration AI model, the method further includes: Calculate the consistency parameters and error parameters between the output parameters of the artificial intelligence model and the labeling parameters, including the coefficient of determination, mean absolute error, and Euclidean distance error; Artificial intelligence models that meet the evaluation criteria are selected as candidate models based on preset threshold standards; The registration results of the candidate models are visually evaluated to assess the overlap between the edges of the myocardium and lungs, as well as the overlap between the SPECT edge and the fat on the surface of the heart on CT. The best artificial intelligence model is selected as the final artificial intelligence model for application.

5. The artificial intelligence automatic registration method for nuclear medicine cardiac images according to claim 1, characterized in that... Annotate the cardiac regions in SPECT and CT images, including: The location of the heart in SPECT and CT images can be annotated in any order from the coronal, sagittal, and transverse planes using image annotation software, as well as the registration of the heart in SPECT and CT images can be performed in any order from the coronal, sagittal, and transverse planes.

Images