Deep Learning-Based Heart Image Reconstruction Method and System

Abstract

The present invention provides a method and system for cardiac image reconstruction based on deep learning, including: selecting the hearts of multiple subjects for slice scanning and imaging, obtaining three images for each slice, including a normal image, an image in the AP direction, and an image in the FH direction, using the scanned images as a training dataset, and the scanned images including clear images and blurred images; using a ResNet model to classify the images obtained from cardiac scanning, and then using multiple SRN-Deblur sub-models to perform deblurring operations on the blurred images in different directions; using the images in the AP direction and the images in the FH direction to calculate the direction and magnitude of the blood flow vector in three-dimensional space for measuring the deblurring effect of the simulated blurred images. Using velocity-encoded magnetic resonance imaging for four-dimensional flow magnetic resonance imaging has great potential in cardiovascular blood flow analysis, and deep learning can reconstruct defective images and eliminate motion blur in velocity-encoded magnetic resonance imaging.

Description

Technical Field

[0001] The present invention relates to the field of image processing and reconstruction, and particularly to a method and system for cardiac image reconstruction based on deep learning. Background Art

[0002] Magnetic Resonance Image (MRI) technology for the heart can provide high-resolution cardiac soft tissue images in a non-invasive manner to obtain anatomical information of the subject's heart. Doctors can obtain decision-making information required in the diagnosis and treatment of heart diseases and their pathological analysis through cardiac MRI. However, due to reasons such as the accuracy limitation of imaging equipment and the difficulty of patient cooperation, inevitably, low-quality MRIs are produced, such as low resolution, motion blur, etc. Therefore, the automation of cardiac MRI reconstruction is of great significance clinically. Summary of the Invention

[0003] The present invention provides a method and system for cardiac image reconstruction based on deep learning, aiming to solve the problem that in velocity-encoded magnetic resonance imaging, defective images can be reconstructed through deep learning to eliminate motion blur.

[0004] To achieve the above object, the present invention provides a method for cardiac image reconstruction based on deep learning, including:

[0005] Step 1, select the hearts of multiple subjects for slice scanning and imaging. Each slice obtains three images, including a normal image, an image in the AP direction, and an image in the FH direction. The images obtained by scanning are used as a training data set, and the images obtained by scanning include clear images and blurred images;

[0006] Step 2, use the ResNet model to classify the images obtained by cardiac scanning, and then use multiple SRN-Deblur sub-models to perform deblurring operations on the blurred images in different directions;

[0007] Step 3, use the images in the AP direction and the images in the FH direction to calculate the direction and magnitude of the blood flow vector in three-dimensional space for measuring the deblurring effect of the simulated blurred images.

[0008] Wherein, the specific content of the step 1 includes:

[0009] The slice scanning and imaging adopts magnetic resonance imaging technology, is performed through the atrial short-axis direction, adopts retrospective gating, and each slice has 25 phases or time frames;

[0010] The magnetic resonance imaging parameters include: echo time TR: 47.1 ms, repetition time TE: 1.6 ms, field of view FOV: (298340) mm2, (134256) mm2 pixel matrix, in-plane resolution of 1.54 mm / pixel, determined by pixel pitch, and through-plane resolution of 6 mm based on slice interval.

[0011] Among them, step 2 includes:

[0012] Determine the blurring direction of the image, classify the training images required for the deblurring model in the blurring direction, and feed the classified images back to the corresponding deblurring sub-model for training;

[0013] During training, use the cross-entropy function as the loss function and set the number of epochs to 50.

[0014] Among them, step 3 includes:

[0015] Calculating the direction and magnitude of the blood flow vector in three-dimensional space requires calculating the absolute value of the pixel difference and the distance of the vector; where FHG, FHB, APG, APB represent FH ground truth, FH blurred image, AP ground truth, AP blurred image respectively, and i, j represent the position of the image;

[0016]

[0017] ω PSNR Calculate using the following method, where MAX represents the sum of the maximum vector distances in the useful area;

[0018]

[0019] Since there is no blurred and clear mapping pair in true slice scanning imaging, two scans at different heartbeats at the same moment need to be compared. At the same moment, the scans of the two heartbeats are also different. Use ω PSNR , instead, compare with vorticity. The mathematical expression of two-dimensional vorticity is as follows:

[0020]

[0021] Among them, the positive and negative signs of ω have different meanings. Among them, the positive value represents a CCW cycle, the negative value represents the fluid rotating clockwise CW, and the magnitude of the value represents the rotation speed;

[0022] Calculate the circulation Γ using the line integral of the CCW closed loop C and write it in the form of an area integral as follows

[0023]

[0024] The present invention also provides a cardiac image reconstruction system based on deep learning, including:

[0025] A dataset acquisition module, which is used to select the hearts of multiple subjects for slice scanning imaging. Each slice obtains three images, including a normal image, an image in the AP direction, and an image in the FH direction. The scanned images are used as the training dataset, and the scanned images include clear images and blurred images;

[0026] An image processing module, which is used to classify the images obtained by cardiac scanning using the ResNet model, and then use multiple SRN-Deblur sub-models to perform deblurring operations on the blurred images in different directions;

[0027] An evaluation module, which is used to calculate the direction and magnitude of the blood flow vector in the three-dimensional space using the images in the AP direction and the images in the FH direction, and is used to measure the deblurring effect of the simulated blurred images.

[0028] The above solution of the present invention has the following beneficial effects:

[0029] The cardiac image reconstruction method and system based on deep learning provided by the embodiments of the present invention use an evaluation criterion consistent with PC MRI. Using the vortex map vector distance in the atrial region as the evaluation criterion is more persuasive and performs better in both visual detection and mathematical evaluation. After removing the blur factor, VENC MRI can help radiologists and clinicians make better clinical judgments and improve the diagnostic accuracy. The ResNet model is used for blur classification, and the classification accuracy rate exceeds 99%.

[0030] Other beneficial effects of the present invention will be described in detail in the subsequent specific implementation part. Brief Description of the Drawings

[0031] Figure 1 It is a flowchart of the cardiac image reconstruction method based on deep learning of the present invention;

[0032] Figure 2 It is a histogram of vorticity measurement and its calculation diagram of the present invention;

[0033] Figure 3 It is a structural diagram of the blur classification model of the present invention;

[0034] Figure 4 It is a diagram of the ResNet training process for blur image classification of the present invention;

[0035] Figure 5 It is a visual comparison diagram of the deblurring results of different models and low-quality images of the present invention;

[0036] Figure 6 This is the true VENC MRI deblurring visualization result diagram of the present invention. Detailed implementation manners

[0037] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0039] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "installed", "connected", "connected" should be understood in a broad sense. For example, it can be a locking connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0040] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0041] Such as Figure 1As shown in the figure, an embodiment of the present invention provides a cardiac image reconstruction method based on deep learning, including: Step 1, select the hearts of multiple subjects for slice scanning imaging. Each slice obtains three images, including a normal image, an image in the AP direction, and an image in the FH direction. The images obtained by scanning are used as the training dataset, and the images obtained by scanning include clear images and blurred images; Step 2, use the ResNet model to classify the images obtained by cardiac scanning, and then use multiple SRN-Deblur sub-models to perform deblurring operations on the blurred images in different directions; Step 3, use the images in the AP direction and the images in the FH direction to calculate the direction and magnitude of the blood flow vector in three-dimensional space for measuring the deblurring effect of the simulated blurred images.

[0042] Among them, Step 1 specifically includes: The slice scanning imaging uses magnetic resonance imaging technology, which is performed in the short axis direction of the atrium and uses retrospective gating. Each slice has 25 phases or time frames; The magnetic resonance imaging parameters include: echo time TR: 47.1 ms, repetition time TE: 1.6 ms, field of view FOV: (298 340) mm2, (134 256) mm2 pixel matrix, in-plane resolution of 1.54 mm / pixel, determined by the pixel pitch, and the through-plane resolution based on the slice interval is 6 mm.

[0043] Among them, Step 2 includes: Determine the blur direction of the image, classify the training images required for the deblurring model in the blur direction, and feed the classified images back to the corresponding deblurring sub-model for training; During the training, use the cross-entropy function as the loss function and set the epoch to 50.

[0044] Among them, Step 3 includes: The calculation of the direction and magnitude of the blood flow vector in three-dimensional space requires calculating the absolute value of the difference between pixels and the distance of the vector; Among them, FHG, FHB, APG, APB respectively represent FH ground truth, FH blurred image, AP ground truth, AP blurred image, and i, j represent the positions of the images;

[0045]

[0046] ω PSNR Calculate using the following method, where MAX represents the sum of the maximum vector distances in the useful area;

[0047]

[0048] Since there are no blurred and clear mapping pairs in true slice-scan imaging, two scans at different heartbeat cycles at the same moment need to be used for comparison. At the same moment, the scans of the two heartbeat cycles are also different. Using ω PSNR , instead, vorticity is used for comparison. The mathematical expression of two-dimensional vorticity is as follows:

[0049]

[0050] Among them, the positive and negative signs of ω have different meanings. The positive value represents the CCW cycle, the negative value represents the clockwise CW rotation of the fluid, and the magnitude of the value represents the rotation speed;

[0051] The circulation Γ is calculated using the line integral of the CCW closed loop C and written in the form of an area integral as follows

[0052]

[0053] The present invention also provides a cardiac image reconstruction system based on deep learning, including: a data set acquisition module for selecting the hearts of multiple subjects for slice-scan imaging. Each slice obtains three images, including a normal image, an image in the AP direction, and an image in the FH direction. The images obtained by the scan are used as the training data set, and the images obtained by the scan include clear images and blurred images; an image processing module for classifying the images obtained by the cardiac scan using the ResNet model, and then using multiple SRN-Deblur sub-models to perform deblurring operations on the blurred images in different directions; an evaluation module for calculating the direction and magnitude of the blood flow vector in three-dimensional space using the images in the AP direction and the images in the FH direction to measure the deblurring effect of the simulated blurred images.

[0054] The present invention uses velocity-encoded magnetic resonance imaging for four-dimensional flow magnetic resonance imaging, which has great potential in cardiovascular blood flow analysis. However, sometimes patients dynamically adjust their heart positions during imaging, resulting in motion blur. Such blurred images lead to inaccurate analysis. Deep learning can reconstruct defective images and eliminate motion blur in velocity-encoded magnetic resonance imaging. A flow image reconstruction model based on SRN-Deblur (a model of Tencent Youtu team accepted by CVPR in 2018) is proposed, and the accuracy of flow analysis is evaluated.

[0055] In actual medical image processing, low-quality low-resolution (LR) images have too few texture details, which is not conducive to the accuracy of heart disease diagnosis. With super-resolution, there is a relevant mapping relationship between low-resolution images and high-resolution images. If these mapping relationships can be learned by training a large number of images through a deep learning model, then the low-resolution image can be used to reconstruct a true high-resolution image. Using a three-layer convolutional neural network (CNN) to learn the mapping relationship between low-resolution images and high-resolution images, high-quality images can be obtained. However, when MSE is used as a loss function, when the image input resolution is large, the high-frequency texture detail features of the image will be lost.

[0056] Therefore, the LapSRN method is used to achieve more stable and efficient model training, thereby providing higher perceptual quality for super-resolution of MRI images. Based on the SR architecture and Laplace pyramid structure, a new LapSRN model is used, which enables the network to perform super-resolution processing on the original low-resolution and noisy MRI images, and select the resolution of the original high-resolution MRI.

[0057] During the imaging process, cardiac MRI imaging equipment requires patients to hold their breath, which is difficult for some patients to do. Similarly, the blur caused by camera motion also makes some originally good photos look bad. There are many deep learning methods for removing motion blur from ordinary images, such as SRN-Deblur, a model of the Tencent Youtu team included in CVPR in 2018.

[0058] The model uses a Laplacian pyramid framework to build a network. The model takes the LR image as input (rather than an upscaled version of the LR image) and gradually predicts the residual image at the log2Spyramid level, which is the upsampling scale factor. For example, the network consists of 3 pyramid levels for super-resolving LR images with a scale factor of 8. The model consists of two branches: (1) feature extraction and (2) image reconstruction.

[0059] Among them, the feature extraction branch consists of (1) a feature embedding sub-network for embedding high-dimensional non-linear feature maps, (2) a transposed convolutional layer for upsampling the extracted features by a factor of 2, and (3) a convolutional layer (Conv_res) for predicting the sub-band residual image. The first pyramid level has an additional convolutional layer (Conv_in) to extract high-dimensional feature maps from the input LR image. At other levels, the feature embedding sub-network can directly transform the features in the high-level feature maps of the previous pyramid level. In the image reconstruction branch, at each level, the input image is upsampled by a factor of 2 through a transposed convolutional layer and initialized with a 4×4 bilinear kernel. Then, the upsampled image (using element-wise summation) is combined with the predicted residual image to generate a high-resolution output image. Then, the reconstructed HR image is used as the input to the +1 level image reconstruction branch. The entire network is a cascaded CNN with the same structure at each level. The upsampling layer is jointly optimized with all other layers to better learn the upsampling function.

[0060] VENC MRI is the result of cardiac slice scanning, and three images are obtained for each slice, including a normal image, an image in the anterior-posterior (AP) direction, and an image in the foot-head (FH) direction. Figure 1 For the VENC MRI schematic diagram, the maximum blood flow velocity is set to 100 cm / sec, showing the correspondence between velocity and phase. It is performed through the atrial short-axis direction. All these images are retrospectively gated, with 25 phases or time frames for each slice. The MRI imaging parameters include: echo time TR: 47.1 ms, repetition time TE: 1.6 ms, field of view FOV: (298 340) mm2, (134 256) pixel matrix. The in-plane resolution is 1.54 mm / pixel, determined by the pixel pitch, and the through-plane resolution based on the slice interval is 6 mm. 500 cardiac VENC MRIs of 10 subjects are used as the training data set, including 250 FH and 250 AP-directed images. One group is scanned twice, once to obtain 50 clear VENC images by moving the body, and the other time to obtain 50 blurred VENC images by moving the body.

[0061] For training purposes, the method used is to translate the image, stack, and average the pixel values. The translation direction, translation step size, and the number of stacks are all randomly generated. 450 original clear images are used, and 7200 blurred images are generated in this way. When generating the blurred images, the blur direction of the image is recorded and classified as 0°, 45°, 90°, and 135° according to the nearest angle principle for subsequent training of the blur direction classification model.

[0062] The main focus of VENC MRI of the heart is related to the main blood flow situation in the heart, and the other parts of the image are almost random noise and are actually meaningless. Therefore, the evaluation model of the training results uses the most important part of the image. In the subsequent evaluation process, only the important part of the image is concerned, and the noise area around the image is ignored. In order to obtain blurred images from the ground truth images for training, a method of copying the translated image and then superimposing and averaging the pixel values is used to generate blurred images.

[0063] At the same time, multiple SRN-Deblur sub-models are used to deblur the motion-blurred VENC MRI in different directions. Before that, ResNet is used to classify the images to determine which sub-model should be used. Different from traditional algorithms, neural networks can combine feature extraction and learning. In the field of image processing, convolution operations can identify the relationships between adjacent pixels. Therefore, the convolutional neural network (CNN) has been widely used in image processing. CNN provides an end-to-end deep learning model. A well-trained CNN can extract image features and classify images. The depth of the CNN model plays a crucial role in image classification, which has led to the depth of the participating models in the ImageNet competition. When pursuing the depth of the network, the problem of degradation appears. ResNet solves this problem through the residual framework.

[0064] The concept of residual representation commonly used in the field of traditional computer vision is applied to the construction of the CNN model to form the basic block of residual learning. It uses multiple parameter layers to learn the residual representation between the input and the output, rather than using parameter layers to directly attempt to learn the mapping between the input and the output like a general CNN network. Experiments show that compared with directly learning the mapping relationship between the input and the output, it is simpler and more effective to directly use a general reference layer to learn the residual.

[0065] SRN-Deblur is a more effective multi-scale image deblurring network structure, where SRN is the initial letter of Scale Recursive Network. The SRN-Deblur model is based on two structures: the scale cyclic structure and the encoder-decoder ResBlock network. The SRN-Deblur technique is based on network weight sharing at different scales, which greatly reduces the training difficulty and improves stability. This method has two advantages. First, it can significantly reduce the number of trainable parameters and improve the training speed. Second, the structure of SRN-Deblur utilizes the loop module to transmit useful information at different scales throughout the network, helping with image deblurring. In the field of computer vision, the encoder-decoder structure is often used in deep learning. Image deblurring is a computer vision task, and srn deblurring also adopts this structure. Instead of directly using the encoding-decoding structure, it combines the encoding-decoding structure with ResBlock, which is called the encoder-decoder ResBlock network. Experimental results show that this structure can make the training speed faster and at the same time make the network more effective in image deblurring, which is why this structure is called the Scale Recursive Network (SRN).

[0066] Vorticity can be used to measure the angular velocity of a fluid at a certain point and can be calculated based on the velocity gradient of the fluid. As Figure 2 shown, the vorticity ω is the circulation area divided by the circulation area and is equal to the line integral of the tangential velocity along a counterclockwise (CCW) loop enclosing the point of interest. For image reconstruction (such as super-resolution, deblurring, etc.), the peak signal-to-noise ratio and the structural similarity index are generally used as evaluation criteria. In this article, two VENC MRIs with FH and AP directions indicate a single part of the heart. This means that there is a significant correlation between the two MRIs, and the evaluation index should be able to combine the FH vector image and the AP vector image for comprehensive evaluation. On the other hand, since the cardiac VENC MRI contains a large amount of useless random noise, only the part of the image containing the blood flow information in the heart cavity is really important. Therefore, the evaluation index should only depend on the part of the image related to the cardiac flow.

[0067] The ResNet model was implemented for blur classification with a classification accuracy exceeding 99%. According to the classification results, 4 SRN-Deblur sub-models were trained to deblur images. Finally, the deblurring results of the model were compared with those of SRN-Deblur alone. The results show that on the dataset, the method is superior to SRN-Deblur. The method is more suitable for complex situations. Different models can be used for different types of images and can handle images that SRN-Deblur cannot handle. The method is more targeted and performs better than SRN-Deblur on specific types of images.

[0068] Through experiments, it is found that each trained model is only suitable for certain types of blurs and cannot remove other types of blurs. This means that the SRN-Deblur network does have the ability to eliminate VENC MRI blurs, but a single SRN-Deblur model is not sufficient to handle all types of blurs. To solve this problem, ResNet is used to pre-classify the blurred images, and multiple SRN-Deblur sub-models are introduced to deblur different types of blurred images.

[0069] As Figure 3 shown, the deblurring architecture distributes the image according to the blur direction of the image and then uses the corresponding deblurring sub-model to process it. First, the blur direction classification model is trained, which can determine the blur direction of the image. Then, the training images required for the deblurring model are classified according to the blur direction, and the classified images are fed back to the corresponding deblurring sub-model for training. The proposed architecture is the first to consider the blur direction as a sub-problem.

[0070] The structure of the classification model used is based on ResNet. As Figure 3 shown, a single residual block contains two convolutional layers, two batch normalization layers, and one Relu layer. Specifically, the beginning part of the network consists of a convolutional layer, a batch normalization layer, and a Relu layer. The input size of the convolutional layer is 64×64, the kernel size is 3×3, the padding is 1, and the stride is 1. Next, there are 8 residual blocks with different parameters, and the tensor output by the last residual block is 512×11×16. At the end of the network, average pooling and fully connected are used to output the classification of the input image. The specific parameters of the 8 ResNets are shown in Table 1.

[0071] Table 1 Parameter values of ResNet

[0072]

[0073] During training, the cross-entropy function is used as the loss function, and the epoch is set to 50. For ResNet, classifying images with different blur directions is a simple task, and the changes in loss and accuracy during the training process are as Figure 4 shown. After a short period of training, the accuracy is close to 100%, and on the other hand, the loss function is close to 0. This lays a solid foundation for the next training model and deblurring images in different blur directions. In other words, since the blur is divided into different categories before deblurring, the subsequent deblurring model does not need to handle this task, making the function of the srn-deblurring model more specific. A more specific function also means that this task requires a smaller model capacity, and the model is better trained.

[0074] When creating the dataset, "black edges" of the segmented images are used to create blurred images by translating and averaging pixels. For different images, the width of the black edges is not fixed and is equal to the number of translation steps. Additionally, due to different combinations of translation directions and positions for cutting the black edges, the images in the training dataset have many different sizes.

[0075] Training is carried out using the open-source SRN-Deblur code. To train the model, Adam is used and β1 = 0.9, β2 = 0.999, ∈ = 10 -8 , the learning rate is set to the initial value of exponential decay, starting from 1×10 -4 to 1×10 -6 after 4000 iterations, and the weight is equal to 0.3; before inputting the blurred images into the neural network for training, the images are randomly cropped to a size of 128×128. The network parameters are initialized using the gloot method, and these parameters are fixed in all experiments.

[0076] Under the same hardware conditions, the training time of SRN-Deblur is approximately 3 hours. The reason for the lower time cost is that the training data is less and the size of the training image slices is smaller.

[0077] The change curve of the validation accuracy of ResNet during training is as Figure 4 shown. The validation accuracy of ResNet rises rapidly in the first 800 cycles, from 0.25 (the validation accuracy of random guessing) to above 0.9. During more than 1000 cycles after 800 cycles, the validation accuracy continues to rise steadily and finally exceeds 0.99. The loss function reaches a very good value in a very short time. Blurred images are used to test the deblurring effects of different methods.

[0078] Figure 5 Shows the differences in the deblurring results of two different deblurring methods, sharp and blur. It can be seen that multiple SRN-Deblur sub-models (methods) and the SRN-Deblur sub-model can all significantly deblur. A careful observation of the processing results of the method and SRN-Deblur shows that the method is clearer and closer to the original image. However, visual inspection cannot fully explain this problem. The differences of the arrows or pixels can be better explained by using ω PSNR mathematical evaluation.

[0079] Motion artifacts can degrade image quality and may interfere with interpretation, especially in magnetic resonance imaging (MRI) applications with low signal-to-noise ratio, such as functional MRI or diffusion tensor imaging, and when imaging small lesions. High-resolution images are highly sensitive to motion artifacts and typically require longer scan times, which can exacerbate the motion artifacts. During scanning, fast imaging techniques and sequences, optimal receiver coils, careful patient positioning, and coaching may minimize motion artifacts. Physiological noise sources include respiration, blood flow, and pulsatile motion coupled to the cardiac cycle, swallowing reflexes, and small spontaneous head movements. For example, during rest, functional MRI spontaneous neuronal activity increases the signal variation by 12%, and even under optimal conditions, the signal contribution from physiological noise still accounts for a substantial proportion. Motion tracking during imaging may allow for prospective correction or post-processing steps to separate the signal and noise.

[0080] Three-dimensional magnetic resonance imaging of the coronary arteries has the potential to provide high resolution and high signal-to-noise ratio, but it is highly susceptible to respiratory artifacts, especially respiratory blurring. The resolution loss caused by respiratory blurring in three-dimensional coronary imaging was theoretically analyzed and verified experimentally. Under normal respiration, the width of any Gaussian point spread function increases to a new value, at least several millimeters (about 3-4 mm). In vivo studies were conducted to compare respiratory pseudo-gated 3D acquisitions and breath-hold 2D acquisitions. On average, the overall quality of the pseudo-gated 3D images was worse than that of the corresponding breath-hold 2D images (P = 0.005). In most cases, respiratory blurring resulted in lower coronary artery resolution in the pseudo-gated three-dimensional data than in the respiratory two-dimensional data.

[0081] To verify the performance of the model in actual images, the true performance of the model was tested using the two scans of the same person mentioned previously. Figure 6 The contrast between the low-quality image and the enhanced image is shown. Two clear and blurred images with the same phase in two cardiac cycles were selected because the blood flow in these two sets of images is consistent. A vortex arrow diagram was made with the blood flow. It can be seen from the color of the vortex that the blood flow has mixed together in the blurred image, and the difference in blood flow in the de-blurred image is obvious. In addition, the arrows in the de-blurred image are parallel to the tangent of the atrial edge, while the arrows in the blurred image form a large angle with the tangent of the atrial edge, proving that the method can de-blur the actual imaging scan well. The average vorticity values are: sharp image: -49.99ωs -1 ,blurred image: -58.25ωs -1 ,de-blurred image: -51.79ωs -1 .

[0082] The above are the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A cardiac image reconstruction method based on deep learning, characterized in that, Including: Step 1: Select the hearts of multiple subjects for slice scanning and imaging. Three images are obtained for each slice, including a normal image, an image in the AP direction, and an image in the FH direction. The scanned images are used as the training dataset, and the scanned images include clear images and blurred images; Step 2: Use the ResNet model to classify the images obtained from the heart scans, and then use multiple SRN-Deblur sub-models to perform deblurring operations on the blurred images in different directions; Step 3: Utilize the images in the AP direction and the images in the FH direction to calculate the direction and magnitude of the blood flow vector in three-dimensional space for measuring the deblurring effect of the simulated blurred images.

2. The method for reconstructing a cardiac image based on deep learning according to claim 1, wherein The specific content of Step 1 includes: The slice scanning and imaging uses magnetic resonance imaging technology, is carried out in the short axis direction of the atrium, and uses retrospective gating. Each slice has 25 phases or time frames; The magnetic resonance imaging parameters include: echo time TR: 47.1 ms, repetition time TE: 1.6 ms, field of view FOV: (298 340) mm 2 , (134 256) mm 2 Pixel matrix, in-plane resolution is 1.54 mm / pixel, determined by pixel pitch, and through-plane resolution based on slice interval is 6 mm.

3. The method for reconstructing a cardiac image based on deep learning according to claim 1, wherein Step 2 includes: Determine the blur direction of the image, classify the training images required for the deblurring model in the blur direction, and feed the classified images back to the corresponding deblurring sub-models for training; During training, use the cross-entropy function as the loss function and set the epoch to 50.

4. The method for reconstructing a cardiac image based on deep learning according to claim 1, wherein Step 3 includes: When calculating the direction and magnitude of the blood flow vector in three-dimensional space, it is necessary to calculate the absolute value of the difference between pixels and the distance of the vector; where FHG, FHB, APG, APB represent FH ground truth, FH blurred image, AP ground truth, AP blurred image respectively, and i, j represent the positions of the images; ω PSNR Calculate using the following method, where MAX represents the sum of the maximum vector distances in the useful area; Since there is no blurred and clear mapping pair in true slice-scanning imaging, two scans at different heartbeats at the same moment need to be compared. At the same moment, the scans of the two heartbeats are also different. Use ω PSNR , instead, compare with vorticity. The mathematical expression of two-dimensional vorticity is as follows: Among them, the positive and negative signs of ω have different meanings, where the positive value represents a CCW cycle, the negative value represents the fluid rotating clockwise CW, and the magnitude of the value represents the rotation speed; Calculate the circulation Γ using the line integral of the CCW closed loop C, and write it in the form of an area integral as follows 5. A cardiac image reconstruction system based on deep learning, characterized in that, Including: A dataset acquisition module for selecting the hearts of multiple subjects for slice scanning and imaging. Three images are obtained for each slice, including a normal image, an image in the AP direction, and an image in the FH direction. The scanned images are used as the training dataset, and the scanned images include clear images and blurred images; An image processing module for using the ResNet model to classify the images obtained from the heart scans, and then using multiple SRN-Deblur sub-models to perform deblurring operations on the blurred images in different directions; An evaluation module for utilizing the images in the AP direction and the images in the FH direction to calculate the direction and magnitude of the blood flow vector in three-dimensional space for measuring the deblurring effect of the simulated blurred images.

Images