WO2024108438A1 - 一种速度编码磁共振成像的运动伪影校正方法 - Google Patents

一种速度编码磁共振成像的运动伪影校正方法 Download PDF

Info

Publication number
WO2024108438A1
WO2024108438A1 PCT/CN2022/133813 CN2022133813W WO2024108438A1 WO 2024108438 A1 WO2024108438 A1 WO 2024108438A1 CN 2022133813 W CN2022133813 W CN 2022133813W WO 2024108438 A1 WO2024108438 A1 WO 2024108438A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
deblurring
images
model
blur
Prior art date
Application number
PCT/CN2022/133813
Other languages
English (en)
French (fr)
Inventor
黄建龙
贾富仓
李聪
Original Assignee
深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Priority to PCT/CN2022/133813 priority Critical patent/WO2024108438A1/zh
Publication of WO2024108438A1 publication Critical patent/WO2024108438A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/269—Analysis of motion using gradient-based methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects

Definitions

  • the present invention relates to the technical field of medical image processing, and more particularly to a motion artifact correction method for velocity encoding magnetic resonance imaging.
  • VENC MRI Velocity-encoded magnetic resonance imaging
  • PC phase contrast
  • VENC MRI provides quantitative information on blood flow without the need to introduce contrast agents into the body.
  • Each pixel in the image corresponds to the blood flow velocity at that location.
  • Cardiac chamber segmentation will further define the boundaries of blood flow and isolate regions of interest for blood flow analysis.
  • the KCNN model can estimate the motion amplitude and direction of the blurred part and use multiple image blocks to describe complex motion. It can handle heterogeneous blur, but the computational complexity is high.
  • the GOPRO dataset is used to train an end-to-end deblurring model without considering the blur kernel. This scheme avoids errors caused by insufficient blur kernel estimation.
  • the object of the present invention is to overcome the defects of the prior art and provide a motion artifact correction method for velocity encoding magnetic resonance imaging, the method comprising:
  • a corresponding deblurring sub-model is selected to perform deblurring processing to obtain a corrected image, wherein the deblurring sub-model is obtained through training, and the number of the deblurring sub-models is the same as the number of blur direction types, and each deblurring sub-model corresponds to a type of blur direction.
  • the advantage of the present invention is that the motion artifact correction method for velocity-encoded magnetic resonance imaging provided is carried out in two stages. First, a residual network is used to determine the category of the blur direction of the image, and then the corresponding deblurring sub-model is dispatched to calibrate the artifacts in the image.
  • This architecture of multiple deblurring sub-models composed of classification rather than end-to-end models can significantly improve blurred and defective medical images, and can be used for deblurring, reconstructing blood flow images, and evaluating blood flow analysis, etc., which is helpful in assisting cardiologists in clinical analysis.
  • FIG1 is a flow chart of a method for correcting motion artifacts in velocity encoding magnetic resonance imaging according to an embodiment of the present invention
  • FIG2 is a schematic diagram of visual blood flow reconstruction in cardiac velocity encoding magnetic resonance imaging according to an embodiment of the present invention
  • FIG3 is a schematic diagram of acquiring cardiac image data in the short-axis direction of the atrium according to an embodiment of the present invention
  • FIG4 is a vortex measurement and a histogram thereof according to an embodiment of the present invention.
  • FIG5 is a schematic diagram of a deblurring model architecture according to an embodiment of the present invention.
  • FIG6 is a schematic diagram of a fuzzy classification model structure according to an embodiment of the present invention.
  • FIG7 is a schematic diagram of a ResNet training process for classifying blurred images according to an embodiment of the present invention.
  • FIG8 is a visual comparison diagram of deblurring effects of different models and low-quality images according to an embodiment of the present invention.
  • FIG9 is a visual result of real VENC MRI deblurring according to an embodiment of the present invention, including visualization of velocity vectors and vorticity scalar maps;
  • FIG10 is a vorticity quantification and vorticity distribution histogram of a blurred image and a deblurred image according to an embodiment of the present invention
  • FIG. 11 is a comparison of vorticity distribution histograms and quantized vorticity of blurred images and deblurred images of five time frames according to an embodiment of the present invention.
  • the motion artifact correction method for velocity encoding magnetic resonance imaging includes the following steps:
  • Step S110 collecting a data set, and then obtaining a one-to-one mapping from a blurred image to a clear image through blur processing, and marking the blur direction type.
  • VENC Pixels in MRI represent blood flow velocity, with maximum blood flow velocity corresponding to a 180° phase shift, and VENC is inversely proportional to the magnitude of these gradients. The larger the gradient, the more phase loss, and the corresponding smaller the VENC value.
  • the image deblurring method adopted in the present invention assumes that the blur kernel is unknown. Therefore, in order to learn the mapping relationship between the blurred image and the clear image, it is necessary to first collect a data set.
  • the result obtained by VENC MRI is a cardiac tomography scan, and three images are obtained for each layer, including a normal image, an image in the anterior-posterior (AP) direction, and an image in the foot-head (FH) direction.
  • an imaging device is used to scan the heart to generate images of each layer (or each slice), including a normal image, AP VENC MRI, and FH VENC MRI, through which the vortex (or vortex) of the blood can be visualized.
  • the corresponding relationship between velocity and phase is also shown in Figure 2, and the maximum blood flow velocity is 100cm/s.
  • VENC MRI was acquired in the short axis direction of the atrium. All of these images were acquired using retrospective gating, with 25 phases or time frames per slice, and the image data is shown in Figure 3.
  • MRI imaging parameters include: echo time TR is 47.1ms, repetition time TE is 1.6ms, field of view FOV is (298 ⁇ 340) mm2 , and pixel matrix is 134 ⁇ 256. Planar resolution is determined by pixel pitch, which is 1.54mm/pixel, and the through-plane resolution based on slice pitch is 6mm.
  • 500 cardiac VENC MRIs of 10 subjects were used as training data sets, including 250 images in the FH direction and 250 images in the AP direction. It should be noted that one of the subjects underwent two scans, one to obtain 50 clear VENC images and the other to obtain 50 blurred VENC images by moving the body.
  • a blurred image can be obtained by translating images, superimposing images, and calculating the average pixel value, wherein the translation direction, translation step size, and number of superpositions are all randomly generated. In this way, 7,200 blurred images are generated from 450 original clear images. After the blurred image is generated, the blur direction of the image is recorded and divided into four categories of blur directions of 0°, 45°, 90°, and 135° according to the nearest angle principle for subsequent training of a blur direction classification model.
  • Cardiac VENC MRI focuses on the main blood flow related to the heart, and the rest of the image is almost full of random noise. Therefore, the most important part of the image will be used to evaluate the model training results in the future. For this reason, all 500 images were manually segmented by medical experts in the field of cardiology. After filtering the noisy data, the important parts of the image were evaluated in the subsequent evaluation process. In order to obtain a blurred image for training from a real image, a blurred image is generated by superimposing and calculating the average pixel value after copying the translated image.
  • Step S120 training a classification model and a deblurring sub-model, wherein the classification model is used to identify the blur direction type of the image, and the deblurring sub-model performs deblurring processing on the image of the corresponding blur direction.
  • the present invention aims to deblur VENC MRI and try to restore the original velocity encoding information.
  • deblurring models such as SRN (scale-recurrent network) networks.
  • SRN scale-recurrent network
  • the following is an example of a deblurring model based on SRN, also known as SRN-Deblur network.
  • the SRN-Deblur network is trained to directly deblur VENC MRI. In most cases, the network deblurs well. But in some special cases, the deblurring process makes the image blurrier. Experiments have found that each deblurring model is only suitable for certain types of blur, but cannot deblur other types. This means that the SRN-Deblur network is able to deblur VENC MRI, but a single SRN-Deblur model is not sufficient to handle all types of blur. To address this problem, it is preferred that a classification model is used to pre-classify the blurred images, and multiple SRN-Deblur sub-models are introduced to deblur different types of blurred images.
  • VENC MRI images with motion blur in different directions are deblurred using multiple SRN-Deblur sub-models.
  • a classification model is used to classify the image to determine which sub-model should be used.
  • the classification model can use various types of neural network models, such as convolutional neural networks. Neural networks can integrate feature extraction and learning. In the field of image processing, convolution operations can confirm the relationship between adjacent pixels, so convolutional neural networks (CNNs) are widely used in image processing.
  • CNNs convolutional neural networks
  • the CNN provides an end-to-end deep learning model, and a trained CNN can extract image features and classify images.
  • the depth of the CNN model plays a vital role in image classification, but the pursuit of network depth will cause degradation problems. Therefore, preferably, the residual network ResNet is used as a classification model to solve this problem.
  • the basic block of residual learning uses multiple parameter layers to learn the residual representation between input and output, rather than using parameter layers to directly try to learn the mapping between input and output like a general CNN network. Experiments have shown that it is easier and more effective to learn the residual directly using a common reference layer rather than a mapping between input and output.
  • SRN-Deblur is a more effective multi-scale image deblurring network structure.
  • the SRN-Deblur model is based on two structures, namely the scale cycle structure and the encoder-decoder ResBlock network.
  • the SRN-Deblur technology is based on the use of shared network weights of different scales, which can significantly reduce the difficulty of training and increase stability. This approach has two major advantages. First, SRN-Deblur can significantly reduce the number of trainable parameters and speed up training. Second, the structure of SRN-Deblur uses a cycle module, and useful information at each scale can be transmitted throughout the network, which helps to deblur the image.
  • Image deblurring is a computer vision task.
  • SRN-Deblur uses an encoder-decoder structure, but the encoder-decoder ResBlock network does not directly use the encoding and decoding structure, but combines the encoding and decoding structure with ResBlock. According to experimental results, this structure can speed up training and make the network more effective in image deblurring, so it is named Scaled Recurrent Network (SRN).
  • SRN Scaled Recurrent Network
  • the optimization parameters of the model such as weights and biases, can be obtained.
  • the training samples of the classification model reflect the correspondence between the blurred image and the blurred direction type
  • the training samples of the deblurring sub-model reflect the correspondence between the blurred image and the blurred direction type.
  • Step S130 using the trained classification model and the deblurring sub-model to correct velocity encoding magnetic resonance imaging artifacts.
  • the trained classification model can be used to identify the type of blur direction of the velocity-encoded magnetic resonance image to be processed; then, the corresponding deblurring sub-model is selected according to the identified type of blur direction to perform deblurring processing to obtain a corrected image. Furthermore, the corrected image can be used to identify vortices in the heart and calculate the vorticity, etc.
  • 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.
  • the fluid rotation can be represented using finite elements and flow vectors.
  • the vorticity calculation is based on the velocity curl at a certain point.
  • the numerical calculation is performed using the vector of the contour around a certain point in the flow field.
  • the vorticity ⁇ is the component of the angular velocity of the rotation in the direction of the plane normal vector, which is equal to the tangential velocity line integral of the counterclockwise (CCW) loop containing the target point.
  • CCW counterclockwise
  • the loop ⁇ is calculated using the line integral of the CCW closed loop C, which can be written as the following surface integral:
  • v represents the rotational linear velocity
  • S represents the area of the closed surface
  • Vx represents the component of the linear velocity in the x-direction
  • Vy represents the component of the linear velocity in the y-direction
  • the positive and negative signs of ⁇ have different meanings.
  • a positive value indicates that the fluid rotates counterclockwise (CCW)
  • a negative value indicates that the fluid rotates clockwise (CW).
  • the size of the value indicates the rotation speed.
  • Entropy can be defined as an indicator of the degree of disorder in a system, so the histogram also provides information about the complexity of the image in the form of an entropy descriptor. The higher the entropy, the more complex and chaotic the image is.
  • the mathematical formula for entropy is defined as follows:
  • the entropy value is used to verify the existence of vortices.
  • the tendency of blood flow to form vortices is not obvious, and the quantified blood flow direction is affected by motion blur. This makes everything tend to the direction of motion blur, and most of the vorticity is concentrated near 0, and the smaller the gradient, the fewer the number of colors. Therefore, the lower the image complexity, the lower the entropy value.
  • the blood flow direction is obviously vortex-shaped and more complex, the vorticity distribution is more discrete, the gradient is more, and more colors are presented.
  • PSNR peak signal-to-noise ratio
  • SSMI structural similarity index
  • two VENC MRIs related to 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.
  • cardiac VENC MRI contains a large amount of useless random noise, only the part of the image with blood flow information in the heart cavity is really important. Therefore, the evaluation index should depend only on the part of the image related to cardiac blood flow.
  • ⁇ PSNR vorticity ⁇ PSNR .
  • Medical experts are invited to help manually segment the important parts of the heart, and ⁇ PSNR will only consider the pixels of this part of the image.
  • ⁇ PSNR the direction and size of the blood flow vector in 3D space are calculated.
  • the absolute value of the difference between pixels needs to be calculated.
  • ⁇ PSNR the distance of the vector needs to be calculated.
  • FHG, FHB, APG and APB represent FH real image, FH blurred image, AP real image and AP blurred image respectively, and i and j represent the position of the image.
  • the vorticity ⁇ PSNR is calculated using the following formula, where MAX represents the sum of the maximum vector distances in the useful area.
  • ⁇ PSNR is used to measure the deblurring effect of the simulated blurred image. Since there is no mapping pair of blurred and clear images in the actual VENC MRI, only two scans at different heartbeats at the same time can be used for comparison. Even at the same time, there are differences in the scan results of two heartbeats, so it is impossible to use an evaluation standard based on pixel-by-pixel correspondence (such as ⁇ PSNR ).
  • the SRN-Deblur network was trained to directly deblur VENC MRI. Experiments found that each trained model was only suitable for certain types of blur, but could not deblur other types. This means that the SRN-Deblur network is indeed able to deblur VENC MRI, but a single SRN-Deblur model is not sufficient to handle all types of blur. To address this issue, ResNet was used to pre-classify the blurred images, and multiple SRN-Deblur sub-models were introduced to deblur different types of blurred images.
  • Training the deblurring sub-model consists of the following steps:
  • the remaining 20% of the images are input into the trained ResNet for classification, and then each classification result is input into the corresponding SRN-Deblur model for deblurring.
  • the training set is divided into four sub-training sets, each of which is used to train an SRN-Deblur sub-model;
  • the trained ResNet model is used to classify the test cases, and the corresponding SRN-Deblur sub-model is used for deblurring based on the classification results.
  • a blur direction classification model is first trained, which can determine the blur direction of the image. Then, the training images required by the deblur model are classified in terms of blur direction, and the classified images are input into the corresponding deblur sub-model for training. Compared with the existing technology, this architecture considers blur direction as a sub-problem for the first time.
  • the classification model structure is built based on ResNet.
  • the classification model contains 8 residual blocks, each of which contains two convolutional layers, two batch normalization layers and a Relu layer.
  • the beginning of the classification model contains a convolutional layer, a batch normalization layer and a Relu layer.
  • the input size of the convolutional layer is 64 ⁇ 64
  • the convolution kernel size, feature map padding width and convolution step size are 3 ⁇ 3, 1 and 1 respectively.
  • there are 8 residual blocks with different parameters and the last residual block outputs a 512 ⁇ 1 ⁇ 16 ⁇ 16 tensor.
  • average pooling and fully connected layers are used to output the classification results.
  • the parameters of the 8 residual blocks are shown in Table 1.
  • each ResNet block has two 2D convolutional layers, each input or output has its own size (marked as s.) and channel (marked as c.), and the corresponding convolution parameters are listed.
  • the cross entropy function is used as the loss function, and the Epoch is set to 50.
  • ResNet the changes in loss and accuracy during the training process are shown in Figure 7, where the curve generally located above represents the training accuracy, and the curve generally located below represents the loss value.
  • the loss drops to an extremely low level and the accuracy is close to 1.0.
  • the accuracy is close to 100% and the loss value is close to 0.
  • the subsequent deblurring model does not need to handle this task, which can make the function of the SRN-Deblur model more specific. More specific functions also mean that the model capacity required for this task is smaller and the model training results are better.
  • blurred images are created by translating and calculating the average pixel and "black edges" of the segmented images.
  • the width of the black edges is not fixed for different images and is equal to the translation step size.
  • the images of the training dataset have many different sizes.
  • the training time of SRN-Deblur is about three hours.
  • the reason for the lower time cost is that there is less training data and the size of the training image layer is smaller.
  • FIG8 shows the difference in deblurring results of two different deblurring methods.
  • multiple SRN-Deblur sub-models (the present invention) and SRN-deblur have obvious deblurring effects, and the deblurred image of the present invention is clearer and closer to the original image.
  • naked eye observation cannot fully explain this problem.
  • Mathematical evaluation can be performed using ⁇ PSNR to better explain the difference in pixels.
  • Table 3 lists the ⁇ PSNR of the blurred image, the image deblurred by SRN-Deblur and the image deblurred by the present invention.
  • the ⁇ PSNR of the image deblurred by the present invention is the highest, and blur in all directions of the image can be removed.
  • Figure 9 shows the comparison of low-quality images and enhanced images. Clear images and blurred images of the same phase of two heartbeat cycles were selected because the blood flow in these two sets of images is consistent.
  • the vortex arrow diagram was drawn using the blood flow. From the vortex color, it can be seen that the blood flow in the blurred image has been mixed together, while the blood flow distinction in the deblurred image is very obvious.
  • the arrow of the deblurred image is parallel to the tangent of the atrial edge, while the arrow of the blurred image forms a large angle with the tangent of the atrial edge, proving that the method of the present invention has a good deblurring effect on the actual imaging scan, and the average vorticity values are as follows: clear image: -49.99 ⁇ (s -1 ), blurred image: -58.25 ⁇ (s -1 ), deblurred image: -51.79 ⁇ (s -1 ).
  • FIG. 11 is a comparison of the vorticity distribution histogram and the quantized vorticity of the blurred image and the deblurred image of 5 time frames.
  • the vorticity distribution histogram the vorticity distribution of the blurred image and the deblurred image is illustrated respectively.
  • ⁇ b and ⁇ r are the vorticity distribution standard deviations of the blurred image and the deblurred image, respectively
  • h b and h r represent the entropy values of the blurred image and the deblurred image, respectively. It can be seen that the deblurred image can accurately identify the vortex position.
  • the vorticity distribution corresponding to the deblurred image is more discrete, the color represents more, and the entropy value is higher, while the blurred image cannot identify the vortex position at all, the vorticity distribution is close to 0, the entropy value is low and the color represents less. This shows that the measurement method used in the present invention is very effective.
  • the present invention proposes a new model, which is a deblurring model that combines ResNet and multiple SRN-Deblur models into one.
  • the ResNet model is used for fuzzy classification, and the classification accuracy exceeds 99%.
  • four SRN-Deblur sub-models are trained to deblur the image, and the four trained SRN-Deblur sub-models can output high-quality deblurred images.
  • the deblurring results of the model are compared with the results of deblurring using SRN-Deblur alone. The results show that the present invention is more suitable for complex situations. Different models can be used for different types of images, and images that SRN-Deblur cannot process can be processed.
  • the difference between the deblurred image and the actual image is significantly smaller than the difference between the blurred image and the actual image.
  • the experimental results show that the vorticity of the deblurred image is closer to the clear image than the blurred image.
  • the model of the present invention is superior to the existing technology in both visual inspection and mathematical evaluation. After removing the blur factor, the clear VENC MRI obtained can help radiologists and clinicians make better clinical judgments and improve diagnostic accuracy, and can be used in a wider range of fields.
  • the present invention may be a system, a method and/or a computer program product.
  • the computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.
  • Computer readable storage medium can be a tangible device that can keep and store the instructions used by the instruction execution device.
  • Computer readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof.
  • Non-exhaustive list of computer readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, for example, a punch card or a convex structure in a groove on which instructions are stored, and any suitable combination thereof.
  • RAM random access memory
  • ROM read-only memory
  • EPROM or flash memory erasable programmable read-only memory
  • SRAM static random access memory
  • CD-ROM compact disk read-only memory
  • DVD digital versatile disk
  • memory stick a floppy disk
  • mechanical encoding device for example, a punch card or a convex structure in a groove on which instructions are stored, and any suitable combination thereof.
  • the computer readable storage medium used here is not interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic waves, an electromagnetic wave propagated by a waveguide or other transmission medium (for example, a light pulse by an optical fiber cable), or an electrical signal transmitted by a wire.
  • the computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing/processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and/or a wireless network.
  • the network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and/or edge servers.
  • the network adapter card or network interface in each computing/processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing/processing device.
  • the computer program instructions for performing the operation of the present invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, Python, etc., and conventional procedural programming languages, such as "C" language or similar programming languages.
  • Computer-readable program instructions may be executed entirely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or entirely on a remote computer or server.
  • the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet).
  • an electronic circuit such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be personalized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present invention.
  • These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions/actions specified in one or more boxes in the flowchart and/or block diagram is generated.
  • These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and/or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions/actions specified in one or more boxes in the flowchart and/or block diagram.
  • Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions/actions specified in one or more boxes in the flowchart and/or block diagram.
  • each box in the flowchart or block diagram can represent a part of a module, program segment or instruction, and the part of the module, program segment or instruction contains one or more executable instructions for realizing the specified logical function.
  • the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved.
  • each box in the block diagram and/or flowchart, and the combination of the boxes in the block diagram and/or flowchart can be implemented by a dedicated hardware-based system that performs the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions. It is well known to those skilled in the art that it is equivalent to implement it by hardware, implement it by software, and implement it by combining software and hardware.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

本发明公开了一种速度编码磁共振成像的运动伪影校正方法。该方法包括：利用经训练的分类模型确定速度编码磁共振图像的模糊方向类型；根据所述模糊方向类型选择对应的去模糊子模型进行去模糊处理，获得校正后的图像，其中，所述去模糊子模型经训练获得，且所述去模糊子模型的数量与所述模糊方向类型的数量相同，每个去模糊子模型对应一类模糊方向。本发明能够有效去除图像中的伪影或噪声，获得更清晰的校正图像，进而准确识别心脏内的涡旋位置。

Description

一种速度编码磁共振成像的运动伪影校正方法 技术领域
本发明涉及医学图像处理技术领域，更具体地，涉及一种速度编码磁共振成像的运动伪影校正方法。
背景技术
速度编码磁共振成像(VENC MRI)是一种基于相位对比(PC)测量血流速度的成像模式，因此也称为PC-MRI。VENC MRI无需将造影剂引入人体即可提供血流的定量信息。图像中的每个像素都与相应位置的血流速度相对应。通过将二维心脏血流叠加在相应的MR图像上，即可参考心脏解剖结构的血流情况。心腔分割将进一步明确血液流动的边界，并隔离出感兴趣区域以进行血流分析。
然而，在VENC MRI扫描过程中，成像设备和患者之间的相对运动会导致成像效果较差。例如，患者在成像过程中移动身体会导致成像模糊。此外，成像过程中往往要求患者屏住呼吸，但有些患者很难长时间屏住呼吸，而即便是最轻微的呼吸，也会导致运动模糊。经研究，在成像过程中，时而发生的心电门控误触发、患者心律失常、无法控制的运动以及轻微的呼吸等都会造成心脏位置的动态调整。这些情况会导致运动模糊并引发伪影，从而影响患者的心血管图像诊断，因此去除VENC MRI中的运动模糊极具实用价值。
随着人工智能(AI)在图像处理领域的发展，与临床诊断相关的深度学习模型在医学图像处理中的应用正日益增加。现有的图像去模糊方案中，盲去模糊(即模糊核未知)的研究成果丰硕。例如，KCNN模型可估计模糊部分的运动幅度和方向，并使用多个图像块描述复杂的运动，可以处理异质模糊，但计算复杂度较高。又如，使用GOPRO数据集训练端到端的去模糊模型，而不考虑模糊核，这种方案避免了模糊核估计不足导致的错 误。
然而，目前的去模糊模型基本都是端到端模型，这意味着，任何类型的运动模糊图像输入模型，模型将输出没有运动模糊的图像。如果模型能够通过端到端框架获得良好的性能，就能发挥简单快速的优势。但在一些应用领域，两相模型可提供更为实用的效果。与普通的自然图像不同，医学图像的轮廓和结构并不明显，也没有丰富的色彩，且噪声较多。尤其是VENC MRI，其图像中的大部分像素都是黑白像素的噪声。VENC MRI的这一特点显著增加了学习特征的难度。
发明内容
本发明的目的是克服上述现有技术的缺陷，提供一种速度编码磁共振成像的运动伪影校正方法，该方法包括：
利用经训练的分类模型确定速度编码磁共振图像的模糊方向类型；
根据所述模糊方向类型选择对应的去模糊子模型进行去模糊处理，获得校正后的图像，其中，所述去模糊子模型经训练获得，且所述去模糊子模型的数量与所述模糊方向类型的数量相同，每个去模糊子模型对应一类模糊方向。
与现有技术相比，本发明的优点在于，所提供的速度编码磁共振成像的运动伪影校正方法分两阶段进行，首先使用残差网络判断图像的模糊方向的类别，进而调度相应的去模糊子模型校准图像中的伪影，这种基于分类而非端到端模型构成多个去模糊子模型的架构可以显著改善模糊和有缺陷的医学图像，用于去模糊、重建血流图像以及评价血流分析等，有利于辅助心脏科医生进行临床分析。
通过以下参照附图对本发明的示例性实施例的详细描述，本发明的其它特征及其优点将会变得清楚。
附图说明
被结合在说明书中并构成说明书的一部分的附图示出了本发明的实施例，并且连同其说明一起用于解释本发明的原理。
图1是根据本发明一个实施例的速度编码磁共振成像的运动伪影校正方法的流程图；
图2是根据本发明一个实施例的心脏速度编码磁共振成像中的视觉血流重建示意图；
图3是根据本发明一个实施例的在心房的短轴方向获取心脏图像数据的示意图；
图4是根据本发明一个实施例的涡量测量及其直方图；
图5是根据本发明一个实施例的去模糊模型架构示意图；
图6是根据本发明一个实施例的模糊分类模型结构示意图；
图7是根据本发明一个实施例的对模糊图像进行分类的ResNet训练过程示意图；
图8是根据本发明一个实施例的不同模型去模糊效果与低质量图像的目视比较图；
图9是根据本发明一个实施例的真实VENC MRI去模糊的视觉结果，包括速度向量和涡量标量图的可视化；
图10是根据本发明一个实施例的模糊图像和去模糊图像的涡量量化和涡量分布直方图；
图11是根据本发明一个实施例的5个时间帧的模糊图像和去模糊图像的涡量分布直方图和量化涡量的比较。
具体实施方式
现在将参照附图来详细描述本发明的各种示例性实施例。应注意到：除非另外具体说明，否则在这些实施例中阐述的部件和步骤的相对布置、数字表达式和数值不限制本发明的范围。
以下对至少一个示例性实施例的描述实际上仅仅是说明性的，决不作为对本发明及其应用或使用的任何限制。
对于相关领域普通技术人员已知的技术、方法和设备可能不作详细讨论，但在适当情况下，所述技术、方法和设备应当被视为说明书的一部分。
在这里示出和讨论的所有例子中，任何具体值应被解释为仅仅是示例 性的，而不是作为限制。因此，示例性实施例的其它例子可以具有不同的值。
应注意到：相似的标号和字母在下面的附图中表示类似项，因此，一旦某一项在一个附图中被定义，则在随后的附图中不需要对其进行进一步讨论。
参见图1所示，所提供的速度编码磁共振成像的运动伪影校正方法包括以下步骤：
步骤S110，采集数据集，进而通过模糊处理获得模糊图像到清晰图像的一对一映射，并标注模糊方向类型。
VENC MRI中的像素代表血流速度，最大血流速度对应于180°相移，并且VENC与这些梯度的大小成反比。梯度越大，失相越多，对应的VENC值越小。
本发明采用的图像去模糊方法的假设是，模糊核未知，因此，为了学习模糊图像与清晰图像之间的映射关系，需要首先采集数据集。在一个实施例中，VENC MRI所获结果为心脏断层扫描，每层获得三幅图像，包括正常图像、前后(AP)方向的图像和足头(FH)方向的图像。参见图2所示，其中使用成像设备扫描心脏，生成各层(或者各切片)的图像，包括正常图像、AP VENC MRI和FH VENC MRI，可以通过这些图像实现血液的涡旋(或称漩涡)可视化。图2中也显示了速度和相位的对应关系，最大血流速度为100cm/s。
例如，VENC MRI在心房的短轴方向获取。所有这些图像均采用回顾性门控获得，每个切片25个相位或时间帧，图像数据参见图3。MRI成像参数包括：回波时间TR是47.1ms，重复时间TE是1.6ms，视场FOV为(298×340)mm 2，像素矩阵为134×256。平面分辨率通过像素间距确定，为1.54mm/像素，基于切片间距的通过平面的分辨率为6mm。使用10例受试者的500次心脏VENC MRI作为训练数据集，包括250幅FH方向和250幅AP方向的图像。应注意的是，其中1例受试者接受了两次扫描，一次获得50幅清晰的VENC图像，另一次通过移动身体获得50幅模糊的VENC图像。
单幅图像去模糊领域的一大问题是，难以获得训练数据集。在现实世界中，无法获得模糊图像到清晰图像的一对一映射对。而为了训练去模糊模型，需要寻求适当的技术来模拟模糊图像到清晰图像的映射。在一个实施例中，可通过平移图像、叠加图像和计算平均像素值来获得模糊图像，其中平移方向、平移步长和叠加次数均随机生成。通过这种方式，将450幅原始清晰图像生成7200幅模糊图像。在生成模糊图像后，记录图像的模糊方向，并根据最近角度原则划分为0°、45°、90°和135°四类模糊方向，用于后续训练模糊方向分类模型。
心脏VENC MRI主要关注与心脏相关的主要血流情况，图像的其他部分几乎充满随机噪声。因此，后续将采用图像中最重要的部分进行模型训练结果的评价。出于这一原因，500幅图像均由心脏领域的医学专家手动分割。过滤噪声数据后，在后续评价过程中对图像的重要部分进行评价。为从真实图像获得训练用模糊图像，采用在复制平移图像后叠加和计算平均像素值的方法生成模糊图像。
步骤S120，训练分类模型和去模糊子模型，其中分类模型用于识别图像的模糊方向类型，去模糊子模型针对相应模糊方向的图像进行去模糊处理。
本发明旨在对VENC MRI进行去模糊处理，并尝试恢复原始的速度编码信息。去模糊模型可采用多种类型，例如，SRN(scale-recurrent network，尺度循环网络)网络等。下文以基于SRN的去模糊模型为例进行说明，也称为SRN-Deblur网络。
训练SRN-Deblur网络，以直接对VENC MRI进行去模糊。在大多数情况下，该网络去模糊效果良好。但在某些特殊情况下，去模糊过程会使图像更加模糊。实验发现，每个去模糊模型都仅适用于某些类型的模糊，但无法对其他类型进行去模糊。这意味着SRN-Deblur网络能够对VENC MRI进行去模糊，但单一的SRN-Deblur模型不足以处理所有类型的模糊。为解决这一问题，优选地，使用分类模型对模糊图像进行预分类，并引入了多个SRN-Deblur子模型，以对不同类型的模糊图像进行去模糊。
使用多个SRN-Deblur子模型对具有不同方向运动模糊的VENC MRI 图像进行去模糊。在此之前，使用分类模型对图像进行分类，以确定应该使用哪个子模型。分类模型可采用多种类型的神经网络模型，例如，卷积神经网络等。神经网络可以集成特征提取和学习。在图像处理领域，卷积运算可以确认相邻像素之间的关系，因此卷积神经网络(CNN)在图像处理方面运用广泛。
CNN提供端到端的深度学习模型，经训练的CNN可以提取图像特征并对图像进行分类。CNN模型的深度在图像分类中起着至关重要的作用，但追求网络深度会引发退化问题。因此，优选地，采用残差网络ResNet作为分类模型解决了这一问题。残差学习的基本块使用多个参数层学习输入和输出之间的残差表示，而不是像一般CNN网络那样使用参数层直接尝试学习输入和输出之间的映射。实验表明，直接使用通用参考层而非输入和输出之间的映射来学习残差更加容易且有效。
SRN-Deblur是一种更有效的多尺度图像去模糊网络结构。SRN-Deblur模型基于两种结构，即尺度循环结构和编码器-解码器ResBlock网络。SRN-Deblur技术以使用不同尺度的共享网络权重为基础，可显著降低训练难度并增加稳定性。这种方法有两大优势。首先，SRN-Deblur能够大幅减少可训练参数的数量，并加快训练速度。其次，SRN-Deblur的结构采用循环模块，各尺度的有用信息可在整个网络中传输，有助于图像去模糊。
图像去模糊是一种计算机视觉任务，SRN-Deblur采用编码器-解码器结构，但编码器-解码器ResBlock网络没有直接使用编码和解码结构，而是将编码和解码结构与ResBlock相结合。根据实验结果，这种结构可以加快训练速度，同时可以使网络在图像去模糊方面更有效，因而被命名为尺度循环网络(SRN)。
采用构建的训练数据集训练分类模型和去模糊子模型后，能够获得模型的优化参数，如权重和偏置等。分类模型的训练样本反映模糊图像与模糊方向类型之间的对应关系，去模糊子模型的训练样本反映模糊图像与模糊方向类型之间的对应关系。
步骤S130，利用经训练的分类模型和去模糊子模型校正速度编码磁共振成像的伪影。
在实际应用过程中，可利用经训练的分类模型识别待处理的速度编码磁共振图像的模糊方向类型；进而根据所识别出的模糊方向类型选择对应的去模糊子模型进行去模糊处理，获得校正后的图像。进一步地，可利用校正图像识别心脏内涡旋，并计算涡量等。
涡量可用于测量流体在某一点的角速度，可根据流体的速度梯度计算。如图4所示，流体旋转可以使用有限元和流向量来表示，涡量计算基于某一点的速度旋度，数值计算使用流场中某一点周围轮廓的向量进行。涡量ω是旋转角速度在平面法向量方向上的分量，等于包含目标点的逆时针(CCW)回路的切向速度线积分。
首先，根据斯托克斯定理(Stokes'theorem)，使用CCW闭环C的线积分计算循环Γ，将其写为下列面积分：
Figure PCTCN2022133813-appb-000001
其中，二维涡量的数学表达式如下：
Figure PCTCN2022133813-appb-000002
其中，v表示旋转线速度，S表示封闭曲面的面积，V x表示线速度在x方向上的分量，V y表示线速度在y方向上的分量，ω的正负号含义不同，正值表示流体逆时针(CCW)旋转，负值表示流体顺时针(CW)旋转，值的大小表示旋转速度。
熵可以定义为系统中混乱程度的指标，因此直方图还以熵描述符的形式提供了有关图像复杂性的信息。熵越高，图像越复杂和混乱。熵的数学公式定义如下：
Figure PCTCN2022133813-appb-000003
其中，p i表示概率质量函数。
在本发明中，熵值用于验证涡旋的存在。在模糊图像中，血流形成涡旋的趋势不明显，量化的血流方向受运动模糊的影响。这使得一切都趋向于运动模糊的方向，且大部分涡量都集中在0附近，梯度越小代表颜色数量越少。因此，图像复杂度越低，熵值越低。去模糊后，血流方向明显呈 涡旋状，且更加复杂，涡量分布也更加离散，梯度更多，呈现出更多颜色。
对于图像重建(如超分辨率、去模糊等)，通常将峰值信噪比(PSNR)和结构相似性指数(SSMI)作为评价标准。在本发明中，与FH和AP方向有关的两种VENC MRI指示心脏的单一部分。这意味着两种MRI之间存在显著相关性，评价指标应能结合FH向量图像和AP向量图像，以进行综合评价。另一方面，由于心脏VENC MRI中包含大量无用的随机噪声，仅心腔内具有血流信息的图像部分才真正重要。因此，评价指标应仅取决于图像中与心脏血流相关的部分。
根据上述要求，提出了心脏VENC MRI的评价标准：涡量ω PSNR。邀请医学专家帮助手动分割心脏的重要部位，ω PSNR将仅考虑这部分图像的像素。使用FH和AP图像，计算3D空间中血流向量的方向和大小。在普通PSNR中，需计算的是像素之间差异的绝对值。而在ω PSNR中，需计算的是向量的距离。FHG、FHB、APG和APB分别代表FH真实图像、FH模糊图像、AP真实图像和AP模糊图像，i和j表示图像的位置。
Figure PCTCN2022133813-appb-000004
使用以下公式计算涡量ω PSNR，式中，MAX表示有用区域的最大向量距离之和。
Figure PCTCN2022133813-appb-000005
然后，使用ω PSNR衡量模拟模糊图像的去模糊效果。由于实际VENC MRI并无模糊和清晰图像的映射对，因此仅能使用同一时刻不同心跳周期的两次扫描结果进行比较。而即使是在同一时刻，两个心跳周期的扫描结果也存在差异，因此无法使用基于逐像素对应的评价标准(如ω PSNR)。
为进一步验证本发明的效果，进行了实验，实验过程和详细信息如下。
首先，训练SRN-Deblur网络，以直接对VENC MRI进行去模糊。实验发现，每个训练模型都仅适用于某些类型的模糊，但无法对其他类型进行去模糊。这意味着SRN-Deblur网络的确能够对VENC MRI进行去模糊，但单一的SRN-Deblur模型不足以处理所有类型的模糊。为解决这一问题， 使用ResNet对模糊图像进行了预分类，并引入了多个SRN-Deblur子模型，以对不同类型的模糊图像进行去模糊。
训练去模糊子模型包括以下步骤：
通过分层抽样法采集40％的图像，用于训练ResNet，并使用经训练的ResNet模型对剩余60％的图像进行分类；
从60％的分类图像中分层采集40％的图像，用于训练多个SRN-Deblur子模型；
将剩余20％的图像输入经训练的ResNet中进行分类，然后将各分类结果输入到对应的SRN-Deblur模型中进行去模糊。
在训练ResNet时，发现ResNet的分类准确度可以超过99％，甚至达到100％。另一方面，由于数据集中的图像并不多，因此对过程进行了以下改进：
通过分层抽样法，将所有图像的80％和20％分别分入训练集和测试集；
根据图像标签，将训练集分为四个子训练集，每个子训练集用于训练一个SRN-Deblur子模型；
在测试中，使用经训练的ResNet模型对测试用例进行分类，并根据分类结果使用对应的SRN-Deblur子模型进行去模糊。
如图5所示，首先训练一个模糊方向分类模型，该模型能够确定图像的模糊方向。然后，将去模糊模型所需的训练图像在模糊方向上进行分类，并将分类后的图像输入相应的去模糊子模型进行训练。与现有技术相比，这种架构首次将模糊方向作为子问题考虑。
在实验中，分类模型结构基于ResNet构建。如图6错误！未找到引用源。所示，分类模型包含8个残差块，每个残差块包含两个卷积层、两个批归一化层和一个Relu层。具体而言，分类模型的开始部分包含一个卷积层、一个批归一化层和一个Relu层。卷积层的输入尺寸为64×64，卷积核大小、特征图填充宽度和卷积步长分别为3×3、1和1。接下来，是8个不同参数的残差块，最后一个残差块输出512×1×16×16的张量。在分类模型的末端，使用平均池化和全连接层来输出分类结果。8个残差块的参数参见表1。
表1：残差块的参数值
Figure PCTCN2022133813-appb-000006
在表1中，有8个依次连接的ResNet块，每个ResNet块都有两个2D卷积层，每个输入或输出都有各自的尺寸(标记为s.)和通道(标记为c.)，并列出了相应的卷积参数。
在训练分类模型过程中，将交叉熵函数作为损失函数，并将Epoch设置为50。对于ResNet而言，训练过程中的损失和准确度变化参见图7所示，其中总体位于上方的曲线表示训练准确度，总体位于下方的曲线表示损失值。由图7可以看出，在2500次循环后，损失下降至极低水平，准确度接近1.0。也就是说，经过短暂的训练(完成训练过程所需时间约为1500s)，准确度接近100％，并且损失值接近0。这为后续训练去模糊模型以及对具有不同模糊方向的图像进行去模糊奠定了坚实的基础。换言之，由于在去模糊之前将模糊度划分为不同类别，因此后续的去模糊模型无需处理此任务，这可以使SRN-Deblur模型的功能更加具体。更具体的功能也意味着此任务需要的模型容量更小，且模型训练结果更优。
在构建训练数据集时，通过平移和计算平均像素以及分割图像的“黑边”来创建模糊图像。对于不同的图像，黑边的宽度不固定，等于平移步长。此外，由于平移方向不同，以及被切割黑边的位置组合不同，训练数据集的图像有很多不同的尺寸。
在实验过程中，为训练去模糊模型，采用Adam参数，设为β 1＝0.9、β 2＝0.999、∈＝10 -8。学习率设为指数衰减的初始值，从1×10 -4到1×10 -6，Epoch等于4000，幂等于0.3。在将模糊图像输入神 经网络训练之前，将图像随机切割成128×128的尺寸。使用Glorot法初始化网络参数，这些参数在所有实验中都是固定参数。
实验所用设备参见表2。
表2：平台配置
Figure PCTCN2022133813-appb-000007
在相同的硬件情况下，SRN-Deblur的训练耗时在三个小时左右。时间成本较少的原因是训练数据较少，且训练图像层的尺寸较小。
训练过程中ResNet的验证准确度变化曲线如图7所示，在前800次循环中，ResNet的验证准确度迅速上升，从0.25(随机猜测的验证准确度)上升至0.9以上。在800次循环后的1000多次循环中，验证准确度继续稳步上升，直至最终超过0.99。此外，在ResNet训练过程中实际损失函数值的变化在短时间内趋于稳定。
进一步地，使用模糊图像来测试不同方法的去模糊效果。图8显示了两种不同去模糊方法的去模糊结果差异。由图8可以看出，多个SRN-Deblur子模型(本发明)和SRN-deblur去模糊效果明显，本发明去模糊的图像更清晰，且更接近原始图像。然而，裸眼观察并不能完全解释这一问题。可通过使用ω PSNR进行数学评价，以更好地解释像素的差异。
表3列出了模糊图像、SRN-Deblur去模糊的图像和本发明的去模糊的图像的ω PSNR。总体而言，利用本发明去模糊后的图像ω PSNR最高，并且可以去除图像所有方向的模糊。
表3：ω PSNR对比结果
Figure PCTCN2022133813-appb-000008
为验证模型对实际图像的性能，使用前述同一受试者的两次扫描结果测试了模型的真实性能。图9显示了低质量图像和增强图像的比较。选择 了两个心跳周期同一相位的清晰图像和模糊图像，因为这两组图像中的血流一致。使用血流绘制涡旋箭头图，从涡旋颜色可以看出，模糊图像中的血流已混合在一起，去模糊图像中的血流区分则十分明显。此外，去模糊图像的箭头与心房边缘的切线平行，而模糊图像的箭头与心房边缘的切线构成较大角度，证明本发明的方法对实际成像的扫描具有良好的去模糊效果，平均涡量值如下：清晰图像：-49.99ω(s -1)，模糊图像：-58.25ω(s -1)，去模糊图像：-51.79ω(s -1)。
如图9所示，在模糊图像中很难识别出漩涡，但在清晰图像中却容易识别出漩涡。用于进行比较的模糊图像和清晰图像的涡量量化图和涡量分布直方图参见图10所示，其中，横轴和纵轴分别表示涡量的大小区间和每个区间内的涡量数量。e是图像的熵值。
图10中，在AP和FH方向，模糊图像的平均涡量测量值ω＝0.2860，标准差σ＝1.4681，熵值e＝23.6731。值得注意的是，正常图像的平均涡量ω＝-0.5519，标准差σ＝3.3258，熵值e＝26.7137。通过涡量分布直方图可得出结论，模糊图像的标准差小于正常图像，因此涡量分布更集中，熵值更低，且图像复杂度更低。由图10可以看出，大部分涡量集中在0附近，说明图像中没有涡量区域。相比之下，正常图像的涡量分布更离散，熵值更高，图像表现更复杂，且图中有两个不同极性的涡量区域。由于涡量平均值为负，涡量更多地落在0的左侧，因此顺时针方向的涡量在图10中占主导地位。
此外，测量了心跳在五个不同时刻的涡量分布。比较模糊图像和去模糊图像的结果参见图11，其是5个时间帧的模糊图像和去模糊图像的涡量分布直方图和量化涡量的比较。在涡量分布直方图中，分别示意了模糊图像和去模糊图像的涡量分布。σ b和σ r分别是模糊图像和去模糊图像的涡量分布标准差，h b和h r分别表示模糊图像和去模糊图像的熵值。可以看出，去模糊图像可准确识别涡旋位置。此外，去模糊图像对应的涡量分布更加离散，颜色表示更多，且熵值更高，而模糊图像完全无法识别涡旋位置，涡量分布接近于0，熵值较低且颜色表示较少。这表明本发明使用的测量方法非常有效。
综上，运动模糊图像中存在明显不同的方向，通过在去模糊之前使用模糊方向的信息进行预分类，可以简化去模糊模型的目标问题，即只需要对相同模糊方向的模糊图像进行去模糊处理，从而减小了解空间的大小，使模型更容易拟合。此外，通过引入多个SRN-Deblur子模型，在训练数据少、训练图像模糊条件不同的情况下，可以校正不同清晰度的图像，增强了模型鲁棒性。
综上所述，本发明提出了一种新模型，是将ResNet和多个SRN-Deblur模型组合成一体的去模糊模型。使用ResNet模型进行模糊分类，分类准确度超过99％。根据分类结果，训练了四个SRN-Deblur子模型，以对图像进行去模糊，训练后的四个SRN-Deblur子模型可输出优质的去模糊图像。最后，将该模型的去模糊结果与单独使用SRN-Deblur去模糊的结果进行了比较。结果表明，本发明更适用于复杂的情况。可将不同的模型用于不同类型的图像，并且可以处理SRN-Deblur无法处理的图像。在VENC MRI中，去模糊图像与实际图像的差异明显小于模糊图像与实际图像的差异。实验结果表明，去模糊图像的涡量比模糊图像更接近清晰图像。此外，本发明模型在目视检查和数学评价方面均优于现有技术，去除模糊因素后，所获得的清晰VENC MRI可以帮助放射科医生和临床医生做出更好的临床判断并提高诊断准确度，并可用于更广泛的领域。
本发明可以是系统、方法和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质，其上载有用于使处理器实现本发明的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是但不限于电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、 以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备，或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令，并转发该计算机可读程序指令，以供存储在各个计算/处理设备中的计算机可读存储介质中。
用于执行本发明操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码，所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++、Python等，以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中，通过利用计算机可读程序指令的状态信息来个性化定制电子电路，例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA)，该电子电路可以执行计算机可读程序指令，从而实现本发明的各个方面。
这里参照根据本发明实施例的方法、装置(系统)和计算机程序产品的流程图和/或框图描述了本发明的各个方面。应当理解，流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合，都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器，从而生产出一种机器，使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时，产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算机可读程序指令存储在计算机可读存储介质中，这些指令使得计算机、可编程数据处理装置和/或其他设备以特定方式工作，从而，存储有指令的计算机可读介质则包括一个制造品，其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设备上，使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤，以产生计算机实现的过程，从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本发明的多个实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分，所述模块、程序段或指令的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。对于本领域技术人员来说公知的是，通过硬件方式实现、通过软件方式实现以及通过软件和硬件结合的方式实现都是等价的。
以上已经描述了本发明的各实施例，上述说明是示例性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原 理、实际应用或对市场中的技术改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。本发明的范围由所附权利要求来限定。

Claims (10)

  1. 一种速度编码磁共振成像的运动伪影校正方法，包括以下步骤：
    利用经训练的分类模型确定速度编码磁共振图像的模糊方向类型；
    根据所述模糊方向类型选择对应的去模糊子模型进行去模糊处理，获得校正图像，其中，所述去模糊子模型经训练获得，且所述去模糊子模型的数量与所述模糊方向类型的数量相同，每个去模糊子模型对应一类模糊方向。
  2. 根据权利要求1所述的方法，其特征在于，根据以下步骤训练所述分类模型：
    采集心脏断层扫描的速度编码磁共振图像，每层获得三幅图像，包括正常图像、前后方向图像和足头方向的图像；
    对所采集的图像进行模糊处理，并标注图像的模糊方向类型，包括0°、45°、90°和135°；
    构建训练数据集，该训练数据集反映经模糊处理图像与模糊方向类型之间的对应关系；
    以设定的损失函数最小化为优化目标，利用所述训练数据集训练分类模型，获得优化参数。
  3. 根据权利要求1所述的方法，其特征在于，所述模糊处理包括：
    针对采集的图像，通过图像平移、复制平移图像后叠加以及计算平均像素值的方式生成模糊图像，其中平移方向、平移步长和叠加次数均随机生成。
  4. 根据权利要求1所述的方法，其特征在于，所述分类模型是残差网络，依次包括第一卷积层、批归一化层、激活层、多个残差块、第二激活层、平均池化层和全连接层。
  5. 根据权利要求1所述的方法，其特征在于，所述速度编码磁共振图像的成像参数设置为：回波时间TR是47.1ms，重复时间TE是1.6ms，视场FOV是298×340mm 2，像素矩阵为134×256，平面分辨率是1.54mm/像素，基于层间距的通过平面的分辨率为6mm。
  6. 根据权利要求1所述的方法，其特征在于，所述去模糊子模型是尺 度循环结构和编码器-解码器残差块网络。
  7. 根据权利要求1所述的方法，其特征在于，还包括：利用所述校正图像识别心脏内涡旋并计算涡量。
  8. 根据权利要求1所述的方法，其特征在于，所述损失函数是交叉熵函数。
  9. 一种计算机可读存储介质，其上存储有计算机程序，其中，该计算机程序被处理器执行时实现根据权利要求1至8中任一项所述的方法的步骤。
  10. 一种计算机设备，包括存储器和处理器，在所述存储器上存储有能够在处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现权利要求1至8中任一项所述的方法的步骤。
PCT/CN2022/133813 2022-11-23 2022-11-23 一种速度编码磁共振成像的运动伪影校正方法 WO2024108438A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/133813 WO2024108438A1 (zh) 2022-11-23 2022-11-23 一种速度编码磁共振成像的运动伪影校正方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/133813 WO2024108438A1 (zh) 2022-11-23 2022-11-23 一种速度编码磁共振成像的运动伪影校正方法

Publications (1)

Publication Number Publication Date
WO2024108438A1 true WO2024108438A1 (zh) 2024-05-30

Family

ID=91194979

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/133813 WO2024108438A1 (zh) 2022-11-23 2022-11-23 一种速度编码磁共振成像的运动伪影校正方法

Country Status (1)

Country Link
WO (1) WO2024108438A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210125037A1 (en) * 2019-10-28 2021-04-29 Ai4Medimaging - Medical Solutions, S.A. Artificial intelligence based cardiac motion classification
CN114565711A (zh) * 2021-12-28 2022-05-31 中南大学 基于深度学习的心脏图像重建方法及系统
CN115063309A (zh) * 2021-06-07 2022-09-16 上海联影智能医疗科技有限公司 运动伪影模拟器及相应方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210125037A1 (en) * 2019-10-28 2021-04-29 Ai4Medimaging - Medical Solutions, S.A. Artificial intelligence based cardiac motion classification
CN115063309A (zh) * 2021-06-07 2022-09-16 上海联影智能医疗科技有限公司 运动伪影模拟器及相应方法
CN114565711A (zh) * 2021-12-28 2022-05-31 中南大学 基于深度学习的心脏图像重建方法及系统

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LIU, FENG: "Simulation And Application of Inverse Filter Eliminating Motion Artifacts in Magnetic Resonance Images", SCIENCE AND TECHNOLOGY INNOVATION HERALD, no. 19, 1 July 2013 (2013-07-01) *
ZWART NICHOLAS R, PIPE JAMES G.: "Multidirectional high‐moment encoding in phase contrast MRI", MAGNETIC RESONANCE IN MEDICINE, WILEY-LISS, US, vol. 69, no. 6, 1 June 2013 (2013-06-01), US , pages 1553 - 1563, XP093173041, ISSN: 0740-3194, DOI: 10.1002/mrm.24390 *

Similar Documents

Publication Publication Date Title
US9968257B1 (en) Volumetric quantification of cardiovascular structures from medical imaging
Shaw et al. MRI k-space motion artefact augmentation: model robustness and task-specific uncertainty
Sander et al. Automatic segmentation with detection of local segmentation failures in cardiac MRI
CN109949349B (zh) 一种多模态三维图像的配准及融合显示方法
US8682051B2 (en) Smoothing of dynamic data sets
Piccini et al. Deep learning to automate reference-free image quality assessment of whole-heart MR images
JP2002539870A (ja) 画像処理の方法と装置
US20230394670A1 (en) Anatomically-informed deep learning on contrast-enhanced cardiac mri for scar segmentation and clinical feature extraction
US9196049B2 (en) Method and system for regression-based 4D mitral valve segmentation from 2D+t magnetic resonance imaging slices
CN114565711A (zh) 基于深度学习的心脏图像重建方法及系统
Laumer et al. Weakly supervised inference of personalized heart meshes based on echocardiography videos
Beache et al. Fully automated framework for the analysis of myocardial first‐pass perfusion MR images
CN110197472B (zh) 一种用于超声造影图像稳定定量分析的方法和系统
Sander et al. Autoencoding low-resolution MRI for semantically smooth interpolation of anisotropic MRI
Ossenberg-Engels et al. Conditional generative adversarial networks for the prediction of cardiac contraction from individual frames
Wright et al. Complete fetal head compounding from multi-view 3D ultrasound
CN113222985B (zh) 一种图像处理方法、装置、计算机设备和介质
WO2024108438A1 (zh) 一种速度编码磁共振成像的运动伪影校正方法
Zhang et al. Multi-scale network with the deeper and wider residual block for MRI motion artifact correction
CN110930394B (zh) 测量肌肉肌纤维束线斜率和羽状角的方法及终端设备
Arega et al. Using polynomial loss and uncertainty information for robust left atrial and scar quantification and segmentation
WO2014106747A1 (en) Methods and apparatus for image processing
CN116385572A (zh) 一种速度编码磁共振成像的运动伪影校正方法
JP2016171961A (ja) 画像処理装置、画像処理方法、及びプログラム
Paul et al. EchoTrace: A 2D Echocardiography Deep Learning Approach for Left Ventricular Ejection Fraction Prediction