WO2021087659A1 - 多模态图像配准的方法、装置、电子设备及存储介质 - Google Patents

多模态图像配准的方法、装置、电子设备及存储介质 Download PDF

Info

Publication number
WO2021087659A1
WO2021087659A1 PCT/CN2019/115311 CN2019115311W WO2021087659A1 WO 2021087659 A1 WO2021087659 A1 WO 2021087659A1 CN 2019115311 W CN2019115311 W CN 2019115311W WO 2021087659 A1 WO2021087659 A1 WO 2021087659A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
trained
intensity
corrected
modal
Prior art date
Application number
PCT/CN2019/115311
Other languages
English (en)
French (fr)
Inventor
王珊珊
郑海荣
黄纬键
刘新
梁栋
Original Assignee
中国科学院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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Priority to PCT/CN2019/115311 priority Critical patent/WO2021087659A1/zh
Publication of WO2021087659A1 publication Critical patent/WO2021087659A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods

Definitions

  • This application relates to the field of pattern recognition technology, and in particular to a method, device, electronic device, and storage medium for multi-modal image registration.
  • Image registration is the positioning and conversion of images. For example, in the field of medical image processing, by looking for a spatial transformation, the corresponding points of the two images can reach the same spatial position and anatomical structure.
  • the purpose of image registration is to compare or fuse images acquired under different conditions for the same object.
  • the registration technology for single-modal images cannot adapt to the difference of multi-modal images, resulting in low registration accuracy. Therefore, a registration method for multi-modal images is required.
  • One of the objectives of the embodiments of the present application is to provide a method, device, electronic device, and storage medium for multi-modal image registration, aiming to solve the problem of multi-modal image registration.
  • a method for multi-modal image registration including:
  • a device for multi-modal image registration including:
  • An image acquisition module for acquiring a first image of a source modality and a second image of a target modality paired with the first image
  • An intensity correction module configured to correct the intensity of the first image to obtain a first corrected image, so that the intensity of the first corrected image matches the intensity distribution of the second image;
  • a deformation field acquisition module configured to acquire the deformation field of the first correction image registered to the target modality according to the first correction image and the second image;
  • the registration module is configured to obtain a registered image of the first image registered to the target modality according to the first image and the deformation field.
  • an electronic device which includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor.
  • the processor executes the computer program, the first The method described in the aspect.
  • a computer-readable storage medium including: the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in the first aspect are implemented.
  • a computer program product is provided.
  • the computer program product runs on an electronic device, the electronic device executes the method steps described in the first aspect.
  • the beneficial effect of the method for multi-modal image registration is that the intensity of the first image to be registered is corrected in advance to match the intensity distribution with the intensity of the target mode, and according to the first correction
  • the image and the second image acquire the deformation field from the first image to the target modal, which reduces the influence of image intensity characteristics, and can avoid the large difference in the distance between the same object in the first image to be corrected and the ideal registration result, resulting in a decrease in accuracy , Thereby improving the registration accuracy.
  • FIG. 1 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
  • FIG. 2 is a schematic flowchart of a multi-modal image registration method provided by an embodiment of the present application
  • 3a is a partial schematic diagram of the strength correction of the data processing architecture of the multi-modal image registration method provided by an embodiment of the present application;
  • 3b is a schematic diagram of the registration part of the data processing architecture of the multi-modal image registration method provided by an embodiment of the present application;
  • 4a is a partial schematic diagram of the strength correction of the data processing architecture of the multi-modal image registration method provided by another embodiment of the present application.
  • 4b is a schematic diagram of the registration part of the data processing architecture of the multi-modal image registration method provided by another embodiment of the present application.
  • FIG. 5 is a schematic diagram of a cyclic generation confrontation network provided by an embodiment of the present application.
  • FIG. 6 is a schematic flowchart of a multi-modal image registration method provided by another embodiment of the present application.
  • FIG. 7 is a schematic diagram of a codec network provided by an embodiment of the present application.
  • FIG. 8 is a schematic structural diagram of a multi-modal image registration device provided by an embodiment of the present application.
  • the purpose of image registration is to compare or fuse images acquired under different conditions for the same object. Especially it plays an important role in many applications in the field of medical image analysis.
  • medical image analysis such as Magnetic Resonance Imaging (MRI) image analysis
  • MRI Magnetic Resonance Imaging
  • the T1 sequence is suitable for observing anatomical structures
  • the T2 sequence is more focused on observing tissue lesions.
  • Accurate disease analysis often requires comparing and analyzing multi-modal images of the same area (object), such as ultrasound images, MRI images, Computed Tomography (CT) images, or fusion processing of multi-modal images to obtain more Fusion image of clinical guidance value.
  • CT Computed Tomography
  • CNN Convolutional Neural Network
  • the problem with the CNN model for image registration is that the template must be universal. For example, in the application of medical imaging, in some images with lesions that have occurred, tissue deformations that can be seen everywhere are often accompanied by the lesions. If the severely deformed images are registered to a fixed standard template at this time, it will often lead to registration. Inaccurate, if the damaged tissue returns to normal after registration, this will affect the accuracy of subsequent diagnostic analysis.
  • Another problem with the image registration method using the CNN model is that the method cannot be applied to medical image processing in different modalities.
  • this method cannot be optimized by a simple loss function.
  • T1 and T2 sequences of MRI as an example.
  • T1 shows low signal to cerebrospinal fluid
  • T2 shows high signal.
  • both modalities show high signal. Therefore, the commonly used mean square error or cross entropy loss function cannot be used for unified processing.
  • the embodiment of the present application provides a first image obtained by acquiring a source modality and a target modality paired with the first image. State of the second image; correcting the intensity of the first image to obtain a first corrected image; the intensity of the first corrected image matches the intensity distribution of the second image; according to the first corrected image and the The second image acquires the deformation field of the first corrected image registered to the target modality; and the registration image of the first image registered to the target modality is acquired according to the first image and the deformation field.
  • the first image to the target mode is acquired according to the first corrected image and the second image.
  • the deformation field reduces the influence of image intensity characteristics, and can avoid the excessive difference in the distance between the same object of the first image to be corrected and the ideal registration result resulting in a decrease in accuracy, thereby improving the accuracy of registration.
  • a trained recurrent generation confrontation network is used to correct the intensity of the first image according to the first image and the second image to obtain the first corrected image.
  • the first corrected image can be made The intensity of the object is close to the second image while retaining the characteristics of the first image, thereby improving the accuracy of the deformation field obtained from the second image and the first corrected image.
  • the use of a cyclic generation confrontation network to correct the image of the source mode according to the image of the source mode and the image of the target mode of the same object can avoid errors caused by the use of a fixed template, thereby improving the registration accuracy.
  • using the first loss function including deformation loss to train the recurrent generation confrontation network to be trained can correct the intensity of the first image while retaining more shape features of the first image in the first corrected image. Therefore, the accuracy of obtaining the deformation field from the first corrected image to the target modality according to the first corrected image and the second image can be further improved.
  • Figure 1 shows an electronic device D10 provided by an embodiment of the present application, including: at least one processor D100, a memory D101, and a computer program stored in the memory D101 and capable of running on the at least one processor D100 D102.
  • the processor D100 implements at least one of the multi-modal image registration methods provided in the embodiment of the present application when the computer program D102 is executed by the processor D100.
  • the above-mentioned electronic devices may be computing devices such as desktop computers, notebooks, palmtop computer servers, server clusters, distributed servers, and cloud servers.
  • the electronic device D10 may include, but is not limited to, a processor D100 and a memory D101.
  • FIG. 1 is only an example of the electronic device D10, and does not constitute a limitation on the electronic device D10. It may include more or less components than shown in the figure, or a combination of certain components, or different components. , For example, can also include input and output devices, network access devices, and so on.
  • the so-called processor D100 may be a central processing unit (Central Processing Unit, CPU), and the processor D100 may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), and application specific integrated circuits (Application Specific Integrated Circuits). , ASIC), ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the memory D101 may be an internal storage unit of the electronic device D10, such as a hard disk or a memory of the electronic device D10.
  • the memory D101 may also be an external storage device of the electronic device D10, for example, a plug-in hard disk equipped on the electronic device D10, a smart memory card (Smart Media Card, SMC), and a secure digital (Secure Digital, SD) card, flash card (Flash Card), etc.
  • the memory D101 may also include both an internal storage unit of the electronic device D10 and an external storage device.
  • the memory D101 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program.
  • the memory D101 can also be used to temporarily store data that has been output or will be output.
  • the above-mentioned electronic devices are collectively referred to as image processing devices in the following embodiments, and it should be understood that they do not constitute a specific limitation on the electronic devices of the present application.
  • FIG. 2 shows a method for multi-modal image registration provided by an embodiment of the present application, which is applied to the electronic device shown in FIG. 1, hereinafter referred to as an image processing device, which can be used by the software/hardware of the image processing device achieve.
  • the method includes steps S110 to S140.
  • the specific implementation principles of each step are as follows:
  • S110 Acquire a first image of a source modality and a second image of a target modality paired with the first image.
  • the image processing device acquires a first image of the source modality, for example, a CT image of the head of a certain subject; and a second image of the target modality paired with the first image, For example, an MRI T1 sequence image of the subject's head.
  • the paired images here refer to images of different modalities of the same object. It is understandable that image registration technology is widely used in the field of medical imaging. Most of the examples in this application take medical image processing as an example, but the image registration method provided in the embodiments of this application can also be applied to other images. Processing fields, such as machine vision field, virtual/enhanced display field, and other fields that require image comparison and fusion, will not be repeated here.
  • S120 Correct the intensity of the first image to obtain a first corrected image, so that the intensity of the first corrected image matches the intensity distribution of the second image.
  • the image processing device corrects the intensity of the first image to obtain the first corrected image.
  • the image processing device corrects the intensity of the first image by using a preset statistical model; for another example, after the image processing device recognizes the region of interest in the first image, after segmenting the region of interest, the intensity of the perceptual region The intensity is corrected; for another example, the image processing device corrects the intensity of the first image through the trained neural network model; thereby obtaining the first corrected image.
  • At least one of the above example methods and other methods that can adjust the image intensity can be used to correct the intensity of the first image so that the intensity of the first image matches the intensity distribution of the second image, that is,
  • the overall image intensity distribution of the first modified image is similar to the intensity distribution of the second image, or the image intensity distribution of the region of interest of the first modified image is similar to the region of interest distribution of the second image.
  • the intensity distribution of the skeletal part of the corrected cranial anterior CT image obtained after intensity correction of the cranial orthographic CT image is similar to the intensity distribution of the skeletal part of the cranial orthographic image of the target modal MRI T1 sequence.
  • the intensity (grayscale) can be corrected.
  • the intensity of one or more channels can be selected for correction according to the actual situation to facilitate subsequent follow-up. Processing, I won't repeat it here.
  • S130 Acquire a deformation field of the first corrected image registered to the target modality according to the first corrected image and the second image.
  • the image processing device obtains the deformation field from the first corrected image to the target modality according to the first corrected image and the second image through a trained unsupervised or semi-supervised neural network model.
  • S140 Acquire a registered image in which the first image is registered to a target modality according to the first image and the deformation field.
  • the image processing device acquires the registered image of the target modality from the first image before correction and the deformation field according to the first image before correction. For example, through the deformation field obtained in step S130, the head orthographic CT image is registered to the image of the MRI T1 sequence. It can be understood that the obtained registration image is adjusted in the corresponding pixel position of the head, but the intensity remains unchanged or the intensity changes more. Few images.
  • Figures 3a and 3b show a non-limiting example.
  • the first corrected image is obtained by correcting the fore degree of the first image of the image intensity correction model.
  • the deformation field where the first corrected image is registered to the target modality is obtained through the deformation field acquisition model.
  • the first image is registered to the registered image of the target modality through the deformation field.
  • the registered image is the pixel position in the first image adjusted to the corresponding position of the target modality, but the intensity does not change.
  • the first image to the target mode is acquired according to the first corrected image and the second image.
  • the deformation field reduces the influence of image intensity characteristics, and can avoid the excessive difference in the distance between the same object of the first image to be corrected and the ideal registration result resulting in a decrease in accuracy, thereby improving the accuracy of registration.
  • FIGS. 4a and 4b show another multi-modal image registration method provided by an embodiment of the present application.
  • a cyclic generation confrontation network is used to correct the intensity of the first image according to the first image and the second image to obtain the first corrected image, so that the intensity of the first corrected image matches the intensity distribution of the second image.
  • the intensity of the object in the first corrected image is close to that of the second image while retaining the characteristics of the first image, thereby improving the accuracy of the deformation field obtained from the second image and the first corrected image.
  • the use of a cyclic generation confrontation network to correct the image of the source mode according to the image of the source mode and the image of the target mode of the same object can avoid errors caused by the use of a fixed template, thereby improving the registration accuracy.
  • the recurrent generation confrontation network shown in FIG. 5 is composed of two generators and two discriminators.
  • the X domain corresponds to the source mode
  • the Y domain corresponds to the target mode.
  • X and Y are data sets of two modalities
  • G is a generator that generates Y-domain images from X-domain images
  • F is a generator that generates X-domain images from Y-domain images in the opposite direction of G.
  • D Y and D X are the discriminators corresponding to the two generators, and their role is to promote the approximation of the distribution of the generator results to the target domain.
  • the function of the generator is to synthesize a new image, which can be realized by a U-Net network.
  • the role of the discriminator is to judge the credibility of the synthesized image, which can be a VGG classification network. It is understandable that those skilled in the art can select a suitable generator network or discriminator network according to actual implementation conditions under the teaching of the embodiments of the present application.
  • the input is the X-domain image x
  • the gold standard is the Y-domain image y paired with the image x
  • the output is the composite image q, at this time the image q and the image y are regarded as D Y Input, judge the authenticity of image q and image y.
  • the path is that the image y and the image q are obtained by the generator F after q', and then the true or false of x and q'are judged by D X.
  • the learning rate of the network is set to 0.0001, and the optimizer selects Adam.
  • a trained generative adversarial network model is obtained.
  • the cyclic generation confrontation network training is completed, input any image x in the X domain and an image y in the Y domain to obtain an image x with intensity correction.
  • the image x has a similar intensity distribution in the Y domain.
  • the cyclic generation confrontation network shown in Figure 5 Train the cyclic generation confrontation network shown in Figure 5 through the loss function formula (1), and the trained cyclic generation confrontation network will correct the intensity of the X-domain image, and it will have the shape feature of the Y-domain, in order to improve the registration Accuracy, on the basis of the loss function represented by the formula (1), the first loss function formula (5) including the deformation loss is introduced to train the to-be-trained cyclic generation confrontation network.
  • the deformation loss is obtained according to the difference between the input image of the generator of the cyclic generation confrontation network and the corresponding parameter representing the shape feature in the output image of the generator. .
  • ⁇ is a weighting factor coefficient, which determines the proportion of the deformation loss in the overall loss, and can be selected and preset according to actual conditions when implementing this embodiment.
  • the deformation loss only acts on the background of the input image to limit the change of its shape.
  • the image may be preprocessed first to remove noise, or the intensity of the background pixel may be adjusted to zero.
  • the parameter characterizing the shape feature includes at least one of the following parameters: the intensity of the pixel point characterizing the shape feature, the border length of the foreground image, and the area of the foreground image; wherein the pixel point characterizing the shape includes at least the following pixels One: the pixels that characterize the background, the pixels that characterize the edge of the image, and the pixels that characterize the contour of the region of interest in the image.
  • the input is the image x in the X domain
  • the gold standard is the image y paired with the image x in the Y domain.
  • the output is synthesized Image q.
  • the image q and the image y are regarded as the input of D Y , and the authenticity of the image q and the image y is determined.
  • the other branch corresponds to it, the path is that the image y and the image q are obtained by the generator F after q', and then the true or false of x and q'are judged by D X.
  • the learning rate of the network is set to 0.0001, and the optimizer selects the Adam optimizer.
  • a trained generative adversarial network model is obtained.
  • input any image x in the X domain and an image y in the Y domain to obtain an image x with intensity correction.
  • the corrected image x has a similar intensity distribution in the Y domain, but the corrected image x retains the shape characteristics of the image x.
  • the input data are paired MRI T1 sequence images and MRI T2 sequence images of the same patient, and the data is preprocessed into a unified image size of the two sets of data, such as 192 ⁇ 192 ⁇ 1.
  • the output is a T1 corrected image.
  • the T1 corrected image has an intensity distribution similar to that of T2, and the T1 corrected image retains the shape characteristics of the T1 image.
  • training the recurrent generation confrontation network to be trained using the first loss function including deformation loss can correct the intensity of the first image while retaining more shape features of the first image in the first corrected image. Therefore, the accuracy of obtaining the deformation field from the first corrected image to the target mode according to the first corrected image and the second image can be further improved.
  • FIG. 6 shows another multi-modal image registration method provided by an embodiment of the present application, as shown in FIG. 5
  • step S130 acquiring the deformation field from the first corrected image to the target modality according to the first corrected image and the second image, includes:
  • the image processing device uses a trained codec network to obtain the deformation field from the first corrected image to the target modality according to the first corrected image and the second image.
  • the codec network is a codec network formed by connecting a deep convolutional network with a U-Net network as the backbone and a deep deconvolution network as shown in Fig. 7.
  • C in Fig. 7 represents the convolution process
  • U represents the deconvolution process.
  • the input data of the network is the first modified image q generated by the cyclic generation countermeasure network according to the first image x and the second image y, and the output is the estimation of the second image y and the first modified image q by the codec network
  • the first corrected image and the second image are grayscale images with dimensions of 192 ⁇ 192 ⁇ 1
  • the output of the network For an image with a dimension of 192 ⁇ 192 ⁇ 1, a deformation field layer is set in front of the output layer of the network, which provides a pixel shift gradient during back propagation.
  • the function of the deformation field is to offset each pixel of the input image, the dimension is 192 ⁇ 192 ⁇ 2, the first dimension of the last channel is the pixel displacement length, and the second dimension is the pixel displacement direction.
  • the image x of the source mode passes through the application module applying the deformation field to obtain the registered image x′.
  • using a trained codec network to acquire the first corrected image before the deformation field of the target modality according to the first corrected image and the second image further includes:
  • the intensity distribution of the modal image is matched, and the source modal correction image and the target modal image sample set are obtained; the source modal correction image and the target modal image sample set are used to train the codec network to be trained to obtain the trained codec network.
  • a single set of paired samples are MRI T1 sequence (source modality) images and T2 sequence (target modality) images of a specific part of an object, and the training sample set is multiple groups of images of different objects.
  • Training the codec network through paired samples can adjust the parameters of the codec network according to the data of the two modalities of the same object, so as to achieve the effect of not requiring a fixed template when applying the codec network, only the pairing of the same object is required.
  • the two modal images can obtain the deformation field.
  • the optimizer for training the coding network selects the Adam optimizer with a learning rate of 0.0001 and 100 batches of training.
  • the encoding and decoding network to be trained is trained, and the second loss function is used to train the encoding and decoding network to be trained.
  • the loss of the second loss function includes the output image of the encoding and decoding network to be trained and the The difference between the target modal images in the training sample set.
  • the second loss function shown in formula (8) is used to train the codec network.
  • T is the codec network
  • G is the generator of the cyclically generated confrontation network
  • X is the source modal image
  • Y is the target modal image.
  • MSE mean square error
  • the cyclic generation confrontation network generates the first corrected image according to the first image and the second image, and uses the first corrected image and the second image as input to the codec network shown in FIG. 7.
  • FIG. 8 shows a multi-modal image registration device provided by an embodiment of the present application, including:
  • the image acquisition module M110 is used to acquire the first image of the source modality and the second image of the target modality paired with the first image.
  • the intensity correction module M120 is configured to correct the intensity of the first image to obtain a first corrected image, so that the intensity of the first corrected image matches the intensity distribution of the second image.
  • the deformation field acquisition module M130 is configured to acquire the deformation field registered to the target modality by the first correction image according to the first correction image and the second image.
  • the registration module M140 is configured to obtain a registration image in which the first image is registered to a target modality according to the first image and the deformation field.
  • the intensity module M120 is used to correct the intensity of the first image to obtain a first corrected image, so that the intensity of the first corrected image matches the intensity distribution of the second image, including:
  • the cyclically generated confrontation network module M121 is configured to use a cyclically generated confrontation network to correct the intensity of the first image according to the first image and the second image to obtain a first corrected image, so that the first corrected image The intensity of matches the intensity distribution of the second image.
  • the cyclic generation confrontation network module M121 which uses a trained cyclic generation confrontation network, and before correcting the intensity of the first image according to the first image and the second image, further includes:
  • the cyclically generated confrontation network training module M121' is used to train the cyclically generated confrontation network to be trained by using the training sample set composed of the paired source modal image and the target modal image to obtain the trained cyclically generated confrontation network.
  • the recurrent generation confrontation network training module M121’ trains the recurrent generation confrontation network to be trained, and further includes:
  • Cyclic generation confrontation network loss function module M1211 used to train the to-be-trained cyclic generation confrontation network by using the first loss function including deformation loss; wherein the input image of the generator of the cyclic generation confrontation network and the generation The deformation loss is obtained by the difference of the parameter characterizing the shape feature in the output image of the filter.
  • the parameter that characterizes the shape feature includes at least one of the following parameters: the intensity of the pixel point that characterizes the shape feature, the border length of the foreground image, and the area of the foreground image; wherein the pixel point that characterizes the shape includes at least one of the following pixel points: characterization The pixels of the background, the pixels that characterize the edge of the image, and the pixels that characterize the contour of the region of interest in the image.
  • the deformation field acquisition module M130 is configured to acquire the deformation field from the first corrected image to the target modality according to the first corrected image and the second image, including:
  • the codec network module M1301 is configured to use a trained codec network to obtain the deformation field from the first corrected image to the target modality according to the first corrected image and the second image.
  • the codec network module M1301 is configured to use a trained codec network to obtain the first corrected image before the deformation field of the target modality according to the first corrected image and the second image, and further includes:
  • the codec network training module M1301' is used to correct the intensity of the source modal image in the training sample set composed of the paired source modal image and the target modal image, so that the source modal image and the source modal image are combined with the source modal image in the sample set. Matching the intensity distribution of the target modal image of the source modal image pairing to obtain the source modal correction image and the target modal image sample set;
  • the codec network training module M1301' is also used to train the codec network to be trained by using the source modal correction image and the target modal image sample set to obtain a trained codec network.
  • the codec network training module M1301' is further configured to train the codec network to be trained using a second loss function including the difference between the output image of the codec network to be trained and the target modal image in the training sample set .
  • the electronic device shown in FIG. 1 is adopted, and the electronic device includes: a memory, a processor, and a computer program stored in the memory and running on the processor.
  • the processor executes the computer program:
  • the implementation of: correcting the intensity of the first image to obtain the first corrected image includes: using a trained loop to generate a confrontation network, according to all The first image and the second image correct the intensity of the first image to obtain a first corrected image, so that the intensity of the first corrected image matches the intensity distribution of the second image.
  • the processor when the processor executes the computer program, it realizes: adopting a trained loop to generate a confrontation network, according to the intensity of the first image and the second image against the first image Before the correction, it also includes: using a training sample set composed of paired source modal images and target modal images to train the recurrent generation confrontation network to be trained to obtain the trained recurrent generation confrontation network.
  • the processor when the processor executes the computer program, it implements: training the loop generation adversarial network to be trained, and further includes: training the loop generation to be trained using a first loss function including deformation loss A confrontation network; wherein, the deformation loss is obtained according to the difference between the input image of the generator of the cyclically generated confrontation network and the output image of the generator corresponding to the parameter characterizing the shape feature.
  • the pixel point that characterizes the shape includes at least one of the following pixel points: a pixel point that characterizes the background, a pixel point that characterizes the edge of the image, and a pixel point that characterizes the contour of the region of interest in the image.
  • the processor executes the computer program, it is realized that: the deformation field from the first corrected image to the target modality is acquired according to the first corrected image and the second image , Including: acquiring a deformation field from the first corrected image to the target modality according to the first corrected image and the second image by using a trained codec network.
  • a trained codec network is used to obtain the first corrected image to the target according to the first corrected image and the second image.
  • a trained codec network In front of the modal deformation field, it also includes:
  • the codec network to be trained is trained using the source modality correction image and the target modality image sample set to obtain a trained codec network.
  • the processor executes the computer program to implement: training the codec network to be trained, including: adopting the target including the output image of the codec network to be trained and the training sample set The second loss function of the difference of the modal image trains the codec network to be trained.
  • the embodiments of the present application also provide a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in each of the foregoing method embodiments can be realized.
  • the embodiments of the present application provide a computer program product.
  • the computer program product runs on an electronic device, the electronic device can realize the steps in the foregoing method embodiments when the electronic device is executed.
  • the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the computer program can be stored in a computer-readable storage medium.
  • the computer program can be stored in a computer-readable storage medium.
  • the steps of the foregoing method embodiments can be implemented.
  • the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file, or some intermediate forms.
  • the computer-readable medium may at least include: any entity or device capable of carrying computer program code to the photographing device/terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (Random Access Memory, RAM), electric carrier signal, telecommunications signal, and software distribution medium.
  • ROM read-only memory
  • RAM random access memory
  • electric carrier signal telecommunications signal
  • software distribution medium for example, U disk, mobile hard disk, floppy disk or CD-ROM, etc.
  • computer-readable media cannot be electrical carrier signals and telecommunication signals.
  • the disclosed apparatus/network equipment and method may be implemented in other ways.
  • the device/network device embodiments described above are only illustrative.
  • the division of the modules or units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units.
  • components can be combined or integrated into another system, or some features can be omitted or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

本申请公开一种多模态图像配准的方法，该多模态图像配准的方法包括获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像；对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配；根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场；根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。降低了因图像强度特征的影响，可以避免待修正的第一图像的同一对象与理想配准结果的距离差异过大导致精度下降，从而提高了配准精度。

Description

多模态图像配准的方法、装置、电子设备及存储介质 技术领域
本申请涉及模式识别技术领域，具体涉及一种多模态图像配准的方法、装置、电子设备及存储介质。
背景技术
这里的陈述仅提供与本申请有关的背景信息，而不必然构成现有技术。图像配准是对图像的定位和转换，例如，在医学图像处理领域，通过寻找一种空间变换，使两幅图像对应点达到空间位置和解剖结构上的一致。图像配准的目的是比较或融合针对同一对象不同条件下获取的图像。在多模态图像配准场景下，用于单一模态的图像的配准技术无法适应多模态图像的差异，导致配准精度低，因此需要一种针对多模态图像的配准方法。
技术问题
本申请实施例的目的之一在于：提供多模态图像配准的方法、装置、电子设备及存储介质，旨在解决多模态图像配准的问题。
技术解决方案
为解决上述技术问题，本申请实施例采用的技术方案是：
第一方面，提供了一种多模态图像配准的方法，包括：
获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像；对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配；根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场；根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。
第二方面，提供了一种多模态图像配准的装置，包括：
图像获取模块，用于获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像；
强度修正模块，用于对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配；
形变场获取模块，用于根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场；
配准模块，用于根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的 配准图像。
第三方面，提供一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如上述第一方面所述的方法。
第四方面，提供了一种计算机可读存储介质，包括：所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现上述第一方面所述的方法步骤。
第五方面，提供了一种计算机程序产品，当计算机程序产品在电子设备上运行时，使得电子设备执行上述第一方面所述的方法步骤。
本申请实施例提供的一种多模态图像配准的方法的有益效果在于：通过预先修正待配准的第一图像的强度，使其强度分布与目标模态的强度匹配，根据第一修正图像和第二图像获取第一图像到目标模态的形变场，降低了因图像强度特征的影响，可以避免待修正的第一图像的同一对象与理想配准结果的距离差异过大导致精度下降，从而提高了配准精度。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例或示范性技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1是本申请一实施例提供的电子设备的结构示意图；
图2是本申请一实施例提供的多模态图像配准方法的流程示意图；
图3a是本申请一实施例提供的多模态图像配准方法的数据处理架构强度修正部分示意图；
图3b是本申请一实施例提供的多模态图像配准方法的数据处理架构配准部分示意图；
图4a是本申请另一实施例提供的多模态图像配准方法的数据处理架构强度修正部分示意图；
图4b是本申请另一实施例提供的多模态图像配准方法的数据处理架构配准部分示意图；
图5是本申请一实施例提供的循环生成对抗网络示意图；
图6是本申请另一实施例提供的多模态图像配准方法的流程示意图；
图7是本申请一实施例提供的编解码网络示意图；
图8是本申请一实施例提供的多模态图像配准装置的结构示意图。
本发明的实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本发明，并不用于限定本申请。
需说明的是，当部件被称为“固定于”或“设置于”另一个部件，它可以直接在另一个部件上或者间接在该另一个部件上。当一个部件被称为是“连接于”另一个部件，它可以是直接或者间接连接至该另一个部件上。术语“上”、“下”、“左”、“右”等指示的方位或位置关系为基于附图所示的方位或位置关系，仅是为了便于描述，而不是指示或暗示所指的装置或元件必须具有特定的方位、以特定的方位构造和操作，因此不能理解为对本申请的限制，对于本领域的普通技术人员而言，可以根据具体情况理解上述术语的具体含义。术语“第一”、“第二”仅用于便于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明技术特征的数量。“多个”的含义是两个或两个以上，除非另有明确具体的限定。
图像配准的目的是比较或融合针对同一对象不同条件下获取的图像。尤其在医学图像分析领域的许多应用中扮演重要角色。在医学影像分析中，例如磁共振成像(Magnetic Resonance Imaging，MRI)影像分析，具有不同侧重的成像序列，如T1序列适合观察解剖结构，而T2序列更侧重观察组织病变。准确的病情分析往往需要比对分析同一区域(对象)的多模态影像，例如超声影像、MRI影像、电子计算机断层扫描(Computed Tomography，CT)影像，或融合处理多模态影像以获得更有临床指导价值的融合图像。
随着深度学习技术的发展，出现了采用卷积神经网络(Convolutional Neural Network，CNN)实现从集合中学习参数化的配准函数。CNN网络接受两个输入，两个输入分别为待配准图像和一个固定的模板，该网络输出一个体素到另一个体素的映射场。通过共享的参数，该过程学习一种通用的表示法，该表示法可以对齐来自同一分布的新数据。
采用CNN模型进行图像配准方法存在的问题是，模板必须具有普遍性。例如，在医学影像应用中，在一些已经发生病变的影像中，病变周围往往伴随着随处可见的组织形变，若此时再将形变严重的影像配准到一个固定的标准模板往往会导致配准不准确，如已经被破坏的组织在配准后恢复正常，这将影响后续诊断分析的准确性。
采用CNN模型进行图像配准方法存在的另一个问题是，该方法也无法应用于不同模态的医学影像处理。例如，在医学影像应用中，由于MRI不同模态影像之间同一组织的成像强度可能不同，该方法无法通过简单的损失函数优化。以MRI的T1及T2序列为例，在脑部MR影像中，T1对脑脊液表现为低信号，而T2表现为高信号，而在对骨骼的成像中，两种模态都表现为高信号，因此无法使用常用的均方误差或交叉熵损失函数做统一处理。
为了解决多模态图像之间同一对象的成像强度不同，造成配准精度低的问题，本申请 实施例提供了通过获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像；对所述第一图像的强度进行修正得到第一修正图像；所述第一修正图像的强度与所述第二图像的强度分布匹配；根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场；根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。
可以理解的是，通过预先修正待配准的第一图像的强度，使其强度分布与目标模态的强度匹配或称对齐，根据第一修正图像和第二图像获取第一图像到目标模态的形变场，降低了因图像强度特征的影响，可以避免待修正的第一图像的同一对象与理想配准结果的距离差异过大导致精度下降，从而提高了配准精度。
可选的，采用经训练的循环生成对抗网络，根据第一图像和第二图像对第一图像的强度进行修正获得第一修正图像，相对于其他神经网络模型，可以使第一修正图像中的对象的强度接近第二图像的同时保留第一图像的特征，进而提高根据第二图像和第一修正图像获取的形变场的精度。另一方面，采用循环生成对抗网络根据同一个对象的源模态的图像和目标模态的图像对源模态的图像进行修正，可以避免采用固定模板造成的误差，从而提高配准精度。
可选的，采用包括形变损失的第一损失函数训练所述待训练的循环生成对抗网络，可以在修正第一图像强度的同时，在第一修正图像中保留更多第一图像的形状特征，从而根据第一修正图像和第二图像获得第一修正图像到目标模态的形变场的精度可以进一步的提高。
为了说明本申请所述的技术方案，以下结合具体附图及实施例进行详细说明。
图1示出的是本申请一实施例提供的一种电子设备D10，包括：至少一个处理器D100、存储器D101以及存储在存储器D101中并可在所述至少一个处理器D100上运行的计算机程序D102，所述处理器D100执行所述计算机程序D102时实现本申请实施例提供的多模态图像配准方法至少之一。
可以理解的是，上述电子设备，可以是桌上型计算机、笔记本、掌上电脑服务器、服务器集群、分布式服务器及云端服务器等计算设备。该电子设备D10可包括，但不仅限于，处理器D100、存储器D101。本领域技术人员可以理解，图1仅仅是电子设备D10的举例，并不构成对电子设备D10的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如还可以包括输入输出设备、网络接入设备等。
所称处理器D100可以是中央处理单元(Central Processing Unit，CPU)，该处理器D100还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电 路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器D101在一些实施例中可以是所述电子设备D10的内部存储单元，例如电子设备D10的硬盘或内存。所述存储器D101在另一些实施例中也可以是所述电子设备D10的外部存储设备，例如所述电子设备D10上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字(Secure Digital,SD)卡，闪存卡(Flash Card)等。进一步地，所述存储器D101还可以既包括所述电子设备D10的内部存储单元也包括外部存储设备。所述存储器D101用于存储操作系统、应用程序、引导装载程序(BootLoader)、数据以及其他程序等，例如所述计算机程序的程序代码等。所述存储器D101还可以用于暂时地存储已经输出或者将要输出的数据。
为了阐述方便，以下实施例中将上述电子设备统称为图像处理设备，可以理解的是，其并不构成对本申请的电子设备的具体限定。
图2示出了本申请一实施例提供的一种多模态图像配准的方法，应用于上述图1所示的电子设备，以下称图像处理设备，可由所述图像处理设备的软件/硬件实现。如图2所示，该方法包括步骤S110～S140。各个步骤的具体实现原理如下：
S110，获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像。
在一个非限定性的示例中，图像处理设备获取源模态的第一图像，例如，某对象的头颅正位的CT影像；和与所述第一图像配对的目标模态的第二图像，例如，该对象的的头颅正位的MRI T1序列的影像。这里的配对的图像指的是同一对象的不同模态图像。可以理解的是，图像配准技术在医学影像领域的应用较为广泛，本申请大部分示例以医学影像处理的为例进行说明，但本申请实施例提供的图像配准方法同样可以应用于其他图像处理领域，例如机器视觉领域、虚拟/增强显示领域等需要对图像进行比对和融合的领域，这里不再赘述。
S120，对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配。
在一个非限定性的示例中，图像处理设备对第一图像的强度进行修正得到第一修正图像。例如，图像处理设备通过预设的统计学模型修正第一图像的强度；又例如，图像处理设备通过对识别第一图像中的感兴趣区域后，对感兴趣区域进行分割后，对感性区域的强度进行修正；再例如，图像处理设备通过经训练的神经网络模型对第一图像的强度进行修正；从而获得第一修正图像。可以理解的是，可以采用以上示例方法至少之一以及其他可 以调整图像强度的方法对第一图像的强度进行修正，使所述第一图像的强度与所述第二图像的强度分布匹配，即第一修正图像的整体图像强度的分布与第二图像的强度分布近似，或第一修正图像的感兴趣区域的图像强度分布与第二图像的感兴趣区域分布近似。例如，对头颅正位CT影像进行强度修正后获得的修正头颅正位CT影像的骨骼部分的强度分布和目标模态MRI T1序列的头颅正位影像的骨骼部分的强度分布近似。可以理解的是，对单通道的灰度图像，修正其强度(灰度)即可，对于多通道的图像，例如RGB图像，可根据实际情况选取一个或多个通道的强度进行修正以方便后续处理，这里不再赘述。
S130，根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场。
在一个非限定性的示例中，图像处理设备通过经训练的无监督或半监督神经网络模型根据第一修正图像和第二图像获得第一修正图像到目标模态的形变场。
S140，根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。
在一个非限定性的示例中，图像处理装置根据修正前的第一图像和所述形变场获取修正前的第一图像配准到目标模态的配准图像。例如，通过步骤S130得到的形变场，将头颅正位CT影像配准到MRI T1序列的影像，可以理解的是，得到的配准图像是头颅对应像素位置调整，但是强度不变或强度改变较少的图像。
图3a和图3b示出的是一个非限定性的示例。如图3a所示，通过图像强度修正模型第一图像的前度进行修正获得第一修正图像。根据第一修正图像和第二图像，通过形变场获取模型得到第一修正图像配准到目标模态的形变场。如图3b所示，第一图像通过形变场配准到目标模态的配准图像，该配准图像是第一图像中的像素位置调整到目标模态对应位置，但是强度不变。可以理解的是，通过预先修正待配准的第一图像的强度，使其强度分布与目标模态的强度匹配或称对齐，根据第一修正图像和第二图像获取第一图像到目标模态的形变场，降低了因图像强度特征的影响，可以避免待修正的第一图像的同一对象与理想配准结果的距离差异过大导致精度下降，从而提高了配准精度。
在如上述图2所述的实施例的基础上，图4a和图4b示出了本申请一实施例提供的另一种多模态图像配准方法。图如4a所示，采用循环生成对抗网络，根据第一图像和第二图像对第一图像的强度进行修正得到第一修正图像，使第一修正图像的强度与第二图像的强度分布匹配。相对于其他神经网络模型，可以在第一修正图像中的对象的强度接近第二图像的同时保留第一图像的特征，进而提高根据第二图像和第一修正图像获取的形变场的精度。另一方面，采用循环生成对抗网络根据同一个对象的源模态的图像和目标模态的图像对源模态的图像进行修正，可以避免采用固定模板造成的误差，从而提高配准精度。
参见图5，在一个具体的非限定性的示例中，图5所示的循环生成对抗网络分别由两个生成器(Generator)和两个鉴别器(Discriminator)构成。其中，X域对应为源模态，Y域对应为目标模态。X、Y为两种模态的数据集，G表示为从X域图像生成Y域图像的生成器，F为与G方向相反的从Y域图像生成X域图像的生成器。D Y、D X分别为两个生成器对应的鉴别器，其作用为促进生成器结果与目标域的分布近似。
生成器的作用是合成新的图像，可以是U-Net网络实现。鉴别器的作用是判断合成图像的可信度，可以是VGG分类网络。可以理解的是，本领域技术人员可以在本申请实施例的教导下根据实际实施情况选用合适的生成器网络或鉴别器网络。
首先训练生成器，输入为X域图像x，金标准为Y域的与图像x配对的图像y，经过生成器G后，输出为合成图像q，此时将图像q与图像y当做D Y的输入，判断图像q与图像y的真假。另一条支路与之对应，路径为图像y和图像q经生成器F得到q′后，经D X判断x和q′真假。非限定性的，网络的学习率设置为0.0001，优化器选择Adam，通过反向传播过程训练约100批次后，得到训练好的生成对抗网络模型。该循环生成对抗网络训练完成后，输入任意一张X域的图像x，和Y域的图像y，将得强度修正后的图像x。图像x具有Y域相似的强度分布。
本示例中，反向传播过程采用的损失函数一种可能的实现方式为：
Figure PCTCN2019115311-appb-000001
其中，
Figure PCTCN2019115311-appb-000002
为生成器G损失，
Figure PCTCN2019115311-appb-000003
为生成器F损失，
Figure PCTCN2019115311-appb-000004
为循环损失。
Figure PCTCN2019115311-appb-000005
Figure PCTCN2019115311-appb-000006
Figure PCTCN2019115311-appb-000007
Figure PCTCN2019115311-appb-000008
通过损失函数公式(1)训练图5所示的循环生成对抗网络，得到的经训练的循环生成对抗网络对X域图像的强度进行修正后，会带有Y域的形状特征，为了提高配准精度，在公式(1)表示的损失函数基础上引入了包含形变损失的第一损失函数公式(5)训练所述待训练的循环生成对抗网络。
Figure PCTCN2019115311-appb-000009
其中，
Figure PCTCN2019115311-appb-000010
Figure PCTCN2019115311-appb-000011
如公式(6)和公式(7)所示根据所述循环生成对抗网络的生成器的输入图像与所述生成器的输出图像中各自对应的表征形状特征的参数的差值获得所述形变损失。
其中，α是权重因子系数，确定了形变损失在整体损失中的占比，可以在实施本实施例时根据实际情况选取并预先设置。x i＝0，y j＝0，表示表征背景的像素点，该形变损失仅在输入图像的背景发生作用，限制其形状的改变。在一种可能的实施方式中，可以先对图像进行预处理以去除噪声，或将背景像素的强度调整为0。
可选的，表征形状特征的参数包括以下参数至少之一：表征形状特征的像素点的强度、前景图像的边界长度和前景图像的面积；其中，所述表征形状的像素点包括以下像素点至少之一：表征背景的像素点、表征图像边缘的像素点和表征图像中感兴趣区域轮廓的像素点。本领域技术人员可以在本申请的实施例的教导下选取合适的参数应用于实际的实施方案中。
可以理解的是，公式(6)和公式(7)为一种可能的实施方式，本应于技术人员应该 能遵循本申请实施例的教导选取方差、均差和交叉熵等符合实际实施条件的差值。
通过第一损失函数公式(5)训练图5所示的循环生成对抗网络，输入为X域图像x，金标准为Y域的与图像x配对的图像y，经过生成器G后，输出为合成图像q，此时将图像q与图像y当做D Y的输入，判断图像q与图像y的真假。另一条支路与之对应，路径为图像y和图像q经生成器F得到q′后，经D X判断x和q′真假。非限定性的，网络的学习率设置为0.0001，优化器选择Adam优化器，通过反向传播过程训练约100批次后，得到训练好的生成对抗网络模型。该循环生成对抗网络训练完成后，输入任意一张X域的图像x，和Y域的图像y，将得强度修正后的图像x。修正后的图像x具有Y域相似的强度分布，但是修正后的图像x保持图像x的形状特征。
输入数据为同一病人的配对的MRI T1序列的影像和MRI T2序列的影像，数据预处理为统一两组数据的图像大小如192×192×1。输出为T1修正影像，T1修正影像具有和T2相似的强度分布，并且T1修正影像保持T1影像的形状特征。
可以理解的是，采用包括形变损失的第一损失函数训练所述待训练的循环生成对抗网络，可以在修正第一图像强度的同时，在第一修正图像中保留更多第一图像的形状特征，从而根据第一修正图像和第二图像获得第一修正图像到目标模态的形变场的精度可以进一步的提高。
在上述图2所示的多模态图像配准方法的实施例的基础上，图6示出了本申请一实施例提供的另一种的多模态图像配准的方法，如图5所示，步骤S130，根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场，包括：
S130’，采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场。
在一个非限定性的示例中，图像处理设备采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场。一种可能的实施方式为，编解码网络为图7所示的以U-Net网络为骨干的深度卷积网络和深度反卷积网络连接构成的编解码网络，图7中C表示卷积过程，U表示反卷积过程。该网络的输入数据为循环生成对抗网络根据第一图像x和第二图像y生成的第一修正图像q，输出为编解码网络对第二图像y和第一修正图像q的估计
Figure PCTCN2019115311-appb-000012
例如第一修正图像和第二图像为维度192×192×1的灰度图像，该网络的输出
Figure PCTCN2019115311-appb-000013
为维度192×192×1的图像，在该网络的输出层前设置形变场层，该层在反向传播时提供像素偏移梯度。形变场的作用是对输入图像的每一个像素进行偏移，则维度为192×192×2，最后一个通道的第一维为像素位移长度，第二维为像素位移方向。如图7所示，在取得形变场后，源模态(X域)的图像x经过应用形变场的应用模块得到 配准图像x′。
在一个可能的实施例中，采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场前，还包括：
对配对的源模态图像和目标模态图像组成的训练样本集中的源模态图像的强度进行修正，使所述样本集中所述源模态图像和与所述源模态图像配对的目标模态图像的强度分布匹配，获得源模态修正图像和目标模态图像样本集；采用源模态修正图像和目标模态图像样本集训练待训练的编解码网络获得经过训练的编解码网络。例如单组配对的样本为一个对象的的特定部位的MRI T1序列(源模态)影像和T2序列(目标模态)影像，训练样本集为多组不同对象的的图像。通过配对的样本训练编解码网络可以使编码网络参数根据同一对象的两个模态的数据进行调整，以达到应用该编解码网络时不需要固定的模板的效果，只需要同一个对象的配对的两个模态的影像即可获得形变场。在一种可能的示例中，训练编码网络的优化器选择Adam优化器，学习率为0.0001，训练100批次。
在一个非限定性的示例中，训练待训练的编码解码网络，采用第二损失函数训练所述待训练的编解码网络，损失第二损失函数包括所述待训练编解码网络输出图像与所述训练样本集中的目标模态图像的差值。在一个可能的实施方式中采用公式(8)所示的第二损失函数训练编解码网络。
Figure PCTCN2019115311-appb-000014
其中，T为编解码网络，G为循环生成对抗网络的生成器，X为源模态图像，Y为目标模态图像。这里采用的是均方差(MSE)损失，可以理解的是在本申请实施例的指引下，本领域技术人员可以根据实际实施的需要选用交叉熵或绝对差等损失。
循环生成对抗网络根据第一图像和第二图像生成的第一修正图像，将第一修正图像和第二图像作为输入如图7所示的编解码网络。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
对应于上述的多模态图像配准方法，图8示出的是本申请一实施例提供的一种多模态图像配准装置，包括：
图像获取模块M110，用于获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像。
强度修正模块M120，用于对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配。
形变场获取模块M130，用于根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场。
配准模块M140，用于根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。
强度模块M120，用于对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配，包括：
循环生成对抗网络模块M121，用于采用循环生成对抗网络，根据所述第一图像和所述第二图像对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配。
循环生成对抗网络模块M121，采用经训练的循环生成对抗网络，根据所述第一图像和所述第二图像对所述第一图像的强度进行修正前，还包括：
通过循环生成对抗网络训练模块M121’，用于采用配对的源模态图像和目标模态图像组成的训练样本集，训练待训练的循环生成对抗网络，得到经训练的循环生成对抗网络。
循环生成对抗网络训练模块M121’训练待训练的循环生成对抗网络，还包括：
循环生成对抗网络损失函数模块M1211，用于采用包括形变损失的第一损失函数训练所述待训练的循环生成对抗网络；其中，根据所述循环生成对抗网络的生成器的输入图像与所述生成器输出图像中表征形状特征的参数的差值获得所述形变损失。
表征形状特征的参数包括以下参数至少之一：表征形状特征的像素点的强度，前景图像的边界长度和前景图像的面积；其中，所述表征形状的像素点包括以下像素点至少之一：表征背景的像素点、表征图像边缘的像素点和表征图像中感兴趣区域轮廓的像素点。
形变场获取模块M130，用于根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场，包括：
编解码网络模块M1301，用于采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场。
编解码网络模块M1301，用于采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场前，还包括：
编解码网络训练模块M1301’，用于对配对的源模态图像和目标模态图像组成的训练样本集中的源模态图像的强度进行修正，使所述样本集中所述源模态图像和与所述源模态图像配对的目标模态图像的强度分布匹配，获得源模态修正图像和目标模态图像样本集；
编解码网络训练模块M1301’，还用于采用源模态修正图像和目标模态图像样本集训练待训练的编解码网络获得经过训练的编解码网络。
编解码网络训练模块M1301’，还用于采用包括所述待训练编解码网络输出图像与所述训练样本集中的目标模态图像的差值的第二损失函数训练所述待训练的编解码网络。
可以理解的是，以上实施例中的各种实施方式和实施方式组合及其有益效果同样适用于本实施例，这里不再赘述。
需要说明的是，上述装置/单元之间的信息交互、执行过程等内容，由于与本申请方法实施例基于同一构思，其具体功能及带来的技术效果，具体可参见方法实施例部分，此处不再赘述。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。实施例中的各功能单元、模块可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中，上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。另外，各功能单元、模块的具体名称也只是为了便于相互区分，并不用于限制本申请的保护范围。上述系统中单元、模块的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请的一些实施例中，采用图1所示的电子设备，该电子设备包括：包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现：
获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像；
对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配；
根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场；
根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。
在一个非限定性的示例中，所述处理器执行所述计算机程序时实现：对所述第一图像的强度进行修正得到第一修正图像，包括：采用经训练的循环生成对抗网络，根据所述第一图像和所述第二图像对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配。
在一个非限定性的示例中，所述处理器执行所述计算机程序时实现：采用经训练的循环生成对抗网络，根据所述第一图像和所述第二图像对所述第一图像的强度进行修正前， 还包括：采用配对的源模态图像和目标模态图像组成的训练样本集，训练待训练的循环生成对抗网络，得到经训练的循环生成对抗网络。
在一个非限定性的示例中，所述处理器执行所述计算机程序时实现：训练待训练的循环生成对抗网络，还包括：采用包括形变损失的第一损失函数训练所述待训练的循环生成对抗网络；其中，所述形变损失为根据所述循环生成对抗网络的生成器的输入图像与所述生成器的输出图像各自对应的表征形状特征的参数的差值获得。
表征形状特征的参数包括以下参数至少之一：
表征形状特征的像素点的强度，前景图像的边界长度和前景图像的面积；
其中，所述表征形状的像素点包括以下像素点至少之一：表征背景的像素点、表征图像边缘的像素点和表征图像中感兴趣区域轮廓的像素点。
在一个非限定性的示例中，所述处理器执行所述计算机程序时实现：所述根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场，包括：采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场。
在一个非限定性的示例中，所述处理器执行所述计算机程序时实现：采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场前，还包括：
对配对的源模态图像和目标模态图像组成的训练样本集中的源模态图像的强度进行修正，使所述样本集中所述源模态图像和与所述源模态图像配对的目标模态图像的强度分布匹配，获得源模态修正图像和目标模态图像样本集；
采用源模态修正图像和目标模态图像样本集训练待训练的编解码网络获得经过训练的编解码网络。
在一个非限定性的示例中，所述处理器执行所述计算机程序时实现：训练待训练的编码解码网络，包括：采用包括所述待训练编解码网络输出图像与所述训练样本集中的目标模态图像的差值的第二损失函数训练所述待训练的编解码网络。
本申请实施例还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现可实现上述各个方法实施例中的步骤。
本申请实施例提供了一种计算机程序产品，当计算机程序产品在电子设备上运行时，使得电子设备执行时实现可实现上述各个方法实施例中的步骤。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实现上述实施例方法中 的全部或部分流程，可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一计算机可读存储介质中，该计算机程序在被处理器执行时，可实现上述各个方法实施例的步骤。其中，所述计算机程序包括计算机程序代码，所述计算机程序代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述计算机可读介质至少可以包括：能够将计算机程序代码携带到拍照装置/终端设备的任何实体或装置、记录介质、计算机存储器、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、电载波信号、电信信号以及软件分发介质。例如U盘、移动硬盘、磁碟或者光盘等。在某些司法管辖区，根据立法和专利实践，计算机可读介质不可以是电载波信号和电信信号。
在上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述或记载的部分，可以参见其它实施例的相关描述。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
在本申请所提供的实施例中，应该理解到，所揭露的装置/网络设备和方法，可以通过其它的方式实现。例如，以上所描述的装置/网络设备实施例仅仅是示意性的，例如，所述模块或单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通讯连接可以是通过一些接口，装置或单元的间接耦合或通讯连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。
以上仅为本申请的可选实施例而已，并不用于限制本申请。对于本领域的技术人员来 说，本申请可以有各种更改和变化。凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的权利要求范围之内。

Claims (20)

  1. 一种多模态图像配准的方法，其特征在于，包括：
    获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像；
    对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配；
    根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场；
    根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。
  2. 如权利要求1所述的方法，其特征在于，对所述第一图像的强度进行修正得到第一修正图像，包括：
    采用经训练的循环生成对抗网络，根据所述第一图像和所述第二图像对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配。
  3. 如权利要求2所述的方法，其特征在于，采用经训练的循环生成对抗网络，根据所述第一图像和所述第二图像对所述第一图像的强度进行修正前，还包括：
    采用配对的源模态图像和目标模态图像组成的训练样本集，训练待训练的循环生成对抗网络，得到经训练的循环生成对抗网络。
  4. 如权利要求3所述的方法，其特征在于，训练待训练的循环生成对抗网络，还包括：
    采用包括形变损失的第一损失函数训练所述待训练的循环生成对抗网络；
    其中，所述形变损失为根据所述循环生成对抗网络的生成器的输入图像与所述生成器的输出图像各自对应的表征形状特征的参数的差值获得。
  5. 如权利要求4所述的方法，表征形状特征的参数包括以下参数至少之一：
    表征形状特征的像素点的强度、前景图像的边界长度和前景图像的面积；
    其中，所述表征形状的像素点包括以下像素点至少之一：表征背景的像素点、表征图像边缘的像素点和表征图像中感兴趣区域轮廓的像素点。
  6. 如权利要求1至5任一项所述的方法，其特征在于，所述根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场，包括：
    采用经训练的编解码网络根据所述第一修正图像和所述第二图像，获取所述第一修正图像到目标模态的形变场。
  7. 如权利要求6所述的方法，其特征在于，采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场前，还包括：
    对配对的源模态图像和目标模态图像组成的训练样本集中的源模态图像的强度进行修正，获得源模态修正图像和目标模态图像样本集，所述样本集中所述源模态图像和与所述源模态图像配对的目标模态图像的强度分布匹配；
    采用源模态修正图像和目标模态图像样本集训练待训练的编解码网络获得经过训练的编解码网络。
  8. 如权利要求7所述的方法，其特征在于，训练待训练的编码解码网络，包括：
    采用第二损失函数训练所述待训练的编解码网络，所述第二损失函数包括所述待训练编解码网络输出图像与所述训练样本集中的目标模态图像的差值。
  9. 如权利要求4所述的方法，其特征在于，所述形变损失还包括预设的权重因子系数。
  10. 一种多模态图像配准的装置，其特征在于，包括：
    图像获取模块，用于获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像；
    强度修正模块，用于对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配；
    形变场获取模块，用于根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场；
    配准模块，用于根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。
  11. 一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述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如下步骤：
    获取源模态的第一图像，和与所述第一图像配对的目标模态的第二图像；
    对所述第一图像的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配；
    根据所述第一修正图像和所述第二图像获取所述第一修正图像配准到目标模态的形变场；
    根据所述第一图像和所述形变场获取所述第一图像配准到目标模态的配准图像。
  12. 如权利要求11所述的电子设备，其特征在于，对所述第一图像的强度进行修正得到第一修正图像，包括：
    采用经训练的循环生成对抗网络，根据所述第一图像和所述第二图像对所述第一图像 的强度进行修正得到第一修正图像，使所述第一修正图像的强度与所述第二图像的强度分布匹配。
  13. 如权利要求12所述的电子设备，其特征在于，采用经训练的循环生成对抗网络，根据所述第一图像和所述第二图像对所述第一图像的强度进行修正前，还包括：
    采用配对的源模态图像和目标模态图像组成的训练样本集，训练待训练的循环生成对抗网络，得到经训练的循环生成对抗网络。
  14. 如权利要求13所述的电子设备，其特征在于，训练待训练的循环生成对抗网络，还包括：
    采用包括形变损失的第一损失函数训练所述待训练的循环生成对抗网络；
    其中，所述形变损失为根据所述循环生成对抗网络的生成器的输入图像与所述生成器的输出图像各自对应的表征形状特征的参数的差值获得。
  15. 如权利要求14所述的电子设备，其特征在于，表征形状特征的参数包括以下参数至少之一：
    表征形状特征的像素点的强度、前景图像的边界长度和前景图像的面积；
    其中，所述表征形状的像素点包括以下像素点至少之一：表征背景的像素点、表征图像边缘的像素点和表征图像中感兴趣区域轮廓的像素点。
  16. 如权利要求11至15任一项所述的电子设备，其特征在于，所述根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场，包括：
    采用经训练的编解码网络根据所述第一修正图像和所述第二图像，获取所述第一修正图像到目标模态的形变场。
  17. 如权利要求16所述的电子设备，其特征在于，采用经训练的编解码网络根据所述第一修正图像和所述第二图像获取所述第一修正图像到目标模态的形变场前，还包括：
    对配对的源模态图像和目标模态图像组成的训练样本集中的源模态图像的强度进行修正，获得源模态修正图像和目标模态图像样本集，所述样本集中所述源模态图像和与所述源模态图像配对的目标模态图像的强度分布匹配；
    采用源模态修正图像和目标模态图像样本集训练待训练的编解码网络获得经过训练的编解码网络。
  18. 如权利要求17所述的电子设备，其特征在于，训练待训练的编码解码网络，包括：
    采用第二损失函数训练所述待训练的编解码网络，所述第二损失函数包括所述待训练编解码网络输出图像与所述训练样本集中的目标模态图像的差值。
  19. 如权利要求14所述的电子设备，其特征在于，所述形变损失还包括预设的权重因子 系数。
  20. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1至9任一项所述的方法。
PCT/CN2019/115311 2019-11-04 2019-11-04 多模态图像配准的方法、装置、电子设备及存储介质 WO2021087659A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/115311 WO2021087659A1 (zh) 2019-11-04 2019-11-04 多模态图像配准的方法、装置、电子设备及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/115311 WO2021087659A1 (zh) 2019-11-04 2019-11-04 多模态图像配准的方法、装置、电子设备及存储介质

Publications (1)

Publication Number Publication Date
WO2021087659A1 true WO2021087659A1 (zh) 2021-05-14

Family

ID=75848683

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/115311 WO2021087659A1 (zh) 2019-11-04 2019-11-04 多模态图像配准的方法、装置、电子设备及存储介质

Country Status (1)

Country Link
WO (1) WO2021087659A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180165554A1 (en) * 2016-12-09 2018-06-14 The Research Foundation For The State University Of New York Semisupervised autoencoder for sentiment analysis
CN109377520A (zh) * 2018-08-27 2019-02-22 西安电子科技大学 基于半监督循环gan的心脏图像配准系统及方法
CN110021037A (zh) * 2019-04-17 2019-07-16 南昌航空大学 一种基于生成对抗网络的图像非刚性配准方法及系统
CN110084863A (zh) * 2019-04-25 2019-08-02 中山大学 一种基于生成对抗网络的多域图像转换方法与系统
CN110163897A (zh) * 2019-04-24 2019-08-23 艾瑞迈迪科技石家庄有限公司 一种基于合成超声图像的多模态图像配准方法
CN110298871A (zh) * 2019-06-10 2019-10-01 东软医疗系统股份有限公司 图像配准方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180165554A1 (en) * 2016-12-09 2018-06-14 The Research Foundation For The State University Of New York Semisupervised autoencoder for sentiment analysis
CN109377520A (zh) * 2018-08-27 2019-02-22 西安电子科技大学 基于半监督循环gan的心脏图像配准系统及方法
CN110021037A (zh) * 2019-04-17 2019-07-16 南昌航空大学 一种基于生成对抗网络的图像非刚性配准方法及系统
CN110163897A (zh) * 2019-04-24 2019-08-23 艾瑞迈迪科技石家庄有限公司 一种基于合成超声图像的多模态图像配准方法
CN110084863A (zh) * 2019-04-25 2019-08-02 中山大学 一种基于生成对抗网络的多域图像转换方法与系统
CN110298871A (zh) * 2019-06-10 2019-10-01 东软医疗系统股份有限公司 图像配准方法及装置

Similar Documents

Publication Publication Date Title
CN111047629B (zh) 多模态图像配准的方法、装置、电子设备及存储介质
US10088544B2 (en) Tractography framework with magnetic resonance imaging for brain connectivity analysis
CN110570426B (zh) 使用深度学习的图像联合配准和分割
US10304158B2 (en) Image processing apparatus, image processing method and non-transitory computer-readable medium with calculation of information representing direction of target tissue and with estimating of depicting appearance of target tissue
US8498459B2 (en) System and method for verifying registration accuracy in digital medical images
US8861891B2 (en) Hierarchical atlas-based segmentation
US9142030B2 (en) Systems, methods and computer readable storage media storing instructions for automatically segmenting images of a region of interest
CN110246580B (zh) 基于神经网络和随机森林的颅侧面影像分析方法和系统
Wu et al. Registration of longitudinal brain image sequences with implicit template and spatial–temporal heuristics
US20210158515A1 (en) Method and apparatus for determining mid-sagittal plane in magnetic resonance images
US20150324994A1 (en) Method of Automatically Calculating Linking Strength of Brain Fiber Tracts
Kuijf et al. Registration of brain CT images to an MRI template for the purpose of lesion-symptom mapping
WO2022086910A1 (en) Anatomically-informed deep learning on contrast-enhanced cardiac mri
CN109961435B (zh) 脑图像获取方法、装置、设备及存储介质
WO2021097595A1 (zh) 图像的病变区域分割方法、装置及服务器
Bergeest et al. A comparison of five methods for signal intensity standardization in MRI
Mazaheri et al. Hybrid Pixel‐Based Method for Cardiac Ultrasound Fusion Based on Integration of PCA and DWT
CN117911432A (zh) 图像分割方法、装置及存储介质
CN109671131B (zh) 影像校正方法、装置、医疗影像设备及存储介质
US9224229B2 (en) Process and apparatus for data registration
EP4167184A1 (en) Systems and methods for plaque identification, plaque composition analysis, and plaque stability detection
WO2021087659A1 (zh) 多模态图像配准的方法、装置、电子设备及存储介质
Ruiz‐España et al. Automatic segmentation of the spine by means of a probabilistic atlas with a special focus on ribs suppression
Woo et al. Geometric feature‐based multimodal image registration of contrast‐enhanced cardiac CT with gated myocardial perfusion SPECT
CN113850710B (zh) 一种跨模态医学影像精准转换方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19951825

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19951825

Country of ref document: EP

Kind code of ref document: A1