阅读说明：本技术 一种无伴随ct辐射的单床pet延迟成像方法 (Single-bed PET (positron emission tomography) delayed imaging method without concomitant CT (computed tomography) radiation ) 是由 朱闻韬 饶璠 杨宝 韩璐 叶宏伟 王瑶法 于 2021-09-09 设计创作，主要内容包括：本发明公开了一种无伴随CT辐射的单床PET延迟成像方法,首先利用一能将PET BP图像转换成更接近于真实PET图像的Pseudo PET图像的图像重建网络,将正常扫描和延迟扫描得到的PET BP图像转换成Pseudo PET图像,然后利用一CT图像生成网络,输入包含正常扫描的Pseudo PET图像和CT图像,以及延迟扫描的Pseudo PET图像,输出获得正常扫描和延迟扫描间的变形场和延迟扫描时刻的CT图像,该CT图像最后用于延迟扫描PET图像重建中的衰减校正,得到SUV定量准确的PET图像并用于肿瘤检测。本发明的方法能消除延迟扫描中病人接受的CT辐射,减轻病人生理和心理上的压力,推动PET延迟成像的应用。(The invention discloses a single-bed PET delayed imaging method without concomitant CT radiation, which comprises the steps of firstly utilizing an image reconstruction network capable of converting a PET BP image into a Pseudo PET image closer to a real PET image, converting the PET BP image obtained by normal scanning and delayed scanning into the Pseudo PET image, then utilizing a CT image generation network, inputting the Pseudo PET image and the CT image containing the normal scanning and the Pseudo PET image of the delayed scanning, outputting the CT image for obtaining a deformation field between the normal scanning and the delayed scanning and a delay scanning moment, and finally using the CT image for attenuation correction in the delayed scanning PET image reconstruction to obtain the PET image with accurate SUV quantification and using the PET image for tumor detection. The method of the invention can eliminate CT radiation received by a patient in delayed scanning, reduce the physiological and psychological pressure of the patient and promote the application of PET delayed imaging.)

1. A single bed PET delay imaging method without concomitant CT radiation, comprising the steps of:

the method comprises the following steps: acquiring PET data needing to be subjected to delayed PET reconstruction for normal scanning of a patient and CT data needing to be subjected to normal scanning, generating a PET BP image and a CT image, acquiring the PET data during delayed scanning, and generating the PET BP image;

step two: respectively converting the PET BP images of the normal scanning and the delayed scanning obtained in the step one into Pseudo PET images of the normal scanning and the delayed scanning which are close to the PET AC images by utilizing a trained image reconstruction network;

step three: generating a network by utilizing a trained CT image, and taking the CT image of the normal scanning obtained in the step one and the Pseudo PET image of the normal scanning and the Pseudo PET image of the delayed scanning obtained in the step two as input to obtain the CT image of the delayed scanning;

step four: converting the CT image acquired in the third step into a mu-map image required by attenuation correction by adopting a bilinear method, and generating a quantitative and accurate reconstructed PET image by combining the PET data acquired in the first step;

the image reconstruction network is obtained by training with the acquired PET BP image data as input, the Pseudo PET image as output and the PET AC image as a label and with a minimum loss function as a target; the CT image generation network takes the acquired PET BP image data of normal scanning and delayed scanning and the acquired CT image of normal scanning as input, the CT image of delayed scanning as output and the acquired CT image of delayed scanning as a label, and performs training by taking a minimum loss function as a target.

2. The single bed PET delay imaging method without concomitant CT radiation of claim 1, wherein the image reconstruction network is selected from UNet.

3. The single bed PET delay imaging method without concomitant CT radiation of claim 1, wherein the CT image generation network comprises an image registration network, a sampler and a deformation network, wherein the image registration network outputs a low resolution deformation field between the normal scan and the delay scan according to the input PET BP image data of the normal scan and the delay scan, the sampler converts the low resolution deformation field into a high resolution deformation field, and the deformation network generates the CT image of the delay scan according to the input CT image of the normal scan and the high resolution deformation field.

4. The single bed PET delay imaging method without concomitant CT radiation of claim 3, wherein the image registration network is selected from the group consisting of Unet, voxelmorphh, ConvNet, and DIRNet.

5. The single-bed PET delayed imaging method without concomitant CT radiation according to claim 3, wherein if the delayed-scan PET BP image is a single bed, the single-bed delayed-scan PET BP image or Pseudo PET image is changed into a whole-body delayed-scan image by zero-filling and a corresponding mask image is acquired as an input of the CT image generation network.

6. The single-bed PET delayed imaging method without concomitant CT radiation according to claim 5, characterized in that the loss function during the CT image generation network training is specifically:

f = D ₁ (CT2, M ◦STN(CT1)) + D ₂ (PET2 , M ◦STN(PET1)) + R(DVF)

in the formula (I), the compound is shown in the specification,CT1andCT2CT images representing a normal scan and a delayed scan respectively,PET1andPET2PET images representing a normal scan and a delayed scan, respectively, the PET image being a PET AC image or a Pseudo PET image,STN(*)in order to be the output of the morphable network,R(DVF)for the regularization term of the deformation field,D ₁ (*)、D ₂ (*)for the similarity metric function, M is a mask image of the PET BP image of the whole-body delayed scan obtained by zero padding.

Technical Field

The invention relates to attenuation correction for PET delay scanning image reconstruction, in particular to a single-bed PET delay imaging method without concomitant CT radiation.

Background

The delayed PET (Positron Emission Tomography) scan is an important basis for early diagnosis of lung cancer, and has very important significance for improving the cure rate of lung cancer and improving the life quality of patients. According to the statistics of researchers, if the lung cancer patients can be diagnosed and treated early, the 5-year survival rate of the lung cancer patients can be improved from 14 percent to 49 percent. In clinic, about 85% of lung cancer cases are advanced cases when diagnosed, and the 5-year survival rate of the lung cancer in the middle and advanced stages is less than 5%, so that early discovery, early diagnosis and early treatment are the key points for improving the cure rate of the lung cancer. PET imaging is a medical image that can provide biochemical and quantitative information of a living body, is the mainstream imaging technology for performing functional, metabolic and receptor imaging in an anatomical form manner at present, and is widely applied to cancer detection.

Generally, the cells in the malignant tumor have more glucose transporters and hexosaminidases, and FDG is phosphorylated by hexosaminidases after entering cancer cells and is therefore trapped in these cells, so that the uptake of most malignant tumors will not change significantly over time, but the uptake of most normal tissues will decrease, ultimately enhancing contrast, and therefore PET delayed scanning can increase the sensitivity of malignant tumor diagnosis. Hou and other researches show that when SUVmax is between 2.5 and 8.0, benign or malignant lesions exist, but SUV values in most malignant tumor cells need 4h to reach a peak value, cells caused by inflammation generally can reach an uptake peak value within about 30min, and therefore, malignant tumors and inflammatory lesions can be effectively identified through delayed imaging; huang et al analyzed and counted the normal and time-lapse PET scan images of 50 isolated lung nodule or tumor patients, and found that the SUV value after about 3h of drug injection had better diagnostic efficacy; chen et al used neighborhood grayscale difference matrix texture features in normal and delayed PET/CT images to identify benign and malignant tumors, and experiments showed that the texture features of dual-time-point PET/CT images can be used as a predictor of malignant solitary lung nodules. The delayed imaging has great advantages in primary lung cancer, and when the CT image is atypical and the conventional PET/CT imaging is difficult to identify the benign and malignant lung nodules, the accuracy of tumor identification can be improved by adding one time of delayed scanning.

Raw data required by PET imaging in PET/CT delayed scanning is derived from gamma rays generated by isotope drugs, the delayed scanning is increased without injecting additional isotope drugs, but because attenuation correction is included in a delayed scanning reconstruction algorithm, CT images of a patient still need to be acquired once, the radiation dose received by the patient is increased, the risk of cancer of the patient is increased, a larger medical ethical problem is faced, and the method is an important factor for restricting PET delayed scanning imaging. To address this problem, clinicians and researchers have proposed different methods of deep learning, which fall into two main categories. One is to use image registration methods to establish spatial correspondence between the patient in the normal scan and the delayed scan, including voxelmorphh and DIRNet. The second method is to adopt a medical image modality conversion method to convert an image of one modality into an image of another modality, which is common on MRI/PET imaging equipment, and scholars have proposed a method for converting an MRI image into a CT image and using the CT image for PET image reconstruction attenuation correction, wherein a typical network comprises pix2pix and CycleGAN.

Disclosure of Invention

The invention provides a single-bed PET delayed imaging method based on image registration, which considers that the anatomical structure of a patient does not change violently between normal scanning and delayed scanning, and comprises the steps of firstly establishing an image reconstruction network, converting a PET BP image into a Pseudo PET image which is closer to a real PET image, converting the PET BP image obtained by the normal scanning and the delayed scanning into the Pseudo PET image, then establishing an image generation network, inputting the network, including the Pseudo PET image and the CT image obtained by the normal scanning and the Pseudo PET image obtained by the delayed scanning, outputting a deformation field between the normal scanning and the delayed scanning and the CT image at the delayed scanning moment, and finally using the CT image for attenuation correction in the delayed scanning PET image reconstruction to obtain the PET image with accurate SUV quantification and using the PET image for tumor detection. The method of the invention can eliminate CT radiation received by a patient in delayed scanning, reduce the physiological and psychological pressure of the patient and promote the application of PET delayed imaging.

The purpose of the invention is realized by the following technical scheme: a single bed PET delay imaging method without concomitant CT radiation, comprising the steps of:

the method comprises the following steps: acquiring PET data needing to be subjected to delayed PET reconstruction for normal scanning of a patient and CT data needing to be subjected to normal scanning, generating a PET BP image and a CT image, acquiring the PET data during delayed scanning, and generating the PET BP image;

step two: respectively converting the PET BP images of the normal scanning and the delayed scanning obtained in the step one into Pseudo PET images of the normal scanning and the delayed scanning which are close to the PET AC images by utilizing a trained image reconstruction network;

step three: generating a network by utilizing a trained CT image, and taking the CT image of the normal scanning obtained in the step one and the Pseudo PET image of the normal scanning and the Pseudo PET image of the delayed scanning obtained in the step two as input to obtain the CT image of the delayed scanning;

step four: converting the CT image acquired in the third step into a mu-map image required by attenuation correction by adopting a bilinear method, and generating a quantitative and accurate reconstructed PET image by combining the PET data acquired in the first step;

the image reconstruction network is obtained by training with the acquired PET BP image data as input, the Pseudo PET image as output and the PET AC image as a label and with a minimum loss function as a target; the CT image generation network takes the acquired PET BP image data of normal scanning and delayed scanning and the acquired CT image of normal scanning as input, the CT image of delayed scanning as output and the acquired CT image of delayed scanning as a label, and performs training by taking a minimum loss function as a target.

Further, the image reconstruction network is selected from UNet.

Further, the CT image generation network comprises an image registration network, a sampler and a deformation network, wherein the image registration network outputs a low-resolution deformation field between the normal scanning and the delay scanning according to the input PET BP image data of the normal scanning and the delay scanning, the sampler converts the low-resolution deformation field into a high-resolution deformation field, and the deformation network generates a delay scanning CT image according to the input CT image of the normal scanning and the high-resolution deformation field.

Further, the image registration network is selected from the group consisting of Unet, VoxelMorph, ConvNet, and DIRNet.

Further, if the delayed scanning PET BP image is a single bed, the single bed delayed scanning PET BP image or Pseudo PET image is changed into a whole body delayed scanning image by zero padding, and a corresponding mask image is acquired as an input of the CT image generation network.

Further, the loss function during the network training of the CT image generation specifically is:

f = D ₁ (CT2, M ◦STN(CT1)) + D ₂ (PET2 , M ◦STN(PET1)) + R(DVF)

in the formula (I), the compound is shown in the specification,CT1andCT2CT images representing a normal scan and a delayed scan respectively,PET1andPET2PET images (PET AC images or Pseudo PET images) respectively representing the normal scan and the delayed scan,STN(*)in order to be the output of the morphable network,R(DVF)for the regularization term of the deformation field,D ₁ (*)、D ₂ (*)for the similarity metric function, M is a mask image of the PET BP image of the whole-body delayed scan obtained by zero padding.

Further, the number of delayed scans may be one or a plurality of consecutive scans.

The invention has the beneficial effects that: the invention provides a single-bed PET (positron emission tomography) delayed imaging method without concomitant CT (computed tomography) radiation, which comprises two neural networks: an image reconstruction network and a CT image generation network. The image reconstruction network converts the PET BP image into a Pseudo PET image close to a real PET AC image, and solves the problem that the PET BP is difficult to be directly used for image registration due to uneven gray scale (the gray scale on the surface of a body is higher, and the gray scale inside the body is lower). The image generation network can generate the CT images required for PET delayed imaging attenuation correction. The method of the invention can eliminate CT radiation received by a patient in delayed scanning, reduce the physiological and psychological pressure of the patient and promote the application of PET delayed imaging.

Drawings

FIG. 1 is an overall flow chart for single bed PET delayed imaging without concomitant CT radiation;

FIG. 2 is an image reconstruction network architecture;

FIG. 3 is an image generation network architecture;

FIG. 4 is another image generation network architecture;

fig. 5 is a schematic diagram of the gradation information and the mask information.

Detailed Description

The present invention will be described in detail below with reference to examples and the accompanying drawings.

Example 1

Fig. 1 is a schematic diagram of a single-bed PET delayed imaging method based on unconnected CT radiation, the method comprising the steps of:

the method comprises the following steps: PET/CT image data for a large number of clinical normal and delayed scans, including in particular, normal and delayed PET BP (Back projection) and PET AC (attenuation correction) images, and normal and delayed CT images, are acquired.

Step two: designing and training a PET image reconstruction supervision learning network;

in this embodiment, the image reconstruction network structure is as shown in fig. 2, a PET BP image (including image data of normal scan and delayed scan) is input, and a predicted corresponding Pseudo PET image is output, and the Pseudo PET image uses a PET AC image as a label, so that the image reconstruction network can convert the PET BP image into a Pseudo PET image close to the PET AC image. The trained objective function is represented as follows;

min:f= D(PB , PA)

in the formula (I), the compound is shown in the specification,PBrepresenting a PET BP image of the ultrasound image,PArepresents a PET AC image;D(PB , PA) The image distance between the two is realized by adopting NCC.

And (4) training the network designed in the second step by using the PET BP and PET AC images obtained in the first step until the target function is converged and stable, and finishing training the image reconstruction network.

Step three: designing and training a CT image generation network;

in the present embodiment, the CT image generation network structure is as shown in fig. 3, and the PET image and the CT image for the normal scan and the PET image for the delayed scan are input, and the deformation field between the normal scan and the delayed scan and the estimated CT image for the delayed scan are output. UNet is used as a registration Network, a deformation Network (STN) is connected behind the registration Network, a deformation field is output after registration of the registration Network, and the deformation field and a normally scanned CT image are subjected to deformation Network to obtain a deformed image. And during model training, the delayed scanning CT image obtained in the step one is used as a label. The trained objective function includes CT image similarity measure D1 (such as NCC measure), PET image similarity measure D2 (such as NCC measure), and deformation field regularization term (L2 norm of first derivative) as specifically expressed below;

min:f= D ₁ (CT2, STN(CT1)) + D ₂ (PET2 , STN(PET1)) + R(DVF)

in the formula (I), the compound is shown in the specification,CT1andCT2CT images representing a normal scan and a delayed scan respectively,PET1andPET2PET images (PET AC images or Pseudo PET images) respectively representing the normal scan and the delayed scan,STN(*)in order to be the output of the morphable network,R(DVF)is a regularization term for the deformation field.

And (3) training the network designed in the fourth step by adopting the PET AC image and the CT image of the normal scanning acquired in the first step and the PET AC image of the delayed scanning until the target function is converged and stable, and finishing training the image reconstruction network.

Step four: acquiring a PET BP image and a CT image of a patient needing delayed PET reconstruction in a normal scanning process, and acquiring a sinogram (PET original information) of the delayed scanning process and generating a PET BP image.

Step five: inputting the PET BP image of the normal scanning and the PET BP image of the delayed scanning in the fourth step into the image reconstruction network trained in the second step to obtain a Pseudo PET image of the normal scanning and a Pseudo PET image of the delayed scanning;

step six: and inputting the normal scanning Pseudo PET image and the delayed scanning Pseudo PET image obtained in the step five and the CT image obtained in the step four in the step three to train to obtain a CT image generation network, wherein the network can output the spatial corresponding relation between the normal scanning and the delayed scanning and the estimated delayed scanning CT image.

Step seven: and converting the CT acquired in the step six into a mu-map image required by attenuation correction by adopting a bilinear method, and generating a quantitative and accurate reconstructed PET image by combining sinogram data acquired in the step four.

The present embodiment includes two networks: the image reconstruction network and the image generation network can be trained independently in practice.

Example 2

As a preferable scheme, in view of the fact that there are only one or more beds of PET data in delayed scanning frequently occurring in clinic, the present embodiment proposes a single-bed PET delayed imaging method without accompanying CT radiation, in which zero padding is performed on a PET image in delayed scanning, a mask image is generated and a level information image is set for registration to generate a CT image for attenuation correction, fig. 5 is a schematic diagram of the mask image and the level information image, gray areas in a normal scanning image and the delayed scanning image are coordinate intervals of the images along the Z axis, a scanning area in normal scanning is generally wider than the delayed scanning area, and a corresponding coordinate interval is also larger; the mask image is an image composed of 0 and 1, the elements of the Z-axis coordinate interval of the delayed scanning image are set as 1, other areas are set as 0, and the mask image can change along with the position and the image size of the delayed scanning image; the gradation information image is an image whose image gradation is set to Z-axis coordinate, and it contains Z-axis coordinate information of the normal scan image.

The method specifically comprises the following steps:

a method of single bed PET delayed imaging without concomitant CT radiation, the method comprising the steps of:

the method comprises the following steps: acquiring digital human technology to generate a large amount of PET/CT image data of normal scanning and delayed scanning, specifically including PET BP (Back projection) images of normal scanning and delayed scanning, PET AC (attenuation correction) images of normal scanning and delayed scanning and CT images of normal scanning and delayed scanning, and recording bed position information.

Step two: designing and training a PET image reconstruction supervision learning network;

in this embodiment, a single bed PET BP image (image data including normal scan and delayed scan) is input, and a corresponding predicted single bed Pseudo PET image using a PET AC image as a label is output, so that the image reconstruction network can convert the PET BP image into a Pseudo PET image close to the PET AC image. The trained objective function is represented as follows;

min:f= D(PB , PA)

in the formula (I), the compound is shown in the specification,PBrepresenting a PET BP image of the ultrasound image,PArepresents a PET AC image;D(PB , PA) The image distance between the two is realized by adopting NCC.

And (4) dividing the PET BP and PET AC image data acquired in the step one into different single bed data according to the bed position information, inputting the different single bed data into the network designed in the step two for training until the target function is converged and stable, and finishing the training of the image reconstruction network.

Step three: designing and training a CT image generation network;

in this embodiment, as shown in fig. 4, the CT image generation network structure is configured to input a normal scan PET (PET AC or Pseudo PET) image and a CT image, and a zero-filled single-bed delayed scan PET (PET AC or Pseudo PET) image and a mask image, and the network itself includes hierarchical information and outputs a distortion field between the normal scan and the delayed scan and an estimated delayed scan CT image.

The ConvNet is used as a registration Network to generate a Low Resolution Deformable field (LR-DVF), and then a sampler is used to convert the Low Resolution Deformable field into a High Resolution Deformable field (HR-DVF), which is then followed by a Spatial deformation Network (STN). And during model training, delayed scanning CT images are used as labels. The trained objective function comprises CT image similarity measure (such as NCC measure), PET image similarity measure (such as NCC measure) and deformation field regularization term (L2 norm of first derivative) which are specifically expressed as follows;

min:f= D ₁ (CT2, M ◦STN(CT1)) + D ₂ (PET2 , M ◦STN(PET1)) + R(DVF)

in the formula (I), the compound is shown in the specification,CT1andCT2CT images representing a normal scan and a delayed scan respectively,PET1andPET2PET images (PET AC images or Pseudo PET images) respectively representing the normal scan and the delayed scan,Mto delay the mask image of the scanned PET image,STN(*)in order to be the output of the morphable network,R(DVF)is a regularization term for the deformation field.

And (3) preprocessing the normal scanning PET AC image and the CT image acquired in the step one and the delayed scanning PET AC image according to the corresponding method in the step two, inputting the preprocessed images into the designed CT image generation network for training until the objective function is converged and stable, and finishing the training of the image reconstruction network.

Step four: acquiring a PET BP image and a CT image of a patient needing to be subjected to delayed PET reconstruction and normal scanning, and single bed sinogram (PET original data) data of delayed scanning and generating a PET BP image;

step five: dividing the normal scanning PET BP image in the fourth step into different single-bed data according to the bed position information, inputting the single-bed data into the image reconstruction network trained in the second step, and splicing the data output by the network to obtain a Pseudo PET image of the normal scanning of the whole body; filling zero in the delayed scanning single-bed PET BP image and inputting the image into the image reconstruction network trained in the step two to obtain a delayed scanning single-bed Pseudo PET image; obtaining a mask image of the delayed scanning PET image according to the bed information;

step six: inputting the normal scanning Pseudo PET image, the delayed scanning Pseudo PET image and the mask image obtained in the step five and the normal scanning CT image obtained in the step four into the step three to be trained to obtain a CT image generation network, wherein the network can output the spatial corresponding relation between the normal scanning and the delayed scanning and the estimated delayed scanning CT image.

Step seven: and converting the CT acquired in the step six into a mu-map image required by attenuation correction by adopting a bilinear method, and generating a quantitative and accurate reconstructed PET image by combining the PET sinogram data acquired in the step four.

It should be understood that the above examples are only for clarity of illustration and are not intended to limit the embodiments. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. This need not be, nor should all embodiments be exhaustive. And obvious variations or modifications of the invention may be made without departing from the scope of the invention.