PET low-dose imaging method and device based on deep learning and medium

Abstract

The invention relates to a PET low-dose imaging method and device based on deep learning and a medium. The method comprises the following steps: extracting two-dimensional image slices from paired low-dose PET data and full-dose PET data along a sagittal plane, a coronal plane and a cross section respectively; respectively slicing the two-dimensional PET image extracted from each surface to obtain a slice image of each surface, and training to obtain an image estimation model of each surface through a two-dimensional coding and decoding network; inputting each surface of the low-dose three-dimensional PET into the image estimation model to obtain a full-dose estimation result of each surface, and superposing the estimation results to obtain a full-dose three-dimensional PET of three-surface estimation; and after splicing, inputting into a three-dimensional residual depth fusion network, and eliminating direction artifacts to obtain a final full-dose three-dimensional PET estimation result. Compared with the prior art, the image obtained by the method comprises three-dimensional features and spatial fusion perception information, and has the advantages of low acquisition difficulty, high image quality and the like.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a PET low-dose imaging method, device and medium based on deep learning. Background Art

[0002] Positron emission tomography (PET) is widely used in the diagnosis, staging and treatment evaluation of clinical diseases. PET scanning requires the injection of radioactive tracers, which will produce nuclear radiation to the human body. In order to minimize the radiation risk, one way is to reduce the injection dose, but this will lead to poor image quality. Therefore, studying the problem of recovering high-quality PET images from low-dose data can reduce nuclear radiation to the human body while maintaining good image quality.

[0003] In the existing technology, deep learning-based technology shows great potential in restoring high-quality PET images from low-dose PET images. Unet-based convolutional neural network (CNN) and C-Gan-based adversarial neural network are used to restore high-quality PET images from low-dose PET images. However, these methods only focus on reconstructing a single view along a two-dimensional axial slice, while ignoring the three-dimensional features of PET images.

[0004] Therefore, it is necessary to propose a low-dose PET imaging method that can reflect the three-dimensional characteristics of the image. Summary of the invention

[0005] The purpose of the present invention is to overcome the defect of the above-mentioned prior art that the three-dimensional features of PET images are ignored and to provide a PET low-dose imaging method, device and medium based on deep learning.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A deep learning-based PET low-dose imaging method comprises the following steps:

[0008] Two-dimensional image slices were extracted from the paired low-dose PET data and full-dose PET data along the sagittal plane, coronal plane, and transverse plane;

[0009] The paired low-dose and full-dose 2D PET images extracted from the sagittal plane are sliced ​​to obtain sagittal slice images, and a sagittal image estimation model is trained through a 2D encoder-decoder network;

[0010] The paired low-dose and full-dose two-dimensional PET images extracted from the coronal plane are sliced ​​to obtain coronal plane slice images, and a coronal plane image estimation model is trained through a two-dimensional encoder-decoder network;

[0011] The paired low-dose and full-dose 2D PET images extracted from the cross-section are sliced ​​to obtain cross-section slice images, and a cross-section image estimation model is trained through a 2D encoder-decoder network;

[0012] The sagittal, coronal and transverse planes of a low-dose three-dimensional PET are input into the image estimation models of the sagittal, coronal and transverse planes respectively to obtain the full dose estimation results of the sagittal, coronal and transverse planes, and the estimated full dose estimation results of the sagittal, coronal and transverse planes are superimposed to obtain the full dose three-dimensional PET estimated by three planes;

[0013] The full-dose 3D PET images estimated from the three surfaces are spliced ​​and input into the 3D residual deep fusion network to eliminate the directional artifacts of the 3D PET and obtain the final full-dose 3D PET estimation result.

[0014] Furthermore, the estimated full-dose estimation results of the sagittal, coronal and transverse planes are superimposed. The specific process is as follows: the image features of the full-dose estimation results of the sagittal, coronal and transverse planes are extracted through a three-dimensional residual dense network, and all image features are accumulated using global feature fusion to achieve fusion perception of three-dimensional space, thus obtaining the full-dose three-dimensional PET estimated on three planes.

[0015] Furthermore, the two-dimensional PET image is sliced ​​and normalized after slicing.

[0016] Furthermore, PET images are used 18 PET image with F-FDG as a tracer.

[0017] Furthermore, the PET image is a brain PET image.

[0018] Furthermore, the two-dimensional codec network uses the Adam optimizer for network training optimization.

[0019] Furthermore, the batch size of the 2D codec network is 4, the initial learning rate is 0.0001, and the number of dataset iterations is 20.

[0020] Furthermore, the two-dimensional codec network includes evaluation indicators, and the evaluation indicators include normalized root mean square error, peak signal-to-noise ratio and structural similarity.

[0021] In a second aspect of the present invention, a deep learning-based PET low-dose imaging device comprises a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements any of the deep learning-based PET low-dose imaging methods described above.

[0022] A third aspect of the present invention is a storage medium having a program stored thereon, which, when executed, implements any of the above deep learning-based PET low-dose imaging methods.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1) The present invention estimates the corresponding full-dose three-dimensional PET image from the sagittal, coronal and transverse images of a low-dose three-dimensional PET, and removes image artifacts, thereby reducing the difficulty of obtaining full-dose three-dimensional PET image data and improving image quality. The obtained full-dose PET can accurately provide the three-dimensional features of the image, including spatial fusion perception information, which is helpful for subsequent image analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The figure is a flow chart of the method of the present invention.

[0026] Figure 2 Schematic diagram of the algorithm of the present invention.

[0027] Figure 3 4 is a comparison chart of image quality evaluation results according to an embodiment of the present invention.

[0028] Figure 4 This is a full-dose image comparison diagram of an embodiment of the present invention.

[0029] Figure 5a This is a diagram of clinical evaluation results of an image according to an embodiment of the present invention.

[0030] Figure 5b A statistical chart showing whether the lesion can be detected from the PET image.

[0031] Figure 6a Comparison of tracer uptake in full-dose PET images and true full-dose PET images.

[0032] Figure 6b This is a correlation comparison of tracer uptake in full-dose PET images and true full-dose PET images. DETAILED DESCRIPTION

[0033] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0034] Example 1

[0035] like Figure 1 As shown, the present invention is a PET low-dose imaging method based on deep learning, comprising the following steps:

[0036] Two-dimensional image slices were extracted from the paired low-dose PET data and full-dose PET data along the sagittal plane, coronal plane, and transverse plane;

[0037] The paired low-dose and full-dose 2D PET images extracted from the sagittal plane are sliced ​​to obtain sagittal slice images, and a sagittal image estimation model is trained through a 2D encoder-decoder network;

[0038] The paired low-dose and full-dose two-dimensional PET images extracted from the coronal plane are sliced ​​to obtain coronal plane slice images, and a coronal plane image estimation model is trained through a two-dimensional encoder-decoder network;

[0039] The paired low-dose and full-dose 2D PET images extracted from the cross-section are sliced ​​to obtain cross-section slice images, and a cross-section image estimation model is trained through a 2D encoder-decoder network;

[0040] The sagittal, coronal and transverse planes of a low-dose three-dimensional PET are input into the image estimation models of the sagittal, coronal and transverse planes respectively to obtain the full dose estimation results of the sagittal, coronal and transverse planes, and the estimated full dose estimation results of the sagittal, coronal and transverse planes are superimposed to obtain the full dose three-dimensional PET estimated by three planes;

[0041] The full-dose 3D PET images estimated from the three surfaces are spliced ​​and input into the 3D residual deep fusion network to eliminate the directional artifacts of the 3D PET and obtain the final full-dose 3D PET estimation result.

[0042] The image processing process of the present invention is as follows Figure 2 As shown, through the above settings, this method can restore the low-dose PET image of the brain to the full-dose PET image.

[0043] The application results of this embodiment are as follows:

[0044] In 294 cases, the 18 The experimental verification was conducted on the PET data of patients with F-FDG tracer, including 241 training data and 53 test data, including 16 patients suspected of having brain masses. The network training was optimized by Adam, with 4 training data in each batch, an initial learning rate of 0.0001, and a total of 20 epochs of the data set.

[0045] The method of this embodiment is compared with the two-dimensional network C-Gan and the three-dimensional network U-Net, which have the best effect in full-dose PET reconstruction. The generated full-dose three-dimensional PET images are evaluated from three aspects: image quality, subjective scoring by clinicians, and local analysis of lesions. The results are as follows:

[0046] Image quality evaluation results are as follows: Figure 3 As shown in the figure, three image quality evaluation indicators, namely normalized root mean square error (NRMSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), were evaluated respectively. The method of the present invention has the lowest normalized root mean square error, the highest peak signal-to-noise ratio and the highest structural similarity, indicating that the method is superior in image quality.

[0047] like Figure 4 As shown in Figure 2, the full-dose PET images estimated by different methods were compared from the sagittal, coronal and transverse planes, and it can be found that our method has the best display effect at the lesion.

[0048] As shown in Figure 5, independent qualitative image analysis was performed on the full-dose PET estimated data of 16 patients suspected of having brain masses, and the sagittal, coronal and transverse planes of the estimated full-dose PET images were examined and scored respectively. The subjective quality was evaluated and scored using a 5-point method, and images with a score of 3 or above were defined as meeting clinical needs. In addition, a judgment was given as to whether lesions could be detected from the estimated PET images. The method of the present invention had the best effect in both image quality evaluation and lesion detection.

[0049] As shown in Figure 6, the consistency of the tracer uptake (SUV) of the full-dose PET image of the lesion and the true full-dose PET image was locally analyzed to evaluate the estimation. The SUV histogram analysis was combined to describe the SUV correlation between the estimated full-dose PET image and the true full-dose PET image. The results show that the results estimated by the method of the present invention have the strongest correlation.

[0050] In a second aspect of the present invention, a PET low-dose imaging device based on deep learning includes a memory, a processor, and a program stored in the memory. When the processor executes the program, a PET low-dose imaging method based on deep learning of the present invention is implemented. The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the program code is executed by the processor or controller, the functions / operations specified in the flowchart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as an independent software package and partially on a remote machine, or entirely on a remote machine or server.

[0051] The third aspect of the present invention is a storage medium having a program stored thereon, which implements a deep learning-based PET low-dose imaging method of the present invention when the program is executed. The machine-readable medium may be a tangible medium that may contain or store a program for use by an instruction execution system, device or equipment or used in conjunction with an instruction execution system, device or equipment. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. The machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing.

[0052] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A PET low-dose imaging method based on deep learning, characterized in that: The following steps are involved: Two-dimensional image slices were extracted from the paired low-dose PET data and full-dose PET data along the sagittal plane, coronal plane, and transverse plane; The paired low-dose and full-dose 2D PET images extracted from the sagittal plane are sliced ​​to obtain sagittal slice images, and a sagittal image estimation model is trained through a 2D encoder-decoder network; The paired low-dose and full-dose two-dimensional PET images extracted from the coronal plane are sliced ​​to obtain coronal plane slice images, and a coronal plane image estimation model is trained through a two-dimensional encoder-decoder network; The paired low-dose and full-dose 2D PET images extracted from the cross-section are sliced ​​to obtain cross-section slice images, and a cross-section image estimation model is trained through a 2D encoder-decoder network; The sagittal, coronal and transverse planes of a low-dose three-dimensional PET are input into the image estimation models of the sagittal, coronal and transverse planes respectively to obtain the full dose estimation results of the sagittal, coronal and transverse planes, and the estimated full dose estimation results of the sagittal, coronal and transverse planes are superimposed to obtain the full dose three-dimensional PET estimated by three planes; The full-dose 3D PET images estimated from the three surfaces are spliced ​​and input into the 3D residual deep fusion network to eliminate the directional artifacts of the 3D PET and obtain the final full-dose 3D PET estimation result.

2. A PET low-dose imaging method based on deep learning according to claim 1, characterized in that: The estimated full dose estimation results of the sagittal plane, coronal plane and transverse plane are superimposed. The specific process is: through the three-dimensional residual dense network, the image features of the full dose estimation results of the sagittal plane, coronal plane and transverse plane are extracted, and global feature fusion is adopted to accumulate all image features to realize the fusion perception of three-dimensional space, and obtain the full dose three-dimensional PET estimated by three planes.

3. The method for low-dose PET imaging based on deep learning according to claim 1, characterized in that: The two-dimensional PET image slices are normalized after slicing.

4. The method for low-dose PET imaging based on deep learning according to claim 1, characterized in that: The PET images were taken using 18 PET image with F-FDG as a tracer.

5. The method for low-dose PET imaging based on deep learning according to claim 1, characterized in that: The PET image is a brain PET image.

6. The method for low-dose PET imaging based on deep learning according to claim 1, characterized in that: The two-dimensional codec network uses the Adam optimizer to perform network training optimization.

7. The method for low-dose PET imaging based on deep learning according to claim 1, characterized in that: The batch size of the two-dimensional encoding and decoding network is 4, the initial learning rate is 0.0001, and the number of data set iterations is 20.

8. The method for low-dose PET imaging based on deep learning according to claim 1, characterized in that: The two-dimensional encoding and decoding network includes evaluation indicators, and the evaluation indicators include normalized root mean square error, peak signal-to-noise ratio and structural similarity.

9. A PET low-dose imaging device based on deep learning, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, a deep learning-based PET low-dose imaging method as described in any one of claims 1-8 is implemented.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, a deep learning-based PET low-dose imaging method as described in any one of claims 1-8 is implemented.

Images