PET image reconstruction algorithm for improving spatial resolution uniformity of PET system based on deep learning

An image reconstruction and deep learning technology, applied in the field of biomedical image analysis, to achieve the effect of improving uniformity and solving uneven spatial resolution

Active Publication Date: 2021-06-08
7 Cites 0 Cited by

AI-Extracted Technical Summary

Problems solved by technology

[0006] In view of the above, the present invention provides a PET image reconstruction algorithm based on deep lea...
View more

Method used

In the estimation stage, input the projection data of the radial position closest to the edge of the FOV into the trained ISTA-Net to directly obtain a relativel...
View more


The invention discloses a PET image reconstruction algorithm for improving the uniformity of the spatial resolution uniformity of a PET systembased on deep learning. The algorithm fully utilizes the characteristic that the radial resolution in the FOV space of the PET system is not uniform to solve the mutual depth effect, takes a high-resolution concentration distribution diagram reconstructed by projection data when a phantom is located in the FOV center as a label, adopting a neural network training means to improve the resolution of a reconstructed concentration distribution diagram of the same phantom at an FOV edge position without any novel detector or obtaining any additional information, such as DOI information or PSF information. According to the algorithm, a software means is used for replacing a complex hardware method at the present stage, and the problem that the spatial resolution of the PET system is not uniform is solved.

Application Domain

Reconstruction from projectionGeometric image transformation +2

Technology Topic

Image reconstruction algorithmImage resolution +7


  • PET image reconstruction algorithm for improving spatial resolution uniformity of PET system based on deep learning
  • PET image reconstruction algorithm for improving spatial resolution uniformity of PET system based on deep learning
  • PET image reconstruction algorithm for improving spatial resolution uniformity of PET system based on deep learning


  • Experimental program(1)

Example Embodiment

[0028] The technical solutions of the present invention will be described in detail with reference to the accompanying drawings and specific embodiments.
[0029] Such as figure 1As shown, the image reconstruction algorithm of the PET system spatial resolution uniformity in the depth of the present invention includes the following steps:
[0030] (1) Collect data. PHANTOM is injected into the PET radioactive tracer, and the body film is placed at different positions of the PET device radial distance field (FOV), detects the photon and counts, and is obtained from different radial positions. i Corresponding original projection data matrix Y i.
[0031] (2) Establish measurement equation model according to PET imaging principle:
[0032] Y = GX + R + S
[0033] Where: g is a system matrix, X is a real tracer concentration profile, and R is the number of random photons during the measurement, and s is the number of scattered photons during the measurement process.
[0034] Since there is a depth of intection, the spatial resolution is uneven, that is, the image resolution in the center of the same cross section is high, the closer the radial position, the resolution is significantly reduced, and the solid will be obtained Projection data of different radial positions Y i The reconstruction problem is split into two sub-problems, and the child problem 1 is the original projected data Y0 Y0 Y0 located at the center of the field, and the child problem 2 is the original projection data matrix Y, which handles different radial position I (i 0). i.
[0035] Sub Problem 1 Directly Make Image Reconstruction, Sub Quality 2 Based on the reconstruction image of sub-problem 1, using a means of depth learning (ISTA-NET).
[0036] The network model structure used in the present invention is figure 2 As shown, it is sequentially connected by a plurality of exact same stages (PHASE), and each stage is input from the input to the output by an operator. After soft threshold algorithm and its symmetrical operator Composition, where operator Includes 2 convolution layers, one RELU function and the other convolution layer are sequentially connected, and the output of each convolutional neural network layer in the network is sequentially processed by mass standardization; The result of the output passes soft threshold algorithm, and It is The fully symmetrical structure, first batch the data and then pass through one convolution layer, a RELU function and the other two convolution layers, and one batch of batchization between the two convolution layers (BN) Layer, and will The final output data is output in the form of linear stacking with the input projection data. A total of 9 phasers in the network of the present embodiment are provided with all the convolution layers in each phase of each phase, each of which is 3 * 3, and the step size is 1.
[0037] (3) Training stage.
[0038] First, you need to make a single frame normalization process for the input data (SINOGRAM) and the corresponding tag:
[0040] Where: X min X max The minimum and maximum values ​​of single frame data are respectively.
[0041] Then initialize the parameters of the ISTA-NET, the original projected data matrix of different radial positions i Divided into training sets and test sets, and the SINOGRAM input ISTA-NET of the training set is calculated by forward propagation formulas, which in turn obtains the ISTA-NET final output, calculating the output of ISTA-NET and the label. Loss function:
[0043] Where: X i Is the iSta-NET number, b is the number of total training sets in ISTA-NET, and A is X i Pixel size.
[0044] Finally, the default number of loss functions is determined by the ADAM algorithm to update the parameters learned in ISTA-NET, repeat forward propagation and reverse guidance until the value of the loss function is sufficiently small.
[0045] (4) Estimate stage.
[0046] In the estimated phase, the projection data that is close to the edge of the FOV edge is trained, and the relatively high quality reconstruction map is directly obtained, thereby significantly weakening the influence of the depth effect, improves the problem of uneven spatial resolution.
[0047] We will experiment based on Monte Carlo simulation data to verify the effectiveness of the present embodiment. Monte Carlo Simulation tracer is 18 F-FDG, the body model is Derenzo Phantom (0.70mm, 0.95mm, 1.40mm, 1.95mm, 2.40mm, 2.80mm), the simulated scanner is a 4-layer DOI detector, its DOI accuracy is 5mm and no DOI information Single layer detectors, where a single-layer detector without DOI information is obtained by different radial positions of different radial locations in the center of the field without DOI information is randomly divided into training set (1080sinograms) and test sets (360 Sinograms). The training set is used to learn the parameters of ISTA-NET, and the test set is used to verify the performance of the ISTA-NET of the training.
[0048] image 3 When comparing Derenzo PHANTOM in the edge position, the reconstruction result of the present invention and the 4-layer DOI detector is used, from left to right columns, respectively: PHANTOM, 75mm from the field of view, located under the 4-layer DOI detector, located in the field of view The central PHANTOM has a rebuilt map of DOI information, and the reconstruction map of the PHANTOM without DOI information from the perspective center and the reconstruction map obtained by the algorithm of the present invention. by image 3 It can be seen that PHANTOM in the center location is also very low, even if there is no DOI information, its resolution of its concentration distribution map is also very high, and the resolution of the concentration profile of the reconstructed concentration map of the edge position is very low, There is no distinguishing between a point source having a diameter of 2.8 mm, but the image resolution after obtaining the DOI information can be slightly higher than that of the central location without DOI information. We use the present invention to reconstruct the SINOGRAM in the edge position, and its resolution has a significant increase, and the effect can be close to the 4-layer DOI information, so that the resolution of the FOV edge and the center is closer to consistent, improve the PET system space. The uniformity of the resolution.
[0049] The above description of the embodiments is to facilitate the understanding of the invention to understand and apply the invention. Perspective of those skilled in the art will apparently, various modifications can be easily made, and the general principles described herein are applied to other embodiments without creative labor. Accordingly, the present invention is not limited to the above embodiments, and those skilled in the art will also be within the scope of the invention, in accordance with the present invention, and will be within the scope of the invention.


no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.

Similar technology patents

Preparation method of ultrafine tungsten carbide powder

ActiveCN103408015AImprove uniformityavoid uneven mixing

Multiple-channel transmit magnetic resonance

InactiveCN101283286AImprove uniformityReduce SAR

Classification and recommendation of technical efficacy words

Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products