Teacher-Student Generator for Low Dose Sinogram Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current PET image reconstruction methods face challenges in producing high-quality images at low doses, which increases radiation risk for patients, and existing methods either rely on time-consuming iterative algorithms or ignore sinogram domain information in neural network approaches.
Innovation Solution
A low dose Sinogram denoising and PET image reconstruction method using a teacher-student generator network, which includes a denoising module and a reconstruction module, employing Poisson downsampling, convolutional blocks, and Generative Adversarial Networks (GAN) to effectively restore low dose sinogram information and reconstruct high-quality PET images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If iterative reconstruction algorithms (MLEM, MAP, PWLS, TV) are used to reconstruct low dose PET images, then image quality can be improved, but reconstruction time becomes excessively long
Solution Approach 1:
The method performs denoising on the sinogram data before reconstruction, preparing the data in advance to facilitate faster reconstruction. By preprocessing the low-dose sinogram to remove noise and enhance structural information, the subsequent reconstruction process requires fewer iterations and less computational time while maintaining or improving image quality
Solution Approach 2:
The reconstruction process is divided into two separate modules: a denoising module that processes sinogram data and a reconstruction module that generates images. This segmentation allows each module to specialize in its function, with the denoising module preparing optimized input for the reconstruction module, thereby improving overall efficiency and reducing total processing time
2Productivity
If neural network mapping learning is used to map low dose images to normal-dose images, then reconstruction speed is improved, but sinogram domain information is ignored
Solution Approach 1:
Instead of mapping images to images, the method inverts the approach by mapping sinograms to sinograms through denoising, then reconstructing from the denoised sinogram. This preserves the advantageous sinogram domain information while still utilizing neural network acceleration, achieving both speed and information retention
3Object-affected harmful factors
If low dose data is used for PET imaging, then radiation dose to patients is reduced, but image quality and noise performance deteriorate
Solution Approach 1:
The method converts the harmful noise present in low-dose sinogram data into beneficial structural information through the denoising module. By training the neural network to recognize and preserve true structural signals while removing noise, the harmful noise is transformed into enhanced structural information that improves image quality
Solution Approach 2:
The method changes the parameter domain by performing denoising in the sinogram domain rather than the image domain. This parameter change allows the neural network to operate on raw projection data where noise characteristics are different and more amenable to processing, thereby improving image quality from low-dose acquisitions
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The method effectively denoises low dose Sinogram data and reconstructs high-quality PET images, outperforming other algorithms by maintaining structural information and reducing reconstruction errors, thereby minimizing radiation exposure while improving image quality.
Implementation Method 1
using Poisson distribution to downsampling the normal dose Sinogram projection data to obtain a low dose Sinogram projection data
Data Source
AI summary
The present invention discloses a low dose Sinogram denoising and PET image reconstruction method based on teacher-student generator, the adopted network model is divided into a Sinogram denoising module and a PET image reconstruction module, the entire network needs to be processed in a training stage and a test stage. In the training stage: the present invention uses the denoising module to denoise the low dose Sinogram, and then makes the reconstruction module use the denoised Sinogram to reconstruct, in which the teacher generator is introduced in the training stage to constrain the whole, the denoising module is decoupled from the reconstruction module, and a better reconstructed image is obtained through training. In the testing stage, the present invention only needs to input low-dose Sinogram to the denoising module to obtain the denoised Sinogram, and then input the denoised Sinogram to the student generator to get the final reconstruction image.