Deep Learning CT Reconstruction from Sparse Projections
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional computed tomography (CT) imaging techniques require multiple angular sampling to avoid aliasing artifacts, limiting imaging time and increasing radiation dose, and have yet to achieve high-quality 3D image reconstruction with ultra-sparse sampling.
Innovation Solution
A residual deep learning network is used for 3D image reconstruction from single-view or few-view 2D projection data, employing an encoder network for feature transformation and a decoder network for volumetric image generation, allowing for high-fidelity reconstruction with sparse sampling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple angular sampling is used to avoid aliasing artifacts, then image quality is improved, but imaging time increases and radiation dose increases
Solution Approach 1:
The patent transforms the problem from the spatial/angular domain to the frequency domain using Fourier transform relationships. By working in the frequency domain and utilizing the projection-slice theorem, the system can reconstruct 3D images from fewer angular views by properly sampling and interpolating in the frequency domain, thus resolving the contradiction between angular sampling density and imaging efficiency
Solution Approach 2:
The patent changes the sampling parameters by using non-uniform angular sampling combined with appropriate frequency domain interpolation. Instead of uniform dense angular sampling, the system uses sparser angular views with optimized sampling patterns and compensates through frequency domain processing, thereby reducing imaging time while maintaining image quality
2Measurement precision
If multiple angular sampling is used to avoid aliasing artifacts, then image quality is improved, but radiation dose increases
Solution Approach 1:
By transitioning to frequency domain processing using the projection-slice theorem, the system can achieve accurate reconstruction with fewer projection views. This dimensional transformation allows efficient use of limited projection data, reducing the number of x-ray exposures needed while maintaining diagnostic image quality, thus lowering radiation dose
Solution Approach 2:
The patent implements partial sampling in the angular domain by using fewer projection views than traditional methods require. Through frequency domain interpolation and reconstruction algorithms, the system recovers complete 3D information from this partial sampling, achieving adequate image quality with reduced radiation exposure
3Loss of time
If sparse angular sampling is used to reduce imaging time, then imaging time is reduced, but image quality deteriorates due to aliasing artifacts
Solution Approach 1:
The patent optimizes sampling parameters by using non-uniform angular spacing combined with frequency domain interpolation. The sampling pattern is specifically designed to minimize aliasing while reducing the number of views, and the frequency domain processing parameters are adjusted to properly reconstruct the image from this optimized sparse sampling pattern
Solution Approach 2:
The frequency domain serves as an intermediary space where sparse angular sampling data can be properly processed. By transforming projection data to the frequency domain, applying appropriate interpolation and filtering, then transforming back, the system eliminates aliasing artifacts that would be present in direct spatial domain reconstruction from sparse views
4Object-affected harmful factors
If ultra-sparse sampling is used to further reduce radiation dose, then radiation dose is reduced, but reconstruction quality becomes insufficient
Solution Approach 1:
The patent exploits the Fourier relationship between projection space and frequency space to enable ultra-sparse angular sampling. By working in the frequency domain where the projection-slice theorem provides a direct relationship between 2D projections and 3D frequency content, the system can accurately reconstruct 3D images from extremely few angular views that would be insufficient in traditional spatial domain methods
Solution Approach 2:
The system optimizes ultra-sparse sampling by carefully selecting angular positions and using frequency domain interpolation parameters that maximize information recovery. The sampling density and distribution are specifically tuned to capture sufficient frequency information for high-quality reconstruction with minimal projections, enabling radiation dose reduction while maintaining diagnostic quality
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
This method enables real-time high-quality 3D image reconstruction from a single or few 2D projections, reducing the need for multiple angular sampling and providing effective imaging for medical and interventional procedures with reduced radiation exposure.
Implementation Method 1
transforming by the encoder network the set of one or more 2D projection images to 2D features
Implementation Method 2
mapping by the transform module the 2D features to 3D features
Implementation Method 3
generating by the decoder network the 3D volumetric image from the 3D features
Implementation Method 4
acquiring a set of one or more 2D projection images, e.g., with a computed tomography x-ray scan
Data Source
AI summary
A method for tomographic imaging comprising acquiring [200] a set of one or more 2D projection images [202] and reconstructing [204] a 3D volumetric image [216] from the set of one or more 2D projection images [202] using a residual deep learning network comprising an encoder network, a transform module and a decoder network, wherein the reconstructing comprises: transforming [206] by the encoder network the set of one or more 2D projection images [202] to 2D features [208]; mapping [210] by the transform module the 2D features [208] to 3D features [212]; and generating [214] by the decoder network the 3D volumetric image [216] from the 3D features [212]. Preferably, the encoder network comprises 2D convolution residual blocks and the decoder network comprises 3D blocks without residual shortcuts within each of the 3D blocks.