Image Anti-aliasing via Domain Transformer for Sparse View Tomography
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Sparse-view tomographic reconstructions suffer from strong aliasing artifacts due to incomplete projection data, leading to reduced image quality and diagnostic value.
Innovation Solution
An image processing system that uses a trained machine learning model to compute output images from input images, transforms these images from the image domain into the projection domain to generate new projection data, and combines this data with initial projection data for artifact-reduced reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If sinogram completion methods are used to synthesize missing projection views, then aliasing artifacts are reduced, but image detail is lost due to smoothing
Solution Approach 1:
The patent transforms the problem from projection domain to image domain by using a U-Net architecture that operates on reconstructed images rather than directly on sinogram data. This dimensional shift allows the model to learn artifact removal patterns in the image domain while preserving structural details, avoiding the smoothing effect that occurs when operating directly in the projection domain.
Solution Approach 2:
The patent introduces a learned transformation model as an intermediary between the incomplete projection data and the final reconstructed image. This model first generates intermediate images from available projections, then uses the U-Net architecture to remove artifacts from these intermediates before producing the final output, thereby preserving detail while reducing artifacts.
2Productivity
If fast kVp-switching technique is used for spectral imaging, then acquisition time is reduced, but projection data sampling is insufficient leading to artifacts
Solution Approach 1:
The patent creates virtual copies of projection data by using the U-Net model to generate artifact-free images from the sparse kVp-switching projections. The model learns to synthesize the missing information that would have been present with full sampling, effectively creating virtual projection views without requiring additional physical measurements, thus maintaining fast acquisition while eliminating artifacts.
Solution Approach 2:
The patent changes the processing approach from direct projection domain operations to image domain operations using a deep learning model. By transforming the problem into the image domain and using learned parameters from training data, the system can handle the undersampled kVp-switching data effectively, removing artifacts while preserving the fast acquisition advantage.
3Loss of information
If projection domain sinogram completion is performed, then missing views are synthesized, but unwanted smoothing occurs reducing image quality
Solution Approach 1:
The patent fundamentally changes the domain of operation from projection domain to image domain. Instead of completing the sinogram, the U-Net architecture directly processes reconstructed images to remove artifacts. This dimensional change allows the model to preserve sharp edges and fine details while filling in missing information, avoiding the smoothing inherent in projection domain methods.
Solution Approach 2:
The patent replaces traditional mechanical interpolation methods in the projection domain with a data-driven machine learning approach in the image domain. The U-Net architecture learns from training data to directly synthesize artifact-free image content, substituting the mechanical sinogram completion process with an intelligent image processing system that preserves detail while recovering missing information.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image processing system (SYS) and related method. The system (SYS) may comprise an interface (IN) through which is receivable two input images (m1, m2), reconstructable from initial projection data (λ) including sets of projection data (λ1, λ2) acquirable by a tomographic imaging apparatus (IA). The two sets are complementary. System (SYS)'s trained machine learning model (M) computes, based on the two input images (m1,m2), two output images (m1',m2'). A domain transformer (DT) transforms the two output images (m1', m2') from image domain into projection domain, to so obtain new projection data (λ'). An output interface (OUT) outputs the new projection data (λ') and the initial projection data (λ) for reconstruction into at a further image (m1", m2"). The system (SYS) may be used in spectral imaging, in particular of the kVp-switching type.