Learned Multiscale Image Reconstruction for Low-Dose CBCT
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
CBCT image quality is poor due to incomplete data reconstruction, photon starvation, and scattering artifacts, which are challenging to address with existing learned iterative schemes, especially in high-attenuation areas and low-dose scans, limiting its application in adaptive radiotherapy.
Innovation Solution
A computer-implemented method using a learned iterative scheme with multiscale reconstruction and invertible operators, applying dual-space and primal-space operators iteratively on primal and dual patches, reducing memory requirements and improving inference speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If learned iterative schemes are used for CBCT reconstruction, then image quality can be improved, but memory requirements increase significantly
Solution Approach 1:
The patent divides the 3D volume into multiple 2D slices that are processed independently. Each slice is reconstructed separately using the learned iterative scheme, which reduces the memory footprint from O(N³) to O(N²) where N is the linear dimension. This segmentation allows the use of sophisticated learned iterative methods without requiring prohibitive amounts of memory.
Solution Approach 2:
The patent transitions from processing the entire 3D volume at once to processing 2D slices individually. This dimensionality reduction from 3D to 2D space enables the application of memory-efficient algorithms while preserving the essential reconstruction quality through subsequent synthesis of the sliced results.
2Measurement precision
If high-resolution reconstruction is performed, then image quality improves, but inference time increases
Solution Approach 1:
By dividing the reconstruction task into independent 2D slice operations, the patent enables parallel processing of multiple slices. This segmentation reduces the overall inference time while maintaining high resolution, as each slice can be processed more quickly and independently rather than requiring sequential processing of the entire 3D volume.
Solution Approach 2:
The patent applies the learned iterative scheme to 2D slices rather than the full 3D volume, performing a partial reconstruction that is then synthesized. This partial action on smaller units reduces computational burden and inference time while still achieving the desired high-resolution result through the combination of slice reconstructions.
3Measurement precision
If complete data is used for reconstruction, then reconstruction accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the complete dataset into multiple 2D slice subsets that can be processed independently. Each slice uses only the relevant projection data for that particular slice, reducing the computational complexity from O(N³) to O(N²) per slice while collectively utilizing the complete dataset for accurate reconstruction of the entire volume.
Data Source
AI summary
A computer-implemented method for image reconstruction using a learned iterative scheme, the method may comprise processing measured dual space data to reduce the resolution, initiating a reconstructed image in the primal space, determining dual and primal parameters defined in the dual and primal spaces, and identifying primal patches and dual patches. The method may further include iteratively updating the reconstructed image by applying a dual-space learned invertible operator to the dual parameters, applying a primal-space learned invertible operator to the primal parameters, updating the reconstructed image based on the updated primal parameters and, when necessary, based on upsampled primal parameters, upsampled dual parameters, and the upsampled reconstructed image. The method may also include outputting the updated reconstructed image.