Learned Invertible Reconstruction for CBCT Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CBCT reconstruction methods face challenges such as high memory requirements for 3D data processing, limited generalization to different regions of interest, and poor imaging quality due to artifacts like streaking and beam hardening.
Innovation Solution
A learned iterative scheme using reversible residual neural networks and patch-wise computations is employed for image reconstruction, allowing for efficient training and reduced memory usage by dividing data into primal and dual patches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning reconstruction methods are used for CBCT, then reconstruction quality is improved, but memory requirements become prohibitively expensive
Solution Approach 1:
The patent divides the 3D volume data into multiple 2D slices along the z-axis, processing each slice independently through the neural network. This segmentation reduces the memory footprint from handling complete 3D volumes to processing manageable 2D slices, while still achieving 3D reconstruction quality through iterative refinement across multiple slices and iterations.
2Measurement precision
If deep learning reconstruction methods are trained on specific region data, then performance on that region is improved, but generalization to other organs deteriorates
Solution Approach 1:
The patent employs a universal neural network architecture that can be trained on data from one anatomical region and applied to reconstruct other regions. The network learns general reconstruction patterns and artifact correction strategies that transfer across different organs and imaging conditions, achieving both specialized performance and broad generalization capability.
Solution Approach 2:
The patent adapts the reconstruction process by adjusting network parameters and processing configurations based on the specific imaging scenario, organ type, and acquisition parameters. This allows the same base model to generalize across different applications while maintaining optimized performance for each specific case through parameter adaptation rather than retraining.
3Productivity
If classical reconstruction algorithms are used, then computational efficiency is maintained, but imaging quality deteriorates due to artifacts
Solution Approach 1:
The patent introduces a neural network as an intermediary post-processing step that takes the output from classical reconstruction algorithms (such as FBP or iterative methods) and refines it by removing artifacts and enhancing image quality. This intermediary network corrects streaking, beam hardening, and scatter-induced artifacts while preserving the computational efficiency of the underlying classical algorithm.
Solution Approach 2:
The patent merges classical reconstruction algorithms with deep learning techniques in a hybrid framework. The classical algorithm provides the base reconstruction and handles the computationally intensive forward/backward projections, while the neural network component handles artifact correction and quality enhancement, combining the strengths of both approaches for efficient high-quality reconstruction.
Data Source
AI summary
A system for image reconstruction comprises a memory for storing data and instructions and a processor system. The system obtains measured data y defined in a dual space, the measured data relating to a primal space. The system determines dual parameters defined in the dual space based on the measured data and primal parameters defined in the primal space based on back-projection of the dual parameters. The system updates the primal parameters and the dual parameters, by obtaining auxiliary dual parameters by forward-projecting the primal parameters, applying a dual-space learned invertible operator to the dual parameters, based on the auxiliary dual parameters, obtaining auxiliary primal parameters by back-projecting the dual parameters, and applying a primal-space learned invertible operator to the primal parameters, based on the auxiliary primal parameters. The system determines an image in the primal space based on the updated primal parameters.