Tomographic Reconstruction Using Wavelet Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional tomographic reconstruction algorithms face challenges in balancing computational efficiency, patient dose, scanning speed, image quality, and artifacts, particularly due to the limitations of convolutional neural networks (CNNs) in handling space-variant tasks and the computational feasibility of fully connected deep neural networks for high-dimensional image reconstruction problems.
Innovation Solution
The use of deep learning techniques involving neural networks with layers based on wavelets, wavelet frames, or sparsifying transforms, which take tomographic transforms of measured data as inputs to overcome the limitations of CNNs and conventional algorithms, reducing complexity and dimensionality, and enabling efficient large-scale, space-variant tomographic reconstruction and correction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If convolutional neural networks (CNNs) with weight sharing are used for tomographic reconstruction, then training time and memory requirements are reduced, but the network becomes limited to space-invariant operations and cannot handle space-variant tasks effectively
Solution Approach 1:
The patent segments the reconstruction process into multiple stages: a CNN performs initial reconstruction to generate a preliminary image, then a separate processing stage applies space-variant corrections using the precomputed Fisher information matrix and point spread function. This segmentation allows each component to specialize - the CNN handles the bulk reconstruction efficiently while the correction stage handles space-variant nuances
Solution Approach 2:
The patent introduces an intermediary representation (the preliminary reconstructed image from the CNN) that bridges between the space-invariant CNN operations and the space-variant correction requirements. This intermediary allows the system to leverage the efficiency of CNNs while still achieving space-variant accuracy through subsequent processing
2Adaptability or versatility
If fully connected deep neural networks are used for tomographic reconstruction, then space-variant tasks can be handled, but the approach is not computationally feasible for high-dimensional image reconstruction problems
Solution Approach 1:
The patent divides the reconstruction workload into two segments: a CNN that handles the high-dimensional initial reconstruction efficiently, and a separate correction module that applies space-variant adjustments. This segmentation avoids the computational infeasibility of fully connected networks while preserving space-variant capability where needed
Solution Approach 2:
The patent changes the operational parameters of the network by using a CNN with shared weights for the main reconstruction task, then applies space-variant corrections through precomputed matrices rather than through additional network layers. This parameter change maintains computational feasibility while achieving space-variant accuracy
3Device complexity
If conventional reconstruction algorithms are used, then computational simplicity is maintained, but image quality, noise reduction, and artifact suppression are insufficient
Solution Approach 1:
The patent introduces an intermediary CNN-based reconstruction step that produces a preliminary image, which then serves as input for the correction stage. This intermediary representation enables the system to achieve superior image quality and noise reduction compared to conventional algorithms while maintaining reasonable computational complexity
Data Source
AI summary
The present approach relates to the use of machine learning and deep learning systems suitable for solving large-scale, space-variant tomographic reconstruction and/or correction problems. In certain embodiments, a tomographic transform of measured data obtained from a tomography scanner is used as an input to a neural network. In accordance with certain aspects of the present approach, the tomographic transform operation(s) is performed separate from or outside the neural network such that the result of the tomographic transform operation is instead provided as an input to the neural network. In addition, in certain embodiments, one or more layers of the neural network may be provided as wavelet filter banks.