3D Tomographic Reconstruction With Multi-Directional ML Regularizers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tomographic reconstruction techniques face challenges in extending two-dimensional image reconstruction to three dimensions, particularly in medical imaging, due to the computationally expensive nature of handcrafted regularizers and the limited applicability of these techniques to three-dimensional image reconstruction, leading to issues such as noise sensitivity and reconstruction artefacts.
Innovation Solution
A computer-implemented method using iterative reconstruction techniques with trained machine learning models as regularizers, specifically trained neural networks, to process tomographic projection data from multiple orthogonal directions, minimizing a cost function to reconstruct high-quality three-dimensional images, including high contrast and resolution, while being less sensitive to noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If handcrafted regularizers are used for tomographic reconstruction, then the reconstruction process can be controlled, but the computational cost becomes expensive and the techniques are limited to two-dimensional image reconstruction
Solution Approach 1:
The patent replaces handcrafted mathematical regularizers with a machine learning-based regularizer that uses a neural network to model prior knowledge about images. This substitution transforms the computational approach from explicit mathematical formulations to implicit learning-based models, achieving superior 3D reconstruction quality with reduced computational complexity compared to traditional handcrafted methods
Solution Approach 2:
The patent changes the parameter space from handcrafted mathematical operators to learned neural network parameters. By training the neural network on large datasets of 3D images and their 2D projections, the system learns optimal reconstruction parameters that generalize to new data, enabling high-quality 3D reconstruction without the computational burden of handcrafted 3D regularizers
2Reliability
If iterative reconstruction algorithms are used to improve image quality, then noise sensitivity is reduced, but the computation time increases significantly
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network regularizer on large datasets of 3D images and their projections before actual reconstruction tasks. This pre-training phase captures essential image priors and relationships, allowing the system to perform rapid inference during reconstruction without requiring time-consuming iterative optimization, thus achieving both noise insensitivity and computational efficiency
3Volume of moving object
If two-dimensional image reconstruction techniques are extended to three dimensions, then volumetric imaging is achieved, but the handcrafted regularizers become computationally expensive and less applicable
Solution Approach 1:
The patent achieves universality by training a single neural network regularizer on diverse 3D image data that can handle multiple reconstruction scenarios and anatomical structures. This universal 3D-aware regularizer replaces multiple specialized 2D regularizers, enabling consistent high-quality reconstruction across different volumetric datasets without requiring handcrafted adjustments for each specific application
Data Source
AI summary
A computer-implemented method of tomographic reconstruction, the method comprising: receiving an input dataset comprising tomographic projection data of an object; and reconstructing an image of the object using an iterative reconstruction technique, wherein the iterative reconstruction technique seeks to minimize a cost function, the cost function including a first regularizer and a second regularizer, wherein the first regularizer is a first trained machine learning model, trained using image data associated with the object and extracted from a first direction and the second regularizer is a second trained machine learning model, trained using image data associated with the object and extracted from a second direction.