Sparse Tomographic Reconstruction Using Geometry and Physics Priors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current tomographic image reconstruction methods, both traditional and data-driven, struggle with ultra-sparse sampling scenarios due to susceptibility to noise, motion artifacts, and the lack of incorporation of prior knowledge, leading to severe artifacts and limited generalizability, especially in CT and MRI imaging.
Innovation Solution
A deep learning-based framework that integrates physics and geometry priors into the reconstruction process, utilizing dual-domain learning to bridge the dimensionality gap between 2D projection and 3D image domains through a geometric back-projection operator and deep neural networks, synthesizing novel-view projections or k-space samples to enhance image reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional iterative reconstruction methods are used, then the reconstruction process can incorporate prior knowledge, but the methods fail to yield high-fidelity images in sparse sampling scenarios
Solution Approach 1:
The patent introduces a geometry module as an intermediary between the data-driven deep learning network and the reconstruction process. This module incorporates physics and geometry priors (such as geometric relationships between projections and images) to guide the reconstruction, enabling high-fidelity image recovery from ultra-sparse sampling while maintaining the benefits of data-driven approaches
Solution Approach 2:
The patent creates a composite reconstruction framework that combines three elements: (1) data-driven deep learning networks for feature extraction, (2) physics and geometry priors for constraint guidance, and (3) iterative optimization for solution refinement. This composite approach leverages the strengths of each component to overcome the limitations of individual methods in sparse sampling scenarios
2Measurement precision
If deep learning methods are used for image reconstruction, then impressive performance is achieved, but the models lack transparency and interpretability
Solution Approach 1:
The geometry module serves as an interpretable intermediary that bridges the black-box deep learning network and the reconstruction output. It incorporates explicit physics and geometry priors (such as geometric relationships between projections and images) that provide transparency and interpretability to the reconstruction process while maintaining high accuracy
Solution Approach 2:
The patent segments the reconstruction process into distinct functional modules: a data-driven network for feature extraction, a geometry module for physics-constrained transformation, and an optimization component for refinement. This segmentation makes each component's role transparent and interpretable while preserving overall reconstruction accuracy
3Reliability
If more views are acquired to improve reconstruction quality, then image fidelity increases, but imaging dose and data acquisition complexity increase
Solution Approach 1:
The patent applies partial action by using only a small subset of views (ultra-sparse sampling) rather than acquiring complete view sets. The geometry module and deep learning network compensate for the missing information, enabling high-quality reconstruction from fewer projections, thereby reducing imaging dose and acquisition complexity while maintaining image fidelity
4Measurement precision
If data-driven approaches are used, then reconstruction performance improves, but geometric misalignment occurs in multi-view image processing
Solution Approach 1:
The geometry module acts as an intermediary that corrects geometric misalignment by incorporating explicit geometry priors. It ensures that the transformation between different views and the reconstructed image respects the underlying geometric relationships, preventing the misalignment issues that arise in purely data-driven multi-view processing
Data Source
AI summary
A method for medical imaging performs a sparse-sampled tomographic imaging acquisition by an imaging system to produce acquired sparse imaging samples; synthesizes by a first deep learning network unacquired imaging samples from the acquired imaging samples to produce complete imaging samples comprising both the acquired imaging samples and unacquired imaging samples; transforms by a physics module the complete imaging samples to image space data based on physics and geometry priors of the imaging system; and performs image refinement by a second deep learning network to produce tomographic images from the image space data. The physics and geometry priors of the imaging system comprise geometric priors of a physical imaging model of the imaging system, and prior geometric relationships between the sample and image data domains.