Invertible Network Image Correction via Physics Prior Regularization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image correction methods face challenges in accurately removing artifacts from images due to simplistic assumptions, lack of sufficient training data, and poor generalization, especially when the imaging protocol changes, leading to suboptimal reconstruction quality and complexity balancing issues.
Innovation Solution
The use of an invertible deep generative model as a regularizer, combined with a physics model of the imaging system, provides a closed-form expression for the prior probability, enabling more accurate and versatile probability predictions, and allowing for joint optimization of imaging and weights, thereby minimizing artifacts in images acquired from medical scanners.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If manually crafted prior models are used, then algorithm complexity is reduced, but reconstruction quality deteriorates due to simplistic assumptions
Solution Approach 1:
The patent replaces manually crafted prior models with a deep generative model that learns image priors automatically from data. This substitution of mechanical/manual model crafting with a learned generative model enables the system to capture complex image structures without requiring manual simplification, thereby maintaining high reconstruction quality while reducing the need for manual intervention.
Solution Approach 2:
The patent changes the parameters of the prior model from fixed manually-designed constraints to learned parameters from a deep generative model. By training the generative model on large datasets, the system adapts the prior model parameters to match the specific characteristics of the imaging system and artifacts, improving reconstruction quality without proportionally increasing algorithm complexity.
2Manufacturing precision
If supervised machine learning is used, then reconstruction quality improves, but adaptability deteriorates when imaging protocols change requiring retraining
Solution Approach 1:
The patent employs a deep generative model that can be dynamically adapted to different imaging protocols. Rather than requiring complete retraining when protocols change, the system can fine-tune the generative model or adjust its parameters to accommodate new protocols, maintaining both high reconstruction quality and adaptability through dynamic model adjustment.
Solution Approach 2:
The patent segments the imaging system into a physics-based forward model and a data-driven generative prior model. This segmentation allows the physics model to handle protocol-specific transformations while the generative model captures general image structures, enabling protocol changes to be handled by adjusting only the physics model parameters rather than retraining the entire system.
3Manufacturing precision
If data-driven direct mapping methods are used, then reconstruction quality improves, but device complexity increases due to coupling of prior model and acquisition protocol
Solution Approach 1:
The patent segments the image restoration system into two independent components: a physics-based forward model that handles acquisition protocol specifics and a deep generative prior model that learns general image structures. This segmentation decouples the complexity, allowing each component to be optimized independently while maintaining high overall reconstruction quality.
Solution Approach 2:
The deep generative model serves as a universal prior that can be applied across multiple imaging protocols and artifact types. By learning general image structures from diverse training data, the generative model provides protocol-agnostic priors that work across different acquisition settings, reducing the need for protocol-specific model customization and thereby reducing overall system complexity.
Data Source
AI summary
For correction of an image from an imaging system, an inverse solution uses an imaging prior as a regularizer and a physics model of the imaging system. An invertible network is used as the deep-learnt generative model in the regularizer of the inverse solution with the physics model of the degradation behavior of the imaging system. The prior model based on the invertible network provides a closed-form expression of the prior probability, resulting in a more versatile or accurate probability prediction.