PET Image Reconstruction Using Uncertainty-Weighted Bayesian DIP
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based image reconstruction methods, such as those using deep image prior (DIP) for PET imaging, suffer from overfitting issues due to the lack of clear strategies to prevent overfitting, and do not effectively utilize uncertainty information for improving image quality across iterations.
Innovation Solution
A modified Bayesian deep image prior (BDIP) is integrated within a Plug-and-Play (PnP) framework for PET image reconstruction, incorporating an uncertainty-weighted loss term that penalizes regions with higher uncertainty more than those with lower uncertainty, using aleatoric and epistemic uncertainty estimation to generate risk maps and improve reconstruction quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If unsupervised denoisers like deep image prior are used for image reconstruction, then the ability to match supervised methods performance is improved, but overfitting occurs and strategies to prevent overfitting are unclear
Solution Approach 1:
The patent implements a feedback mechanism by computing uncertainty estimates from the reconstructed image and using this uncertainty information to guide the reconstruction process. The uncertainty loss function provides feedback about regions with high uncertainty, allowing the system to adjust reconstruction parameters and prevent overfitting in those regions while maintaining performance in reliable areas.
Solution Approach 2:
The patent changes the loss function parameters by introducing an uncertainty-weighted loss term that dynamically adjusts the weighting of different regions in the reconstruction objective. This allows the system to reduce the influence of uncertain regions during optimization, effectively controlling overfitting without sacrificing overall reconstruction quality.
2Reliability
If deep learning-based methods are used for image reconstruction, then image quality and robustness are improved, but hallucination of artifacts occurs on unseen data
Solution Approach 1:
The patent converts the harmful effect of uncertainty into a beneficial constraint by using uncertainty estimates to weight the loss function. Regions with high uncertainty (which could lead to artifacts) are down-weighted, transforming the uncertainty information from a source of potential harm into a mechanism for preventing artifact generation, especially in unseen data scenarios.
3Reliability
If iterative reconstruction methods are used, then image quality is improved through multiple iterations, but computational time and complexity increase
Solution Approach 1:
The patent applies partial action by focusing computational resources on regions with high uncertainty rather than uniformly processing the entire image. The uncertainty-weighted loss function automatically identifies and prioritizes problematic regions for correction, reducing the overall computational burden compared to uniform iterative refinement across all image regions.
Data Source
AI summary
Model-based image reconstruction (MBIR) methods using convolutional neural networks (CNNs) as priors have demonstrated superior image quality and robustness compared to conventional methods. Studies have explored MBIR combined with supervised and unsupervised denoising techniques for image reconstruction in magnetic resonance imaging (MRI) and positron emission tomography (PET). Unsupervised methods like the deep image prior (DIP) have shown promising results and are less prone to hallucinations. However, since the noisy image is used as a reference, strategies to prevent overfitting are unclear. Recently, Bayesian DIP (BDIP) networks that model uncertainty tend to prevent overfitting without requiring early stopping. However, BDIP has not been studied with data-fidelity term for image reconstruction. Present disclosure provides systems and method that implement a MBIR framework with a modified BDIP. Specifically, an uncertainty-based penalty is included to the BDIP to improve reconstruction across iterations.