Norm Ball Projection for Noise-Adaptive Deep Learning Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image reconstruction techniques, particularly in medical imaging, face stability issues due to the risk of hallucinating structures and unpredicted behavior when input data significantly differs from training data, and hard data consistency operations fail to account for noise levels, leading to reduced denoising capabilities and low signal-to-noise ratio outputs.
Innovation Solution
The implementation of a norm ball projection method for data consistency, which adjusts the balance between reconstruction network predictions and input data based on noise level information, using a trainable weighting parameter and L2 or L∞ norm balls to enforce data consistency, thereby accounting for varying noise levels and improving image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If hard data consistency operation is used to force output consistency with input measurements, then stability and robustness are improved, but noise propagation occurs and signal-to-noise ratio deteriorates
Solution Approach 1:
The patent changes the parameter of data consistency enforcement from a binary hard constraint to a continuous soft constraint by introducing a differentiable projection operation that takes noise level as input. This allows the system to adapt the strength of data consistency enforcement dynamically based on the measured noise level, preventing noise propagation while maintaining stability.
Solution Approach 2:
The patent introduces dynamic adaptability by making the data consistency operation noise-level dependent. The projection radius is determined by the noise level estimate, allowing the system to automatically adjust between enforcing strict data consistency (low noise) and allowing more flexibility (high noise), thereby preventing noise propagation while maintaining stability.
2Manufacturing precision
If deep learning reconstruction networks are used to improve image quality, then image quality is improved, but hallucination of structures and unpredicted behavior occur
Solution Approach 1:
The patent introduces a differentiable projection operation as an intermediary between the deep learning reconstruction network and the data consistency constraint. This intermediary guides the network output toward physically plausible solutions by projecting onto a norm ball defined by the measurement consistency constraint, preventing hallucinations while preserving the denoising and quality improvement capabilities of deep learning.
Solution Approach 2:
The patent implements feedback by using the measurement residual (difference between predicted and actual measurements) to guide the projection operation. The network output is projected based on how well it satisfies the data consistency constraint, providing continuous feedback that prevents unrealistic structures while maintaining image quality.
Data Source
AI summary
A computer-implemented method includes, based on scan data defining an input image, determining a reconstructed image using a reconstruction algorithm, and executing a data consistency operation for enforcing consistency between the input image and the reconstructed image. The data consistency operation includes using a norm ball projection that takes into account the available noise level information in order to automatically adjust the balance between the network prediction and the input measurements.