Iterative CT Image Reconstruction with Deep Learning Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Non-invasive imaging technologies face challenges in minimizing image noise and maximizing spatial resolution while maintaining computational efficiency, particularly in CT images, which are limited by factors like finite focal spot size and detector cell size, and are exacerbated by increased radiation dose or spatial resolution.
Innovation Solution
The method involves iterative image reconstruction using a sinogram input, with iterative processing that includes a datafit operation and a denoising operation performed using trained artificial neural networks, allowing for noise reduction and improved spatial resolution without increasing computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If radiation dose is increased to reduce image noise, then image noise is reduced, but patient dose increases
Solution Approach 1:
The patent replaces the traditional mechanical approach of increasing radiation dose to improve image quality with a computational approach using deep learning denoising algorithms. The neural network model processes the noisy low-dose images to produce high-quality images without requiring additional radiation exposure, thus substituting the physical mechanism (increasing dose) with an information processing mechanism (denoising algorithm).
Solution Approach 2:
The patent performs denoising operations during the iterative reconstruction process itself, rather than as a separate post-processing step. By integrating the deep learning denoising into the reconstruction loop, the system proactively removes noise as images are being formed, preventing noise accumulation and achieving high-quality results from low-dose data without requiring additional radiation.
2Measurement precision
If spatial resolution is increased to improve image detail, then spatial resolution is improved, but image noise increases
Solution Approach 1:
The patent replaces the traditional trade-off between spatial resolution and noise with a computational solution. The deep learning model is trained to preserve high-frequency spatial details while simultaneously suppressing noise patterns. This allows the system to achieve high spatial resolution without the proportional increase in visible noise that would normally occur, by using the neural network to distinguish between meaningful high-frequency signal and noise.
Solution Approach 2:
The denoising operation is applied locally during iterative reconstruction, allowing different regions of the image to be processed with appropriate denoising strength. The neural network learns to apply different levels of denoising to different image regions based on local characteristics, preserving edges and fine details where needed while removing noise in homogeneous regions, thus maintaining high spatial resolution without uniform noise increase.
3Reliability
If iterative reconstruction with denoising is performed to reduce image noise, then image noise is reduced, but computational complexity increases
Solution Approach 1:
The patent performs denoising operations during the iterative reconstruction process itself, rather than as a separate post-processing step. By integrating the deep learning denoising into the reconstruction loop, the system proactively removes noise as images are being formed, preventing noise accumulation and achieving high-quality results from low-dose data without requiring additional radiation.
Solution Approach 2:
The patent applies denoising operations selectively during the iterative reconstruction process, performing denoising at specific iteration steps rather than continuously. This partial application of denoising reduces the overall computational burden while still achieving effective noise suppression, balancing image quality improvement with computational efficiency.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to image reconstruction with favorable properties in terms of noise reduction, spatial resolution, detail preservation and computational complexity. The disclosed techniques may include some or all of: a first-pass reconstruction, a simplified datafit term, and/or a deep learning denoiser. In various implementations, the disclosed technique is portable to different CT platforms, such as by incorporating a first-pass reconstruction step.