Neural CT Reconstruction via Signed Distance Function
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
X-ray computed tomography (CT) image quality is compromised by object motion during data acquisition, particularly when imaging fast-moving structures like the heart, leading to significant blurring of key structures and limitations in current methods to correct motion artifacts.
Innovation Solution
A neural computed tomography (CT) algorithm that uses a neural network to represent object boundaries with a signed distance function (SDF), enabling time-resolved image reconstruction without explicit motion estimation, by optimizing the SDF representation to match acquired sinogram data and reduce motion artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional CT reconstruction methods are used, then acquisition speed is maintained, but image quality deteriorates due to motion artifacts
Solution Approach 1:
The patent replaces conventional mechanical reconstruction algorithms with a neural network-based system. The neural network is trained to recognize and correct motion artifacts in CT images, substituting the traditional mathematical reconstruction process with a learned model that can identify and compensate for motion-related distortions, thereby improving image quality without requiring slower acquisition speeds
Solution Approach 2:
The patent transforms the reconstruction problem by changing the parameter space from direct image reconstruction to latent space representation. The neural network learns to map from corrupted image space to corrected image space through training on paired datasets, effectively changing the reconstruction parameters from fixed algorithmic operations to adaptive learned transformations that can correct motion artifacts
2Measurement precision
If motion correction algorithms are applied, then image quality improves, but processing time increases
Solution Approach 1:
The patent performs motion correction in advance during the training phase of the neural network. The network is pre-trained on large datasets of motion-corrupted and motion-corrected image pairs, so that when deployed, it can rapidly apply the learned corrections without requiring time-consuming post-processing. The heavy computational work is done beforehand, enabling fast real-time or near-real-time correction during actual use
Solution Approach 2:
The patent creates a learned model (copy of the correction knowledge) during training that can be repeatedly applied without reprocessing the training data. The neural network captures the essence of motion correction patterns and stores them as weights and biases, allowing rapid application of corrections to new images without repeating the full correction computation each time
3Measurement precision
If explicit motion estimation is performed, then motion correction can be applied, but system complexity increases
Solution Approach 1:
The patent extracts the motion correction capability directly from the neural network output without requiring separate motion estimation steps. Instead of taking out motion parameters as intermediate results, the network directly produces corrected images by learning the transformation from corrupted to clean images, eliminating the need for explicit motion estimation algorithms and reducing overall system complexity
Solution Approach 2:
The neural network serves multiple functions simultaneously: it performs denoising, deblurring, and motion correction all in a single unified model. This multi-functional approach replaces what would traditionally require multiple separate algorithms for different correction tasks, simplifying the system while maintaining comprehensive correction capabilities across various types of degradation
Data Source
AI summary
Methods and systems that pertain to an image reconstruction of motion-corrupted images are disclosed. In some embodiments of the disclosed technology, an image reconstruction method includes obtaining an initial estimate of object boundaries from motion-corrupted images, creating an implicit representation of the motion-corrupted images, updating the implicit representation of the motion-corrupted images using acquired imaging data to generate an updated implicit representation of the motion-corrupted images, and converting the updated implicit representation of the motion corrupted images to an explicit set of motion-corrected images.