3D Tomographic Reconstruction Artifact Correction Using Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
3D tomographic reconstruction from phaseless images suffers from high information loss due to limited illumination angles and the absence of phase data, leading to inaccurate optical index measurements and reconstruction artifacts.
Innovation Solution
A method using a trained neural network to correct reconstruction artifacts by comparing simulated and acquired 3D tomographic images, employing a numerical model with deformation noise to train the network and applying it to correct the acquired images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If phaseless imaging is used to acquire only intensity information, then the acquisition process is simplified and faster, but the phase information is lost leading to inaccurate optical index measurements
Solution Approach 1:
The patent creates a synthetic complete dataset by combining phaseless images with simulated phase information. The neural network learns to map from the simplified phaseless images to the complete complex images that contain both phase and amplitude information, effectively copying the missing phase data without requiring actual phase measurements during acquisition.
Solution Approach 2:
The neural network acts as an intermediary that processes the incomplete phaseless image data and transforms it into complete complex image data. This intermediary component recovers the lost phase information by learning the relationship between phaseless images and their corresponding complete images from training data.
2Device complexity
If limited illumination angles are used (e.g., maximum angle ≤ 20°), then the acquisition system is simpler and less complex, but information loss increases causing lengthening of reconstructed images
Solution Approach 1:
The patent generates synthetic complete datasets by combining limited-angle phaseless images with simulated full-angle illumination data. The neural network learns to map from the limited-angle inputs to full-angle reconstructed images, copying the missing angular information through the trained model rather than requiring actual full-angle acquisition.
Solution Approach 2:
The neural network is pre-trained on synthetic data that simulates the complete imaging process. This preliminary training allows the network to compensate for the limited acquisition angles by having learned the relationship between limited-angle measurements and the expected complete image structure before actual reconstruction occurs.
3Reliability
If neural network training uses simulated three-dimensional numerical models with deformation noise, then the training process is more realistic and robust, but the training data generation becomes more complex
Solution Approach 1:
The patent introduces deformation noise as a parameter variation in the synthetic training data. By randomly deforming the ground truth structures and adjusting the corresponding phaseless images, the training process learns to handle realistic variations and reconstruction artifacts, making the model more robust to similar variations in actual applications.
Solution Approach 2:
The patent converts the harmful effect of deformation noise into a beneficial training feature. Instead of using perfect, noise-free synthetic data, the introduction of deformation noise makes the training more challenging and realistic, causing the neural network to learn better generalization capabilities and become more robust to real-world variations.
Data Source
AI summary
A method is provided for correcting a reconstruction artefact of a three-dimensional tomographic image. The method includes the steps of providing an acquired three-dimensional tomographic image from a cell group, and applying the image to a neural network trained in advance to determine a corrected tomographic image.