X-Ray CT High-Quality Image Generation From Single-Scan Data

The patent changes the fundamental parameter of noise reduction from traditional filtering methods to machine learning-based methods. Specifically, it uses denoising autoencoders and other neural network models that learn optimal noise reduction parameters from training data, allowing for noise reduction while preserving structural information that traditional filters blur.

The patent replaces the mechanical filtering process (adaptive filters) with an information-processing approach using machine learning models. The denoising autoencoder and other neural networks process the projection data or reconstructed images through learned transformations rather than mathematical filtering operations, thereby reducing noise while maintaining structural fidelity.

The patent introduces multiple parameter changes: using denoising autoencoders instead of iterative reconstruction, applying total variation regularization, and using deep learning models that learn optimal parameters. These changes allow noise reduction while preserving spatial resolution and natural noise texture by learning from training data rather than applying fixed iterative algorithms.

The patent uses training data that copies the statistical properties and structural characteristics of actual imaging data. By training denoising models on pairs of noisy and clean images (or projection data), the models learn to copy the essential features and textures of the original data while removing noise, thereby preserving spatial resolution and natural appearance.

The patent segments the learning data preparation into manageable components: generating synthetic projection data with controlled noise levels, creating corresponding ground truth images, and organizing them into training sets. This segmentation of the data preparation process makes the complex task of creating machine learning training data systematic and manageable.

The patent uses copying to create training data by generating synthetic projection data that mimics real imaging conditions. By copying the statistical properties and structural characteristics of actual imaging data into synthetic datasets, the patent creates realistic training data without requiring extensive manual annotation or complex data collection protocols.

The patent uses copying to create multiple versions of training data from a single imaging session. By generating synthetic projection data with different noise levels and applying various reconstruction algorithms to create corresponding ground truth images, the patent copies the essential imaging characteristics without requiring repeated physical exposures of the subject.

The patent segments the acquisition of learning data into computational steps rather than physical imaging steps. By dividing the projection data and creating multiple datasets through computational processing and synthetic generation, the patent obtains diverse training data without increasing the actual X-ray exposure to the subject.

The patent replaces the physical approach of increasing exposure (more X-ray photons) with an information-processing approach using machine learning. The denoising autoencoders and other neural networks process the low-dose projection data or reconstructed images to remove noise computationally, thereby achieving noise reduction without increasing the physical exposure to the subject.

The patent fundamentally changes the parameter of noise reduction from exposure amount to computational processing. By using deep learning models that learn optimal denoising parameters from training data, the patent achieves noise reduction through information processing rather than physical dose increase, enabling low-dose imaging with high image quality.