Synthetic CT Image Generation for Neural Network Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for reducing noise in CT images using convolutional neural networks (CNNs) are limited in their ability to distinguish noise from anatomical features, resulting in noise-reduced images that still contain undesirable noise levels.
Innovation Solution
The method involves creating synthetic CT images from high-resolution reference images, such as MR images acquired using optimized pulse sequences, and histological images, to train a neural network model that can more accurately differentiate noise from anatomical features, thereby improving noise reduction performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a CNN is trained on real CT images to reduce noise, then the model can process real images, but the model cannot accurately distinguish noise from anatomical features resulting in insufficient noise reduction
Solution Approach 1:
The patent creates synthetic CT images that copy the anatomical structures and noise characteristics of real CT images, but with ground truth labels indicating noise locations. These synthetic copies serve as training data, allowing the CNN to learn noise patterns without the ambiguity present in real images, thereby improving noise distinction accuracy and reduction effectiveness
Solution Approach 2:
The patent introduces synthetic images as an intermediary training medium between real CT images and the CNN model. These synthetic images act as a bridge, providing labeled noise annotations that are difficult to obtain from real images, thus enabling the model to learn effective noise discrimination before being applied to real clinical images
2Manufacturing precision
If high-resolution reference images are used to generate synthetic CT images, then the training data quality improves, but the complexity of the image generation process increases
Solution Approach 1:
The patent segments the image generation process into distinct steps: tissue segmentation of reference images to identify anatomical structures, synthetic image generation by assigning CT values to segmented regions, and noise addition to create training pairs. This segmentation of the complex generation process into manageable steps improves synthetic image quality while making the overall process more controllable and less complex
Solution Approach 2:
The patent performs preliminary tissue segmentation and anatomical structure identification on high-resolution reference images before generating synthetic CT images. This preliminary action ensures that the synthetic images have accurate anatomical foundations, improving quality while organizing the complex generation process into preparatory and execution phases
Data Source
AI summary
The current disclosure provides methods and systems to reduce an amount of noise in image data. In one example, a method for creating synthetic computed tomography (CT) images for training a model to reduce an amount of noise in acquired CT images is proposed, comprising performing a tissue segmentation of reference images of an anatomical region of a subject to determine a set of different tissue types of the reference images; and generating synthetic CT images of the reference images by assigning CT image values of the synthetic CT images based on the different tissue types.