Neural Network CT Denoising via Synthetic Reference Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for reducing noise in low-dose CT images, such as deep neural network-based approaches, require high-dose reference images or repeated scans, which are challenging due to radiation risks and practical limitations.
Innovation Solution
A method that generates both training inputs and labels from the same existing CT scans, eliminating the need for high-dose CT images or repeated scans by splitting full-dose data into independent partial-dose scans, which are then used to train a machine learning-based system like a deep neural network for denoising low-dose CT images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural network-based denoising methods are used, then image noise is reduced, but high-dose reference images or repeated scans are required which increase radiation exposure
Solution Approach 1:
The patent creates synthetic training data by copying and transforming existing low-dose CT images through various operations (adding noise, adjusting intensity, geometric transformations) to generate artificial high-dose reference images without requiring actual high-dose scans. This allows the neural network to learn denoising patterns from synthesized pairs rather than real radiation-exposure pairs.
Solution Approach 2:
The patent changes the parameters of existing low-dose images by adjusting intensity values, adding controlled noise, and applying transformations to create synthetic high-dose versions. This parameter manipulation enables the generation of training data with varying dose levels without actual radiation exposure, allowing the network to learn across a range of conditions using only low-dose input images.
2Measurement precision
If iterative reconstruction techniques are used, then image quality is improved, but computational time increases significantly
Solution Approach 1:
The patent replaces the iterative mechanical reconstruction process with a trained neural network model. Instead of repeatedly solving complex optimization equations, the pre-trained network performs denoising through forward propagation, which is computationally much faster. The heavy computational work is moved to the training phase, allowing rapid inference during clinical use.
3Productivity
If image processing methods are used, then computational speed is improved, but image resolution decreases due to blurred edges
Solution Approach 1:
The patent replaces traditional image processing filters with a deep neural network that has learned optimal denoising operations. The network preserves edges and fine details by learning from training data, avoiding the blurring effects of conventional filters while maintaining computational efficiency through the learned transformation.
4Adaptability or versatility
If Ld2Ld training is used, then training without high-dose images is achieved, but independent reference measurements are required which need repeated scans
Solution Approach 1:
The patent creates synthetic reference images by copying and transforming the same low-dose input image through various operations. Instead of requiring independent reference measurements from repeated scans, the system generates artificial references from the single input image, eliminating the need for additional scans while providing sufficient training variability through transformations.
Solution Approach 2:
The system uses the low-dose CT images themselves to generate both training inputs and reference labels without external references. The same input images serve dual purposes: as the noisy input and as the basis for creating synthetic reference images through transformations, making the training process self-sufficient.
Data Source
AI summary
First and second substantially independent identically distributed half scans are obtained; the first substantially independent identically distributed half scan is used as training data to train a machine learning-based system, and the second substantially independent identically distributed half scan is used as label data to train a machine learning-based system. This produces a trained machine learning-based system.