Disentangled Feature Denoising for Low-Dose CT Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing denoising models for medical images, particularly in CT scans, are ineffective when applied to images acquired with different acquisition parameters, such as reduced radiation doses, due to their specificity to training noise levels, limiting their generalizability and applicability.
Innovation Solution
A system and method that utilize domain-agnostic learning with disentangled feature representation to train a neural network model, allowing it to denoise images by separating anatomy and noise features, enabling the model to reconstruct images from varying noise levels and parameters, including lower radiation doses, through a standard and reduced quality image module with a loss calculation module for tuning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a denoising model is trained on standard quality images with specific noise levels, then it achieves good denoising performance on training noise levels, but it becomes ineffective when applied to images with different acquisition parameters and noise levels
Solution Approach 1:
The patent segments the noise feature from the anatomy feature using separate encoder networks. The standard noise encoder and reduced quality noise encoder independently process noise components, allowing the model to handle different noise levels separately while preserving anatomical information. This segmentation enables the system to adapt to various acquisition parameters without retraining the entire model.
Solution Approach 2:
The patent creates a universal denoising model that can process both standard quality and reduced quality images through a shared generator architecture. The generator receives features from different encoders and can reconstruct images at multiple quality levels, making the model versatile across different noise conditions and acquisition parameters without requiring separate models for each scenario.
2Object-affected harmful factors
If radiation dose is reduced in CT scans to lower patient exposure, then patient safety is improved, but noise and artifacts in reconstructed images significantly increase
Solution Approach 1:
The patent introduces an intermediary standard noise encoder that processes reduced quality images and transforms their noise characteristics into standard noise features. This intermediary component acts as a bridge between reduced quality and standard quality representations, allowing the generator to reconstruct images with reduced noise while preserving the benefits of lower radiation dose acquisition.
Solution Approach 2:
The patent changes the noise level parameter by encoding reduced quality noise features through a standard noise encoder, effectively transforming the noise characteristics. This parameter transformation allows the model to operate on low-dose images while producing outputs with noise levels appropriate for diagnostic quality, thereby maintaining information quality despite reduced radiation exposure.
3Measurement precision
If separate models are created for different imaging parameters and noise levels, then each model is optimized for its specific condition, but the overall system complexity and number of models required increases
Solution Approach 1:
The patent merges multiple specialized encoders (standard anatomy encoder, standard noise encoder, reduced quality anatomy encoder, reduced quality noise encoder) into a single integrated model architecture. The shared generator combines features from all encoders, unifying the processing of different image types and noise levels into one model, thereby reducing system complexity while maintaining specialized processing capabilities for various imaging conditions.
Data Source
AI summary
A system and method are provided for denoising images. A standard image module is configured to generate a standard anatomy feature and a standard noise feature from a standard image and reconstruct the standard image from the standard anatomy feature and the standard noise feature. A reduced quality image module is configured to generate a reduced quality anatomy feature and a reduced quality noise feature from a reduced quality image, and reconstruct the reduced quality image from the reduced quality anatomy feature and the reduced quality noise feature. A loss calculation module is provided for calculating loss metrics at least partially based on a comparison between 1) the reconstructed standard image and the standard image, and 2) the reconstructed reduced quality image and the reduced quality image. Upon providing the standard image module with the reduced quality anatomy feature, the standard image module outputs a reconstructed standard transfer image.