Medical Image Denoising Without Clean References Using Image Sequences
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current denoising techniques for medical images, particularly in CT imaging, require clean reference images for training, which are difficult to obtain and result in limited generalization across varying acquisition parameters, leading to sub-optimal performance and over-smoothing or incomplete noise removal.
Innovation Solution
A method that trains a machine-learning model using sequences of connected images, leveraging convolutional memory units and noise models based on acquisition parameters, allowing denoising without clean reference images and adapting to different noise levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If clean reference images are used for training, then denoising performance is improved, but obtaining such images requires higher radiation dose and longer acquisition time
Solution Approach 1:
The system uses the noisy image itself to generate the clean reference image through an identity transformation branch, eliminating the need for separate clean images acquired with high radiation dose. The noisy image passes through a clean image generator that produces the reference without requiring additional patient exposure
Solution Approach 2:
The training image is segmented into multiple branches: one branch processes the noisy image through denoising, while another branch applies identity transformation to generate the clean reference. This segmentation allows the system to create paired training data from single noisy images without requiring separate clean image acquisitions
2Measurement precision
If denoising model is trained on images with specific acquisition parameters, then performance on those parameters is improved, but generalization to different acquisition parameters deteriorates
Solution Approach 1:
The system incorporates acquisition parameters as additional input channels to the neural network. By feeding parameters such as tube voltage, tube current, and slice thickness directly into the model, the network learns to adapt its denoising behavior based on the specific acquisition conditions, enabling generalization across different parameter settings
Solution Approach 2:
The denoising model is designed to handle multiple acquisition scenarios within a single unified architecture. By integrating parameter-based conditioning, the same model can process images acquired with different radiation doses, tube voltages, and other parameters, eliminating the need for separate models for each acquisition protocol
3Quantity of substance
If synthetic noise is added to high-quality images for training, then training data is obtained, but the noise model fails when mathematical assumptions do not match acquisition setup
Solution Approach 1:
Instead of relying on synthetic noise models that make mathematical assumptions, the system extracts the actual noise characteristics directly from the noisy input image itself. The noise is treated as an inherent property of the input data rather than something to be artificially generated, ensuring the noise model matches the actual acquisition conditions
Solution Approach 2:
The system uses acquisition parameters to condition the noise modeling process. By adjusting the noise characteristics based on the actual acquisition parameters (tube voltage, tube current, etc.), the noise model accurately reflects the real noise present in the image rather than relying on generic synthetic noise assumptions
Data Source
AI summary
A method for training a machine-learning model for denoising is provided, including retrieving a target image data frame. The target image data frame is one image data frame of a sequence containing imaging data of a subject. The method further includes retrieving at least one prior image data frame and at least one following image data frame of the sequence. Contents of the prior and following image data frames each overlap partially with contents of the target image data frame. The method further includes retrieving acquisition parameters associated with the image data frames of the sequence and generating a prediction for a denoised target image data frame based on the prior and following image data frame. The method trains a machine-learning algorithm based on the prediction and a noise model based on the acquisition parameters. Also provided are a system and denoising method.