Medical Image Denoising With Double Over-Parameterization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep image prior (DIP) methods for image denoising, such as those used in PET imaging, face challenges with overfitting and require early stopping of the training process, which can affect image quality and detail preservation.
Innovation Solution
A double over-parameterized (DOP) approach utilizing CT or MRI images from the same patient as anatomical prior information, with discrepant learning rates for vectors g and h, to model image noise separately, avoiding overfitting and improving noise reduction and detail preservation in PET images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If early stopping is used to prevent overfitting in DIP-based methods, then overfitting is avoided, but image quality and detail preservation deteriorate
Solution Approach 1:
The patent segments the single noise vector into two separate noise vectors (g and h), creating a double over-parameterized model. This segmentation allows the model to learn noise patterns more effectively without overfitting, as the two vectors can capture different aspects of the noise distribution independently, thereby maintaining image quality while preventing overfitting.
Solution Approach 2:
The patent changes the parameter configuration by introducing two noise vectors instead of one, effectively doubling the parameter count for noise modeling. This parameter change enables the model to achieve better generalization by capturing more complex noise patterns, resolving the contradiction between preventing overfitting and maintaining image quality.
2Device complexity
If traditional DIP methods are used for noise reduction, then computational simplicity is maintained, but noise reduction effectiveness and detail preservation worsen
Solution Approach 1:
By segmenting the noise modeling into two separate vectors with discrepant learning rates, the patent achieves better noise reduction effectiveness while maintaining relative computational simplicity. The segmentation allows each vector to specialize in different noise characteristics, improving overall denoising quality without requiring complex architectural changes.
Solution Approach 2:
The patent employs parameter changes by using discrepant learning rates for the two noise vectors, allowing them to converge at different speeds and capture different temporal aspects of noise patterns. This parameter adjustment significantly improves noise reduction quality while keeping the optimization process relatively simple.
3Device complexity
If single noise vector modeling is used in DIP, then model simplicity is maintained, but noise modeling accuracy deteriorates
Solution Approach 1:
The patent segments the noise modeling task into two separate vectors, each capable of capturing different aspects of the noise distribution. This segmentation improves noise modeling accuracy by allowing each vector to specialize in specific noise patterns, while the overall model structure remains relatively simple and manageable.
Solution Approach 2:
By changing the parameter configuration from a single noise vector to two noise vectors with different learning rates, the patent achieves superior noise modeling accuracy. The discrepant learning rates allow the vectors to learn at different paces, capturing both short-term and long-term noise patterns effectively.
Data Source
AI summary
A method, apparatus, and non-transitory computer-readable storage medium for image denoising whereby a deep image prior (DIP) neural network is trained to produce a denoised image by inputting the second medical image to the DIP neural network and combining a converging noise and an output of the DIP network during the training such that the converging noise combined with the output of the DIP network approximates the first medical image at the end of the training, wherein the output of the DIP network represents the denoised image.