Medical Image Denoising via Frequency-Decomposed Self-Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical imaging techniques such as spectral CT and MRI face challenges in denoising due to increased noise from reduced dose acquisitions and material decomposition, with existing methods like bilateral and guided filtering being time-consuming or unreliable, and trained machine learning models requiring difficult-to-acquire training pairs.
Innovation Solution
A frequency decomposition method is applied to two imaging datasets with different parameters, optimizing a trainable denoising algorithm using an objective function that depends on denoised high-frequency datasets to remove noise without losing relevant details, without the need for training images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If strong denoising is applied to reduce noise in medical imaging data, then noise suppression is improved, but fine potentially relevant details are lost
Solution Approach 1:
The image is segmented into different frequency components using a decomposition model, separating the image into a low-frequency component (capturing overall structure and intensity) and a high-frequency component (capturing fine details and edges). This segmentation allows different denoising strategies to be applied to different frequency bands, reducing noise while preserving important high-frequency details.
2Object-affected harmful factors
If bilateral filtering is used for denoising, then noise suppression is improved, but processing time increases and artifacts may occur
Solution Approach 1:
The patent replaces traditional iterative bilateral filtering (a mechanical filtering process) with a learned denoising model based on neural networks. This substitution leverages the power of deep learning to achieve denoising in a single pass or few passes, dramatically reducing processing time while avoiding the artifacts and intensity drifts associated with bilateral filters.
3Object-affected harmful factors
If trained machine learning models are used for denoising, then denoising performance is improved, but training data acquisition becomes difficult
Solution Approach 1:
The denoising model is trained to be self-supervised using the input noisy image itself as both training data and reference. The model learns to denoise by comparing its own predictions across different frequency components and iterations, eliminating the need for external ground truth images or manual training data acquisition. This self-service approach makes the model applicable to any medical imaging scenario without requiring specific training datasets.
Data Source
AI summary
For denoising medical imaging data, a first imaging dataset and a second imaging dataset are decomposed according to spatial frequency bands to generate high-frequency datasets corresponding to a high-frequency band and low-frequency datasets corresponding to a low-frequency band. A trainable denoising algorithm is trained by carrying out an optimization that uses at least one parameter of the denoising algorithm as an optimization variable and an objective function that depends on a denoised high-frequency dataset and the high-frequency dataset of the second imaging dataset. The denoised high-frequency dataset is generated by applying the denoising algorithm to the high-frequency dataset of the first imaging dataset. The trained denoising algorithm is applied to the high-frequency dataset of the first imaging dataset to generate a final denoised high-frequency dataset.