Residual Noise Weighting for Partial Medical Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical imaging modalities suffer from noise due to statistical signal formation and data acquisition, leading to impaired diagnostic value, and existing machine-learning denoising methods often result in underestimating noise levels, causing regression-to-mean behavior and loss of valuable image information.
Innovation Solution
A computer-implemented method using a convolutional neural network to generate a residual noise image, which is weighted and combined with the original image to produce a partially denoised image, utilizing predetermined weighting factors based on reference image noise characteristics to avoid over-smoothing and maintain natural noise texture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine-learning denoising methods are used to reduce noise in medical images, then image quality is improved, but over-denoising or regression-to-mean behavior occurs
Solution Approach 1:
The patent applies partial denoising by controlling the denoising strength to be less than complete removal of noise. The system intentionally retains some noise characteristics by adjusting denoising parameters, thereby avoiding over-denoising while still improving image quality. This partial action allows preservation of natural noise textures that contain diagnostic information.
Solution Approach 2:
The patent changes the parameter of noise level from complete removal to controlled retention. By adjusting denoising parameters to achieve a target noise level that matches reference images, the system prevents regression-to-mean behavior and preserves authentic noise patterns while removing excessive noise.
2Reliability
If fully denoised images are generated using elaborate machine-learning approaches, then noise level is reduced, but computational complexity and processing time increase
Solution Approach 1:
The patent extracts only the necessary denoising operation from complex machine-learning approaches. Instead of using elaborate generative models, the system isolates and applies a simpler denoising algorithm that achieves sufficient noise reduction without the excessive computational complexity of full machine-learning pipelines.
Solution Approach 2:
The patent employs computationally inexpensive denoising methods rather than expensive, complex machine-learning models. The approach uses simpler algorithms that can be applied quickly and discarded, avoiding the high computational cost and long processing times associated with elaborate machine-learning approaches.
3Measurement precision
If machine-learning models are optimized for mean squared error, then denoising performance is improved, but regression-to-mean behavior occurs
Solution Approach 1:
The patent introduces feedback by comparing the denoised image against a reference image with known noise characteristics. The system uses this feedback to adjust denoising parameters, ensuring that the output maintains appropriate noise levels rather than exhibiting regression-to-mean behavior. This closed-loop control preserves natural noise textures while achieving accurate denoising.
Data Source
AI summary
A mechanism for generating a partially denoised image. A residual noise image, obtained by processing an image using a convolutional neural network, is weighted. The blending or combination of the weighted residual noise image and the (original) image generates the partially denoised image.