CNN CT Denoising With User-Controlled Noise-Bias Tradeoff
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods for low-dose CT image denoising provide a single level of performance, lacking user-tunable control over the trade-off between noise reduction and image properties, leading to potential misrepresentations and loss of structural features.
Innovation Solution
A convolutional neural network (CNN) with a hyper-parameter σ for controlling the noise-bias trade-off, allowing flexible selection of spatial resolution and bias for specific tasks, using a loss function to manage mean squared error between denoised and ground truth images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If traditional image filters (nonlocal means, block-matching 3D) are used for noise reduction, then noise is reduced to a great extent, but structural features are over-smoothed and lost
Solution Approach 1:
The patent applies parameter changes by modifying the loss function parameters in the neural network training process. Specifically, it uses a parameterized loss function that incorporates a regularization term with a controllable parameter λ, allowing dynamic adjustment of the trade-off between noise reduction and structural feature preservation during training and inference
Solution Approach 2:
The patent replaces traditional mechanical filtering approaches (nonlocal means, block-matching 3D) with a neural network-based approach. This substitution enables the system to learn optimal denoising strategies from data while maintaining structural features through learned representations rather than fixed filtering operations
2Object-affected harmful factors
If deep learning techniques (RED-CNN) are used for denoising, then noise is suppressed and structural features are maintained, but there is no parametric control over the aggressiveness of noise reduction
Solution Approach 1:
The patent implements dynamics by introducing a learnable parameter λ that can be dynamically adjusted during training and inference. This parameter controls the strength of the regularization term in the loss function, enabling adaptive control over the denoising aggressiveness based on specific diagnostic tasks or user preferences
Solution Approach 2:
The patent changes parameters by incorporating a parameterized loss function with controllable hyperparameters. The loss function includes a regularization term weighted by parameter λ, which can be tuned to achieve different levels of noise reduction while maintaining structural integrity, providing versatility for various diagnostic scenarios
3Measurement precision
If machine learning methods are used for image denoising, then dramatic improvements over traditional methods are achieved, but only a single level of performance is provided with no user-tunable control
Solution Approach 1:
The patent makes the system dynamic by introducing learnable parameters that can be adjusted after training. The parameter λ in the loss function can be modified to provide different levels of denoising strength, allowing users to tune the system according to specific diagnostic needs while maintaining the high image quality achieved through machine learning
Solution Approach 2:
The patent achieves universality by designing a flexible loss function with controllable parameters that can adapt to different diagnostic tasks. The same trained network can provide multiple performance levels by adjusting the parameter λ, making the system versatile for various applications from aggressive denoising to subtle enhancement
Data Source
AI summary
Provided herein are methods of producing processed images having user-controlled image quality levels. In some embodiments, the methods include receiving a selected input value of a control parameter from a user in a trained electronic neural network in which the selected input value determines an image quality level of the processed test image. The image quality level typically comprises relative amounts of a noise measure and a bias measure in the processed test image. The methods also generally include passing test image data through the trained electronic neural network, and outputting from the trained electronic neural network the processed test image. Related systems and computer program products are also provided.