Octave Convolution Neural Network for Low-Dose CT Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional CT denoising techniques face challenges in producing high-quality images due to difficulties in reflecting signal characteristics across different frequency bands, leading to inefficient removal of artifacts and loss of detailed structures in low-dose CT images, which can result in diagnostic inaccuracies and increased radiation exposure.
Innovation Solution
An octave convolution-based neural network is employed to process CT images, decomposing them into high-frequency and low-frequency components and using inter-convolution and intra-convolution to learn and separate these features, along with loss functions like L1 and edge loss to optimize image quality and reduce spatial redundancy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional deep learning-based CT denoising techniques are used, then the model can process CT images, but the image quality is degraded due to inability to reflect signal characteristics in different frequency bands
Solution Approach 1:
The patent applies segmentation by dividing the feature extraction process into high-frequency and low-frequency bands using separate convolutional neural network branches. This allows the model to process different frequency components independently, capturing signal characteristics that conventional single-branch models miss, thereby improving image quality while maintaining frequency adaptability
Solution Approach 2:
The patent introduces a frequency dimension by processing images through parallel high-frequency and low-frequency convolutional branches. This dimensional separation enables the model to capture multi-frequency signal characteristics simultaneously, resolving the contradiction between image quality and frequency band adaptability
2Productivity
If filter-based reconstruction technique is used, then reconstruction speed is fast, but the technique is vulnerable to noise and artifacts making it difficult to obtain high-quality images
Solution Approach 1:
The patent replaces traditional filter-based mechanical reconstruction methods with a deep learning-based convolutional neural network approach. This substitution maintains fast reconstruction speeds while significantly improving noise resistance and artifact suppression through learned feature extraction and denoising capabilities
Solution Approach 2:
The model performs self-service by automatically learning optimal denoising and reconstruction features from training data, eliminating the need for manual filter design and tuning. This enables the system to adapt to different noise patterns and image characteristics while maintaining high reconstruction speeds
3Manufacturing precision
If model-based reconstruction technology is used, then high-quality images can be obtained with robustness against noise and artifacts, but computational amount is high and reconstruction speed is slow
Solution Approach 1:
The patent segments the complex model-based reconstruction into lighter convolutional neural network operations with separate high-frequency and low-frequency branches. This segmentation reduces computational complexity while preserving the robust noise and artifact suppression capabilities, achieving both high image quality and fast reconstruction speed
Solution Approach 2:
The patent changes the computational parameters by using convolutional neural network operations with fewer parameters compared to traditional model-based reconstruction methods. This parameter reduction maintains denoising effectiveness while significantly improving reconstruction speed and reducing computational burden
Data Source
AI summary
An image processing method of the present disclosure may include receiving a scanned image, and processing the received image through an octave convolution-based neural network to output a high-quality image and an edge image for the received image. The octave convolution-based neural network may include a plurality of octave encoder blocks and a plurality of octave decoder blocks. Each octave encoder block may include an octave convolutional layer, and may be configured to output a high-frequency feature map and a low-frequency feature map for the image.