Unsupervised Neural Network for Low-Dose CT Image Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for reconstructing low-dose X-ray CT images using deep learning face challenges due to the difficulty in obtaining matched low- and routine-dose CT image pairs, especially in coronary CTA, where cardiac motion and contrast agent injection complicate the process, leading to noise and reduced image quality, and existing generative adversarial networks risk generating artificial features.
Innovation Solution
An unsupervised learning-based neural network using a convolutional framelet is employed, learned with predefined identity, cyclic, and adversarial losses to remove noise from low-dose X-ray CT images and reconstruct high-quality images, avoiding the degeneracy problem by using unmatched data and configuring neural networks with local and non-local bases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is used to remove noise from low-dose CT images, then noise removal performance is improved, but matched low- and routine-dose CT image pairs are difficult to obtain due to cardiac motion and contrast agent injection
Solution Approach 1:
Instead of using matched pairs (low-dose as input, routine-dose as target), the patent inverts the approach by using unmatched pairs where routine-dose images are used as input and low-dose images as target, or by using self-supervised approaches where the network learns from the distribution of routine-dose images alone. This inversion resolves the data availability problem while maintaining denoising capability.
Solution Approach 2:
The patent employs self-supervised learning where the network learns to denoise without requiring paired training data. The model uses only routine-dose images or unmatched low-dose images to learn the underlying image statistics and noise characteristics, enabling it to denoise low-dose images independently without external paired data.
2Measurement precision
If generative adversarial network is used to match data distributions, then image quality is improved, but the network may generate artificial features due to degeneracy of mapping
Solution Approach 1:
The patent incorporates feedback mechanisms through the adversarial loss function where the discriminator provides feedback to the generator about the realism of generated images. This feedback loop ensures that the generated images match the true data distribution of routine-dose images while maintaining anatomical accuracy, preventing the generation of artificial features.
Solution Approach 2:
The patent uses a composite loss function that combines multiple loss terms (adversarial loss, reconstruction loss, and regularization terms) to train the GAN. This composite approach balances the competing objectives of generating realistic images while maintaining fidelity to the input low-dose image, preventing the network from generating purely artificial features.
3Measurement precision
If iterative reconstruction methods are used to obtain clear images, then image quality is improved, but reconstruction time is relatively long due to iterative applications of forward and back projections
Solution Approach 1:
The patent replaces the iterative mechanical reconstruction process (forward and back projections) with a deep learning-based approach. The neural network is trained offline to learn the mapping from low-dose to routine-dose images, enabling rapid online reconstruction without iterative computations, thus dramatically reducing reconstruction time while maintaining image quality.
Solution Approach 2:
The patent performs the complex denoising computation in advance during the offline training phase, creating a pre-trained neural network model. During actual clinical use, the pre-trained model rapidly applies the learned denoising function without requiring iterative computations, achieving both high image quality and fast reconstruction speed.
Data Source
AI summary
A method for processing an unmatched low-dose X-ray computed tomography (CT) image using a neural network and an apparatus therefor are provided. The method includes receiving a low-dose X-ray CT image and removing noise from the low-dose X-ray CT image using a unsupervised learning based neural network learned using unmatched data to reconstruct a routine-dose X-ray CT image corresponding to the low-dose X-ray CT image.