Low-dose CT reconstruction with Noise2Noise network and testing-time fine-tuning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning shows promise for low-dose CT (computed tomography) image denoising and reconstruction.
• Existing methods often require normal-dose images for training, which are not always available (e.g., dynamic CT, large patients).

Purpose of the Study:

• To develop a deep learning-based low-dose CT reconstruction algorithm that does not require clean, normal-dose images for training.
• To enable effective low-dose CT reconstruction in scenarios where clean data is inaccessible.

Main Methods:

• Proposed a novel iterative reconstruction algorithm utilizing a Noise2Noise network for image prior.
• Network weights were fine-tuned concurrently with image reconstruction via an alternating optimization scheme.
• A self-consistent loss was created using projection data splitting and mapping FBP (filtered backprojection) results via a deep neural network.

Main Results:

• The Noise2Noise reconstruction method outperformed existing algorithms (TV, Non-local mean, CSC) in RMSE, SSIM, and texture preservation on the 2016 Low-dose CT Challenge dataset.
• Performance was robust across varying noise levels, hyperparameters, and network structures.
• Achieved competitive results even with randomly initialized network weights (no pre-training), demonstrating empirical convergence.

Conclusions:

• The proposed Noise2Noise reconstruction method significantly enhances image quality in low-dose CT.
• The algorithm effectively utilizes only noisy data for training or fine-tuning, offering a versatile solution.
• The method demonstrates applicability with or without network pre-training.