Multi-Level Noise Sampling From Single Image for Low-Dose Tomography Reconstruction.

Area of Science:

• Medical Imaging
• Image Processing
• Radiology

Background:

• Low-dose digital radiography (DR) and computed tomography (CT) reduce radiation exposure but degrade image quality.
• Effective denoising is crucial for low-dose medical imaging due to reduced signal-to-noise ratios.
• Single-image-based denoising methods offer a solution but often lack sufficient training data and utilize limited information.

Purpose of the Study:

• To develop an advanced single-image denoising technique for low-dose DR and CT.
• To address the limitations of existing methods by generating richer multi-perspective denoising clues.
• To improve the effectiveness of denoising in medical imaging applications.

Main Methods:

• A novel Multi-Level Noise Sampling (MNS) method is proposed for low-dose tomography denoising.
• MNS generates multi-level noisy sub-images by partitioning the input space into subspaces.
• An optimization function is introduced to guide training using prior knowledge from multi-level noisy sub-images.

Main Results:

• The MNS method demonstrates theoretical superiority in single-image denoising.
• Extensive experiments on clinical low-dose CT and DR datasets validate its effectiveness.
• MNS outperforms state-of-the-art supervised and self-supervised denoising methods.

Conclusions:

• The MNS approach offers a superior solution for denoising low-dose medical images.
• This method effectively bridges the gap between self-supervised and supervised denoising networks.
• MNS significantly enhances image quality in low-dose radiography and tomography.