Speckle2Self: Self-supervised ultrasound speckle reduction without clean data.

Area of Science:

• Computer Vision
• Medical Imaging
• Signal Processing

Background:

• Speckle noise in ultrasound (US) imaging degrades image quality due to complex wave interference.
• Existing deep learning denoising methods struggle with US speckle noise due to its tissue-dependent and spatially dependent nature.
• Self-supervised learning approaches like Noise2Noise are infeasible for US imaging as they require multiple independent noisy observations.

Purpose of the Study:

• To introduce Speckle2Self, a novel self-supervised algorithm for effective speckle reduction in ultrasound images.
• To address the limitations of existing denoising methods in handling tissue-dependent US speckle noise.
• To enable speckle suppression using only single noisy ultrasound observations.

Main Methods:

• Speckle2Self utilizes a multi-scale perturbation (MSP) operation to introduce scale-dependent variations in speckle patterns.
• The algorithm models the clean image as a low-rank signal and isolates the sparse noise component.
• Self-supervised learning framework designed for single-channel, spatially correlated noise.

Main Results:

• Speckle2Self demonstrates superior speckle reduction performance compared to conventional and state-of-the-art learning-based methods.
• The algorithm shows robust generalization across simulated and real human carotid ultrasound images from multiple machines.
• Validation confirms the effectiveness of the low-rank modeling and sparse noise isolation approach.

Conclusions:

• Speckle2Self offers a viable solution for unsupervised speckle reduction in ultrasound imaging.
• The proposed multi-scale perturbation strategy effectively handles the unique characteristics of US speckle noise.
• This method advances self-supervised learning applications in medical image denoising.