Reproduce “Speckle2Speckle: Unsupervised Learning of Ultrasound Speckle Filtering Without Clean Data” on OpenResearch

About this paper

Rüdiger Göbl, Christoph Hennersperger, Nassir Navab

The "Speckle2Speckle" method introduces an unsupervised deep learning approach for ultrasound speckle filtering, leveraging realistic simulations to generate paired, independent speckle images for training. It achieves despeckling performance comparable to state-of-the-art conventional filters on real phantom and in-vivo data, while significantly reducing processing time from minutes to tens of milliseconds.

Key insights

• Realistic ultrasound simulations can effectively overcome the lack of clean ground truth data, enabling robust training of unsupervised deep learning models for speckle reduction.
• A deep learning model trained exclusively on simulated ultrasound data can generalize effectively to real experimental phantom and in-vivo acquisitions, demonstrating strong domain transfer capabilities.
• Integrating domain-specific knowledge, such as prioritizing interface preservation, into the loss function can guide unsupervised learning to produce clinically more useful image enhancements.

Summary by alphaXiv.