Unsupervised Learning of Speckle Removal from Real Ultrasound Acquisitions without Clean Data

Ultrasound (US) images suffer from speckle noise, a granular pattern that hampers contrast and resolution, making low-contrast structures critically difficult to identify. Albeit traditional filtering and machine learning approaches can handle this problem, both have limitations: such as the need of (hyper-)parameters fine-tuning or the necessity of data collection and annotation. In our study, we explored an unsupervised image filtering method based on blind denoising, so that we can systematically overcome the need of ground truth annotations. Our approach is based on a noise2noise u-net backbone (N2N) fed by a novel image representation approach. Dubbed Emulated Frequency Compound (EFQ), this study is intended to propose and validate it in the small data regime which is compatible with the typical applicative scenario of US imaging. As our experimental validation shows, the adoption of EFQ for N2N results in a favorable performance with respect to a number of state-of-the-art methods and related baselines.