Abstract
Ultrasound imaging (UI) is characterized by the presence of multiplicative speckle noise and various acquisition artefacts. Designing ultrasound (US) similarity measures thus requires a particular attention. In the specific context of motion estimation, incorporating US characteristics does not only benefit traditional methods but also learning-based approaches, which are highly sensitive to the quality of training data. Deriving similarity measures from a maximum likelihood (ML) perspective allows us to take these specificities into account. As opposed to the classical Rayleigh modelling, the proposed similarity measures incorporate more realistic scattering conditions, such as, varying speckle densities and shadowing. Specifically, the deviations from the Rayleigh statistics are modelled using the t -distribution for the complex radio-frequency (RF) signals and the Nakagami-Gamma (NG) compound model for the echo amplitudes. Furthermore, the model parameters are learnt patch-wise, which leads to data-adaptive similarity measures. The proposed criteria are investigated in the context of motion estimation using synthetic, phantom, as well as 2D and 3D in vivo images. The experimental results show an improvement in performance and robustness in comparison to the classical Rayleigh-based approach and state-of-the-art similarity measures.
Original language | English |
---|---|
Pages (from-to) | 1244 - 1254 |
Number of pages | 11 |
Journal | IEEE Journal of Selected Topics in Signal Processing |
Volume | 14 |
Issue number | 6 |
DOIs | |
Publication status | Published - 11 Jun 2020 |
MoE publication type | A1 Journal article-refereed |
Keywords
- similarity measures
- ultrasound
- heavy-tailed distributions
- motion estimation
- t-distribution
- Nakagami-Gamma distribution
- data-adaptive