Computationally-efficient model-based clutter suppression with ADMIRE

Aperture domain model image reconstruction (ADMIRE) algorithm demonstrates off-axis and reverberation clutter suppression while preserving the signal of interest. However, ADMIRE is computationally slow due to the large size of the model matrix. Generating a large in vivo image may require a total serial run-time of over ten hours per frame. Such a lengthy run-time limits the usefulness of ADMIRE in practice. The goal of this study is to accelerate ADMIRE without losing image information. We used a singular value decomposition (SVD) algorithm to reduce the number of predictors in the ADMIRE model matrix, which is necessarily oversampled due to the correlation characteristics of reverberation clutter. To assess the SVD-based reduced model, we simulated 12 realizations of an anechoic cyst phantom with varying levels of clutter. We then measured image quality metrics resulting from the full and reduce models, along with delay-and-sum (DAS). The contrast improvement relative to DAS for the full model was 3.2 ± 0.4 dB, 12.6 ± 1.0 dB, 8.2 ± 0.7 dB and 4.7 ± 1.1 dB for the uncluttered, −10, 0 and 10 dB of signal-to-clutter ratio, while the improvement for the reduced model was 3.6 ± 0.3 dB, 12.7 ± 0.7 dB, 8.2 ± 0.4 dB and 4.7 ± 0.6 dB for each case, respectively. We also measured the total serial run-time of each form of ADMIRE in MATLAB on a 3.40 GHz CPU desktop computer. The full model had 1,320 predictors versus 96 in the reduced model at the focal depth. This resulted in total single-core serial run-times of 1,294 sec and 91 sec for the full versus reduced models, respectively.