Automated parameter selection for super-resolution ultrasound image processing using statistics of fast and slow time sampling

Abstract

Recent advances in the field of contrast-enhanced ultrasound (US) have led to the development of super-resolution US (SRUS) imaging for improved spatial resolution by up to an order of magnitude. Variations in US system settings and microbubble (MB) contrast agent properties can lead to differences in image stacks that force a need to fine tune multiple SRUS image processing parameters. Current SRUS algorithms require expertise to tune the numerous parameters that influence achievement of optimal image results. Thus, there is a need for automated selection of SRUS parameters for both MB detection and localization, which are two critical steps in the image formation process. To that end, the purpose of this research was to develop novel methods to automate selection of spatiotemporal filtering parameters for MB detection and thresholding levels used for subsequent localization. Simulations provided synthetic US data at various signal-to-noise ratios (SNRs) and tissue clutter signal levels. US images with MB flow in crossing channels against a background of tissue were produced. Statistics of both fast and slow time sampling of the US image stacks were used to select cutoffs for singular value decomposition (SVD) filtering and then thresholds for MB localization. In addition, in vivo datasets from liver cancer-bearing rats were used to evaluate SRUS image quality from use of automatic parameter selection. Contrast-enhanced US imaging was performed with a preclinical system (Vevo 3100, FUJIFILM VisualSonics Inc) equipped with a 15 MHz linear array transducer. Raw image data was collected and analyzed offline using custom software. With more challenging levels of SNR and tissue clutter content, automated parameter selection for SRUS resulted in a considerable improvement in positive predictive value (PPV) with similar levels of sensitivity and localization accuracy as compared to conventional fixed parameter settings. For the in vivo US data, our novel image processing methods created highly detailed spatial maps of the tumor microvasculature.