Robust adversarial learning model to segment non-speckle regions in blood flow echo

Abstract In our previous study, we analyzed the contrast of blood flow echo, and non-speckle regions were more frequently detected in the porcine blood with the high flow velocity. However, this contrast method is dependent on the degree of smoothing and threshold for outliers. This study developed a new U-Net model incorporating domain adaptation with both in silico and experimental data. This model segments blood flow echo into speckle and non-speckle regions. The performance of the developed U-Net model with several conditions of scatterer number density from 0.1 to 1.5 scatterers/mm3 and scatterer amplitude from 2 to 50 times against the speckle component was assessed using in silico data and experimental data with blood-mimicking fluid. The results indicated that the developed U-Net model with adversarial learning could stably detect non-speckle regions compared to the model without the adversarial learning and the contrast analysis method, in both in silico and experimental data.