Comparison between U-Net and DeepLabV3 for Crawling Waves Sonoelastography approach

Abstract

Crawling Wave Sonoelastography is an elastography ultrasound-based approach that provides tissue stiffness information. Two mechanical sources generate an interference pattern on a tissue to obtaining a shear wave speed (SWS) map. Many estimators have been developed with important limitations such as the presence of artifacts, border effects or high computational cost. In this study, semantic segmentation approaches based on convolutional neuronal network (CNN) were performed to enhance the crawling waves differentiation. Training and validation were realized with 220 simulated frames in a range from 200 Hz to 480 Hz. Testing was made with real data acquired from experiments on a gelatin phantom with a circular inclusion vibrating on a range between 200 Hz and 360 Hz with steps of 20 Hz. It is found that comparable SWS map were generated using both architectures and model within a certain region of interest in the inclusion and in the background. Finally, CNNs presented lowest coefficient of variation and the highest contrast-to-noise ratio for most frequencies, which allows better discrimination between regions.