Synthetic elastography from B-mode ultrasound through deep learning

Tissue elasticity can be locally estimated using shear-wave elastography (SWE), an advanced technique that measures the speed of laterally-traveling shear waves induced by a sequence of acoustic radiation force "push" pulses. However, SWE is not available on all ultrasound machines due to e.g. power, equipment, and procedural requirements; in particular, wireless devices would face challenges delivering the required power. Here, we propose a fully-convolutional deep neural network for the synthesis of an SWE image given the corresponding B-mode (side-by-side-view) image. Fifty patients diagnosed with prostate cancer underwent a transrectal SWE examination with SWE imaging regions chosen such that they covered the entire or parts of the prostate. The network was trained with the images of 40 patients and subsequently tested using 30 image planes from the remaining 10 patients. The neural network was able to accurately map the B-mode images to sSWE images with a pixel-wise mean absolute error of 4.8 kPa in terms of Young's modulus. Qualitatively, tumour sites characterized by high stiffness were mostly preserved (as validated by histopathology). Despite the need for further validation, our results already suggest that deep learning is a viable way to retrieve elasticity values from conventional B-mode images and can potentially provide valuable information for cancer diagnosis using devices on which no SWE imaging is available.