Knowledge Distillation for Mobile Quantitative Ultrasound Imaging

In this paper, we present a mobile-friendly quantitative ultrasound imaging network (MQI-net) that can provide high-quality quantitative ultrasonic images in resource-limited portable environment. Quantitative ultrasound imaging has been studied to identify biomechanical properties of tissue on an absolute scale and it now demonstrates clinical potential in differential cancer diagnosis and hepatic steatosis. Recently, learning-based approaches demonstrated the real-time reconstruction of quantitative ultrasound images while using conventional ultrasound systems. However, such methods require extensive computations which hinders rapid real-world adoption. To overcome the limitation, the MQI-net is proposed. It reduces the neural network parameters by 96% compared to the conventional convolutional neural network. In addition, a knowledge distillation scheme is employed in MQI-net to enhance reconstruction accuracy. The performance of the system is assessed through numerical simulations and in-vivo breast patient measurements. The proposed knowledge distillation scheme enhances the accuracy of the MQI-net by 33% and 17% for attenuation coefficient and speed of sound reconstruction, respectively