Deep learning for ultrasound image reconstruction on a realtime system

Ultrasound speed-of-sound reconstruction has been proven to provide images of higher quality in some specific clinical situations. However, commercial available ultrasound machines usually do not support speed-of-sound image reconstruction. Jetson AGX Xavier and Nvidia Clara are real-time embedded systems that provides computing power for running deep neural network applications. This thesis report decribes the process of training deep learning neural network models for speed-of-sound reconstruction using datasets of diverging wave transmission scheme and virtual source transmission scheme respectively. The integration of the image reconstruction pipeline into the embedded systems is also covered in the report. Multiple models have been trained using both datasets respectively. The performance of the models are measured and compared with models using the same dataset and with models of different dataset. The experiment result shows that models trained with virtual source transmission scheme perform better when objects are closer to the scanner while models with diverging wave are better at detecting object far away. The result also shows the possibility and difficulties when integrating the software and hardware of the image reconstruction pipeline into the embedded systems