Sensorless End-to-End Freehand Ultrasound with Physics Inspired Network

Three-dimensional ultrasound (3D US) imaging finds broad applications in clinical practice. Compared to using 'wobbler' (motor embedded) or 'matrix array' transducers which suffers from a restricted field of view, freehand US offers more flexibility in mapping the 3D volume along the scanning path. Nevertheless, current methods for reconstructing 3D US volumes using freehand scanning are challenged by significant elevational shifts along the scanning trajectory. Previous research explored the integration of motion sensors to resolve transducer drift, yet frequent motion artifacts often compromise the benefits brought by these sensors. Recent work turned to deep neural networks (DNNs) for estimating the relative pose of imaging planes between frames in the scanning trajectory, allowing for sensorless freehand US. However, these data-driven models fall short in accuracy as they lack physical constraints, hence limiting their use in clinical practice. Inspired by the physical properties of US, we design a physics inspired DNN architecture that can better leverage the contextual cues between elevational frames for sensorless freehand 3D US reconstruction, and demonstrate substantially improving reconstruction accuracy.