Localizing 2D Ultrasound Probe from Ultrasound Image Sequences Using Deep Learning for Volume Reconstruction

Abstract

This paper presents an ultrasound (US) volume reconstruction method only from US image sequences using deep learning. The proposed method employs the convolutional neural network (CNN) to estimate the position of a 2D US probe only from US images. Our CNN model consists of two networks: feature extraction and motion estimation. We also introduce the consistency loss function to enforce. Through a set of experiments using US image sequence datasets with ground-truth motion measured by a motion capture system, we demonstrate that the proposed method exhibits the efficient performance on probe localization and volume reconstruction compared with the conventional method.