Motion estimation for ultrasound image sequences using deep learning

Abstract

Increased number of medical images being used for medical praxis makes automatic processing of images a necessity. Conventional techniques of motion estimation in Ultrasound images such as exhaustive search-based block matching (ES-BM) are known to be computationally expensive and are unsuitable for portable devices. On the other hand, research in computer vision has helped the development of deep learning-based techniques for real-time motion estimation of day-to-day non-medical objects. In this paper, we propose to adopt one such deep neural network-based Fully Convolutional Siamese tracker along with Linear Kalman Filter (SiamFC-LKF) to track regions of interest in ultrasound image sequences. Siamese networks use two convolutional neural networks to create a feature map of the given reference block and the possible candidate blocks of subsequent frames. The candidate block of the subsequent frame with the maximum correlation value is considered as the tracked output. Since SiamFC does not consider any motion model, we introduce a linear Kalman filter to track the wall of the carotid artery. SiamFC-LKF and ES-BM were tested on five different image sequences of the longitudinal section of the carotid artery. Our experiments showed that SiamFC-LKF was 6 times faster and performed better than ES-BM in most of the cases.