Segmentation of Vessels in Ultra High Frequency Ultrasound Sequences using Contextual Memory

Abstract

High resolution images provided by Ultra High Frequency Ultrasound (UHFUS) scanners permit the vessel-based measurement of the Intimal-Media Thickness (IMT) in small vessels, such as those in the hand. However, it is challenging to precisely determine vessels in UHFUS sequences due to severe speckle noise obfuscating their boundaries. Current level set-based approaches are unable to identify poorly delineated boundaries and are not robust against varying speckle noise. While recent neural network-based methods, including recurrent neural networks, have shown promise at segmenting vessel contours, they are application specific and do not generalize to datasets acquired from different scanners, such as a traditional High Frequency Ultrasound (HFUS) machine, with different scan settings. Our goal for a segmentation approach was the accurate localization of vessel contours, and generalization to new data within and across biomedical imaging modalities. In this paper, we propose a novel ultrasound vessel segmentation network (USVS-Net) architecture that assimilates features extracted at different scales using Convolutional Long Short Term Memory (ConvLSTM) and segments vessel boundaries accurately. We show the results of our approach on UHFUS and HFUS sequences. To show broader applicability beyond US, we also trained and tested our approach on a Chest X-Ray dataset. To the best of our knowledge, this is the first learning-based approach to segment deforming vessel contours in both UHFUS and HFUS sequences.