Automatic Vessel Recognition and Segmentation: a novel Deep Learning Architecture with Transfer Learning Approach

Abstract

Ultrasound (US) imaging stands as a valid alternative to X-rays based methodologies for navigation and intraoperative tracking of vascular probes, thanks to its non-ionizing nature. However, US images quality is highly operator dependent, being subject to probe’s orientation and contact force. In recent years, researchers have worked to develop Robotic US Systems (RUSS), granting the acquisition of good quality real-time US images, without the need of an expert operator [1]. Besides, to facilitate US images analysis, deep learning strategies have been developed. Applications in the field include the automatic segmentation of vessels, which is fundamental during endovascular procedures. Intraoperatively, an automatic method to classify images based on the presence of vessels and selectively segment only vascular images would be valuable. For example, during hand-held probe procedures it would increase the quality of information feedback. In RUSS, it would enable automatic adjustment of probe positioning serving as alternative to manual positioning by highly trained sonographers. Additionally, a method for precisely discriminating the presence of vessels in the image plane could increase safety in visual-servoing platforms, by preventing possible control instabilities generated by imaging artifacts. However, segmentation architectures typically assume that the processed image contains vessels to be segmented [2], but this is not granted in real intraoperative settings especially at the beginning of the procedure when the imaging probe is not yet optimally positioned. To address these unmet needs, in this paper we propose a multi-task convolutional neural network (CNN) architecture able to distinguish between vessel and no vessel images, in addition to segmenting them. The goal of such architecture is to enable robust and automatic US images analysis in real intraoperative settings.