A long short-term memory network for vessel reconstruction based on laser doppler flowmetry via a steerable needle

Abstract

Hemorrhage is one risk of percutaneous intervention in the brain that can be life-threatening. Steerable needles can avoid blood vessels thanks to their ability to follow curvilinear paths, although knowledge of vessel pose is required. To achieve this, we present the deployment of laser Doppler flowmetry (LDF) sensors as an in-situ vessel detection method for steerable needles. Since the perfusion value from an LDF system does not provide positional information directly, we propose the use of a machine learning technique based on a Long Short-term Memory (LSTM) network to perform vessel reconstruction online. Firstly, the LSTM is used to predict the diameter and position of an approaching vessel based on successive measurements of a single LDF probe. Secondly, a "no-go" area is predicted based on the measurement from four LDF probes embedded within a steerable needle, which accounts for the full vessel pose. The network was trained using simulation data and tested on experimental data, with 75 % diameter prediction accuracy and 0.27 mm positional Root Mean Square (RMS) Error for the single probe network, and 77 % vessel volume overlap for the 4-probe setup.