Evaluation of Tracking Algorithms for Contrast Enhanced Ultrasound Imaging Exploration
Cristina Laura Sîrbu, Ciprian Seiculescu, Ghita Adrian Burdan, Tudor Voicu Moga, Cătălin Daniel Căleanu · 2022
Deep Neural Networks (DNN) and Deep Learning (DL) are providing efficient solutions to many Computer Vision (CV) problems. As state-of-the-art solutions, DNNs have been successfully applied to automate the detection and classification of lesions in medical ultrasound images for various diseases. Classification of focal liver lesions (FLLs) in Contrast Enhanced Ultrasound Imaging (CEUS) recording is becoming a topic of interest as it could efficiently assist healthcare practitioners in the diagnosis process. Like all DNN-based solutions, model performance is highly dependent on the quality of the training set. To produce quality training datasets, specialists have to manually label thousands of frames, which is very labor intensive and prone to errors. In this paper, we investigate various tracking algorithms that can be utilized to efficiently generate training images using just a few annotated frames. Unlike the vast majority of use cases for which most of the tracking algorithms have been proposed and tested against, CEUS recordings are characterized as being very noisy and having low contrast, thus previous established rankings should be re-evaluated: based on experimental results, we conclude that the Kernelized Correlation Filter tracking algorithm emerged as one of the best candidates for tracking in CEUS imagery.