A methodology for peripheral nerve segmentation using a multiple annotators approach based on Centered Kernel Alignment

Peripheral Nerve Blocking (PNB) is a technique commonly used to perform regional anesthesia and for pain management. The success of PNB procedures depends on the accurate location of the target nerve. Recently, ultrasound imaging has been widely used to locate nerve structures to carry out PNB, due to it enables a non-invasive visualization of the target nerve and the anatomical structures around it. However, the ultrasound images are affected by several artifacts making difficult the accurate delimitation of nerves. In the literature, several approaches have been proposed to carry out automatic or semi-automatic segmentation. Nevertheless, these methods are designed assuming that the gold standard is available, and for this segmentation problem this gold standard can not be obtained considering that it corresponds to subjective interpretation. In this sense, for building those segmentation models, we do not have access to the actual label but an amount of subjective annotations provided by multiple experts. To deal with this drawback we use the concepts of a relatively new area of machine learning known as “Learning from crowds”, this area deals with supervised learning problems considering the case when the gold standard is not available. In this project, we develop a nerve segmentation system that includes: a preprocessing stage, feature extraction methodology based on adaptive methods, and a Centered Kernel Alignment (CKA) based representation to measure the annotators performance for building a classifier with multiple annotators in order to support peripheral nerve segmentation. Our approach to classification with multiple annotators based on CKA is tested on both simulated data and real data; similarly, the methodology of automatic segmentation proposed in this work was tested over ultrasound images labeled by a set of specialists who give their opinion about the location of nerve structures. According to the results, we conclude that our methodology can be used to locate nerve structures in ultrasound images even if the gold standard (the actual location of nerve structures) is not available in the training stage. Moreover, we determine that the approach proposed in this work could be implemented as a guiding tool for the anesthesiologist to carry out PNB procedures assisted by ultrasound imaging