Nerve segmentation of ultrasound images with Bayesian U-net models

Abstract

Ultrasound imaging is a widely adopted method for non-invasive examination of internal structures, valued for its cost-effectiveness, real-time imaging capability, and absence of ionizing radiation. Its applications, including Peripheral Nerve Blocking (PNB) procedures, benefit from the direct visualization of nerve structures. However, the inherent distortions in ultrasound images, arising from echo perturbations and speckle noise, pose challenges for accurate localization of nerve structures even for experienced practitioners. Computational techniques, particularly Bayesian inference, offer a promising solution by providing uncertainty estimates in model predictions. This study focused on the development and implementation of an optimal Bayesian U-net for nerve segmentation in ultrasound images, presented through a user-friendly application. Utilizing Bayesian Convolution layers, and the Monte Carlo Dropout method, were the two Bayesian techniques explored and compared, with a specific emphasis on facilitating medical professionals’ decision-making processes. The research revealed that integrating the Monte Carlo Dropout technique for Bayesian inference yields the most optimal results. The Bayesian model demonstrates an average binary accuracy of 98.99%, an average dice coefficient score of 0.72, and an average IOU score of 0.57, when benchmarked against a typical U-net. The culmination of this work is an application designed for practical use by medical professionals, providing an intuitive interface for Bayesian nerve segmentation in ultrasound images. This research contributes to the broader understanding of Bayesian techniques in medical imaging models, offering a comprehensive solution that combines advanced methodology with user-friendly accessibility.