Machine learning based bone segmentation in ultrasound

Abstract

Ultrasound (US) guidance is of increasing interest for minimally invasive procedures in orthopedics due to its safety and cost benefits. However, bone segmentation from US images remains a challenge due to the low signal to noise ratio and artifacts that hamper US images. We propose to learn the appearance of bone-soft tissue interfaces from annotated training data, and present results with two classifiers, structured forest and a cascaded logistic classifier. We evaluated the proposed methods on 143 spinal images from two datasets acquired at different sites. We achieved a segmentation recall of 0.9 and precision 0.91 for the better dataset, and a recall and precision of 0.87 and 0.81 for the combined dataset, demonstrating the potential of the framework.