Spatial Resolution Enhancement in Ultrasound Images from Multiple Annotators Knowledge

Abstract

Enhancement of spatial resolution for medical images improves clinical procedures such as diagnosis of different diseases, image registration, and tissue segmentation. Although different methods have been proposed in the literature to tackle this problem, each of them comes with their own strengths and their own weaknesses. In this work, we present a novel approach for the enhancement of spatial resolution in ultrasound images that aims at improving resolution enhancement by combining different interpolation methods. The methodology is based on learning from multiple annotators, also known as learning from crowds, a recent development in supervised learning to incorporate the diverse levels of knowledge that different experts can have on a prediction problem, in order to leverage the prediction performance in a single model. In particular, we consider each pixel intensity value in each new high resolution image as a corrupted version of a gold standard. Each of the single interpolation algorithms acts as an expert that provides a level of intensity for a particular pixel. We then use a regression scheme for multiple annotators based on Gaussian Processes with the aim of computing an estimate of the actual image from the noisy annotations given by the interpolation algorithms. We compare our approach against two super resolution schemes based on Gaussian process regression. This comparison is performed using the mean square error (MSE) for the interpolation validation and the Dice coefficient (DC) for the morphological validation. Results obtained show that our approach is a promising methodology for enhancing spatial resolution in ultrasound images.