A web-based application for difference assessment of medical image segmentations

In this work we address the problem of dfference assessment of the medical volume segmentations (binary images), performed by inexperienced users, by automatically comparing them to the ground truth segmentations, performed by a doctor. The comparison of two volumes is done with the Jaccard's (similarity) coefficient, computed on a discrete voxel grid. To this end, we propose a web application, which supports the process of segmented volume pairwise correlation of MR images. The backend of the tool converts and persists segmented images as objects in a document-based database, and applies aggregation pipeline in order to perform pairwise correlations. The frontend of the tool is doted with Direct Volume Rendering to visualize the outliers, introduced by the segmentation algorithm, providing thus, visual means to assessing the dfferences of the segmentations. Interactive colored table (heatmap) visualizations provide overviews of the variability of the correlated segmentations. These visual analytics techniques (frontend) complement the proposed data management scheme with automated volume similarity computation (backend), andtogether constitute the proposed pipeline of difference assessment of the medical segmentations. Finally, we present an early evaluation of the proposed pipeline on a real clinical data set, and discuss the future development of the application.