Quantification of guidance strategies in online interactive semantic segmentation of glioblastoma MRI

Abstract

Interactive segmentation promises to combine the speed of automatic approaches with the reliability of manual techniques. Its performance, however, depends largely on live iterative inputs by a human supervisor. For the task of glioblastoma segmentation in MRI data using a Random Forest pixel classifier we quantify the benefit in terms of speed and segmentation quality of user inputs in falsely classified regions as opposed to guided annotations in regions of high classifier uncertainty. The former results in a significantly higher area under the curve of the Dice score over time in all tumor categories. Exponential fits reveal a significantly higher final Dice score for larger tumor regions (gross tumor volume and edema) but not for smaller regions (necrotic core, non-enhancing abnormalities and contrast-enhancing tumor). Time constants of the exponential fits do not differ significantly.