An alternative approach of evaluating Dice Similarity Index for Pre-segmented blood vessels

Abstract

Modeling and simulating blood vessels in order to predict their behavior prior to surgical interventions is an emerging research field. These three dimensional models at their region of interest are generated from corresponding computed tomography (CT) or magnetic resonance imaging (MRI) scans and is often referred as the segmentation task. A critical point in similar segmentation and related post processing tasks is to measure the accuracy of the applied method with `Dice Similarity Index'. The standard procedure of evaluating this index is to compare the segmented label masks in their two dimensional phase. However, more cases are coming forward where only the finalized three dimensional structures are available. This work focuses on developing an alternative framework where the dice index of two three dimensional models are computed. The objective has been achieved by slicing the solid structures at multiple planes and for five sets with varying slice numbers. The observed Dice indices vary from 0.67 to 0.9 with a median slightly above 0.8, which is seemingly a fairly reasonable value. The results also indicate that even though dice index for a singular plane of the whole structure is not enough, increasing the slice numbers to a fairly high value does not affect the result as well.