General framework for unsupervised evaluation of quality of segmentation results

Evaluation of segmentation algorithms is clearly important, but despite many years of research, no consensus on approach has been reached. Supervised approaches (comparing outputs with ground truth) are labour intensive and of uncertain reliability, while unsupervised approaches (judging quality without ground truth knowledge) are usually demonstrated on synthetic data sets, rarely agree with each other, and usually put serious constraints on image properties. This work aims to deliver a general measure which can deal with synthetic, real- life and medical imagery and provide comprehensive information about the segmentation. In this paper, we present a new metric, compare its performance against existing unsupervised and supervised approaches and demonstrate its reliability for automated segmentation evaluation.