Automatic Design of Realistic Multiple Sclerosis Lesion Phantoms

Abstract

The segmentation and quantification of multiple sclerosis (MS) lesions is an important issue in medical image analysis. To reach clinical acceptance, a careful evaluation of each algorithm is required. Today, the standard approach is a comparison with results of patient data sets generated by domain experts. Unfortunately, the underlying ground truth is unknown in these data sets, and the results of expert analyses suffer from intra- and inter-rater variabilities. In this work, we present an automatic approach to develop digital MS lesion phantoms. The algorithm combines a statistical map of lesion positions with a lesion model extracted from actual patient data. A standard brain phantom is used as reference data set. Instead of creating one best phantom, our approach allows to parametrically generate a large range of different phantoms. This way, we can capture the variability of MS lesions encountered in practice without the need of manual interactions during the phantom design process. To evaluate our approach, a visual assessment is performed by a clinical expert. Furthermore, a published MS lesion segmentation algorithm is used to segment the phantom data. The results indicate the applicability of our approach.