Introducing spatial neighbourhood in Evidential C-Means for segmentation of multi-source images: Application to prostate multi-parametric MRI

Abstract

In this paper we introduce an evidential multi-source segmentation scheme for the extraction of prostate zonal anatomy using multi-parametric MRI. The Evidential C-Means (ECM) classifier was adapted to a segmentation scheme by introducing spatial neighbourhood-based relaxation step in its optimisation process. In order to do so, basic belief assignments on voxels membership were relaxed using distance-weighted combination of belief from spatial neighbours. For the application on prostate tissues, geometric a priori was modelled and used as an additional data source. Our method was first experimented on simulated images to prove the improvement brought to the ECM. A validation study of the segmentation method was then conducted on 31 patients MRI data. In order to take into account inter-observer variability, each MRI was manually segmented by three independent expert radiologists, and an estimated truth was computed using STAPLE algorithm. This validation proved that segmentation obtained with our method is accurate and comparable to expert segmentation. We also show that our segmentation scheme enables to detect and highlight outliers, which could be interpreted by physicians as irregular tissues. The use of belief functions also provides additional information on borders between structures. We do believe these are sources of evidence that could help physicians/algorithms in characterising tissues and structures. The method that is introduced in this paper is a step forward to the use of belief functions theory in the context of multi-source image segmentation. Nevertheless, a full comparison to both baseline (e.g. Gaussian Mixture Models) and recent (e.g. Graph Cut) segmentation methods is needed to assess its performance.