A Comparative Study of Segmentation Techniques used for MR Brain Images

In this paper we present a comparative study of MR brain image segmentation techniques. The aim of this study is to assess the robustness and accuracy of three most commonly used unsupervised segmentation methods k-means (KM), FCM and EM. KM is a well known hard segmentation method for quicker processing whereas FCM and EM are popularly used soft segmentation methods particularly for brain tissue models. Most of the neuroimage analyzing software employ one of these techniques as a preprocessing tool to enrich their subsequent processes. Based on the parameters similarity index and processing time, the performance of the methods on brain tissue segmentation are measured and compared. The similarity index calculation was done by using the 20 normal volumes and their gold standards available from IBSR. The rest of parameters are tested on the normal and abnormal brain volumes collected from scan centers and brain image pools. For normal and noiseless volumes all the three methods produced comparable results. FCM has an excellent performance over the anisotropic nature of volumes that are affected by partial volume effects (PVE) as well as diseased sets and EM was suitable for noisy sets, particularly volumes affected by intensity non-uniformity (INU) artifacts. KM, FCM and EM are rated in the given order in terms of processing time. This study is useful to choose a suitable method for computerizing brain diagnostic system.