A Multiatlas Approach for Segmenting Subcortical Brain Structures using Local Patch Distance

Abstract

In the diagnosis and treatment of various diseases, often segmenting the brain structures from MRI data is the key step.  Since there are larger variations in the anatomical structures of the brain, segmentation becomes a crucial process. Using only the intensity information is not enough to segment structures since two or more structures may share the same tissues.  Recently, the use of multiple pre-labeled images called atlases or templates are used in the process of segmentation of image data. Both single atlas and multiple atlases can be used.  However, using multiple atlases in the segmentation process proves a dominant method in segmenting brain structures with challenging and overlapping structures. In this paper, we propose two multi atlas segmentation methods: Local Patch Distance Segmentation (LPDS) and Weighted Local Patch Distance Segmentation (WLPDS). These methods use local patch distance in the label fusion step.  LPDS uses local patch distance to find the best patch match for label propagation.  WLPDS uses local patch distance to calculate local weights. The brain MRI images from the MICCAI 2012 segmentation challenge are chosen for experimental purposes. These datasets are publicly available and can be downloaded from MIDAS. The proposed techniques are compared with existing fusion methods such as majority voting and weighted majority voting using the similarity measures such as Dice overlap (DC), Jaccard coefficient (JC) and Kappa statistics. For 20 test data sets, LPDS gives DICE=0.95±0.05, JACCARD=0.91±0.04 and KAPPA=0.94±0.07. WLPDS gives DICE=0.98±0.02, JACCARD=0.92±0.03 and KAPPA=0.95±0.04.