Development of an ensemble of extreme learning machines for 3D medical object segmentation and classification

Abstract

This report presents a semi-automatic approach to segmentation of liver parenchyma from 3D computed tomography (CT) images. More specifically, liver segmentation is formalized as a pattern recognition problem, where a given voxel is to be assigned a correct label - either a liver or a non-liver class. Based on the extracted texture features, an Extreme Learning Machine (ELM) classifier is employed to perform the voxel classification. Since preliminary voxel segmentation tends to be less accurate at the boundary, and there are other non-liver tissue voxels with similar texture characteristics as liver parenchyma, morphological smoothing and 3D level set refinement are applied to enhance the segmentation. Our approach is validated on a set of CT data. The experimental result shows that the proposed ELM method is capable of delivering reasonably good performance for liver parenchyma segmentation. The proposed ELM method is demonstrated to have comparable classification accuracy compared with support vector machine (SVM), but with a much faster training speed. Our work has been submitted as a conference paper to the 34th Annual International Conference of IEEE Engineering in Medicine and Biology Society (EMBC'12) in San Diego, California, USA.