Accurate Non-rigid Object Segmentation in Medical Images by Fusing Statistical Features with Structural Features

Abstract

This paper presents an accurate non-rigid object segmentation method that fuses both statistical features and structural features. In particular, the approach is detailed and applied on liver segmentation. It consists of three main components. First, an image texture analysis is done to derive pixel level features. It efficiently fuses the statistical features with structural features to achieve better segmentation. Then, a trained classifier based on support vector machine (SVM) is applied to classify the image into liver pixels or non-liver pixels. Finally, composite morphological operations are used to remove small wrongly classified areas and delineate the liver region. The approach is unique in two aspects: it states and provides experimental data to demonstrate that the fusion of the two classes of features does improve segmentation rate, comparing to the cases where only statistical features or structural features are used; it shows that an accurate segmentation can be achieved by combing regional morphological operations with pixel-wised SVM classifier. The algorithm can be applied to general non-rigid object segmentation which is a crucial part of an automatic surgical training and planning system.