Automatic knee segmentation from multi-contrast MR images

This thesis is devoted to developing methods for automatic knee segmentation from multi-contrast MR images which provide different contrasts between joint structures and help the separation of different structures. By exploiting the combined information of FS SPGR and IDEAL GRE water & fat image...

Full description

Saved in:
Bibliographic Details
Main Author: Zhang, Kunlei
Other Authors: Pina Marziliano
Format: Theses and Dissertations
Language:English
Published: 2013
Subjects:
Online Access:http://hdl.handle.net/10356/52482
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:This thesis is devoted to developing methods for automatic knee segmentation from multi-contrast MR images which provide different contrasts between joint structures and help the separation of different structures. By exploiting the combined information of FS SPGR and IDEAL GRE water & fat images, a simple but reliable technique for automatic bone segmentation is proposed by using a threshold-based method followed by connected component labeling and distance transform. Next an automatic cartilage segmentation scheme is developed by using supervised classification with the incorporation of spatial dependencies. The choice of the classifier can be some powerful classification models such as SVM and ELM. The advantages of the developed scheme are achieved via effective incorporation of both a useful feature set and the ELM (or SVM) classification with spatial dependencies between neighboring voxels via a DRF framework. Also, the proposed cartilage segmentation scheme can be applied to the segmentation of other joint structures. For example, we perform the proposed method for automatic meniscus segmentation with meniscal searching region in a volume of interest. The automatic segmentations of knee bones, cartilages and menisci are evaluated on a comprehensive multi-contrast MRI database. The developed segmentation methods achieve good performance compared with gold standard segmentations. They also outperform the ones based on independent classifiers in terms of segmentation accuracy, and compare favorably with other state-of-the-art automatic knee segmentation methods.