| HBM 2005 Abstracts: Previous | Next | Click here for a PDF of this poster. |
1Department of Electrical and Computer Engineering, University of Minnesota, USA, 2Department of Radiology, University of Minnesota, USA, 3Department of Neurology, University of Minnesota, USA |
Objective:
Automatic segmentation of MRI brain volumes into anatomical compartments, e.g., left cerebral hemisphere (LCH), right cerebral hemisphere (RCH), and cerebellum+brain stem (CBB), is important for many biomedical and neuroscience applications, and various techniques [1,2,3] have been proposed to accomplish this; however, most methods require human interaction to obtain satisfactory results [2]. Fully automatic methods are desirable but may not perform well because of the limited spatial and tissue resolution of MRI images and the physical connectivity between brain structures. For example, region-growing-based techniques [1] are sensitive to initialization and frequently fail by leaking through weak object boundaries or callosal WM. Although registration to a template [2,3] may help to solve large leakage problems, it relies heavily on the quality of the registration, which is frequently inadequate. We propose a simple method for creating anatomical compartments from T1-weighted MRI brain volumes based on graph cuts [4].
Methods: T1-weighted MRI brain volume is segmented into white matter (WM), gray matter (GM) and CSF [5], and the segmented image is represented by a graph where each voxel in the image represents a node. Each node links to its neighboring nodes, and the weights of the links control the bonding between neighboring nodes. Based on anatomical considerations we let CSF nodes bond weakly with their neighboring nodes, and let WM and GM nodes bond strongly. Once some nodes are specified as "source" or "sink", a standard graph-cuts method [4] will find an optimal cut (minimum aggregated weights of severed links) separating source and sink. Source and sink nodes are obtained automatically using a template volume and a registration procedure [5] to initialize the nodes belonging to LCH, RCH or CBB (Fig. 1a). The registration need only be moderately accurate. We begin by using one set of component nodes, e.g., LCH nodes, as source and RCH and CBB nodes as sink; we then apply the graph-cuts algorithm to separate LCH. (Fig. 1b). After removing LCH (Fig. 1c) we reconstruct a graph from an image containing only RCH and CBB and re-apply the graph-cuts algorithm to separate RCH and CBB.
Results & Discussion:
Preliminary compartmentalization results using 1.5T T1-weighted whole-brain MRI volumes are very encouraging (Fig. 1d). The graph-cuts method is non-parametric, runs fast—in less than one minute on a 2GHz Pentium PC workstation, and produces more accurate results than the prevailing approach based on region growing. The graph-based technique appears better than competing methods at overcoming the effects of physical connectivity between structures as it always locates the correct cut in WM. However, owing to the simple optimization function it can not completely solve problems associated with weak boundaries.
Conclusions:
The graph-cuts method automatically generates a globally optimal cerebral hemisphere segmentation, and preliminary results appear promising.
References & Acknowledgements: |
|
