Retrospective techniques for segmentation of structural and functional MR brain images

• Main Author: Suresh Anand Sadananthan
• Other Authors: Vitali Zagorodnov
• Format: Theses and Dissertations
• Language: English
• Published: 2010
• Subjects: DRNTU::Engineering::Computer science and engineering::Computer applications::Life and medical sciences; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/42523
• Institution: Nanyang Technological University

In this thesis, we focus on several segmentation problems arising in the area of structural and functional neuroimaging. Our solutions to these problems are based on a retrospective framework, wherein a complex segmentation procedure is divided into two simpler steps, initial segmentation and incorporation of prior information. While combining these steps might potentially lead to an optimal solution, we show that the simpler two-step approach can either be made equivalent to a combined procedure or achieve superior performance, due to simpler optimization. We first consider the problem of detecting activated regions in functional Magnetic Resonance Imaging (fMRI). While activated regions are typically large, smaller spurious activations caused by noise are likely to appear in the segmentation. These can be removed by cluster size thresholding. Though cluster size thresholding can be regarded as a method to reduce false positives, it affects the smaller true activations. We show that in the context of Markov Random Field (MRF) based segmentation, simple removal of small regions after the segmentation is not optimal. We propose a retrospective correction approach that allows the regions to grow before they are eliminated based on the cluster size. This approach finds the best modification (removal or growth) and achieves superior performance to that of the standard MRF-based segmentation. The second and the main problem considered in this thesis is skull stripping, i.e., the problem of separating the brain tissues (white matter, grey matter, cerebral spinal fluid) from the non-brain tissues (skull, scalp, eye sockets, neck tissues, etc.). Skull stripping performance suffers from the problem of narrow connections between brain and non-brain structures which usually results in preservation of significant amount of non-brain tissues. Many popular methods rely on iterative surface deformation to fit the brain boundary and tend to leave residual dura. We first approach the problem in a general framework of narrow connection removal. Here, we show that of all existing approaches, isoperimetric algorithm performs the best but can be sensitive to initialization. Instead, we propose a novel approach based on graph cuts that has beneficial features - global optimality, speed, and possibility of fully automated implementation for many applications. We also show that incorporating intensity information into graph weight assignment can further improve performance. In the case of skull stripping, compared to the Hybrid Watershed Algorithm (HWA), our approach achieves an additional 10-30% of dura removal without incurring further brain tissue erosion. When used in conjunction with HWA, our approach substantially decreases and often fully avoids cortical surface overestimation in subsequent segmentation. Lastly, we address the problem of skull stripping in multimodal images (combinations of T1-weighted and T2-weighted images). T1W and T2W images have different contrast properties and we show that the combined use of this information can help to further differentiate dura from brain structures, compared to using only T1W modality. We employ the graph cuts based skull stripping approach with suitably modified thresholding and post-processing procedures to suit multimodal images. The results obtained show significant decrease in the amount of dura in the resultant brain mask when using multimodal images.