The human brain is a complex neural system composed of several dozen anatomical structures. To study the functional and structural properties of its deeper sub-cortical regions, three-dimensional image segmentation is a critical step in quantitative brain image analysis and clinical diagnosis. However, segmenting sub-cortical structures is difficult because they are relatively small and have significant shape variations. Moreover, some structure boundaries are subtle or even missing in images. Although manual annotation is a standard procedure for obtaining quality segmentation, it is time-consuming and can suffer from inter- and intra-observer inconsistencies. In recent years, researchers have been focusing on developing automatic atlas-based segmentation methods incorporating...[ Read more ]

The human brain is a complex neural system composed of several dozen anatomical structures. To study the functional and structural properties of its deeper sub-cortical regions, three-dimensional image segmentation is a critical step in quantitative brain image analysis and clinical diagnosis. However, segmenting sub-cortical structures is difficult because they are relatively small and have significant shape variations. Moreover, some structure boundaries are subtle or even missing in images. Although manual annotation is a standard procedure for obtaining quality segmentation, it is time-consuming and can suffer from inter- and intra-observer inconsistencies. In recent years, researchers have been focusing on developing automatic atlas-based segmentation methods incorporating expert prior knowledge about the correspondences between intensity profiles and tissue labels. We introduce some novel methods for brain MR image segmentation in this thesis, which can be categorized into two main parts.

In the first part, several methods relying on non-rigid registration are proposed for the label inference of sub-cortical structures in brain MR images. A united atlas-based segmentation framework is presented, including forward deformation and label refinement. One novel label inference method integrated with registration and patch priors is introduced to help correct the label errors around structural boundaries. Given the significant overlap of the intensity distribution among different tissues, the patch prior based on similarity measurement can be adversely impacted. To deal with this problem, a new label inference method encoded with local and global patch priors is proposed to obtain a more discriminative patch representation.

In the second part, we introduce some advanced label inference methods, which don’t need the non-rigid registration process with expensive computation. A novel network called multi-scale structured CNN is proposed, on top of which label consistency is enforced to refine the preliminary results obtained using deep learning. With multiple features available for image segmentation, Feature Sensitive Label Fusion is presented, which takes the sensitivity among distinct features into consideration. To obtain the comprehensive properties for 3D brain image, both convolutional LSTM and 3D convolution are employed in the Randomized Connection network. Comprehensive experiments have been carried out on publicly available datasets and results demonstrate that our methods can obtain better performance as compared with other state-of-the-art methods.