Nerve Fiber Segmentation from Microscopic Cross Sections Based on Spatially Constrained Registration Strategy

Segmentation of myelinated nerve fiber cross sections (MNFC) from a sequence of microscopic cross-sectional images, which can be used to construct the three-dimensional (3D) structure of nerve fiber, is important for evaluating neurological disorders and nerve regeneration. However, most existing segmentation methods do not take the spatial relationships between inter-frame MNFCs into account,so they usually suffer from several difficulties, such as large intensity variations among frames and blurred MNFC boundaries, in obtaining desirable segmentation results. In this paper, we propose a new segmentation method based on spatially constrained registration (SCR) to automatically segment and reconstruct the 3D structure of nerve fiber from microscopic images. At first, we utilize a multi-scale gradient watershed-based segmentation algorithm to segment the MNFCs from each image frame. Considering the continuity of the 3D fiber structure, the MNFC contexts between adjacent frames are supposed to be dependent. The SCR strategy is then employed to assure the connectivity between the adjacent MNFCs. Based on the connectivity we subsequently recover the missing MNFCs with a compensation mechanism and finish the segmentation process. Two sets of rat tibia nerve images (45 nerve fiber cross-section patches) were used to validate the accuracy of the proposed method. Experiments showed that our method can overcome the aforementioned difficulties and achieve satisfactory segmentation results. The experiments also demonstrated that the proposed SCR method established more reliable correspondences of MNFCs over the conventional iterative closest point (ICP) registration algorithm.