Manual Segmentation of Fiber Tracts with Bundles of Interest
Paolo Avesani (CIMeC Center for Mind/Brain Sciences), Eleftherios Garyfallidis (University of Cambridge), Emanuele Olivetti (CIMeC Center for Mind/Brain Sciences), Stephan Gerhard (ETH Zurich)
In neurological studies and presurgical planning, the segmentation of the network of white matter fibers into known anatomically structures, called fiber bundles or fiber tracts, is a task of growing interest. The current procedures and tools to manually segment a fiber tract are based on the notion of region of interest (ROI). The segmentation of the fiber tract of interest is performed by defining two or more ROIs in order to localize where the related tracks start and end. The anatomical fiber tract is obtained by filtering the streamlines that cross the ROIs.
This approach has some important drawbacks. First, it tends to underestimate the fiber tract geometry since it does not retain the streamlines that are broken in one or more points in their path due to incorrect reconstruction. Second, the manual design of the ROIs is a challenging task which is based on the, possibly inaccurate, alignment of the tracks to a structural scan (e.g T1 or T2). Third, with ROIs the user has to face the complexity of navigating the full cluttered and densely packed tractography.
We propose an alternative approach based on the notion of bundle of interest (BOI). A BOI is defined as a set of tracks sharing similar shape and spatial characteristics. The proposed approach for manual segmentation is based on direct interaction of the user with the tracks, in contrast with the indirect method based on ROIs. The intuitive idea is to provide the user with a summary of the tractography. This summary is defined by clustering the streamlines into a set of representative bundles and then showing one representative track per bundle. The task of manual segmentation is conceived as an iterative process where the user alternates a phase of bundle-representative selection to a phase of re-clustering the selected bundles into smaller bundles. The selection of bundles aims to best approximate the target fiber tract, while the bundle re-clustering step allows the user to work incrementally at finer detail.