Purpose: White matter tractography has become an important tool for neuro-surgical planning and navigation. Generating accurate tracts in a brain deformed by a tumor may prove to be a challenging task. Beside a robust tractography algorithm, significant neuro-anatomical expertise is required to accurately delineate the fiber seeding ROIs. For example, an accurate tractography of the optic radiation requires the delineation of the Calcarine sulcus and the lateral geniculate nucleus (LGN). In addition to the neuro-anatomical knowledge, significant amount of time is required for the manual delineations of complex, non-planar 3-D structures like the Calcarine sulcus or the precentral gyrus in the motor cortex. Considering the limited amount of time often available for planning an urgent brain surgery, the automatic tools are badly needed for the delineation of anatomical ROIs.
Methods: The accurate segmentation of tractography ROIs usually requires information from both the anatomical scan, typically T1w, and the principal direction of diffusion (PDD), that is the RGB color-coded map. For this intrinsically multi-modal segmentation task, we propose two alternative FCNN architectures: the W-net and the Y-net. Each architecture implements a different approach to the combination of anatomical (T1w) and orientation (PDD) information.
Results: Both architectures are successfully validated on 3-D segmentations of the Lateral Geniculate Nucleus (LGN) and the Calcarine sulcus, namely the seeding ROIs (SROI) for the optic radiation tract. A dataset of 90 pre-surgical cases for which manual 3-D segmentation of the ROI was provided. The average Dice overlap coefficient, computed between manual and automatic segmentation, demonstrated the superiority of the proposed methods over state-of-the-art V-net architecture: W-net (Dice=0.79), Y-net(Dice=0.795) and V-net (Dice=0.73).
Conclusions: The presented method demonstrated the feasibility of automatic segmentation of complex neuro-anatomical ROIs jointly using multiple MRI sequences. In future work, the method will be further extended to motor and languages ROIs.
