Abstract
Introduction
We present a framework that enables automatic cortical parcellation based on multi-modal anatomical data. Utilizing cortical surface reconstruction, in one step we derive polygonal representations of the cortex and naturally curved profiles across the cortical ribbon, without additional computations. Our 3D, rather than 2D[1], approach allows the extraction of laminar fingerprints more closely in accordance with local neural organization. Enriching the profiles with contrast information from spatially aligned high-resolution data allows the analysis of local differences in cortical lamination patterns. All computational steps are fully automatic and well suited for pipelined processing of larger populations.
Methods
T1-weighted images were acquired using a standard MP-RAGE sequence (TI=650 ms; TR=1300 ms; TE=3.93 ms; alpha=10°; 1.0mm isotropic), and a Siemens TIM-Trio 3T scanner (Erlangen, Germany). High-resolution data was acquired using a 3D-Flash sequence (TE=10.7ms; TR=23ms; alpha=10°; 0.5mm isotropic) and a 2D-TSE sequence (TE=17ms; TR=16s; alpha=120°; 0.5mm isotropic) and a Siemens MAGNETOM 7T.
Non-uniformity correction, denoising and registration into MNI space were performed, prior to a probabilistic tissue classification [3]. Probabilistic edge maps were derived representing the GM/WM and GM/CSF borders. A topology-constrained level set method was used to represent implicitly an evolving surface, and the level set was initialized with spherical topology close to the target boundary [3,4]. The gradient vector flow approach [5], periodic re-initialization based on tri-cubic interpolation [6,7] and a narrow band implementation [8] plus multi-threading were used to control and improve stability and accuracy of the convergence behavior and to reduce computation time. Inner and outer cortical surfaces were detected separately for each hemisphere and boundary. The polygonal representations were isocontoured using a topology-preserving Marching Cubes algorithm [9]. The cross-cortical profiles were extracted starting from each vertex of the GM/WM boundary mesh and integrating through the final level set function using a 4th order Runge-Kutta method[10], until the GM/CSF boundary was reached. The high-resolution data was registered with the T1-weighted images. Associated scalar values are assigned for all points along a profile.
Results
Polygonal cortical surface representations take only a few hours for each hemisphere (Figure 1). The cortical profiles are curved and perpendicular to the iso-surfaces represented by the level sets. The number of profiles, and thus the density of coverage, can easily be increased by further subdivision of the mesh (Figure 2). Figure 3 shows mapping of laminar contrast variation within an easily recognizable cortical region (V1) to the profiles.
Conclusions
Cortical surface representations can be efficiently calculated using level set methodology. This also allows the creation of realistic cortical profiles, curved to mimic the columnar organisation of the cortex, and therefore more natural than simple straight lines. Enriching these profiles with high field MRI, high-resolution data may facilitate anatomically-based cortical parcellation.
