“Sulcus Sink”: A Compact Binary and Semi-Automated Inverse Dijkstra-based System for Describing Sulcal Trajectories

Abstract

AbstractWe present a description of a system that uses a compact binary representation to describe and trace sulci on a reconstructed human cortical surface, based on a set of human-generated targets. The inputs to the system were manually created on a training set of 20 normal subjects (11 females, 9 males) with ages 22 – 40 years. T1 weighted MPRAGE images were collected on a Siemens 3T Trio scanner, with TR/TE = 2530/3.3, matrix = 256×256, FOV = 256mm, slice thickness 1.33mm. The resultant images were reconstructed with Freesurfer, and 10 sulci on each hemisphere were traced by an expert human operator and independently assessed for accuracy. Presented with these input trajectories in its training phase, the system attempted to determine a compact binary feature vector of each sulcus on each subject using as descriptor a binary parametrized function of several surface-geometry variables (such as mean curvature, sulcal depth, edge length, etc.). This function was optimized in a supervised learning fashion using a Dijkstra-based graph theory formulation, in which the binary weights were used to define graph edge costs. In the setup phase, the system was presented with sulcal trajectories already defined on surfaces, and then adjusted its parametrized weights in a binary fashion to minimize differences between the training input path and its Dijkstra-generated output path. Once the setup phase was complete and sulci had been described in a per-sulcus, per-subject manner, we generalized the per-sulcus description across all the subjects to construct a template binary word for each specific sulcus. The performance of the system for each subject and each sulcus, and for each template sulcus group was measured against the original human reference in both a quantitative and qualitative manner. Individual subjects generally showed very good optimization to their manually traced training samples across all sulci, with 91% average overlap within 4mm of the human target. Generalized group results, as expected, showed less overlap with the original human targets, but still performed on average with 80% overlap. Quantitatively, the group results were nonetheless for the most part quite acceptable to an independent human evaluator. The parametrized binary weight description that drives the Dijkstra path optimization is presented as a mechanism to succinctly and compactly describe individual human sulci and groups of sulci.