Group neuroimaging studies of the cerebral cortex benefit from accurate, surface-based, cross-subject alignment for investigating brain architecture, function and connectivity. There is an increasing amount of high quality data available. However, establishing how different modalities correlate across groups remains an open research question. One reason for this is that the current methods for registration, based on cortical folding, provide sub-optimal alignment of some functional subregions of the brain. A more flexible framework is needed that will allow robust alignment of multiple modalities. We adapt the Fast Primal-Dual (Fast-PD) approach for discrete Markov Random Field (MRF) optimisation to spherical registration by reframing the deformation labels as a discrete set of rotations and propose a novel regularisation term, derived from the geodesic distance between rotation matrices. This formulation allows significant flexibility in the choice of similarity metric. To this end we propose a new multivariate cost function based on the discretisation of a graph-based mutual information measure. Results are presented for alignment driven by scalar metrics of curvature and myelination, and multivariate features derived from functional task performance. These experiments demonstrate the potential of this approach for improving the integration of complementary brain data sets in the future.

No abstract

No abstract

Summary1. Static sensor networks to observe animals are widely used in ecological, management and conservation research, but quantitative methods for designing these networks are underdeveloped. 2. In the context of aquatic systems, we present a method for quasi-optimal network design, which accounts for blocking of detections by obstacles, horizontal and vertical movement behaviour of the target animals, and type of research question (is the network intended for estimation of detailed movement or home range?). Optimal design is defined as the sensor configuration that maximizes the expected number of unique animal detections. As finding the global optimum is generally time consuming, we use a greedy algorithm instead, which places sensors optimally relative to already placed sensors. The design method requires access to topographic data of the study site and knowledge of the sensor detection range. 3. We illustrate the method with real topographic data from a rugose coral reef where network performance is highly influenced by detection shadowing. Network performance is visualized by a coverage map indicating the probability of detection at any location in the study area. The reported unique recovery rate summarizes the expected ability of the network to collect data given the design constraints. Because sensors are placed sequentially, the information gain per sensor can be evaluated and used as a proxy for sensor value. 4. The presented method formalizes important considerations, when designing sensor networks, that were previously often based on heuristics and intuition. The method provides a guide to maximizing the information potential of future monitoring studies as well as a means to improve existing networks. The method is available as an R package and can be tested via an online web tool.

No abstract