Memory-Efficient Analysis of Dense Functional Connectomes.

The functioning of the human brain relies on the interplay and integration of numerous individual units within a complex network. To identify network configurations characteristic of specific cognitive tasks or mental illnesses, functional connectomes can be constructed based on the assessment of synchronous fMRI activity at separate brain sites, and then analyzed using graph-theoretical concepts. In most previous studies, relatively coarse parcellations of the brain were used to define regions as graphical nodes. Such parcellated connectomes are highly dependent on parcellation quality because regional and functional boundaries need to be relatively consistent for the results to be interpretable. In contrast, dense connectomes are not subject to this limitation, since the parcellation inherent to the data is used to define graphical nodes, also allowing for a more detailed spatial mapping of connectivity patterns. However, dense connectomes are associated with considerable computational demands in terms of both time and memory requirements. The memory required to explicitly store dense connectomes in main memory can render their analysis infeasible, especially when considering high-resolution data or analyses across multiple subjects or conditions. Here, we present an object-based matrix representation that achieves a very low memory footprint by computing matrix elements on demand instead of explicitly storing them. In doing so, memory required for a dense connectome is reduced to the amount needed to store the underlying time series data. Based on theoretical considerations and benchmarks, different matrix object implementations and additional programs (based on available Matlab functions and Matlab-based third-party software) are compared with regard to their computational efficiency. The matrix implementation based on on-demand computations has very low memory requirements, thus enabling analyses that would be otherwise infeasible to conduct due to insufficient memory. An open source software package containing the created programs is available for download.

Fast construction of voxel-level functional connectivity graphs.

BACKGROUND: Graph-based analysis of fMRI data has recently emerged as a promising approach to study brain networks. Based on the assessment of synchronous fMRI activity at separate brain sites, functional connectivity graphs are constructed and analyzed using graph-theoretical concepts. Most previous studies investigated region-level graphs, which are computationally inexpensive, but bring along the problem of choosing sensible regions and involve blurring of more detailed information. In contrast, voxel-level graphs provide the finest granularity attainable from the data, enabling analyses at superior spatial resolution. They are, however, associated with considerable computational demands, which can render high-resolution analyses infeasible. In response, many existing studies investigating functional connectivity at the voxel-level reduced the computational burden by sacrificing spatial resolution. METHODS: Here, a novel, time-efficient method for graph construction is presented that retains the original spatial resolution. Performance gains are instead achieved through data reduction in the temporal domain based on dichotomization of voxel time series combined with tetrachoric correlation estimation and efficient implementation. RESULTS: By comparison with graph construction based on Pearson's r, the technique used by the majority of previous studies, we find that the novel approach produces highly similar results an order of magnitude faster. CONCLUSIONS: Its demonstrated performance makes the proposed approach a sensible and efficient alternative to customary practice. An open source software package containing the created programs is freely available for download.

Local topographic influences on vision restoration hot spots after brain damage.

PURPOSE: Vision restoration training (VRT) in hemianopia patients leads to visual field enlargements, but the mechanisms of this vision restoration are not known. To investigate the role of residual vision in recovery, we studied topographic features of visual field charts and determined residual functions in local regions and their immediate surround. METHODS: We analyzed High Resolution Perimetry visual field charts of hemianopic stroke patients (n = 23) before and after 6 months of VRT and identified all local visual field regions with ("hot spots", n = 688) or without restoration ("cold spots", n = 3426). Topographic features of these spots at baseline where then related to (i) their respective local residual function, (ii) residual activity in their spatial neighbourhood, and (iii) their distance to the scotoma border estimated in cortical coordinates following magnification factor transformation. RESULTS: Visual field areas had a greater probability of becoming vision restoration hot spots if they had more residual activity in both local areas and in a spatially limited surround of 5° of visual angle. Hot spots were typically also located closer than 4 mm from the scotoma border in cortical coordinates. Thus, restoration depended on residual activity in both the local region and its immediate surround. CONCLUSIONS: Our findings confirm the special role of residual structures in visual field restoration which is likely mediated by partially surviving neuronal elements. Because the immediate but not distant surround influenced outcome of individual spots, we propose that lateral interactions, known to play a role in perceptual learning and receptive field plasticity, also play a major role in vision restoration.

A treatment outcome prediction model of visual field recovery using self-organizing maps.

Brain injuries caused by stroke, trauma, or tumor often affect the visual system that leads to perceptual deficits. After intense visual stimulation of the damaged visual field or its border region, recovery may be achieved in some sectors of the visual field, but the extent of restoration is highly variable between patients and is not homogeneously distributed in the visual field. We now assess the visual field loss and its dynamics by perimetry, a standard diagnostic procedure in medicine, to measure the detectability of visual stimuli in the visual field. Subsequently, a treatment outcome prediction model (TOPM) has been developed, using features that were extracted from the baseline perimetric charts. The features in the TOPM were either empirically associated with treatment outcomes or were based on findings in the vision-restoration literature. Among other classifiers, the self-organizing map (SOM) was selected because it implicitly supports data exploration. Using a data pool of 52 patients with visual field defects, the TOPM was constructed to predict areas of improvement in the visual field topography. To evaluate the predictive validity of the TOPM, we propose a method to calculate the receiver operating characteristic graph, whereby the SOM is used in combination with a nearest neighbor classifier. We discuss issues relevant for medical TOPMs, such as appropriateness to the patient sample, clinical relevance, and incorporation of a priori knowledge.

Predicting the recognition of natural scenes from single trial MEG recordings of brain activity.

In our daily life we look at many scenes. Some are rapidly forgotten, but others we recognize later. We accurately predicted recognition success with natural scene photographs using single trial magnetoencephalography (MEG) measures of brain activation. Specifically, we demonstrate that MEG responses in the initial 600 ms following the onset of scene photographs allow for prediction accuracy rates up to 84.1% using linear Support-Vector-Machine classification (lSVM). A permutation test confirmed that all lSVM based prediction rates were significantly better than "guessing". More generally, we present four approaches to analyzing brain function using lSVMs. (1) We show that lSVMs can be used to extract spatio-temporal patterns of brain activation from MEG-data. (2) We show lSVM classification can demonstrate significant correlations between comparatively early and late processes predictive of scene recognition, indicating dependencies between these processes over time. (3) We use lSVM classification to compare the information content of oscillatory and event-related MEG-activations and show they contain a similar amount of and largely overlapping information. (4) A more detailed analysis of single-trial predictiveness of different frequency bands revealed that theta band activity around 5 Hz allowed for highest prediction rates, and these rates are indistinguishable from those obtained with a full dataset. In sum our results clearly demonstrate that lSVMs can reliably predict natural scene recognition from single trial MEG-activation measures and can be a useful tool for analyzing predictive brain function.