A heuristic information cluster search approach for precise functional brain mapping

Abstract

Detection of the relevant brain regions for characterizing the distinction between cognitive conditions is one of the most sought after objectives in neuroimaging research. A popular approach for achieving this goal is the multivariate pattern analysis which is currently conducted through a number of approaches such as the popular searchlight procedure. This is due to several advantages such as being automatic and flexible with regards to size of the search region. However, these approaches suffer from a number of limitations which can lead to misidentification of truly informative regions which in turn results in imprecise information maps. These limitations mainly stem from several factors such as the fact that the information value of the search spheres are assigned to the voxel at the center of them (in case of searchlight), the requirement for manual tuning of parameters such as searchlight radius and shape, and high complexity and low interpretability in commonly used machine learning-based approaches. Other drawbacks include overlooking the structure and interactions within the regions, and the disadvantages of using certain regularization techniques in analysis of datasets with characteristics of common functional magnetic resonance imaging data. In this article, we propose a fully data-driven maximum relevance minimum redundancy search algorithm for detecting precise information value of the clusters within brain regions while alleviating the above-mentioned limitations. Moreover, in order to make the proposed method faster, we propose an efficient algorithmic implementation. We evaluate and compare the proposed algorithm with the searchlight procedure as well as least absolute shrinkage and selection operator regularization-based mapping approach using both real and synthetic datasets. The analysis results of the proposed approach demonstrate higher information detection precision and map specificity compared to the benchmark approaches.

Keywords: algorithm design; data mining; functional magnetic resonance imaging; neuroimaging.

Figure 1

A schematic plot of an example region traversed by the proposed algorithm to detect information clusters originating from voxel V ₁ and V ₂. The algorithm first creates the information map for the cluster starting from V ₁ from top to bottom, and then moves to V ₂ and follows the same process. The red voxels are the admitted voxels and the yellow voxels represent the neighbors of the cluster at each step. The online spectral relevance analysis is performed during each search in the neighborhood layer, and the group biased mutual information redundancy analysis is performed before each next search in the neighborhood layer

Figure 2

An interaction matrix after a complete search over 100 voxels. Nonzero elements are depicted in blue color. The right figure illustrates the size of clusters originating from each voxel

Figure 3

(a) An illustration of the steps and output of searchlight procedure compared with the ICS algorithm. The gray voxel is the search sphere center voxel in the searchlight method, and the starting voxel in ICS algorithm. The radius for searchlight in this schematic illustration is one voxel, and the information of the searchlight sphere, denoted by a specific color for each sphere is assigned to the voxel at the center of the sphere, that is, each voxel in the output map has the same color (information) as its search sphere. On the other hand, the output of the ICS method is a set of clusters expanded from the starting voxel through a data‐driven heuristic. The information of each cluster is demonstrated by a specific color. (b) Left: An example illustration of overlapping clusters created by ICS. Right: The same clusters depicted individually. The voxel indicated by black dots are the starting voxel v _s which are expanded based on the discriminant analysis heuristic, resulting in a specific discriminant score for each cluster

Figure 4

Top: A comparison of the above chance accuracy clusters derived using the searchlight process (top row), least absolute shrinkage and selection operator (LASSO) (middle row), and the ICS algorithm (bottom row) on the Autism Brain Imaging Data Exchange (ABIDE) data set. Major differences between the three maps are indicated by the red circles. In case of overlapping clusters generated by ICS, the clusters with the highest predictability were selected for this visualization. Bottom‐left: Classification performance on full‐brain search space for ABIDE dataset based on the above chance clusters as the features. Bottom‐right: Classification performance on full‐brain search space for ABIDE dataset based on the top 50 clusters as the features. The train‐test population for both settings was 546–137, respectively

Figure 5

Comparison of classification area under the curve (AUC) with SVM between ICS, least absolute shrinkage and selection operator (LASSO), and searchlight with voxel radius set to 3 for left Crus II of the cerebellum (region 93 per automated anatomical labeling [AAL]) with 573 voxels

Figure 6

Test area under the curve (AUC) for classification with SVM of the ICS algorithm, least absolute shrinkage and selection operator (LASSO)‐based features, and the searchlight method with different search radii for right and left hippocampus and amygdala from the Autism Brain Imaging Data Exchange (ABIDE) dataset

Figure 7

Area under the receiver operating characteristic (ROC) curve for synthetic datasets of size 100 (a); 500 (b); 10,000 (c); and 30,000 (c) voxels based on the top 50 features according to the three approaches

Figure 8

Detection results obtained by all methods using the dataset with 25% simulated atrophy

Figure 9

Information maps created through searchlight (top row), least absolute shrinkage and selection operator (LASSO) (middle row) and ICS (bottom row) approaches. (a) The voxel‐level map of left Crus II of the cerebellum (Region 93 per automated anatomical labeling (AAL)) and (b) belongs to Cerebellum 4 (Region 97 per AAL). For both regions, two dimensional subsegments (100 by 100 voxels) of the information maps are depicted in this figure to facilitate readable illustrations. Moreover, five sample information clusters derived from these maps through the ICS approach are depicted on the sides of the maps. In case of overlaps, the information clusters with higher quality (darker red) overshadow the ones with lower information

Figure 10

Computation time of ICS on five data set sizes