Interpreting 3D Computerized Tomography (CT) scans poses challenging problems for their immense dimensionality (in the order of one million voxels) and the (typically) small number of samples (in the order of ten to hundreds of samples). In the current case, the brains scans are collected from two groups of guppies under different selection regimes. The work focuses on the problem of identifying Regions Of Interest (ROIs) of the scans which change significantly under the different treatments. Two different techniques are employed to address this task: a parametric and a non-parametric one.
In the parametric approach, expertise knowledge is used to construct a brain-atlas, pointing to the different regions as studied in brain morphology for guppies. This reference image is constructed using a blend of segmentation, smoothing and manual annotation. Then this reference image is aligned ('registered') to the different samples, so that the ROIs are mapped to each individual. This gives us the precise size in volume, shape and morphological information of these ROIs of each sample. In turn, standard univariate statistics are then employed to detect differences.
The non-parametric approach is entirely based on processing of the intensities of the scans using a multivariate pattern analysis approach. For this, a combination of Support Vector Machines (SVMs) and permutation testing is employed. This results in a so-called P-map, organizing P-values per voxel in a brain-like structure. While the multitude of the P-values corrupts direct statistical interpretation, the P-map is used to indicate information useful for interpretation.
The proposed approaches are illustrated on the guppy brains scans. With the parametric method, several significantly different regions in the male and female guppy brains are indicated. The non-parametric method shows us the potential significant/interesting regions through the P-map.