Self organizing maps for class discovery in the quantitative colocalization analysis feature space

Quantitative colocalization analysis in fluorescent microscopy imaging is a promising procedure used to perform functional protein analysis. Images acquired are degraded, and the features extracted are affected by this degradation. Moreover, the classification of the data becomes uncertain. In this paper, we address an application of SOM to a clustering problem formulated via feature extraction from multichannel fluorescence microscopy. First we describe the features that are extracted. Second, we use the PCA/KLT to un-correlate the data in the hyperplane; and Third, SOM network is trained to find and visualize the clusters (classes) in the data. The SOM model shows the existence of two classes, implying it is possible to design a classifier that distinguishes between images with co-localized structures and without them. We provide quantitative proof of the liability and discriminant capabilities of the feature space.