UNIVERSAL k-NN (UNN) CLASSIFICATION OF CELL IMAGES USING HISTOGRAMS OF DoG COEFFICIENTS

Abstract

Cellular imaging is an emerging technology for studying many biological phenomena. Cellular image analysis generally requires to identify and classify cells according to their morphological aspect, staining intensity, subcellular localization and other parameters. Hence, this task may be very time-consuming and poorly reproducible when carried out by experimenters. In order to overcome such limitations, we propose an automatic segmentation and classification software tool that was tested on cellular images acquired for the analysis of NIS phosphorylation and the identification of NIS-interacting proteins. On the algorithmic side, our method is based on a novel texture-based descriptor that is highly discriminative in representing the main visual features at the subcellular level. These descriptors are then used in a supervised learning framework where the most relevant prototypical samples are used to predict the class of unlabeled cells, using a new methodology we have recently proposed, called UNN, which grounds on the boosting framework. In order to evaluate the automatic classification performances, we tested our algorithm on a significantly large database of cellular images annotated by an expert of our group. Results are very promising, providing precision of about 84% on average, thus suggesting our method as a valuable decision-support tool in such cellular imaging applications.