Gabor Features for the Classification and Evaluation of Chromogenic In-Situ Hybridization Images

High-throughput chromogenic in-situ hybridization (CISH) is a brightfield microscopic technique that reveals the spatial distribution of gene expression in animal cells and tissues by an easily detectable coloured precipitate. The "golden standard" for the grading of CISH-stained tissues involves qualitative scoring by a domain expert. This method is biased, suffers from low reproducibility, and lowers the efficiency of high-throughput experiments. A few quantitative image analysis approaches resolve these issues, but the proposed methods are sensitive to experimental conditions or require expert adjustment of multiple parameters. The idea of our research team is to extract textural information from CISH-images that will be used to generate a feature space for semantic segmentation and functional analysis of gene expression. In our current work, we explore the idea by unsupervised classification based on features generated via Gabor energy filters. The tissue was divided into overlapping 150 mu m tiles and processed with a Gabor filter bank (5 wavelengths, 16 directions, bandwidth 1.4). The results for the 16 directions at each wavelength were combined by a maximum superposition into a single image, and the mean grey value, standard deviation and entropy were measured. After appropriate dimensionality reduction, the tiles were classified by a fuzzy C-means algorithm. Four experts without prior knowledge of the classification results evaluated the strength and pattern of gene expression of a set of randomly selected tiles, and independently each class in the original whole-slide images. A comparison between the class-scale and tile-scale evaluations was used to assess the usefulness of the selected features.