Entropy estimation for segmentation of multi-spectral chromosomeimages

摘要

In the early 1990s, the state-of-the-art in commercial chromosomeimage acquisition was grayscale. Automated chromosome classification wasbased on the grayscale image and boundary information obtained duringsegmentation. Multi-spectral image acquisition was developed in 1990 andcommercialized in the mid-1990s. One acquisition method, multiplexfluorescence in-situ hybridization (M-FISH), uses five color dyes. Wepreviously introduced a segmentation algorithm for M-FISH images thatminimizes the entropy of classified pixels within possible chromosomes.In this paper, we extend this entropy-minimization algorithm to work onraw image data, which removes the requirement for pixel classification.This method works by estimating entropy front raw image data rather thancalculating entropy from classified pixels. A successful example imageis given to illustrate the algorithm. Finally, it is determined thatentropy estimation for minimum entropy segmentation adds a heavycomputational burden without contributing any significant increase inclassification performance, and thus not worth the effort