ACCOUNTING FOR NON-IDENTICAL DATA IN BIOIMAGE INFORMATICS

Abstract

Bio-image informatics involves developing computational methods to analyze microscopic images in biological studies. The underlying assumption in these methods is that data are independent and identically distributed. However, in real-world settings, data generally originate from heterogeneous sources even if they do possess a common data-generating mechanism. Since these sources are not identically distributed by necessity, the assumption of identical distribution is inappropriate.

Here, we have proposed a distance measure called Poisson-Binomial Radius (PBR) to account for non-identical data. Results from image classification and recognition experiments demonstrate that PBR outperforms state-of-the-art distance measures. Furthermore, we have used the PBR principle to design an automated pattern recognition system for detecting nasopharyngeal carcinoma using indirect immunofluorescence images. This new framework allows clear-cut cases (which are in the majority) to be machine-evaluated, hence freeing pathology staff to concentrate solely on the difficult cases.