Statistical characterization of tissue images for detection and classification of cervical precancers

TL;DR

This paper presents a statistical analysis of tissue images from cervical biopsies, demonstrating that correlation properties and distribution moments can reliably distinguish normal from precancerous tissues with high accuracy.

Contribution

It introduces a novel statistical characterization method for tissue images that enhances the detection and classification of cervical precancers.

Findings

Correlation domains indicate cellular clustering differences.

Statistical moments effectively differentiate tissue types.

High sensitivity and specificity in classification.

Abstract

Microscopic images from the biopsy samples of cervical cancer, the current "gold standard" for histopathology analysis, are found to be segregated into differing classes in their correlation properties. Correlation domains clearly indicate increasing cellular clustering in different grades of pre-cancer as compared to their normal counterparts. This trend manifests in the probabilities of pixel value distribution of the corresponding tissue images. Gradual changes in epithelium cell density are reflected well through the physically realizable extinction coefficients. Robust statistical parameters in the form of moments, characterizing these distributions are shown to unambiguously distinguish tissue types. These parameters can effectively improve the diagnosis and classify quantitatively normal and the precancerous tissue sections with a very high degree of sensitivity and specificity.