Advances towards fully automated in vivo assessment of oral epithelial dysplasia by nuclear endomicroscopy – A pilot study

Abstract

Background

uncertainties in detection of oral epithelial dysplasia (OED) frequently result from sampling error especially in inflammatory oral lesions. Endomicroscopy allows non-invasive, "en face" imaging of upper oral epithelium but parameters of OED are unknown.

Methods

mucosal nuclei were imaged in 34 toluidine blue stained oral lesions with a commercial endomicroscope. Histopathological diagnosis showed 4 biopsies in "dys-/ neoplastic", 23 in "inflammatory", 7 in "others" disease groups. Strength of different assessment strategies of nuclear scoring, nuclear count and automated nuclear analysis were measured by area under ROC–curve (AUC) to identify histopathological "dys-/neoplastic" group. Nuclear objects from automated image analysis were visually corrected.

Results

best performing parameters of nuclear-to-image ratios were the count of large nuclei (AUC=.986) and 6-nearest-neighborhood relation (AUC=.896) and for parameters of nuclear polymorphism the count of atypical nuclei (AUC=.996) and compactness of nuclei (AUC=.922). Excluding low grade OED nuclear scoring and count reached 100% sensitivity and 98% specificity for detection of dys-/neoplastic lesions. In automated analysis, combination of parameters enhanced diagnostic strength. Sensitivity of 100% and specificity of 87% was seen for distances of 6-nearest-neighbors and aspect ratios even in uncorrected objects. Correction improved measures of nuclear polymorphism only. The hue of background color was stronger than nuclear density (AUC=.779 vs. .687) to detect dys-/neoplastic group indicating that macroscopic aspect is biased.

Conclusions

nuclear-to-image ratios are applicable for automated optical in vivo diagnostics for oral potentially malignant disorders. Nuclear endomicroscopy may promote non-invasive, early detection of dys-/neoplastic lesions by reducing sampling error.

This article is protected by copyright. All rights reserved.

http://ift.tt/2uIkjWi