Bispectrum Analysis of Thermal Images for the Classification of Retinal Vascular Diseases

Abstract

In the field of medicine, thermal imaging provides a non-invasive means to diagnose diseases by displaying temperature variations. Specifically, thermal images of the ocular region are proving effective in diagnosing retinal diseases like diabetic retinopathy (DR), diabetic macular edema (DME), age-related macular degeneration (AMD), and wet AMD. This study explores the use of bispectrum analysis on ocular thermal images. The bispectrum, derived from third-order statistics (cumulants), is estimated using parametric (ARMA) and non-parametric methods (direct and indirect). Various entropy measures are extracted as features from these bispectrum images. The effectiveness of these features is evaluated through experiments employing different machine learning algorithms: discriminant analysis (DA), K-nearest neighbour (KNN), decision tree (DT), and Naïve Bayes Classifier (NBC). The results indicate that the KNN classifier achieves the highest accuracy when using features extracted via the indirect method of bispectrum estimation. Meanwhile, the DT classifier performs exceptionally well with bispectrum features extracted through the direct method, outperforming other feature extraction approaches.