A Neuroimaging ML-Based Framework for Anosmia Grading in Covid-19

Abstract

SARS-CoV-2 is the cause of the acute respiratory illness COVID-19. In order to identify whether a COVID-19 patient has normal, moderate, or severe anosmia, this study suggests a computer-assisted diagnostic system. It utilizes fluid-attenuated inversion recovery Magnetic Resonance Imaging and Diffusion Tensor Imaging to extract the olfactory nerve’s appearance, morphology, and diffusivity markers. The suggested CAD system consists of the following stages: (1) extract appearance markers, morphology markers, and diffusivity markers, (2) apply markers fusion, and (3) select the most promising classifier as the basis for the decision and the relevant performance indicators. The current work is unique in that it uses an ensemble of the trained, tuned, and improved ML classifiers to identify OB anosmia using majority voting. From the fusion tests using different folds, stacking boosted the weight accuracy, weighted precision, weighted specificity, weighted sensitivity, and weighted F1-score to 85.30%, 80%, 89.50%, 78.90%, and 78.30% respectively. Regarding the various performance criteria, the LGBM outperforms other ML classifiers.