Empirical Fusion: Deploying Moment Invariants and Deep Learning in COVID-19 Diagnostics

The ongoing COVID-19 pandemic has engendered a global health emergency. Its swift transmission has resulted in a substantial number of infections and fatalities worldwide. Consequently, there is a pressing demand for rapid and accurate diagnostic methods for COVID-19. Many scholars have made commendable advancements in developing deep learning (DL) models for automated COVID-19 detection utilizing computerized tomography (CT) scans. However, concerns persist about the robustness of these models in the face of minor perturbations and variations in CT imagery. In response to this challenge, this study introduces a methodology that merges a moment invariant (MI) technique with a DL paradigm to extract features, thereby bolstering the resilience of extant COVID-19 diagnostic models. This innovative approach embeds MI-derived features into DL architectures via the cascade fusion mechanism. When assessed on the SARS-CoV-2 dataset, the amalgamation of VGG16 with Hu moments emerges as the most effective, recording a sensitivity of 90% and an accuracy of 93%. These findings signal the promising prospects of enhancing COVID-19 diagnostic efficacy by merging MI and DL features.