Dynamically Weighted Pairwise Cross-Attention Driven Feature Fusion in Hybrid Convolutional Neural Networks for Classification of COVID-19 Variants

The extensive global impact of coronavirus is evident, causing widespread disruption to public health and economies around the world. This disease is caused by the severe acute respiratory syndrome virus. Accurate detection helps control the virus spread, reduces death rates, and lessens the overall impact on communities. Several significant research gaps exist in handling unbalanced datasets and achieving high accuracy with properly balanced data. These issues pose substantial challenges to the development of robust and reliable classification models. Unbalanced datasets, where certain classes are overrepresented, can bias models towards dominant classes, leading to suboptimal performance for underrepresented strains. To address these gaps, this paper introduces a novel and effective method to classify the three dominant variants of severe acute respiratory syndrome. Alpha, Delta, and Omicron. Here, we utilize a balanced dataset of 9000 images and we propose an innovative series of Dynamically Weighted Pairwise Cross-Attention feature fusion models designed to effectively integrate complementary genomic features, delivering robust and accurate performance across diverse genomic classification tasks. To achieve this, we first utilize Genomic Image Processing techniques, such as Frequency Chaos Game Representation and Markov Transition Field, to transform genomic sequences into informative visual representations, enabling more effective feature extraction and fusion. Then the resultant images are used to train our series of models. Our enhanced models outperform state-of-the-art results by achieving a remarkable accuracy of 99.62%.