XVAE-mViT: Explinable Hybrid Artificial Intelligence Framework for Predicting COVID-19 from Chest X-Ray and CT Scans

The COVID-19 virus has rapidly spread as a global pandemic, causing significant impacts on public health, economies, and daily life worldwide. Accurately and quickly predicting COVID-19 is crucial to maintaining stronger healthcare systems. This paper introduces a novel hybrid model of artificial intelligence that combines the benefits of the Variational Auto-Encoder (VAE) with the attention mechanism based on the Vision Transformer (Vi
$T$
). The novel encoder network is structured with four sequential blocks, each involving residual connections of two multiscale kernel depth-wise separable convolution (MKnDSC) modules. The mobile Vi
$T$
is coupled with the V AE to serve as the classification head for predicting COVID- 19 using chest X-ray (CXR) and computed tomography (CT) scan modalities. We achieved promising classification results with overall accuracies of 96.16% and 95.42% using CXR and CT images, respectively. The proposed hybrid AI framework appears to be a practical solution, especially considering its lightweight structure of 2.15 million parameters and 0.68 FLOPs.