Optimizing COVID-19 Detection: A Deep Dive into Autoencoder Architectures Using X-ray Images

COVID-19 has caused serious health issues worldwide in the past few years, thus making it crucial for researchers to explore the data of patients. Autoencoder are neural network architectures that are gaining popularity for identifying meaningful patterns from complex medical images. This study investigates the ability of autoencoder variants in classifying COVID-19 cases based on chest X-ray images. This study compares six common autoencoder variants: Vanilla, Sparse, Deep, Convolutional, Undercomplete, and Denoising, with a dataset of COVID-19 patient and normal chest X-ray images.Empirical analysis suggests that Sparse autoencoders achieved superior performance, with an accuracy of 80% and F1 score of 0.85, reflecting its capability to detect anomaly and feature identification and achieving low reconstruction errors. Convolutional autoencoders is also found to be promising in reconstructing image structural details, but lagged in fine feature extraction. While, the Deep Autoencoder records a very high reconstruction error of 0.54, suggesting its poor ability to reconstruct and highlighting its incompetence with complicated data sets. The research identifies the effectiveness of sparse autoencoders in image-based COVID-19 classification as validating Autoencoder's applicability to early diagnosis.