A COVID Net-predictor: A multi-head CNN and LSTM-based deep learning framework for COVID-19 diagnosis

Abstract

COVID-19 pandemic alerts the necessity of preparing alternate respirational health detective measures that improve time, expense, and prediction performance.  Prevention of COVID-19 spread depends on early identification and precise diagnosis. Since the commonly used real-time Reverse Transcriptase Polymerase Chain Reaction (RT-PCR) swab test is laborious and unreliable, radiography images are still advised for chest screening. Unfortunately, complexities in early detection using traditional approaches urge innovative research in this field. Intending to introduce a novel COVID-19 prediction scheme, this paper employed a COVIDNet-Predictor. This model is built with various stages including preprocessing, segmentation, feature extraction, selection and fusion, prediction and monitoring.  The input images are initially preprocessed to enhance image quality and noise reduction. A U-net segmentation is carried out to find the Region of Interest (ROI). Color, shape and textual features are extracted and are further optimally chosen by a hybrid optimizer EvoNSGA II. Besides, the optimal features are fused through a Hierarchical Attention Network (HAN) and given as input to the COVIDNet-Predictor. The proposed COVIDNet-Predictor is a combination of Multi-Head Convolutional Neural Network (MHCNN), and Long-Short-Term Memory (LSTM)architectures. Additionally, a monitoring and feedback loop is added to make the model fit the real-time applications based on patient data. The efficacy of the proposed COVIDNet -Predictor is evaluated via a comparison with SOTA models and proved its competence by attaining 95.04% accuracy.