Attention-CoviNet: A Deep-Learning Approach to Classify Covid-19 Using Chest X-Rays

Covid-19 pandemic is spreading across the world at a breakneck pace. As of September 2022, JHU CSSE COVID-19 Data and Our World in Data have recorded up to 5.2 M deaths due to Covid-19. One of the essential steps of fighting this pandemic is to detect the disease early. Studies show some abnormalities in the chest X-rays of the Covid-19 patients, and these features help classify the covid-19 positive patients from the negative patients. In recent years in the field of medical x-rays, Deep Learning models have proven to have the ability to learn and identify features that a trained person can only identify. Due to the exponential increase in the covid-19 cases, the waiting time for covid results has increased, resulting in the late diagnosis of patients for Covid-19. Classifying a patient as Covid-19 positive for a genuinely covid positive patient is as essential as classifying a patient Covid-19 negative who is genuinely covid-negative. Here we present a detailed quantitative analysis on the performance of state-of-the-art models like ResNet-50, Dense-Net, Mobile-Net-V2, and MNAS-Net on classifying patients for Covid positive, Pneumonia positive, and Normal by evaluating Accuracy, Precision, Recall, and Jaccard Index. We propose a new deep neural network classification model for low-end devices that uses two Attention mechanisms. Figure 1. shows the attention map with confidence percentage for each classes. Our model uses fewer parameters and FLOPS than other state-of-the-art models and recorded a 2% increase in Accuracy and other evaluation parameters. The dataset used for implementation is a public dataset COVID-DATASET and NIH Chest X-ray from kaggle. It has over 1000 images of Covid Chest X-ray, Pneumothrax, Mass, Pneumonia, Cardiomegaly, Nodule, Effusion, Atelectasis and inflitration. All the model implementations are implemented on PyTorch.