An Uncertainty-Aware Deep Learning-Based Model for COVID-19 Diagnosis

The application of deep-learning(DL) to the processing of medical images has been highly successful. A few groundbreaking DL-based works have achieved substantial progress in computerized screening of COVID-19 using chest X-ray( CXR) pictures in the wake of the current epidemic scenario caused by SARS-CoV-2. While DL models have shown promise, they currently lack a crucial component for use in medical image analysis: the ability to communicate uncertainty in the strategies prediction. Therefore, in this paper, we create a design for automatically detecting COVID19 from CXR images using an uncertainty-aware convolutional-neural-network(UACNN) and estimate the uncertainty in its predictions. In the suggested method, an EfficientNetB3 model is tuned using CXR pictures, and then the model is used in conjunction with Monte Carlo (MC) dropout. The posterior predictive distribution has been obtained after M forward passes of inference using MC dropout. The anticipation and design uncertainty are then determined by calculating the mean and entropy of the generated anticipative distribution. The COVID19CXr, XR-pictures, and Kaggle CXR datasets are used to test the suggested approach. Multi-class classification on the COVID19CXr dataset was completed with an accuracy of 99.81% using the proposed UACNN approach. For the purpose of diagnosing COVID-19 cases from CXR pictures, our suggested methodology demonstrates its superiority over the existing methodologies.