This article was originally published here
Comput Biol Med. Nov 23, 2021; 140: 105047. doi: 10.1016 / j.compbiomed.2021.105047. Online ahead of print.
Deep learning (DL) has been very successful in the field of medical image analysis. In the wake of the current SARS-CoV-2 pandemic situation, some DL-based pioneering work has made significant progress in automated screening for COVID-19 disease from chest x-ray images (CXR). But these DL models have no inherent way of expressing the uncertainty associated with model prediction, which is very important in the analysis of medical images. Therefore, in this article, we develop an uncertainty-aware convolutional neural network model, named UA-ConvNet, for the automated detection of COVID-19 disease from CXR images, with estimation of the associated uncertainty in model predictions. The proposed approach uses the EfficientNet-B3 model and the Monte Carlo (MC) stall, where an EfficientNet-B3 model has been refined on the CXR images. During inference, MC dropout was applied for M forward passes to obtain the posterior predictive distribution. After that the mean and entropy have been calculated on the obtained predictive distribution to obtain the mean prediction and the uncertainty of the model. The proposed method is evaluated on the three different data sets of chest x-ray images, namely the COVID19CXr, x-ray image and Kaggle data sets. The proposed UA-ConvNet model achieves a mean G of 98.02% (with a confidence interval (CI) of 97.99 to 98.07) and a sensitivity of 98.15% for the multiclass classification task on the COVID19CXr dataset. For binary classification, the proposed model achieves a G mean of 99.16% (with a CI of 98.81 to 99.19) and a sensitivity of 99.30% on the x-ray image dataset. Our proposed approach shows its superiority over existing methods to diagnose COVID-19 cases from CXR images.
PMID: 34847386 | DOI: 10.1016 / j.compbiomed.2021.105047