A Transfer Learning-Based Framework Aware of Uncertainty for Diagnosing COVID-19

Preventing and controlling an epidemic of COVID-19 requires rapid and accurate patient detection. There are real issues concerning the reliability and efficacy of the PCR tests used to identify COVID-19, and many of these assays are not accessible in many countries. In light of these constraints, this study introduces a transfer learning approach that takes uncertainty into account while searching for COVID-19 in medical images. ResNet50, VGGI6, DenseNetl21, as well as InceptionResNetV2 are four popular convolution neural networks (CNNs) used to extract deep attributes from chest X-ray and CT scans. To identify COVID 19, the collected attributes are fed into several machine learning and statistical modeling frameworks. In addition, we compute and make available the conceptual uncertainty of classification findings, which may be used to spot situations in which trained models are unsure of their verdicts (out-of-distribution issue). Sensitivity, precision, sensitivity, and ROC-curve area-under-the-curve are all maximised using linear support vector machines along with neural network models for huge simulated X-ray as well as CT scan data sets. It is also shown that the uncertainty in CT image predictions is much higher than in X-ray image predictions.