An Optimized Uncertainty Aware Fine-Tuned Transfer Learning for COVID-19 Diagnosis from Medical Images

Abstract

In worldwide, COVID-19 has had a significant influence on patients and healthcare systems. Earlier stage of COVID-19 diagnosis and identification are the primary problems in the current pandemic condition. The identification of COVID-19 in CT and chest-X-ray (CXR) imaging is essential for diagnosis, treatment, and evaluation. However, radiologists face a foreseeable issue when it comes to coping with analytical ambiguity in medical imaging. In that situation, a paradigm based on convolutional neural network (CNN) with transfer learning (TL) and taking uncertainty into account was suggested to identify COVID-19 from CT and CXR scan images. However, this method was less capable to extract more useful and distinct image attributes. By fine-tuning the TL network design, this issue can be resolved. The fine tuning model can only fine tune specific layers associated with various goal objectives. However, one of the primary issues with such a method is the selection of layers. To solve this issue, this research uses an enhanced spider monkey optimization (ESMO) technique to select layers of ResNet architecture. Every population of an initialized spider monkey (SM) selects layer and parameter for fine tuning architecture. The fitness value of each SM is used to find best optimal solutions. Categorical cross-entropy loss (CCEL) is considered as fitness of SM. The fitness value of each SM is employed to determine the highest optimal solutions. Subsequent processes such as the stance update process, the learning and decision phase for the local and global leaders of ESMO algorithms, iteratively search for near optimal solutions until convergence. The proposed method can automatically estimate the various CNN layers, which can then be fine-tuned to extract more significant and discriminative features for efficient COVID-19 identification. Finally, the results reveal that the proposed ESMO-ResNet model on SARS-CoV-2 CT database achieves 91.23% accuracy, which is 21.46%, 20.2%, 12.4%, and 6% higher than the AlexNet, multi-source deep transfer learning (MSDTL), stacked convolutional neural network (S-CNN) and dynamic mutual training (DMT) models, respectively. Similarly, the ESMO-ResNet model on Covi-19 Radiography dataset achieves 90.06% accuracy, which is 21.34%, 20%, 9.7%, and 3.5% higher than AlexNet, MSDTL, S-CNN, and DMT models, respectively.