A data-augmented approach to transfer learning for Covid-19 detection

Early Covid-19 detection can help with efficient treatment and isolation plans to prevent its spread. Recent transfer learning methods are constrained by the size of labeled datasets, affecting the reliability of the Covid-19 diagnosis. Motivated by the success of data augmentation for various classification problems, in this work, we adopt an approach called contrast limited adaptive histogram equalization (CLAHE) to train the last layer of the various popular convolutional neural network models. This optimization to transfer learning aims to mitigate the bias prevalent in insufficient labeled Covid-19 datasets. Specifically, we transfer learned AlexNet, ZFNet, VGG-16, ResNet-18, and GoogLeNet using the CLAHE-augmented dataset. Experiment results reveal that the CLAHE-based augmentation demonstrates better performance in contrast to models fine-tuned under non-augmented datasets.