Few Shot Learning of COVID-19 Classification Based on Sequential and Pretrained Models: A Thick Data Approach

Classification tasks face several issues when applied to complex data sets and sophisticated images such as CT scans. Long training times are needed to properly train traditional networks to classify images, as well as the need for large amounts of data for these networks to draw accurate conclusions. Even when supplied with large datasets, popular neural networks like VGG and ResNet fail to classify images accurately and consistently for sensitive tasks like identifying COVID-19 in a CT lung scan. To overcome these challenges, we apply Siamese neural network architecture, which has been reported to reduce training times and required training data, to a sequential network. To further empower this network, we incorporate thick data heuristics into the CT image dataset, specifically, we annotate areas of interest in the images that a radiologist would be looking for to make a diagnosis, such as ground glass opacities. Our network outperforms five leading image classification neural networks by about 3% when classifying the same CT lung scan images as positive or negative for COVID-19. By applying data thickening heuristics, we have shown that accuracy is improved, and suspect that the accuracy will continue to increase as more heuristics based on more radiologists and imaging experts are to be added on top of what we have considered in this paper.