Thick Data Analytics for Small Training Samples Using Siamese Neural Network and Image Augmentation

Abstract

AbstractAlthough machine learning and deep learning has provided solutions and effective predictions to a variety of complex tasks, it requires to be trained with large amount of labeled data in order to make the learning models perform with high accuracy. In many applications such as in healthcare and medical imaging, collecting big amount of data is sometimes not feasible. Thick data analytics is an attempt to solve this challenge by incorporating additional qualitative interventions such as involving expert’s heuristics to annotate and augment the training data. In this article, we are embarking on an investigation to involve the heuristics of a human radiologist in identifying COVID-19 few cases of CT-Scans imaging through the use of groups of image annotation and augmentation techniques. The identification of new COVID-19 is carried out utilizing unique structure Siamese network to rank similarity between new COVID-19 CT Scan images and images determined as COVID provided by the radiologist. The Siamese network extracts the features of the augmented images compared to the new CT-Scan image to determine whether the new image is COVID-19 positive using a similarity ratio. The results show that the proposed model of using the augmentation heuristics trained on small dataset outperforms the advanced models that are trained on datasets containing large numbers of samples. This article starts by answering key questions on why we need CT-Scans for COVID-19 diagnosis and what is the notion of Thick Data and the use of image augmentation as heuristics as well as what is the role of Siamese Neural Network in learning from small samples. Based on answering these questions, the analytics method described in this paper will have better justification.KeywordsThick data analyticsImage annotation and augmentationSiamese neural networkCOVID-19 CT-scan imagingFew shot learning