Slicer: Feature Learning for Class Separability with Least-Squares Support Vector Machine Loss and COVID-19 Chest X-Ray Case Study

Abstract

Datasets from real-world applications usually deal with many variables and present difficulties when modeling them with traditional classifiers. There is a variety of feature selection and extraction tools that may help with the dimensionality problem, but most of them do not focus on the complexity of the classes. In this paper, a new autoencoder-based model for addressing class complexity in data is introduced, aiming to extract features that present classes in a more separable fashion, thus simplifying the classification task. This is possible thanks to a combination of the standard reconstruction error with a least-squares support vector machine loss function. This model is then applied to a practical use case: classification of chest X-rays according to the presence of COVID-19, showing that learning features that increase linear class separability can boost classification performance. For this purpose, a specific convolutional autoencoder architecture has been designed and trained using the recently published COVIDGR dataset. The proposed model is evaluated by means of several traditional classifiers and metrics, in order to establish the improvements caused by the extracted features. The advantages of using a feature learner and traditional classifiers are also discussed.