A Large Imaging Database and Novel Deep Neural Architecture for Covid-19 Diagnosis
Αναστάσιος Αρσένος, Dimitrios Kollias, Stefanos Kollias
Abstract
Deep learning methodologies constitute nowadays the main approach for medical image analysis and disease prediction. Large annotated databases are necessary for developing these methodologies; such databases are difficult to obtain and to make publicly available for use by researchers and medical experts. In this paper, we focus on diagnosis of Covid-19 based on chest 3-D CT scans and develop a dual knowledge framework, including a large imaging database and a novel deep neural architecture. We introduce COV19-CT-DB, a very large database annotated for COVID-19 that consists of 7,750 3-D CT scans, 1,650 of which refer to COVID-19 cases and 6,100 to non-COVID19 cases. We use this database to train and develop the RACNet architecture. This architecture performs 3-D analysis based on a CNN-RNN network and handles input CT scans of different lengths, through the introduction of dynamic routing, feature alignment and a mask layer. We conduct a large experimental study that illustrates that the RACNet network has the best performance compared to other deep neural networks i) when trained and tested on COV19-CT-DB; ii) when tested, or when applied, through transfer learning, to other public databases.