Search

COVID-19 has spiked worldwide, having multiple outbreaks even with the production of vaccines. Imaging exams, such as Computer Tomography (CT) and X-ray, are recommended by the World Health Organization after performing RT-PCR tests to enhance COVID-19 diagnosis for serious cases. This work proposes a deep learning methodology to evaluate whether a patient presents COVID-19-related findings in CT images as an auxiliary diagnostic tool. As a CT exam produces many images related to a patient, some are irrelevant for COVID-19 diagnosis (e.g., closed lungs), using the raw information may hinder the model. Hence, we provide a CT scan image selection algorithm to filter the most informative images with two versions: (a) non-sequential, and (b) sequential. Then, online data augmentation is applied before feeding these images to a Convolutional Neural Network (CNN). Moreover, we evaluate the performance of the model for both versions of the CT selection algorithm in different approaches: (i) 'per-image', (ii) 'per-patient majority voting', and (iii) 'per-patient conservative voting (30%)'. We applied the proposed methodology in a Brazilian university hospital, a reference for COVID-19 treatment. For the test set, approaches (a-i), (a-ii), and (a-iii) reached an accuracy of 84.6%, 70%, and 70%, respectively, while approaches (b-i), (b-ii), and (b-iii) reached 90.9%, 80%, and 80%, respectively. Hence, we consider the proposed sequential version the most suitable image selection algorithm for the analyzed data set and useful for supporting decisions in COVID-19 diagnosis. [ABSTRACT FROM AUTHOR]