A Shallow Learning Investigation for COVID-19 Classification

COVID-19, an infectious coronavirus disease, triggered a pandemic that resulted in countless deaths. Since its inception, clinical institutions have used computed tomography as a supplemental screening method to reverse transcription-polymerase chain reaction. Deep learning approaches have shown promising results in addressing the problem; however, less computationally expensive techniques, such as those based on handcrafted descriptors and shallow classifiers, may be equally capable of detecting COVID-19 based on medical images of patients. This work proposes an initial investigation of several handcrafted descriptors well known in the computer vision literature already been exploited for similar tasks. The goal is to discriminate tomographic images belonging to three classes, COVID-19, pneumonia, and normal conditions, and present in a large public dataset. The results show that kNN and ensembles trained with texture descriptors achieve outstanding accuracy in this task, reaching accuracy and F-measure of 93.05% and 89.63%, respectively. Although it did not exceed state of the art, it achieved satisfactory performance with only 36 features, enabling the potential to achieve remarkable improvements from a computational complexity perspective.