Machine learning classification for COVID19 patients performed on small datasets of CT scans.

In this work we evaluated the possibility of carrying out classifications of the outcome of patients with COVID19 disease through machine learning (ML) techniques working on small datasets of computed tomography (CT) images. In fact, one of the most common problems for medical artificial intelligence (AI) applications is the limited availability of annotated clinical data for model training. In the framework of the artificial intelligence in medicine (AIM) project funded by INFN, we analyzed datasets of CT scans of 79 subjects combined with clinical data containing information relating to positive outcome (no need for intensive care) or poor prognosis (admission into intensive care unit and/or death). After segmentation of ground glass opacities related to this pathology, the radiomic features were subsequently extracted from the CTs, selected through various algorithms of dimension reduction or feature selection and used for the training various classifiers. Values of the area under the ROC curve (AUC) of 0.84 were obtained with Gradient Boosting after BORUTA feature selection. Features selected are related to disease characteristics of poor prognosis patients.