Semantic segmentation of CT images for the detection of coronavirus lesions: the case of Algeria

Abstract

Deep learning has acquired a central position in recent times in the medical field, in particular conditions that are delicate to diagnose by the croaker Coronavirus complaint 2019, given its recent appearance and its complexity. The majority of people with COVID- 19 have endured only mild or moderate symptoms and recovered without specific treatment. We propose in this work four purely deep learning infrastructures using classification and semantic segmentation algorithms for the vaticination of coronavirus complaint 2019 grounded on scanographic images from two Algerian medical institutes. After a dataset collection and preprocessing steps, we propose to apply a classic CNN as a point extractor and the XGBoost algorithm as a classifier, as a alternate armature we replace the classic CNN by the pretrained VGG16 as a point extractor. The third armature is grounded on segmentation using a U-Net armature and for the last one is grounded on the U-Net-VGG16 combination. The experimental results show that the combination of U-Net and VGG16 algorithms as a CT image segmentation model yielded better results with an IoU of 97%, a precision of 95%, and a loss of 30%.