Segmentation of COVID-19 Lesions in CT Scans through Transfer Learning

Abstract

Since its emergence at the end of 2019, SARS-CoV-2 has infected millions worldwide, challenging healthcare systems globally. This has prompted many researchers to explore how machine learning can assist clinicians in diagnosing infections caused by SARS-CoV-2. Building on previous studies, we propose a novel deep learning framework designed for segmenting lesions evident in Computed Tomography (CT) scans. For this work, we utilized a dataset consisting of 20 CT scans annotated by experts and performed training, validation, and external evaluation of the deep learning models we implemented, using a 5-fold cross-validation scheme. When splitting data by slice, our optimal model achieved noteworthy performance, attaining a Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) score of 0.8644 and 0.7612 respectively, during the validation phase. In the external evaluation phase, the model maintained strong performance with a DSC and an IoU score of 0.7211 and 0.5641, respectively. When splitting data by patient, our optimal model achieved a DSC score of 0.7989 and an IoU score of 0.6686 during the validation phase. During the external evaluation phase, the model maintained strong performance with a DSC and IoU score of 0.7369 and 0.5837, respectively. The results of this research suggest that incorporating transfer learning along with appropriate preprocessing techniques, can contribute to achieving state-of-the-art performance in the segmentation of lesions associated with SARS-CoV-2 infections.