An enhanced UNet3+ model for accurate identification of COVID-19 in CT images

Abstract

The COVID-19 pandemic has resulted in an out-of-control number of infections worldwide, causing severe and irreparable consequences. Computed Tomography scans show lung damage in patients. In this study, we propose using deep learning techniques, particularly image segmentation techniques on medical data, to facilitate the identification of affected areas, aiding medical professionals in detecting and screening this disease. This research is based on applying the UNet3+ architecture for image segmentation on CT lung images. Additionally, the integration of the UNet3+ architecture with SE-ResNeXt50 and ResNet50 has demonstrated the effectiveness of leveraging the strengths of these architectures together. The proposed methods are evaluated on a dataset that includes 373 out of the total of 829 slices from 9-axis computed tomography images evaluated by experienced radiologists. Experimental results show that the combination of UNet3+ and SE-ResNeXt50 is more effective for identifying COVID-19 infection, with a mean Intersection over Union value of 0.9290 and a mean Dice coefficient value of 0.9619. At the same time, the segmentation efficiency of COVID-19-infected regions achieved quite good results, with the Dice index reaching 0.9111 and IoU reaching 0.8367, which are promising for medical data segmentation and strong support for Healthcare.