Weighted ensemble learning for accurate COVID-19 CT image segmentation

Coronavirus disease (COVID-19) has been spreading rapidly around the world and causing massive human death and substantial economic losses. Computed tomography (CT) images provide a rich information about patient conditions, assisting physicians to diagnose, treat, and evaluate COVID-19 assessment. Automatic segmentation of CT images with COVID-19 faces many challenges such as the diversity of lesion features and the similarity of lesions to normal tissue. To address these issues, we propose a weighted ensemble learning approach that combines several deep learning models assigned with different weights for this segmentation. We trained and evaluated our method on 259 CT volumes with 23115 slices of COVID-19 patients from three public databases, with the experimental results showing that our method outperforms the other deep learning models for COVID-19 CT segmentation, improving the average dice and jaccard from (71.6%, 57.9%) to (72.4%, 59.5%). Particularly, Without creating a new network structure, our method can improve the accuracy and generalization ability of deep learning models.