ABSTRACT
This paper presents an automated COVID-19 lung lesions segmentation method based on a deep three-dimensional convolutional neural network model which automatically detects and extracts multifocal, bilateral and peripheral lung lesions from chest 3D-CT scans. The proposed CNN model is based on a modified 11-layer U-net architecture and employs a loss function that combines Dice coefficient and Cross-Entropy. It has been tested and evaluated on Covid-19-20-v2 training dataset containing a total of 199 3D-CT scans of different subjects with COVID-19 lesions representing different sizes, shapes and locations in CT images. The obtained results have proven to be satisfactory and objective, as well as similar and close to ground truth data provided by medical experts. On these challenging CT data, the proposed CNN obtained average scores of 0.7639, 0.8129 and 0.9986 corresponding to Dice Similarity Coefficient, Sensitivity and Specificity metrics respectively. © 2022 IEEE.