Application of Neural Network-Based U-Net Architecture for COVID-19 Region Detection in CT Scan Images

This study introduces a convolutional neural network (CNN)-based U-Net architecture to automate the segmentation of COVID-19-affected regions from thoracic CT scan images. After expanding the dataset using preprocessing techniques such as rotation and filtering, we designed both <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{U}$</tex> Net and SegNet models to compare their segmentation capabilities. The U-Net model demonstrated high accuracy and F1-scores, effectively identifying normal lung tissue, groundglass opacities, and background areas. Our experiments showed that the approach achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 7. 9 8 \%}$</tex> accuracy on the training set and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5. 4 9 \%}$</tex> on validation. Testing results revealed minimal segmentation errors. The system yielded <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 2 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 7 \%}$</tex> IoU scores in the validation and testing stages, respectively, with MSE values of 0.08 and 0.13. These promising outcomes indicate the potential of our approach to expedite and enhance COVID19 screening in clinical settings.