Robust Chest CT Image Segmentation of COVID-19 Lung Infection based on limited data

The coronavirus disease 2019 (COVID-19) affects billions of<br> lives around the world and has a significant impact on public<br> healthcare. Due to rising skepticism towards the sensitivity of<br> RT-PCR as screening method, medical imaging like computed<br> tomography offers great potential as alternative. For this<br> reason, automated image segmentation is highly desired as<br> clinical decision support for quantitative assessment and<br> disease monitoring. However, publicly available COVID-19<br> imaging data is limited which leads to overfitting of traditional<br> approaches. To address this problem, we propose an innovative<br> automated segmentation pipeline for COVID-19 infected<br> regions, which is able to handle small datasets by utilization as<br> variant databases. Our method focuses on on-the-fly<br> generation of unique and random image patches for training<br> by performing several preprocessing methods and exploiting<br> extensive data augmentation. For further reduction of the<br> overfitting risk, we implemented a standard 3D U-Net<br> architecture instead of new or computational complex neural<br> network architectures. Through a 5-fold cross-validation on 20<br> CT scans of COVID-19 patients, we were able to develop a<br> highly accurate as well as robust segmentation model for lungs<br> and COVID-19 infected regions without overfitting on the<br> limited data. Our method achieved Dice similarity coefficients<br> of 0.956 for lungs and 0.761 for infection. We demonstrated<br> that the proposed method outperforms related approaches,<br> advances the state-of-the-art for COVID-19 segmentation and<br> improves medical image analysis with limited data. The code<br> and model are available under the following link:<br> https://github.com/frankkramer-lab/covid19.MIScnn