Dataset Open Access

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

Dominik Müller; Iñaki Soto Rey; Frank Kramer


Citation Style Language JSON Export

{
  "publisher": "Zenodo", 
  "DOI": "10.5281/zenodo.4279398", 
  "title": "Robust Chest CT Image Segmentation of COVID-19 Lung Infection based on limited data", 
  "issued": {
    "date-parts": [
      [
        2020, 
        6, 
        29
      ]
    ]
  }, 
  "abstract": "<p>The coronavirus disease 2019 (COVID-19) affects billions of<br>\nlives around the world and has a significant impact on public<br>\nhealthcare. Due to rising skepticism towards the sensitivity of<br>\nRT-PCR as screening method, medical imaging like computed<br>\ntomography offers great potential as alternative. For this<br>\nreason, automated image segmentation is highly desired as<br>\nclinical decision support for quantitative assessment and<br>\ndisease monitoring. However, publicly available COVID-19<br>\nimaging data is limited which leads to overfitting of traditional<br>\napproaches. To address this problem, we propose an innovative<br>\nautomated segmentation pipeline for COVID-19 infected<br>\nregions, which is able to handle small datasets by utilization as<br>\nvariant databases. Our method focuses on on-the-fly<br>\ngeneration of unique and random image patches for training<br>\nby performing several preprocessing methods and exploiting<br>\nextensive data augmentation. For further reduction of the<br>\noverfitting risk, we implemented a standard 3D U-Net<br>\narchitecture instead of new or computational complex neural<br>\nnetwork architectures. Through a 5-fold cross-validation on 20<br>\nCT scans of COVID-19 patients, we were able to develop a<br>\nhighly accurate as well as robust segmentation model for lungs<br>\nand COVID-19 infected regions without overfitting on the<br>\nlimited data. Our method achieved Dice similarity coefficients<br>\nof 0.956 for lungs and 0.761 for infection. We demonstrated<br>\nthat the proposed method outperforms related approaches,<br>\nadvances the state-of-the-art for COVID-19 segmentation and<br>\nimproves medical image analysis with limited data. The code<br>\nand model are available under the following link:<br>\nhttps://github.com/frankkramer-lab/covid19.MIScnn</p>", 
  "author": [
    {
      "family": "Dominik M\u00fcller"
    }, 
    {
      "family": "I\u00f1aki Soto Rey"
    }, 
    {
      "family": "Frank Kramer"
    }
  ], 
  "note": "Code: https://github.com/frankkramer-lab/covid19.MIScnn", 
  "version": "1.0", 
  "type": "dataset", 
  "id": "4279398"
}
1,656
441
views
downloads
All versions This version
Views 1,656302
Downloads 44162
Data volume 605.0 GB48.2 GB
Unique views 1,433270
Unique downloads 19126

Share

Cite as