Unveiling COVID-19 from Chest X-ray with deeplearning: a hurdles race with small data

Enzo Tartaglione, Carlo Alberto Barbano, Claudio Berzovini, Marco Calandri, Marco GrangettoZenodo 2020 · DOI

Abstract

The possibility to use widespread and simple chestX-ray (CXR) imaging for early screening of COVID-19 patientsis attracting much interest from both the clinical and the AIcommunity. In this study we provide insights and also raisewarnings on what is reasonable to expect by applying deeplearning to COVID classification of CXR images. We providea methodological guide and critical reading of an extensive set ofstatistical results that can be obtained using currently availabledatasets. In particular, we take the challenge posed by currentsmall size COVID data and show how significant can be thebias introduced by transfer-learning using larger public non-COVID CXR datasets. We also contribute by providing resultson a medium size COVID CXR dataset, just collected by oneof the major emergency hospitals in Northern Italy during thepeak of the COVID pandemic. These novel data allow us tocontribute to validate the generalization capacity of preliminaryresults circulating in the scientific community. Our conclusionsshed some light into the possibility to effectively discriminateCOVID using CXR.
BibTeX
@inproceedings{tartaglione2020unveiling,
  author = {Enzo Tartaglione and Carlo Alberto Barbano and Claudio Berzovini and Marco Calandri and Marco Grangetto},
  title = {{Unveiling COVID-19 from Chest X-ray with deeplearning: a hurdles race with small data}},
  booktitle = {Zenodo},
  year = {2020},
  doi = {10.5281/zenodo.4058309},
}

Loading related content...

Figure