Deep Learning-based COVID-19 diagnostics of low-quality CT images

Abstract

Mass testing of the population is among the most effective measures to combat the COVID-19 pandemic. Among existing diagnostic methods, deep learning-based solutions have the potential to be affordable, quick and accurate. However, these techniques often rely on high-quality datasets, which are not always available in medical scenarios. In this work, we use convolutional neural networks to diagnose COVID-19 on computed tomography images from the COVIDx-CT dataset. The available scans often present noisy artifacts, originated from sensor- and capturing-related errors, that can negatively impact the performance of the model if left untreated. In this sense, we explore several preprocessing strategies to reduce their impact and obtain a more accurate method. Our best model, a ResNet50 fine-tuned with preprocessed images, obtained 97.84% accuracy when prompted with a single image and 99.50% when processing multiple images from the same patient. In addition to achieving high accuracy, interpretability experiments show that the network correctly learned features from the lung and chest area.

@inproceedings{ferber2021bsb,
    author       = "Daniel Ferber and Felipe Vieira and Jo{ã}o Dalben and Mariana Ferraz and Nicholas Sato and Gabriel Oliveira and Rafael Padilha and Zanoni Dias",
    title        = "Deep Learning-based COVID-19 diagnostics of low-quality CT images",
    booktitle    = "Brazilian Symposium on Bioinformatics (BSB)",
    year         = 2021
}