Artificial Intelligence Theory and Applications, cilt.1, sa.3, ss.19-28, 2021 (Hakemli Dergi)
Introduction-Objectives: The high contagiousness of the SARS-COV-2 virus
has resulted in many people being infected worldwide. In many countries, the
capacity of intensive care units has been insufficient and has become unable to
accept new patients. Imaging-based non-invasive methods developed as an
alternative to the RT-PCR technique to control the spread of the virus during
the pandemic process generally focus on the presence or absence of the disease.
However, these methods do not provide information about how severe the
disease is and how it progresses. Therefore, in this study, a deep learning-based
estimation framework with low computational load is proposed to predict
severity scores using chest radiographs.
Materials-Methods: The pre-trained ImageNet models are used as feature
extraction networks to extract generic features. A two-headed estimation
subnetwork each with the same number of layers is created to learn taskspecific features. Eventually, an end-to-end trainable lightweight deep model
is created by connecting the estimation subnetwork to the feature extraction
network.
Results: The proposed model is evaluated on a publicly available Cohen’s
covid-chestxray-data set. The best cross-validation performance in terms of
RMSE, MAE, and R2
in the prediction of lung involvement and opacity is
obtained as 1.39/0.98, 1.1/0.81, 0.65/0.66, respectively.
Conclusions: Although the model has been trained with limited data,
promising results are achieved with an end-to-end framework for estimating
the severity of the COVID-19 disease.