SEnsembleNet: A Squeeze and Excitation based Ensemble Network for
COVID-19 Infection Percentage Estimation from CT-Scans
Abstract
Coronavirus (COVID-19) is a contagious disease caused by SARS-CoV-2
virus. Usually, COVID-19 is diagnosed by PCR test, which requires less
human expertise, but this test’s false-negative ratio is high. COVID can
also be diagnosed from radiographs such as CT-scan and X-ray, but it
requires expert radiologists. So there is a need for an automated way to
interpret chest radiographs using artificial intelligence. Several
labelled datasets and deep learning algorithms are available to diagnose
corona patients using radiographs. These algorithms classify the images
into predefined categories such as healthy or infected. But there is no
way to know how much area of chest radiograph is infected by COVID. This
paper proposed an ensemble network to predict COVID-19 percentage
infection from a chest CT scan. The proposed ensemble network used
squeeze and excitation bock to learn individual models’ weights during
the training process. On validation data and test data, the proposed
approach obtained a mean absolute error of 4.469 and 3.64, respectively.
Implementation is publicly available at
https://github.com/talhaanwarch/Covid-Infection-Estimation