Abstract
Coronavirus is a pandemic that affects the respiratory system causing
cough, shortness of breath, and death in severe cases. Polymerase chain
reaction (PCR) tests are used to diagnose coronavirus. The
false-negative rate of these tests is high, so there needs a supporting
method for an accurate diagnosis. CT scan provides a detailed
examination of the chest to diagnose COVID but a single CT scan
comprises hundreds of slices. Expert and experienced radiologists and
pulmonologists can diagnose COVID from these hundreds of slices, but
this is very time-consuming. So an automatic artificial intelligence
(AI) based method is required to diagnose coronavirus with high
accuracy. Developing this AI-based technique requires a lot of resources
and time, but once it is developed, it can significantly help the
clinicians. This paper used an Automated machine learning (AutoML)
technique that requires fewer resources (optimal architecture trials)
and time to develop, resulting in the best diagnosis. The AutoML models
are trained on 2D slices instead of 3D CT scans, and the predictions on
unknown data (slices of CT scan) are aggregated to form a prediction of
3D CT scan. The aggregation process picked the most occurred case,
whether COVID or non-COVID from all CT scan slices and labeled the 3D CT
scan accordingly. Different thresholds are also used to label COVID or
non-COVID 3D CT scans from 2D slices. The approach resulted in accuracy
and F1-score of 89% and 88%, respectively. Implementation is available
at github.com/talhaanwarch/mia-covid19