Prediction of COVID-19 Diagnosis from Healthy and Pneumonia CT scans using Convolutional Neural Networks

BackgroundCurrent methods of COVID-19 detection from other respiratory illnesses using computed tomography (CT) scans are highly inaccurate. However, understanding pathogen-specific immune responses can help reduce inconsistencies and improve the accuracy of COVID-19 and Pneumonia detection. A deep learning model using Relief-based feature selection (RBAs) was developed to detect COVID-19 and Pneumonia. Patient-specific Class Activation Maps (CAMs) were produced to highlight immunopathogenic differences and identify differences between COVID-19 and Pneumonia on CT scans. MethodsTo examine the effect on lung lesions, a COVIDx CT-2 dataset, containing CT scans from 3,745 patients, was examined. We developed an algorithm to convert the 3-D CT scan of each patient into multiple 2-D slices. Altogether, there were 194,344 2-D slices retrieved from 3,745 CT Scans. The distribution of slices was 67%-20%-17% consisting of COVID-19, Pneumonia, and normal CT scan, respectively. An AlexNet architecture was implemented with additional feature extraction layers (containing RBA) and classification layers to perform deep learning. The 2-D slices were divided into 3 groups: Training, Test, and Validation. The training set consisted of 70% of the data, the test set consisted of 20% of the data, and the validation consisted of 10% of the data. After training, unique CAMs were generated on patient CT scans using the immunopathogenic differences to highlight COVID-19 and Pneumonia related abnormalities. ResultsThe model accurately distinguished hyperinflammation in COVID-19 patients from Pneumonia patients and achieved a validation accuracy of 95.60% and a false-positive rate of 4.65%. Additionally, the segmented lung, shown by the patient-specific CAMs, identified higher levels of inflammation in the lung of COVID scans compared to the other two groups. DiscussionThe use of deep learning in disease diagnosis and prevention has provided many avenues to advance current techniques. Likewise, in this analysis, deep learning was shown to successfully predict COVID-19 via CT scan. By providing patient-specific CAMs, the model can be used to not just aid in diagnosis but potentially also to evaluate serial chest CT scans for treatment.