Deep Learning based Super-Resolution for Accurate COVID-19 diagnosis in Chest CT Images

The pandemic due to COVID-19 led to a massive demand for more accurate and efficient detection of infected people. One such detection method is CT scanning, which can visualize the features of COVID-19, such as COVID pneumonia, visible in CT scans as ground glass opacities in the lungs. But CT scan imaging has a limited resolution, which makes it more challenging to determine lesions and changes in tissues that are small in size. To avoid this, researchers used super-resolution techniques to increase the resolution of CT images, which reconstructs a low-resolution image into a high-resolution image. The resultant super-resolved images have improved image details and clarity. This study proposed two convolutional neural networks, MediSRClassCNN and MediSRClassTransfer, to classify COVID- 19 using super-resolved CT images. The proposed methods combine a MediSR super-resolution network with a classification CNN model to classify COVID-19 disease as positive or negative using CT images. The proposed methods were evaluated on the COVID-CT public dataset. The results show that the proposed approach resulted in better performance than the existing work. It achieves at least a 5% improvement compared to existing works. With its improved Accuracy, Sensitivity, Specificity, F1 Score and AUC, the proposed work enables the detection of COVID-19 earlier, leading to timely treatment and cure of the disease.