Evaluating Deep Learning for CT Scan COVID-19 Automatic Detection


Aimee Putri Hartono; Callista Roselynn Luhur; Clarissa Angelita Indriyani; Claudia Rachel Wijaya; Nunung Nurul Qomariyah; Ardimas Andi Purwita

Aside from Reverse Transcription Polymerase Chain Reaction (RT-PCR), another common method to check for the 2019 novel Coronavirus disease (COVID-19) is by using a chest CT scan. Imaging data is profoundly useful in the diagnosis, detection of complications, and prognostication of COVID-19, displaying various spots in the lungs affected by the viral infection. The complex results often require some time before radiologists can analyze them and are more prone to human errors. Inventions of medical assisting tools, through enhancement of artificial intelligence, are crucial in fighting the COVID-19 pandemic through automation of classifications and the future of medicine. To overcome the above challenges, this paper aims to propose and evaluate the performance between Convolution Neural Network (CNN) and Transfer Learning (TL) in the detection of COVID-19 infections from a Lung CT Scan. Gradient-Weighted Class Activation Mapping (Grad-CAM) will also be utilized to display the infected areas in the lungs for explorative experiments. Transfer-learning using our pre-trained model resulted in a detection accuracy result of 89% while our proposed CNN demonstrated the best result in terms of classification accuracy at 97%. Training time required for the two frameworks are 12 and 22 minutes respectively. By and large, our comparison of using the CNN model versus the pre-trained model gives rise to the conclusion that using the former method proves to be a more effective technique of COVID-19 detection by CT-scan.