Boosting Performance of Transfer Learning Model for Diagnosis of COVID-19 from Computer Tomography Scans

Abstract

Early-stage rapid and accurate diagnosis of COVID-19 pneumonia is of great importance as a measure to the fight against the pandemic. Even if real-time reverse transcription-polymerase chain reaction (RT-PCR) test seems like a gold standard for determining COVID-19, the availability and the accuracy is still a challenge. Thus, alternative diagnostic techniques are required for controlling the spreading of the disease. Amongst the radiodiagnostic methods, the computer tomography (CT) technique is one of the most powerful candidates for screening COVID-19 pneumonia accurately. In this study, it is aimed to develop a reliable transfer learning-based CNN model tailored to detect the COVID-19 from chest CT scans with high accuracy and sensitivity to help to accelerate the application of the required treatment and taking of measures. The CT scan dataset used in the training process of the CNN model was obtained from “SARS-CoV-2 CT-Scan Dataset”. This dataset contains 1252 CT scans for positive COVID-19 case and 1230 CT scans for the non-COVID-19 case, 2482 CT scans in total, all data have been collected from real patients from hospitals in Sao Paulo, Brazil. ResNet18, ResNet50, VGG16, AlexNet, and SqueezeNet1_1 architectures were re-trained to extract COVID-19 classes by transfer learning. The highest classification performance parameters were obtained for ResNet50 architecture and were calculated as 99.80% accuracy, 99.61 % precision, and 100.00% sensitivity. The activation maps were created to highlight the crucial areas of the CT images and improve causality and intelligibility. The developed transfer learning model can be utilized for reliable clinical diagnosis of COVID-19 cases from CT images to accelerate the triaging and save critical time for disease control as well as assisting the radiologist to validate their initial diagnosis.

Ähnliche Arbeiten