Performance Evaluation of CNN-based Transfer Learning for COVID-19 Pneumonia Identification with Various Levels of Layer Partial Freezing

Abstract

Pneumonia is a serious complication of coronavirus that can be fatal, especially among the elderly. Early diagnosis of COVID-19 pneumonia increases the likelihood of recovery and prevents the further spread of the virus. Chest X-ray (CXR) images can be utilized to detect specific signs associated with COVID-19, but this needs well-trained radiologists. Alternatively, deep Convolutional Neural Network (CNN)-based models have been successfully applied to diagnose COVID-19 and the associated pneumonia from CXR using transfer learning. This study explores various levels combining layer fine-tuning and freezing in two popular pretrained CNN-based models, VGG16 and ResNET50, and how these combinations influence the learning transferability of pretrained models to improve the identification of COVID-19 pneumonia from CXR images. We found that robust models can be learned with less labeled data in a shorter training time by applying partial freezing instead of the full network fine-tuning without sacrificing a significant portion of their diagnostic performance.