Computer-Aided Diagnosis of COVID-19 Radiography using deep transfer deep learning model TLDeNet

ABSTRACT

Computer-Aided Diagnosis (CAD) is applied in the medical analysis of X-ray images widely. Due to the COVID-19 pandemic, the speed of COVID-19 detection is slow, and the workforce is scarce. Therefore, we have an idea to use CAD to diagnose COVID-19 and effectively respond to the pandemic. Recent studies show that convolutional neural network (CNN) is an appropriate technique for medical image classification. However, CNN is more suitable for datasets with many images, such as ImageNet. Medical image classification relies on doctors to label medical images, so obtaining large-scale medical image data sets is a time-consuming, costly, and unrealistic task. The method of data augmentation for a limited medical dataset can be used to increase the number of images. However, this technology will produce many repeated images, which will easily lead to the overfitting problem of CNN. In the case of a limited number of radiological images, transfer learning is a practical and effective method which can help us overcome the overfitting problem of ordinary CNN by transferring the pre-Trained models on large datasets to our tasks. The proposed model is DenseNet based deep transfer learning model (TLDeNet) to identify the patients into three classes COVID-19, Normal or Pneumonia. We then analyzed and assessed the performance of our model on COVID-19 X-ray testing images by performing extensive experiments. It is finally demonstrated that the proposed model is superior to other deep transfer learning models according to comparative analyses. The Grad-Cam method is finally applied to interpret the convolutional neural network, revealing that our proposed model focuses on the similar region of the X-ray images as doctors. © 2022 IEEE.