Performance Evaluation of Deep Transfer Learning Models in COVID-19 Chest X-Rays Diagnosis

Abstract

The devastation of the COVID-19 pandemic, or even similar worldwide disease catastrophes in the future, can undoubtedly be scaled down by the rapid testing of these infections. Even though RT-PCR has been the primary method of diagnosis in the current pandemic, many experts have also proven the credibility of chest radiography analysis and suggested the use of this testing method. In turn, a wide array of studies in 2020 attempted to create state-of-the-art models for COVID-19 diagnosis, utilizing established deep learning architectures like ResNeXt, Xception, etc. Of course, the studies used different pre-trained models, datasets, and had varying results. Therefore, we look to measure the performance of all the popular CNN architectures in classifying COVID-19 infected chest x-ray from healthy chest x-rays and using a single dataset as a benchmark to find the best performing pre-trained models in this task. In turn, future studies related to COVID-19 CXR detection can use the results of this study to select the best suited pre-trained models. The architectures we tested are all augmented with one Global Average Pooling layer (and the 2-unit output layer, of course) and are trained using an exponential learning rate scheduler. We also experimented with different hyperparameters in an attempt to fine-tune the model and get the best possible results. Our experiments show that most of the CNN models have a similar performance in this task, and even simpler architectures were able to achieve similar results as the more complex ones while having faster training time. However, ResNet models (particularly ResNet101) were able to consistently, though marginally, outperform all the other architectures we included in the study.