COVID-19 Detection on Chest X-ray Using an Enhanced Neural Network Model: Impact of Network Complexity, Data Augmentation and Transfer Learning

Machine learning (ML) algorithms have potential to rapidly screen COVID-19 from chest x-ray (CXR). Current deep convolutional neural network (DCNN) models for COVID-19 detection are limited by small datasets and overfitting. We hypothesized that less network complexity, heavy data augmentation, and transfer learning would result in the best model. A COVID-19 detection model was developed using the COVIDx public dataset of 16,352 de-identified CXRs associated with known COVID-19 status by reverse transcriptase polymerase chain reaction (RT-PCR). Twenty-four pre-trained DCNNs with various enhancement features were compared using 80/20 split for testing and validation. Among 5 pretrained DCNN’s, the low complexity but deep ResNet18 architecture performed best. Data augmentation using horizontal flip (HF), Gaussian blur (GB), and cutout (CO) improved ResNet18 performance- with the ResNet18- CO/GB model performi¬¬ng best at 1,000 iterations. Although transfer learning using an extrinsic pneumonia detection model did not boost performance, transfer learning from tuberculosis (TB) detection models enhanced performance of ResNet18-HF and ResNet18-CO/HF/GB models. Comparing the top models at 10K iterations, the best model was ResNet18-GB/CO without transfer learning with sensitivity 82.0%, specificity 96.5%, and accuracy 94.5%. Our findings suggest utility for automated COVID-19 detection by CXR using DCNN’s enhanced by data augmentation more so than transfer learning.

Author(s): Himal Bamzai-Wokhlu

Abstract | PDF

Share This Article