Fine-Tuning A Lightweight Convolutional Neural Networks for COVID-19 Diagnosis

Abstract

In this paper, we compare the performance of the deep neural network-based image classifiers and fine-tune with different hyperparameter configurations for an automatic COVID-19 diagnosis from various and limited chest x-ray image dataset provided by Deep Learning and Artificial Intelligence Summer School 3 (DLAI3). We show that high accuracy results can be obtained using the transfer learning technique combined with a well fine-tuned Convolutional Neural Network. Moreover, we seek for not only smaller deep learning architectures with less trainable parameters to reduce the training and inference time of AI applications for mobile and edge devices, but also relatively high performance. The results from the DLAI3 hackathon session show that our model outperforms other submitted models in terms of effectiveness and generalization.