A high accuracy of deep learning based CNN architecture: classic, VGGNet, and RestNet50 for Covid-19 image classification

Ibnu Utomo Wahyu Mulyono (University of Dian Nuswantoro)
Eko Hari Rachmawanto (University of Dian Nuswantoro)
Christy Atika Sari (University of Dian Nuswantoro)
Md Kamruzzaman Sarker (Bowie State University)

TELKOMNIKA (Telecommunication Computing Electronics and Control)
Vol 22, No 5: October 2024

This research paper provides a detailed examination of different convolutional neural network (CNN) structures used in Covid-19 image classification tasks. The study thoroughly investigates the performance of classic CNN, visual geometry group (VGG), and ResNet-50 architectures across a variety of datasets. The analysis focuses on evaluating the efficacy of each architecture by considering metrics such as accuracy, precision, recall, and F1-Score. The experimental results reveal that the ResNet-50 architecture achieves the highest performance with an accuracy rate of 96.63%, outperforming both VGG and classic CNN models. This finding emphasizes the importance of architectural choices and hyperparameter selection in achieving optimal performance in image classification tasks. The combination of the ResNet-50 architecture with the Adam optimizer demonstrates its effectiveness in improving classification accuracy. These findings contribute to the field of deep learning by providing valuable insights into the performance analysis of CNN architectures and highlighting the significance of selecting appropriate hyperparameters for optimal model performance. The selection of VGG and ResNet-50 architectures was based on their strong feature extraction capabilities, proven state-of-the-art performance, and their suitability for transfer learning. VGG and ResNet-50 also have widely available pre-trained models, facilitating their usage and experimentation.