Open Access

ARTICLE

Enhanced COVID-19 and Viral Pneumonia Classification Using Customized EfficientNet-B0: A Comparative Analysis with VGG16 and ResNet50

Williams Kyei^*, Chunyong Yin, Kelvin Amos Nicodemas, Khagendra Darlami

Department School of Computer Science, Nanjing University of Information Science and Technology, Nanjing, 210044, China

* Corresponding Author: Williams Kyei. Email:

Journal on Artificial Intelligence 2026, 8, 19-38. https://doi.org/10.32604/jai.2026.074988

Received 22 October 2025; Accepted 11 December 2025; Issue published 20 January 2026

Abstract

The COVID-19 pandemic has underscored the need for rapid and accurate diagnostic tools to differentiate respiratory infections from normal cases using chest X-rays (CXRs). Manual interpretation of CXRs is time-consuming and prone to errors, particularly in distinguishing COVID-19 from viral pneumonia. This research addresses these challenges by proposing a customized EfficientNet-B0 model for ternary classification (COVID-19, Viral Pneumonia, Normal) on the COVID-19 Radiography Database. Employing transfer learning with architectural modifications, including a tailored classification head and regularization techniques, the model achieves superior performance. Evaluated via accuracy, F1-score (macro-averaged), AUROC (macro-averaged), precision (macro-averaged), recall (macro-averaged), inference speed, and 5-fold cross-validation, the customized EfficientNet-B0 attains high accuracy (98.41% ± 0.45%), F1-score (98.42% ± 0.44%), AUROC (99.89% ± 0.05%), precision (98.44% ± 0.43%), and recall (98.41% ± 0.45%) with minimal parameters (4.0M), outperforming VGG16 (accuracy 84.83% ± 2.24%, F1 84.75% ± 2.33%, AUROC 95.71% ± 1.01%) and ResNet50 (accuracy 93.83% ± 0.59%, F1 93.80% ± 0.59%, AUROC 99.22% ± 0.15%) baselines. It improves over existing methods through compound scaling for efficient feature extraction, reducing parameters by approximately 6x compared to ResNet50 while providing quantitatively assessed explanations via Grad-CAM (average IoU with lung regions: 0.489). In essence, the customized EfficientNet-B0’s integration of compound scaling, transfer learning, and explainable AI offers a lightweight, high-precision solution for differentiating COVID-19 from viral pneumonia in enterprise-level healthcare systems and Internet of Things (IoT)-based remote diagnostics.

---

Keywords

---