Different Transfer Learning Approaches for COVID-19 Detection and Grad-CAM-Based Severity Assessment Using Chest X-Ray and CT Images
Keywords:
assessment, severity, CAM, Grad, CT scanCXR, CT scan, CXR, transfer learning, COVID-19Abstract
Abstract— COVID-19 detection from chest radiological images remains challenging because its imaging patterns may overlap with pneumonia and normal lung variations. This paper presents a comparative transfer learning framework for automatic COVID-19 detection and Grad-CAM-based severity assessment using chest X-ray and CT images. Three pre-trained convolutional architectures—DenseNet-201, VGG-16, and Inception-v3—were modified and evaluated on two balanced datasets, DS_X9K and DS_CT9K, each containing 9,000 images equally distributed among COVID-19, pneumonia, and normal classes. Twelve models were investigated across both imaging modalities using two strategies: frozen feature extraction with trainable classification layers, and full fine-tuning of all network parameters. The results show that full fine-tuning achieved stronger and more stable performance, reaching up to 98% test accuracy on CXR images and approximately 99% on CT images. Grad-CAM heatmaps provided interpretable visual evidence by localizing disease-related lung regions and supporting visual severity assessment.
