Grad-CAM guided transfer learning with optimizer evaluation for multi-disease chest x-ray classification

Abstract

Lung diseases are serious health conditions that affect normal human life and are considered life-threatening. The timely diagnosis of lung infections is essential to ensure treatment and thus improve recovery rates. Clinicians most commonly use chest radiographs to detect conditions such as tuberculosis, lung opacity, COVID-19, lung cancer, and pneumonia. Interpreting chest X-rays manually is both time-consuming and dependent on individual expertise, which can introduce variability. In this context, this study aims to examine the impact of optimizer selection on transfer learning for multiclass classification of chest X-ray images while incorporating explainability through visualization. Five pretrained deep learning network models, including InceptionV3, InceptionResNetV2, ResNet50, DenseNet, and NASNetMobile, were fine-tuned using transfer learning. The COVID QU-Ex dataset, comprising chest X-ray images categorized as COVID-19, normal, and non-COVID (pneumonia), was used to conduct the experiments. The models were trained and evaluated using competing optimizers, including Adaptive Moment Estimation (Adam), Stochastic Gradient Descent (SGD), and Root Mean Squared Propagation (RMSProp), to determine the most effective optimization strategy. Experimental results demonstrate that InceptionResNetV2 with Adam achieved 93.89% accuracy, DenseNet201 with SGD achieved 95.19% accuracy, and InceptionV3 with the RMSProp optimizer achieved 92.39% accuracy. Gradient-Weighted Class Activation Mapping was applied to the best-performing model, DenseNet201, to effectively visualize the chest X-ray regions that contribute to prediction and classification.