Comparative Study of Deep Learning Models for Pneumonia Classification

Abstract

Deep learning is a powerful method for analyzing medical data such as detecting pneumonia and other respiratory  diseases from chest X-rays. This paper presents a comparative analysis of a few of the most prominent convolutional neural network (CNN) architectures. These CNNs include VGG16, VGG19, DenseNet121, DenseNet201, MobileNetV1, MobileNetV2, InceptionV3, and Inception-ResNetV2.This study also explores a hybrid VGG19–Transformer architecture to enhance pneumonia detection by combining CNN based spatial feature extraction with transformer-based global context learning. Each of these models was evaluated on a chest X-ray dataset while measuring a set of prediction performance metrics, namely, accuracy, precision, recall, and the F1 score. The results are heterogeneous with respect to the different models, and the highest levels of test accuracy were, however, 82.71% for VGG19 and 81.53% for MobileNetV1. Other architectures such as Dense Net and Inception variant models were noted to have competitive accuracy, but these models were significantly weak for the more difficult problem of class imbalance, particularly distinguishing bacterial from viral pneumonia. The trade-offs of different architectures are discussed, underscoring the trade-off of merely model accuracy for class-for robustness. These outcomes represent a critical foundation for further work aimed at the improvement of deep
learning systems focused on the practical and validated clinical detection of pneumonia, and other respiratory diseases.