A Review of Lightweight CNN Architectures and Transfer Learning Methods for Automated Pneumonia Detection

Abstract

Pneumonia is a major respiratory disease that requires timely and accurate diagnosis to optimize the outcome for patients. In recent years, the most popular approach to the automated detection of pneumonia on chest X-ray (CXR) images is based on deep learning, particularly Convolutional Neural Networks (CNNs) and Transfer Learning (TL). Here we review 27 research articles, focusing on methodologies, datasets and performance measures, published between 2019 and 2025. Most studies rely on TL-based CNN architectures, with VGG, Reset, DenseNet, Mobile Net and Efficient Net being the most widely used pre-trained models, with the Kermani pediatric CXR dataset being the most commonly used, followed by large datasets from RSNA[2] and NIH ChestX-ray14. Kaggle is an important resource, it is both a database of data and a community where people share reproducible python code. The accuracy of the model is reported to be from 80% to 90% and recently it is reported to be approximately 99% under optimized condition. It highlights that the same methodologies are being used across the board, the importance of using standardized datasets, and the ongoing improvement of the performance of models, giving a full reference for future AI-supported studies on pneumonia diagnosis. In addition, lightweight CNN architectures have gained significant attention because they can reduce computational complexity while maintaining competitive classification performance. Recent studies have also explored hybrid CNN–Transformer models, ensemble learning, and explainable artificial intelligence techniques to improve robustness and clinical interpretability. However, variations in datasets, preprocessing techniques, evaluation metrics, and experimental conditions make direct comparison between different models difficult. The review further emphasizes the need for diverse and high-quality datasets to improve the generalization capability of pneumonia detection systems. Future research should focus on developing efficient, interpretable, and clinically reliable models that can be deployed in resource-constrained healthcare environments. Such advancements can support radiologists and healthcare professionals by providing rapid and reliable computer-aided diagnosis of pneumonia.