LXNet: A lightweight CNN for lung disease classification from Chest X-ray with XAI-based interpretability.
Authors
Affiliations (4)
Affiliations (4)
- Department of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh.
- School of Informatics, Kochi University of Technology, Kami, Kochi, Japan.
- Information Systems Department, College of Computer and Information Sciences, Imam Mohammad Ibn Saud Islamic University (IMSIU), Riyadh, Saudi Arabia.
- Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.
Abstract
The diagnosis of lung diseases such as pneumonia and tuberculosis remains a major global health challenge, especially in resource-limited regions. Artificial Intelligence (AI) has shown strong potential in analyzing Chest X-Rays (CXR) for accurate and timely diagnosis, but most existing models are computationally heavy and lack interpretability, limiting their practical application. In this study, we present LXNet, a lightweight and explainable Convolutional Neural Network (CNN) for nine-class lung disease classification (Normal, Pneumonia, Higher Density, Lower Density, Obstructive Pulmonary Diseases, Degenerative Infectious Diseases, Encapsulated Lesions, Mediastinal Changes and Chest Changes). The model was evaluated on a diverse CXR dataset containing 6,743 images collected from a private imaging center (GRS Imagem, Brazil), enabling comprehensive multiclass assessment. LXNet contains only 0.35 million parameters and employs a no-pooling final block to preserve subtle diagnostic features while maintaining very low computational cost. Robustness was enhanced through adaptive Contrast Limited Adaptive Histogram Equalization (CLAHE), grayscale normalization and stratified class balancing. LXNet was benchmarked against pretrained CNNs (DenseNet201, ResNet50V2 and InceptionV3) under identical settings. Explainable AI (Grad-CAM, Score-CAM and LIME) provided meaningful visualizations. LXNet achieved 96.1% accuracy in 5-fold cross validation, outperforming the baselines (DenseNet201: 90.3%, InceptionV3: 88.9%) by 1-8%, with only 308 seconds of training on standard hardware. Statistical significance was confirmed using Wilcoxon signed-rank tests (p = 0.03125). These results demonstrate LXNet's promising performance and interpretability; however, reduced external performance indicates limited generalizability and its clinical applicability requires further validation.