A Lightweight Deep Learning Framework for Cardiomegaly Detection in Chest X-Rays Using CNN and Visual Attention

Abstract

Cardiomegaly, the abnormal growth of the heart, is a very important indicator of heart disease. Although chest X-rays are often used for diagnosis, early detection remains hard because of subtle signs on the X-rays. This paper presents a lightweight and interpretable deep learning framework for the detection of cardiomegaly in chest radiographs. The proposed framework employs a custom convolutional neural network (CNN) integrated with global average pooling to reduce model complexity, while high performance is maintained. Contrast enhancement is applied using CLAHE, and the training pipeline incorporates normalization and data augmentation to improve robustness under diverse imaging conditions. The model uses the AdamW optimizer with decoupled weight decay to enhance generalization. Evaluated on the ChestX-ray14 dataset using five-fold cross-validation, the framework achieves 91.3% accuracy and a 93.2% AUC-ROC with fewer than 1.5 million parameters. Furthermore, Grad-CAM visualizations offer transparency by highlighting decision-related cardiac regions. The proposed method balances accuracy, efficiency, and interpretability, making it suitable for deployment in mobile clinics and similar low-resource settings.