A radiology-aware fuzzy deep learning framework with entropy-guided feature selection for robust multi-disease chest X-ray classification across multiple magnifications.
Authors
Affiliations (3)
Affiliations (3)
- Department of Industrial Engineering and Management Systems, Amirkabir University of Technology, No. 350, Hafez Ave, Valiasr Square, Tehran, 1591634311, Iran. [email protected].
- Department of Industrial Engineering and Management Systems, Amirkabir University of Technology, No. 350, Hafez Ave, Valiasr Square, Tehran, 1591634311, Iran.
- Department of Industrial Engineering, Sharif University of Technology, P.O. Box 11155-9414, Azadi Ave., Tehran, 1458889694, Iran.
Abstract
Chest X-ray (CXR) imaging remains the most widely used and cost-effective modality for diagnosing thoracic diseases, yet automated multi-disease interpretation remains challenging due to acquisition variability, subtle overlapping pathologies, and multi-class classification complexity. Existing deep learning approaches often lack uncertainty modeling, interpretability, and robustness across heterogeneous image resolutions, limiting clinical adoption. We propose a fuzzy deep learning framework for multi-disease CXR classification, integrating: (i) radiology-aware augmentation to enhance generalization while preserving diagnostic fidelity; (ii) a grayscale-optimized ResNet-50 backbone with spatial-channel attention for improved feature extraction of subtle abnormalities; (iii) entropy-guided recursive feature elimination (RFE) achieving > 85% dimensionality reduction with minimal information loss; and (iv) a hybrid fuzzy-neural classifier with confidence-weighted defuzzification for explicit uncertainty estimation and reliable handling of borderline cases. The framework was evaluated on four public datasets-COVID-19 Radiography, Tuberculosis CXR, CXR Pneumonia, and CXR COVID-19 Pneumonia-across four magnification levels (×1, ×2, ×5, ×20). At ×20 magnification, accuracies reached 0.9593, 0.9859, 0.9831, and 0.9576, with F1-scores up to 0.9889 and recalls up to 0.9755. Even at ×1, performance remained high (accuracy 0.9401; F1-score 0.9564). Compared with the strongest baseline (CNN), the proposed model improved accuracy by 3.5-8.6%, recall by 2.3-6.7%, and F1-score by 3.8-10.4%. The fuzzy-neural integration stabilized borderline predictions, while confidence-weighted defuzzification reduced false positives. Collectively, radiology-aware augmentation, entropy-guided feature selection, and fuzzy-deep integration enable high accuracy, robustness across resolutions, and interpretable predictions, demonstrating the framework's potential for deployment in heterogeneous clinical and portable imaging environments.