Detection of Pneumonia from Chest X-Ray Images using Deep Learning and Ensemble Approaches with Grad-CAM Explainability
1st International Turkish World Artificial Intelligence & Digital Transformation Congress -TURKWAI 2026, Bishkek, Kırgızistan, 18 - 20 Mayıs 2026, ss.252-258, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: Bishkek
- Basıldığı Ülke: Kırgızistan
- Sayfa Sayıları: ss.252-258
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Ankara Üniversitesi Adresli: Evet
Özet
This study introduces RDE-PneuNet(ResNet, DenseNet, EfficientNet Pneumonia Network), a robust, balanced, and explainable deep learning framework for detecting pneumonia from chest X-ray images. The proposed architecture integrates three structurally diverse backbone models—ResNet50, DenseNet121, and EfficientNetB0 utilizing an optimized weighted soft-voting ensemble strategy to maximize clinical reliability. To mitigate inherit class imbalances and prevent overfitting on the Chest X-ray dataset, transfer learning, fine-tuning, data augmentation, and strategic class weighting were implemented. Experimental results demonstrate that the ensemble configuration outperforms standalone models, achieving a highly competitive 93% accuracy alongside balanced precision, recall, and F1-score metrics. Crucially, the system overcomes the common clinical trade-off of prioritizing sensitivity at the severe expense of specificity, yielding a well-balanced 92% specificity and an AUC of 0.975. To ensure qualitative transparency and overcome black-box limitations, Grad-CAM visual explanations were integrated, confirming that the network consistently attends to anatomically meaningful pulmonary regions rather than background artifacts. Ultimately, RDE-PneuNet provides a reproducible, computationally efficient, and highly transparent decision-support tool suitable for real-world clinical workflows.