An Optimized DenseNet-121 Framework for Highly Sensitive and Interpretable Pneumonia Detection from Chest X-Rays
2026 International Conference on Cybersecurity, Digital Forensics, and AI Applications, ICCSDFAI 2026, İstanbul, Türkiye, 25 - 27 Haziran 2026, ss.517-521, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/iccsdfai70505.2026.11647966
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.517-521
- Anahtar Kelimeler: chest x-ray, computer-aided diagnosis, deep learning, densenet-121, pneumonia detection, transfer learning
- Gazi Üniversitesi Adresli: Evet
Özet
The interpretation of chest radiographs (CXR) is fundamental for diagnosing pneumonia, yet it is often hampered by significant inter-observer variability. This study investigates the effectiveness of an optimized DenseNet-121 architecture as a robust clinical decision support system to enhance diagnostic consistency. Utilizing a transfer learning approach on a benchmark dataset of 5,863 pediatric CXR images, we implemented a methodical two-stage fine-tuning protocol combined with comprehensive data augmentation and balanced class weighting. On an independent test set of 624 images, the model achieved a 92.15% accuracy (95% CI: 90.0%-94.3%) and an ROC-AUC of 0.9703 (95% CI: 0.959-0.981), with confidence intervals estimated using the Clopper-Pearson method. Notably, the model demonstrated a clinically vital Sensitivity (Recall) of 97.95%, effectively minimizing the risk of missed pneumonia cases. Furthermore, the system exhibits high computational efficiency with an average inference speed of 85 ms per image on standard hardware, confirming its suitability for real-time clinical integration. The model's decision-making process was qualitatively validated using Grad-CAM, ensuring that classifications were based on relevant pathological features rather than background artifacts. These results confirm that a meticulously tuned DenseNet-121 architecture provides a reliable, explainable, and efficient baseline for automated pneumonia detection in clinical workflows.