VISIBILITY ENHANCEMENT OF LESION REGIONS IN CHEST X-RAY IMAGES WITH IMAGE FIDELITY PRESERVATION

Abstract: Pneumonia diagnosis via chest X-ray (CXR) imaging remains challenging due to low-contrast lesion regions and inter-reader variability. This paper presents an integrated framework combining intelligent image enhancement with deep learning-based pneumonia classification to improve diagnostic accuracy and lesion visibility.The enhancement module employs a hybrid approach: Contrast Limited Adaptive Histogram Equalization (CLAHE) preprocessing followed by a custom Lesion-Aware Enhancement Network (LAEN) built on U-Net architecture. LAEN selectively amplifies pneumonia-indicative opacity and consolidation patterns while preserving structural integrity, optimized via a multi-component loss function combining perceptual loss, SSIM loss, and pixel-wise reconstruction loss.For classification, a modified DenseNet-121 architecture with attention mechanisms classifies enhanced X-rays into three categories: Normal, Bacterial Pneumonia, and Viral Pneumonia. Grad-CAM visualization generates interpretable attention maps to localize diseased regions for radiologist guidance.The system is evaluated on two publicly available datasets: the Kaggle Chest X-Ray dataset (~5,856 images) and the NIH CXR-14 dataset. The enhancement module achieves SSIM of 0.941 and PSNR of 38.5 dB, demonstrating excellent fidelity preservation. The classification model achieves 97.3% accuracy, 98.1% sensitivity, 95.8% specificity, and AUC-ROC of 0.991, substantially outperforming baseline methods including standard DenseNet-121 (94.2% accuracy) and standalone CNN approaches.Results demonstrate that combining intelligent lesion enhancement with attention-based deep learning creates a robust clinical decision-support tool for pneumonia detection, improving both diagnostic accuracy and interpretability for radiologists.

Keywords: Chest X-ray Imaging, Pneumonia Classification, Image Enhancement, Lesion Visibility, DenseNet-121, U-Net Architecture, Attention Mechanisms, Grad-CAM Visualization, SSIM, PSNR, Deep Learning, Automated Diagnosis.