Deep Learning–Driven Chest Radiography: An Integrated Framework for Image Enhancement, Automated Diagnosis, and Fairness Optimization

Abstract

Chest radiography is the most accessible and economical imaging modality for the diagnosis of pulmonary and cardiovascular diseases. However, issues, including low image quality, inter-observer variability, and demographic bias, limit the full therapeutic potential of Artificial Intelligence (AI) driven interpretations. To solve these issues, the current work presents a unified deep learning system that integrates multi-stage image augmentation, automated diagnostics, and fairness optimization. The first stage employs several enhancement techniques, such as adaptive median filtering for noise suppression, Histogram Equalization (HE) using Contrast-Limited Adaptive Histogram Equalization (CLAHE) for contrast refinement, and Drago's tone-mapping for brightness normalization, to improve the perceptual and diagnostic quality of chest X-rays. In the second stage, a deep learning model based on DenseNet is presented for the automated classification of thoracic illnesses. Significant data augmentation and transfer learning optimize it. In order to minimize bias and ensure parity among demographic groupings, Supervised Contrastive Learning (SCL) is integrated in the final stage. Extensive testing on benchmark datasets (NIH ChestX-ray8, CheXpert, and COVIDx) demonstrates notable increases in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), classification accuracy, and fairness metrics when compared to current baselines. The proposed framework improves image quality from 22.37 dB to 30.45 dB PSNR and 0.65 to 0.88 SSIM, while diagnostic performance improves, achieving 92.4% accuracy and 0.937 AUC. Finally, demographic disparity measured using Demographic Parity Difference (DPD) decreases from 0.10 to 0.05.