Pneumothorax Detection in Chest X-rays Using Transfer Learning and AI-Based Image Enhancement

Abstract

Pneumothorax is a clinically significant condition requiring timely diagnosis, commonly assessed using chest X-ray imaging. Manual interpretation of radiographs is challenging due to subtle visual patterns and inter-observer variability, motivating the use of automated machine learning approaches. While deep learning models have been widely applied to pneumothorax detection, limited work has systematically compared traditional machine learning and deep learning models under consistent experimental conditions, particularly with respect to the role of image enhancement. This study presents a controlled comparative evaluation of multiple machine learning and deep learning models using the NIH ChestX-ray14 dataset. A strict patient-wise data splitting strategy was employed to prevent data leakage. Image enhancement techniques, including contrast-limited adaptive histogram equalization and wavelet-based denoising, were applied and evaluated across Random Forest, Support Vector Machine, and XGBoost classifiers using ResNet50-based feature extraction, as well as end-to-end deep learning models including ResNet50 and DenseNet121. Model performance was assessed using macro-averaged evaluation metrics to address class imbalance. Results indicate that image enhancement does not uniformly improve performance and its impact varies across model architectures. KeyWords: Pneumothorax detection, Chest X-ray, Image Enhancement, Random Forest, SVM, XGBoost, DenseNet121, Resnet50, CLAHE, Wavelet Denoising