Optimized Pediatric Pneumonia Detection Using Dimensionality Reduction, Class Balancing and Transfer Learning

Pediatric Pneumonia continues to be a major cause of illness and death globally, underscoring the need for reliable and efficient diagnostic tools. This study proposes a comprehensive Chest X-ray-based pediatric pneumonia detection through deep learning techniques. The methodology incorporates advanced preprocessing techniques, including image resizing, normalization, and noise reduction, to enhance image quality and ensure robust input data for model training. The Uniform Manifold Approximation and Projection (UMAP) approach was adopted for reducing the data’s dimensionality. UMAP effectively reduced the feature space while preserving the essential structure and critical diagnostic features of the data. To address class imbalance in the dataset, class weighting strategies were employed, improving model performance on underrepresented classes. Transfer learning was utilized by fine-tuning a pre-trained ResNet architecture, which facilitated efficient feature extraction and enabled state-of-the-art performance with minimal training time. The proposed system achieved a classification accuracy of 99%. This combination of preprocessing, dimensionality reduction, and transfer learning provides an accurate, reliable, and computationally efficient solution for pediatric pneumonia detection, advancing the development of automated diagnostic tools in medical imaging.