Implementation and Optimization of Saliency Mapping Algorithms in Convolutional Neural Networks (CNN) to Enhance Transparency in Pneumonia Diagnosis

Abstract

This study aims to develop a transparent and reliable artificial intelligence model for pneumonia diagnosis using chest X-ray images by implementing and optimizing Convolutional Neural Networks (CNN) with Saliency Mapping. The research employed a combination of advanced optimization techniques, including aggressive data augmentation, class weight balancing, L2 regularization, dropout, batch normalization, and adaptive learning rate scheduling to address overfitting challenges. A functional prototype was then deployed in a Streamlit-based application to provide an interactive diagnostic tool. The evaluation results demonstrated that the model achieved strong performance, with high training accuracy and competitive testing accuracy, while visualization through Saliency Mapping provided meaningful interpretability by highlighting critical lung regions, particularly the mid-to-lower lung fields and hilar area. This interpretability ensured that the system not only delivered accurate predictions but also supported clinical reasoning by aligning with radiological characteristics of early-stage pneumonia and bronchopneumonia. The integration into a user-friendly application illustrates the potential for practical adoption in healthcare settings, especially in regions with limited access to radiologists. Overall, the study demonstrates that combining CNN-based classification with explainable AI techniques can bridge the gap between advanced machine learning and clinical applicability, offering a strategic pathway to improve pneumonia diagnosis and patient outcomes.