AI-based Chest X-Ray Pneumonia Classification with Optimal Threshold and Uncertainty Detection

Pneumonia continues to be one of the principal causes of morbidity and mortality worldwide, especially among the children and the aged. Early detection and correct diagnosis are essential in reducing patient outcome morbidity. The proposed system leverages a ResNet-50V2 Convolutional Neural Network to perform binary classification of chest X-ray images within an AI-based diagnostic framework. The system was trained using a publicly accessible dataset and achieved a test accuracy of 93.43%, precision of 97.12%, and an AUC-ROC score of 0.9959. To enhance decision-making confidence, an optimal threshold of 0.96 was utilized, resulting in a significant drop in false positives. Furthermore, an uncertainty detection mechanism was added to highlight low-confidence predictions, which flagged 17 borderline cases for review by experts. The proposed method exhibits high diagnostic accuracy, interpretability, and practical applicability for implementation in resource-poor clinical environments, providing a valid instrument for early pneumonia screening.