Multi-Stage Adaptive Hyperparameter Optimization Framework for Pneumonia Detection

Abstract

Pneumonia is a leading cause of childhood mortality worldwide, and automated detection from chest X-ray images using deep learning holds strong clinical promise. However, model performance is critically sensitive to hyperparameter selection, and existing single-strategy optimization methods are insufficient for complex medical imaging tasks. This paper proposes the Multi-Stage Adaptive Hyperparameter Optimization (MASAHO) framework, which sequentially employs random search, Bayesian optimization, genetic algorithms, and reinforcement learning with principled stage-transition criteria. The framework incorporates an adaptive model selection mechanism based on quantified dataset characteristics and a confidence-weighted clinical loss function that explicitly penalizes false negatives. Experiments on three publicly available datasets—CXR14 [21], RSNA [22], and COVID-19/Pneumonia [23]—demonstrate 97.5% accuracy, 96.2% sensitivity, and 98.8% specificity o n C XR14, outperforming single-strategy baselines, AutoML, Neural Architecture Search, and state-of-the-art domain methods while reducing optimization time by 83% compared to grid search. All reported improvements are statistically confirmed (p < 0.01, Bonferroni-corrected paired t-tests).