Hybrid Architectures with Attention Mechanisms and Evolutionary Hyperparameter Optimization for Robust Pneumonia Detection

Quispe M.F.M.; Gonzales L.A.; Mollocondo C.I.M.; Vargas J.C.J. · Proceedings - IEEE CHILEAN Conference on Electrical, Electronics Engineering, Information and Communication Technologies, ChileCon (2025) · DOI: 10.1109/chilecon66915.2025.11476598

Resumen

For medical artificial intelligence, automated X-ray-based pneumonia diagnosis is a major challenge. This paper suggests a thorough and reliable method that combines CNNs with hybrid CNN-Transformer architectures and visual attention mechanisms (Grad-CAM, CBAM). Stratified cross-validation is used to optimize hyperparameters using evolutionary methods (both Bayesian and genetic). Numerous open Kaggle databases are used to evaluate the method, guaranteeing patient demographic variety across imaging modalities and clinical contexts. AUC, sensitivity, and accuracy gains over baseline models are statistically significant, according to the results. The suggested framework aids in the creation of clinical decision support systems that are both explanatory and adaptive, making them appropriate for practical application in a variety of healthcare environments.

Instituciones: Universidad Nacional del Altiplano