Optimized Mobile SE-CNN for Pneumonia Detection Using Chest X-Ray Images

Thu, 02 Jul 2026 00:00:00 +0000

Pneumonia remains one of the leading causes of morbidity and mortality worldwide, particularly in regions with limited access to diagnostic facilities. Chest X-ray (CXR) imaging is widely used for pneumonia detection; however, manual interpretation can be time-consuming and prone to variability among radiologists. This study proposes an optimized Mobile SE-CNN architecture that integrates Mobile Inverted Bottleneck Convolution (MBConv) and Squeeze-and-Excitation (SE) mechanisms to improve feature representation while maintaining computational efficiency. The model was trained and evaluated using the COVID-19 Radiography Database consisting of four classes: COVID-19, Lung Opacity, Viral Pneumonia, and Normal. Experimental results show that the proposed model achieved a test accuracy of 93.58% with a macro-average F1-score of 94.17%. Compared with the baseline CNN model, the proposed architecture improves classification accuracy by 3.69% while reducing the number of parameters by approximately 99.62%, using only 40,606 parameters and a total size of approximately 0.15 MB. These results demonstrate that the proposed Mobile SE-CNN achieves an effective balance between diagnostic performance and computational efficiency, making it suitable for deployment in mobile or embedded medical diagnostic systems.

Safety-Oriented Air Quality Index Classification for Imbalanced Data Using Optimized Boosting Models with Optuna and Oversampling

Made Yudi Dwipayana, Gede Angga Pradipta, Dandy Pramana Hostiadi — Tue, 28 Jul 2026 00:00:00 +0000

Air Quality Index (AQI) classification is essential for communicating environmental health risks. However, hazardous air conditions occur far less frequently than normal conditions, challenging conventional classification models. This study investigates multi-class AQI classification using the "Global Air Quality 2025" dataset, comprising 52,704 observations with an extreme class imbalance ratio of approximately 1:173. Under such conditions, conventional accuracy metrics often mask systemic failures in detecting critical minority classes. To address potential data leakage present in previous approaches, this research implements a rigorous cross-validation architecture combined with an independent 20% hold-out test set. The methodology employs an Ablation Study to systematically isolate the impacts of Optuna hyperparameter tuning guided by Macro F1-Score and oversampling techniques (SMOTE and ADASYN). The results demonstrate that the proposed Hybrid-SMOTE LightGBM configuration successfully balances hazard detection sensitivity with global stability. On the unseen hold-out set, the optimal model achieved a Macro F1-Score of 0.8079, an accuracy of 92.80%, and a ROC-AUC of 0.9847. Crucially, the model delivered a 65.12% recall for the critical Unhealthy minority class, a nearly 40% improvement over the baseline. Error profile analysis confirmed the model's safety-oriented robustness, as 97.6% of peak hazardous events were either accurately classified or safely constrained to the adjacent warning category, minimizing catastrophic misclassifications. These findings prove that reliable detection of environmental hazards requires safety-oriented per-class evaluation and strict validation frameworks, as reliance on aggregate global metrics leads to dangerously misleading performance assessments.

TIERS Information Technology Journal

Optimized Mobile SE-CNN for Pneumonia Detection Using Chest X-Ray Images

Safety-Oriented Air Quality Index Classification for Imbalanced Data Using Optimized Boosting Models with Optuna and Oversampling