EvansKonadu/AI-Powered-Pneumonia-Detection-in-Low-Resource-Settings

Deep learning models for pneumonia detection in chest X-rays, optimised for deployment in resource-constrained clinical environments. # **AI-Powered-Pneumonia-Detection-in-Low-Resource-Settings** **Aim:** Develop a **computationally efficient** and **interpretable** deep learning model for pneumonia detection in chest X-rays, optimised for deployment in **low-resource clinical settings**. ## **Abstract** Pneumonia remains the leading infectious cause of mortality among children under five, claiming over 700,000 lives annually. This research addresses the critical gap between advanced deep learning pneumonia detection capabilities and practical deployment limitations in low-resource clinical environments. **Image source:** *UNICEF, "A child dies of Pneumonia every 43 seconds globally," Instagram, Nov. 17, 2024. [Online]. Available: instagram.com [Accessed September 9, 2025].* -------- ---------- #### Research Design and Overview *Figure1: Methodology Flowchart* ### **Key Achievements:** - 97.12% accuracy on paediatric data (Xception) - 86.10% accuracy on adult data (EfficientNetB4) - 91% model size reduction through quantisation - Sub-second inference on edge devices (Raspberry Pi) - Real-time explainability via optimised Grad-CAM ------ ### Datasets ##### NIH ChestX-ray14 Dataset - Size: 112,120 chest X-ray images from 30,805 patients - Demographics: Adult population (mixed ages) - Class Distribution: Severe imbalance (60,412 Normal vs 1,431 Pneumonia) - Challenge: 42:1 class ratio requiring careful balancing strategies *Figure 2: Distribution of 14 Labels in NIH ChestX-ray14 Dataset - Bar chart with annotated counts and pie chart with percentage annotations showing all 14 labels* *Figure 3: Class Distribution for Pneumonia Detection Task - Bar chart and pie chart highlighting the imbalance (60,412 Normal vs. 1,431 Pneumonia)* ##### Mendeley Paediatric Dataset (Kermany et al.) - Size: 5,856 chest X-ray images - Demographics: Children aged 1-5 years - Class Distribution: 4,672 Pneumonia vs 1,184 Normal cases - Advantage: Higher pneumonia prevalence (79 …