CNN Models for Pediatric Pneumonia Detection in High-Risk Regions
This repository contains our work on developing convolutional neural network (CNN) models for detecting pneumonia in children under five years old in Nigeria and Kenya. The project was designed to address the scarcity of radiologists and high rates of pneumonia-related mortality in these regions. We built and compared several CNN models to identify the one with the highest accuracy and the lowest rate of false negatives.
Key Skills and Tools Used:
Data Analysis and Preprocessing: Managed large datasets of pediatric X-ray images, preparing them for model training.
Convolutional Neural Networks: Developed and optimized multiple CNN models, fine-tuning hyperparameters such as layer depths, kernel sizes, dropout rates, and learning rates.
Performance Evaluation: Utilized various metrics including accuracy, area under the curve (AUC), recall, and precision to evaluate model performance. Special focus on minimizing false negatives to enhance medical safety.
Cost-Benefit Analysis: Conducted comprehensive financial analyses to assess the economic impact of deploying each model, considering the costs associated with false positives and negatives.
Tools Used: Python, TensorFlow, Keras, and several data augmentation techniques to enhance model robustness.
Each folder in this repository includes scripts, model configurations, training logs, and performance evaluations, alongside a presentation of our findings aimed at potential stakeholders like the Gates Foundation.