Convolutional Neural Networks for Pneumonia Detection: A Preliminary Investigation and Development Framework

Abstract

Pneumonia remains the leading infectious cause of death globally, particularly in low-resource settings where delayed diagnosis and limited radiologist availability exacerbate mortality. Existing AI-based radiograph interpretation systems often demand high computational resources and lack robustness across imaging projections. This study presents a proof-of-concept convolutional neural network diagnostic tool optimised for projection-invariant pneumonia detection under constrained conditions. Using a DenseNet-121 backbone trained on 2,000 curated images from the MIMIC-CXR dataset, our model achieved an AUC of 0.7310 and F1-score of 0.6207, with 66.36% validation accuracy. The model’s performance was consistent across posteroanterior and anteroposterior projections, though lateral view evaluation is pending. Preprocessing included CLAHE and DICOM standardisation, while augmentation improved generalisation. Though early-stage, this work shows the potential of lightweight, projection-tolerant CNNs in offline diagnosis pipelines. Future work will validate deployment feasibility on edge devices and expand evaluation across diverse patient demographics.