Optimization of matching networks with transfer learning in few-shot pneumonia detection

Abstract Pneumonia remains a leading cause of death among children under five years of age, with approximately 1.6 million fatalities annually, especially in low-resource settings where diagnostic capacity is limited. Early detection plays a critical role in reducing pneumonia-related mortality; however, most traditional diagnostic methods depend heavily on trained personnel and medical infrastructure, which are often lacking in underserved areas. While machine learning has emerged as a promising approach for automating diagnosis, existing models typically rely on large volumes of annotated data, making them less applicable in real-world clinical scenarios where labeled data is scarce. To address this gap, we propose a hybrid model that combines transfer learning with few-shot learning, using MobileNetV3 as a lightweight feature extractor and Matching Networks for classification based on limited labeled examples. This integration is designed to improve both data efficiency and diagnostic accuracy. The key novelty of this work lies in the fusion of pre-trained deep representations with a metric-learning framework to enable effective pneumonia detection under one-shot or few-shot conditions. The model was evaluated on a publicly available, balanced chest X-ray dataset involving three classes: COVID-19, pneumonia, and normal. Experimental results demonstrated high classification performance, with an accuracy of 93.21%, precision of 93.34%, recall of 93.20%, and F1-score of 93.19%. The proposed model also achieved AUCs of 1.00 for COVID and 0.98 for both pneumonia and normal cases, outperforming traditional machine learning baselines such as Random Forest and Support Vector Machines. An ablation study was conducted to isolate the contributions of each component, confirming that the performance gains are attributable to the integration of MobileNetV3 and Matching Networks. Although external validation was not included in this study, the architecture was designed with adaptability in mind, making it a promising candidate for deployment in real-world, resource-constrained healthcare environments and future expansion to new disease categories.