<![CDATA[Pneumonia continues to be a leading cause of childhood mortality worldwide, with the greatest impact in resource limited regions where access to radiologists is limited. Deep learning has emerged as a promising tool for automated screening, yet conventional object detection models demand high computational resources, restricting their use on low cost edge devices common in rural healthcare settings. To address this challenge, we developed a lightweight pneumonia detection framework designed for realtime inference on standard hardware. Our approach builds on the YOLOv5lu architecture, which incorporates an anchor free decoupled head, and was trained on a balanced dataset of 14,863 chest X-rays from the RSNA Pneumonia Detection Challenge.To enable deployment on edge devices, we applied a sequential compression pipeline. First, 30% of convolutional filters were removed through structured pruning, followed by post training quantization to INT8. These optimizations reduced the model size by nearly half (from 101 MB to 51.72 MB) and improved inference speed to 6.24 ms per image, equivalent to more than 160 frames per second on a standard CPU. Importantly, the quantized model preserved diagnostic performance, achieving a mean Average Precision (mAP@0.4–0.75) of 0.266 compared to the baseline score of 0.287. These findings confirm the practical feasibility of deploying advanced deep learning models in limited resource regions. By effectively balancing efficiency and accuracy, this framework offers a scalable solution for early pneumonia screening and establishes a foundation for extending detection to other diseases, including Tuberculosis and COVID-19.]]>
Copyrights © 2026
