Effective Pneumonia Detection From Chest X-ray Images on an Embedded Platform Using Deep Transfer Learning

Abstract

Pneumonia is a significant health problem worldwide, posing serious threats particularly to children and the elderly. Early diagnosis and accurate treatment play a critical role in improving the quality of life and survival rates of patients. Medical imaging techniques are indispensable tools in diagnosing pneumonia, providing non-invasive methods for visualizing the internal organs of patients. In recent years, artificial intelligence and especially deep learning methods have made significant advancements in medical image analysis. Transfer learning extracts meaningful features from medical images using large datasets and powerful processors, classifying these images with high accuracy. In this study, the pre-trained MobileNetV2 model is used to classify pneumonia and normal chest X-ray images on a personal computer using transfer learning. The dataset includes X-ray images provided by the Indian Institute of Science, PES University, MS Ramaiah Institute of Technology, and Concordia University (doi: 10.17632/9xkhgts2s6.3). The model's performance is evaluated with metrics such as accuracy, precision, sensitivity, and specificity, demonstrating high accuracy with low computational costs. The trained model is then deployed on a low-cost, high-performance, and portable embedded system, the Nvidia Jetson Nano, and its performance on this platform is examined. On the personal computer, the model's prediction time is 36.57 milliseconds, with accuracy at 98.1%, sensitivity at 97.8%, specificity at 98.5%, precision at 98.9%, and F1-score at 98.4%. On the Jetson Nano platform, the prediction time is 111.2 milliseconds, with accuracy at 97.8%, sensitivity at 96.8%, specificity at 99.4%, precision at 99.6%, and F1-score at 98.2%. The findings of the study indicate that the MobileNetV2 model can effectively operate on the Nvidia Jetson Nano embedded platform and provide high performance. This study offers an innovative and effective solution for pneumonia diagnosis by integrating deep learning and embedded systems. Future research may further enhance classification performance and significantly improve the quality and accessibility of healthcare services by using larger datasets and various deep learning models.