On-device Subspace Learning Chest X-Ray Screening

The effort to make consumer healthcare devices more reliable, accurate and automated is generating a growing interest in finding ways to integrate AI on the mobile device. Medical imaging is an area that is steadily gaining traction while clinical valid wearables are still emerging. With FDA-approved portable x-ray handheld units been widely used, there is a need to have an accurate image classifier integrated on-device to support faster diagnoses and reduced privacy concern. In this work, a small-scale and cost-effective neural network model is implemented on Raspberry Pi 3B embedded system to demonstrate the scalability of subspace learning. The classification performance of such embedded system is tested on well-known chest x-ray datasets from Open-I and NIH ChestX-ray14. This paper has two main contributions: (1) a subspace weakly supervised learning x-ray image classification model is developed with better than 90% accuracy. (2) The network is implemented on a small and an affordable single-board computer without any accelerators to form a low-power on-device AI x-ray image classifier.