검색 상세

Deep Learning based Lung Ultrasound Pathology Detection for Mobile Point-of-care Devices

초록 (요약문)

During the recent COVID-19 pandemic, the need to diagnose lung pathologies quickly, safely, and accurately increases. Imaging techniques for diagnosing lung pathologies such as CT, Chest X-ray have the disadvantage of irradiating, high cost, and time consuming. But ultrasound is non-invasive, portable, inexpensive, and safe from radiation compared to other imaging techniques. Therefore, ultrasound imaging is suitable for point-of-care (POC) on the bed side patients or the place where the diagnosis environment is poor. However, since ultrasound including lung ultrasound (LUS) depends on the sonographers, detection of lung pathologies using deep learning is conducted recently. Such a deep learning network for LUS pathologies detection has a large parameters and weights for using in mobile POC ultrasound devices. Therefore, in order to work in mobile devices, lightweight techniques are required. In this paper, the deep learning model for detection of lung pathologies in LUS dataset is proposed. The LUS dataset uses publicly available data, and the dataset is classified COVID-19, bacterial pneumonia, and healthy. Backbone network is a VGG-16 that is compared to MobileNet and is lightweight for application on mobile devices using android studio emulator. In addition, a label smoothing method is added to increase model calibration and prevent overconfidence problem. Also, the result is checked using Grad-CAM to evaluate whether it is judged by looking at the correct area.

more