Detection and Classification of Pneumonia from Lung Ultrasound Images

Abstract

The lungs are the primary organs of the respiratory system in humans. Meanwhile, lungs are also vulnerable and are easily damaged by inflammation or impact lesions during the course of our daily lives. Due to the epidemic of COVID-19 pneumonia, the confirmed and suspected cases often grow rapidly beyond the capabilities of medical institutions, rapid and accurate diagnosis for patients have become the first priority. Hence, ultrasound images have started to be adopted in lung diagnosis as they are more convenient, flexible, cheaper, and without ionizing radiation as compared with CT and CXR. This paper aims to use VGG, ResNet and EfficientNet networks to accurately classify Lung Ultrasound images of pneumonia according to different clinical stages based on self-made LUS datasets. The hyperparameters of the three networks were tuned and their performances were carefully compared. Our results indicate that the EfficientNet model outperformed the others, providing the best classification accuracies for 3 and 4 clinical stages of pneumonia are 94.62% and 91.18%, respectively. The best classification accuracy of 8 imagological features of pneumonia is 82.75%. This result is a proof of the promising potential of the LUS device to be used in pneumonia diagnosis and prove the viability of deep learning for LUS classification of pneumonia.