Detecting the Absence of Lung Sliding in Ultrasound Videos Using 3D Convolutional Neural Networks

Abstract

During recent years, deep learning models proved to be very effective in multiple tasks in medicine and frequently outperformed other traditional machine learning methods. Especially in tasks where the processing of image or video data is necessary, deep networks present a very popular tool. Especially in medicine, image and video data present a frequent source of data. The work presented in this paper focuses on the use of deep learning models to detect specific phenomena from lung ultrasonography data. We focused on the detection of lung sliding, which can be observed from such data and used by clinicians in the diagnostic process of evaluating of patient's health condition. Previous research in this area mostly focused on processing a sequence of static images obtained from ultrasonography. In our work, we focused on the development of deep learning models able to process short video sequences. We used different architectures of a Resnet model and experimentally evaluated them on a real-world dataset. Then, we compared the results of best-performing architectures with the more traditional approach based on static image processing.