Semantic Segmentation of LUS Retraining a Convolutional Neural Network

Abstract

On the last years respiratory diseases have been on the news almost every day. After the Covid- 19 global pandemic, many studies where done in order to detect and avoid spreading the disease. As in many other different diseases, this studies tend to work with techniques that are available mostly on first world countries (using MRI, TC and X-Ray). Lung Ultrasound (LUS) has been increasingly used to diagnose and monitor different lung diseases. It’s a non-invasive and cheaper technique than the others talked previously. The main issue is that acquiring a LUS scan is easier than analysing and understanding the images and characteristics of each respiratory disease. To try solving this issue, over the last decade, many studies had the main purpose to demonstrate that deep learning techniques can be a possible solution. Despite having very promising results in automatic segmentation, there’s still a necessity of available data to train and test this structures. Through this work, a convolutional neural network based on the architecture U-Net has been retrained, validated and tested in order to segment different patterns from lung ultrasound images obtained from papers published around the last years. At the end of this project, a U-Net network has been obtained with 98.1455 % accuracy detecting a mask with all 3 types of lines that can be found in LUS, A-lines, B-lines and Pleural line. And a U-Net Network with 98.6894 % accuracy detecting a B-line mask. Despite being high accuracy values, as it will be seen on this project, the results aren’t good. From this bad results, at the end of this final degree project, it couldn’t be concluded that the use of deep learning techniques in Semantic Segmentation of LUS isn’t a good combination. In order to discard this theory, more tests and further investigation should be done.