COVID-19 Pulmonary Infiltrate Manifestation Segmentation Leveraging Auxiliary Tasks

Abstract

In many recent years, COVID-19 has become one of the biggest concerns all around the world. The patients who suffered from COVID-19 still have post-COVID symptoms and COVID-19 pulmonary infiltrate manifestation in the lungs. For diagnosis of the effect of COVID-19 lung infection, Chest X-ray screening is an efficient method to localize the damaged area. With the growth of deep learning, we aim to use a convolutional neural network with triple-task learning for conducting COVID-19 classification, lung segmentation, and infected area segmentation. When discovering COVID-19 patterns through X-ray screening, radiologists will mostly delineate the damaged area. Following the guidelines of doctors, this paper focuses on leveraging dual auxiliary tasks for enhancing the robustness of COVID-19 pulmonary infiltrate manifestation, specifically through area segmentation. Our work also examined multiple settings of the task weight to find the most optimal one through the mean F1-Score metric. We achieved the best result with 93.37% of the mean F1-Score through three tasks with 87.57%, 78.64%, and 85.90% on Infection Segmentation F1-score, IoU, and Dice metrics, respectively.