- Abstract: Building the gold standard for training a deep-learning-based model is a time-consuming and very expensive task, especially in medicine. In the case of lung segmentation, blurred boundaries from various types of pulmonary diseases make the problem more difficult. In general, the accurate and robustness model needs a large size of dataset, cost and time effective labeling process is vital. Therefore, we proposed a method to generate the gold standard efficiently based on initial segmentation using a conventional method for lung segmentation with various types of diffuse interstitial lung disease (DILD). The accuracy and robustness of the deep-learning-based segmentation method trained with these data were evaluated and compared with the conventional method using different protocols, including high-resolution computed tomography (HRCT) and volumetric CT.
- Keywords: medical image processing, lung segmentation, u-net, DILD
- Author affiliation: University of Ulsang College of Medicine, Asan Medical Center, Seoul, South Korea