Semantic Segmentation of Lung Tissues in HRCT Images by Means of a U-Net Convolutional Network

In this work, a strategy for the segmentation of Interstitial Lung Diseases (ILD) from a High-Resolution Computed Tomography (HRCT) volumetric image by means of a U-Net convolutional network is presented. In particular, the delimitation of Idiopathic Pulmonary Fibrosis (IPF), Pulmonary Emphysema (PE) and Healthy Lung Tissue (HLT) were carried out. The key idea of the proposed strategy is that the U-Net training was performed using three slices from the HRCT. The results of the segmentation of all tissues at once in the whole volumetric image were: 93.6%25 of accuracy; 80%25 for the intersection over union (IoU) metric; and above 0.9 for HLT, around 0.8 for IPF and PE in terms of the DICE similarity coefficient. These results suggest that the proposed approach could be used to properly segment different lung tissues at the same time, using only partial data from the HRCT image instead of a large dataset for the training of the U-Net. © 2020, Springer Nature Switzerland AG.

Convolutional network; Interstitial lung diseases; Semantic segmentation; U-Net Biological organs; Biomedical engineering; Biophysics; Computerized tomography; Convolution; Histology; Large dataset; Semantic Web; Semantics; Tissue; Convolutional networks; High-resolution computed tomography; Idiopathic pulmonary fibrosis; Interstitial lung disease; Pulmonary emphysema; Semantic segmentation; Similarity coefficients; Volumetric images; Image segmentation

VIVO