A Novel 3D Method Based on Region-Growing and Morphology for Lung Segmentation

Abstract

The purpose of this study is to help in the early detection of lung abnormalities by accurately segmenting the volume of interest from the CT scans. For this, we suggest a novel method that loads the volume by interpolating additional images to increase the resolution, thus improving the efficiency of the region-growing application, without any pre-processing, for segmenting voxels in the range of -1000 and -500 HU, and then applying a combination of chain code and region-growing to repair the lacunae due to blood vessels, trachea branches, and/or lesions that have intensities outside the range of interest. The validation of the segmentation result, using metrics, shows how close our method is to the ground truth, with an accuracy of 99.99%, a dice coefficient of 98.99%, an IoU of 98.02%, a recall of 98.44%, a precision of 99.56%, and an F1-score of 98.99%. Compared to UNET, UNET++, and 3D-UNET. Our method presents better results except for recall, which is higher than ours with a minor difference of 0.09–0.85%.