Performance evaluation of automated lung segmentation for High Resolution Computed Tomography (HRCT) thorax images

Abstract

Segmentation is the preliminary steps in developing a computer aided diagnosis (CAD) system. Determining the quality of segmentation will be able to minimize errors in the CAD system. Ninety-six High Resolution Computed Tomography (HRCT) thorax images in DICOM format were obtained from the Department of Diag-nostic imaging of Kuala Lumpur, Malaysia consisting of Interstitial Lung Disease (ILD) cases, other lung related diseases (Non-ILD) cases and healthy (normal) cases. The study utilizes a framework of having five pre-determined levels of HRCT Thorax image slices based on lung anatomy selected by the radiologist. For the purpose of this study only Level 1 is used. The images were automatically segmented and compared with ground truth which the manual tracings done by a radiologist. Polyline distance metric and Euclidean distance were used to determine the quality of segmentation. The quality of the segmentation deteriorates when the polyline and Euclidean distance increases. Generally values above five pixels would yield poor segmentation quality. Using the Bland-Altman method and plot, it can be seen the level of agreement between polyline and Euclidean distance metrics as well as the quality of segmentation.