On segmentation of lung parenchyma in quantitative computed tomography of the lung

Our purpose in this study was to investigate the influence of segmentation threshold and number of erosions on parameters used in quantitative compute d tomography (CT) of the lung (erosions are shrink operations on the segmen ted area). Parameters assessed were mean lung density, area of the segmente d lung, two percentiles, and the pixel index, which is the relative area of the histogram below -905 Hounsfield Units (HU). We analyzed images of ten emphysematous and ten nonemphysematous patients, that had been scanned at c arina level in inspiration and expiration, using sections of 1, 2, 3, 5, an d 10 mm in combination with a standard, a smooth, and an ultrasmooth recons truction kernel. The lungs were segmented using pixel tracing at thresholds of -200, -400, and -600 HU with 0-4 erosions, followed by histogram analys is. The area of the segmented lungs decreased with 0.9%-3.2% per 100 HU dec rease in threshold and with 2.2%-3.1% per erosion, dependent on patient gro up and respiratory status. Estimated mean lung density changed up to 30% by changing the threshold and the number of erosions. The pixel index and the 10th percentile depended only slightly on threshold and number of erosions , but the 90th percentile showed a strong dependence of up to 40%. It is co ncluded that the segmentation protocol can have a large impact on densitome tric parameters and that standardization is mandatory for obtaining compara ble results. Ideally a threshold equal to the average of the densities of l ung and soft tissue should be used, but -400 HU will do in a Limited but co mmon density range (-910 to -790 HU). For densitometry two erosions are rec ommended, for volumetry zero erosions should be used. (C) 2998 American Ass ociation of Physicists in Medicine. [S0094-2405(98)01912-9].