Quantitative CT Analysis of IPF

The CT images used for this study were those typically termed “high resolution”: non-volumetric 1 mm slices with 10 mm spacing with a sharp kernel image reconstruction obtained without the administration of intravenous contrast. Using a previously described automated technique, the lung was segmented from the surrounding tissue [21 (link)]. The axial images were then visually inspected and manually edited as needed to correct inaccurate segmentations.
For the densitometric evaluation, the histogram of distribution of the density of each voxel within the lung was plotted as shown in Fig. 1, and the skewness, kurtosis, and the mean of that distribution (mean lung density, MLD) were measured. In addition, the percentage of the total volume of tissue that had a density between -250 Hounsfield units (HU) and -600 HU was recorded as the percent high attenuation area (HAA%) [22 (link)].Fig. 1

a Representative images from subjects with less severe (Patient 1) and more severe (Patient 2) visual evidence of IPF. b Histograms of distribution of the number of voxels on the y axis for each tissue density in Hounsfield Units on the x axis. c Summary statistics and selected pulmonary function test parameters for each subject. Abbreviations: mean lung density (MLD)

Full details regarding the local histogram based objective quantification of the volume of radiographic feature subtypes are available in the (Additional file 1). Briefly, we used both the properties of the local tissue and the distance from the pleural surface to determine a radiographic feature subtype for every portion of the lung [18 (link), 19 , 23 ]. First, in order to train the subtype identification tool, a single expert placed a total of 3357 fiducials, in 30 randomly selected subjects, on the following radiographic subtypes: normal, interstitial (reticular, centrilobular nodule, linear scar, nodular, subpleural line, ground glass and honeycombing), and emphysematous (centrilobular and panlobular) as shown in Fig. 2.1 This was done to build a library of points to be used as tissue classifiers. Regions of interest consisting of 30 by 30 in-plane voxels were built around these training points, and both the local histogram information and distance from the pleural surface were used to create a tissue classification vector for each region [18 (link), 19 ]. After the training process was completed, the feature vectors of all of the 30 by 30 in-plane voxel regions within the lungs of each of the subjects were classified into tissue subtypes based on their similarity to the training data as shown in Fig. 3.Fig. 2

a Sample slice for CT scan of a subject. b The same sample slice from a subject CT scan showing placement of fiducials for the training of the local histogram based objective method. Abbreviations: ground glass (GG), honeycombing (Hon), reticular (Ret), computed tomography (CT)

Fig. 3

a Representative CT images from subjects with less severe IPF (patient 1) and more severe IPF (patient 2). b Overlay of categorization of lung parenchyma into radiographic subtypes using the local histogram analysis and distance based analysis for each subject. c Legend for radiographic subtypes

Fig. 4

Kaplan Meier survival curves for transplant free survival for the densitometric CT measures. a HAA%, b MLD, c skewness, d kurtosis. Abbreviations: percentage of high attenuation area (HAA%), mean lung density (MLD)

The total percentage of all of the interstitial features (objective interstitial score, interstitial%) was then determined by combining the reticular, centrilobular nodule, linear scar, nodular, subpleural line, ground glass and honeycombing subtype volumes and dividing by the total volume of all tissue types (normal, interstitial and emphysematous). The percentage of interstitial disease made up of by honeycombing (honeycombing%) was determined by dividing the volume of the honeycombing subtype by the total volume of all of the interstitial subtypes. Due to the exploratory nature and small size of this study, subjects used in the training set were not excluded from the final analysis.