From an anonymised image database of patients with emphysema collected in the setting of clinical trials, 1495 CT studies were identified with severe homogeneous emphysema (defined as having an emphysema score at −910 Hounsfield Units (HU) of at least 25% in each lobe, and a difference in emphysema score for the different lobes of less than 25%). A subset of 116 patient CT studies were selected if they met the criteria of having near-isotropic voxel size and slice spacing (axial resolution between 0.6 and 0.8 mm and slice spacing of at most 1 mm). 20 cases were used for developing the computer system and another 96 were read for testing. Table 1 provides descriptive baseline statistics for the 96 CT studies used in this paper. The patient characteristics and emphysema severity were the same for the development set and the reading set. Image acquisition protocols were standardised for specific CT systems at each site to maintain a consistent image quality across the different sites and systems. Eleven different CT systems from four manufacturers were used to obtain the 96 CT studies. Images were obtained at maximum breath hold to total lung capacity after careful breath hold coaching (120 KV, 80 to 110 mAs; reference mAs for regular to large patients). Informed consent and IRB approval for participation in the trials and for the analysis of the images by our imaging laboratory was obtained for each participating centre.
Visual fissural completeness scores
Before analysis, visual assessment of fissural completeness was performed by carefully inspecting the axial, sagittal and coronal views, without any reference to the computer output. For visual scoring a trichotomous scale was used: complete, partial, and absent. The fissure was scored as complete when it was thought to be visually complete and there was no evidence of bronchovascular structures crossing from adjacent lobes. The score absent was used when no appreciable fissure could be seen. In all other cases, the fissures were scored as partial. For all 96 CT studies, two radiologists with four and six years’ experience independently scored the completeness of all lobar fissures. For cases where the two radiologists disagreed, a consensus reading was performed with a chest radiologist who had 14 years’ experience. The two radiologists disagreed in 36 out of 288 (12.5%) fissures, out of which 12 were right major fissures, 13 right minor fissures and 11 left major fissures. The kappa between the readers were 0.75, 0.67 and 0.74, for the right major, right minor and left major fissures respectively
Automatic fissural completeness quantification
The lungs, fissures and lobes were automatically segmented using previously proposed and evaluated methods [8–10]. The lung segmentation was performed using a hybrid method specifically designed to overcome segmentation errors typical in CT studies exhibiting pathological features . The lung segmentation begins with a conventional approach to segmenting the lungs using region-growing and morphology. Segmentation failure was automatically detected based on statistical deviation from a range of volume and shape measurements. To CT studies with segmentation errors, an advanced, multi-atlas-based algorithm using non-rigid registration was applied.
Fissure detection was based on pattern classification using image derivatives, image gradient and Hessian features. The pattern classification was trained based on manually identified examples of fissure and non-fissure voxels, as described in van Rikxoort et al. . For this study, the fissure detection method was trained using manually segmented fissures in the 20 CT studies set aside for development.
Given the fissures detected, the lobe borders were automatically segmented by fitting an atlas with complete fissures (see van Rikxoort et al. ). As a result, the lobe borders are derived, even when the fissures are incomplete. The completeness of a pulmonary fissure is determined from the segmentations of the lobe and fissure by assigning each point on the detected fissure to the closest point on the lobe border; voxels on the lobe border not assigned to a fissure voxel are considered to be non-fissure and therefore incomplete. The percentage of lobar border voxels identified as complete quantifies the completeness of the fissure. The result of the automatic fissural completeness quantification is visually displayed as a colour-coded lobar boundary. The procedure of fissural completeness calculation is illustrated in Fig. 1.
Descriptive statistics on fissural completeness for both the visual scoring and the automatic quantification are provided. Box-plots are provided for the automatic score stratified by the different visual categories for the consensus read. Rank sum tests were performed to test whether the automatic quantification is able to distinguish complete and partial fissures. In case three categories were scored by the visual read (complete, partial, absent), the Kruskal-Wallis test was applied. To test the sensitivity of the automatic quantification, Receiver Operating Characteristic curves (ROC) were constructed for the automatic quantification compared with each visual read, defining a complete fissure as a positive test. The level of statistical significance was set to 0.05. This was implemented using Stata V.10.0 (College Station, TX, USA).