From April 2008 through February 2009, a total of 38 consecutive patients (36 men, 2 women; mean age, 65.9 years; age range, 46-78 years) who met the diagnostic criteria for COPD (1) underwent two-phase xenon ventilation CT, and they were included in this study. These patients were identical to the initially enrolled patients in a previous two-phase xenon ventilation study (7). Predominant lung diseases included emphysema in 32 patients, tuberculosis-destroyed lung in three patients, bronchiectasis in two patients, and postinfectious bronchiolitis obliterans in one patient.

Patients underwent xenon ventilation using tightly-fitting face masks (King Systems Co., Noblesville, IN, USA) designed to deliver positive pressure ventilation. The xenon gas was a mixture of 30% nonradioactive xenon and 70% oxygen. The patients inhaled the xenon gas for approximately 1 minute during the WI period and 100% oxygen for 2 minutes during the WO period with the use of a xenon gas inhalation system (Zetron V; Anzai Medical, Tokyo, Japan). Patients were instructed to breathe normally during the WI and WO periods. The respiratory rate, oxygen saturation, blood pressure, as well as tidal carbon dioxide and xenon concentrations were closely monitored under the supervision of a chest radiologist. After the CT examination, all of the patients were observed for 30 minutes.

All of the patients were imaged on a CT scanner (Somatom Definition; Siemens Medical Solutions, Forchheim, Germany) during breath holding at full inspiration. Pre-xenon CT was performed using a single tube with a tube voltage of 120 kV, tube current-time product of 150 mAs (reference effective milliampere seconds), and collimation of 64 × 0.6 mm that covered the entire thorax in a caudocranial acquisition. After xenon inhalation, two-phase CT scans were performed using the same coverage. The CT scan during the WI period was taken at approximately 1 minute after the initiation of xenon inhalation when the xenon concentration reached 30%. The CT scan during the WO period was usually obtained at 80-90 seconds when xenon reached zero on the tidal xenon concentration curve after stopping xenon inhalation. Scanning parameters for xenon ventilation CT were as follows: a 512 × 512-pixel matrix, 14 × 1.2-mm collimation, 51 mAs (reference effective milliampere seconds) at 140 kV and 213 mAs (reference effective milliampere seconds) at 80 kV, a pitch of 0.45, and a 0.33-second rotation time. Images were reconstructed with 1.5-mm thickness, 1.2-mm increment, and using soft-tissue reconstruction kernels (B30f for pre-xenon images and D30f for xenon images).

Wash-in and WO images were obtained with Syngo Dual Energy (Siemens Medical Solutions, Forchheim, Germany), which was based on the material decomposition theory (19). To assess data of the whole lung, the minimum value was set at -1024 Hounsfield units (HU) and the maximum value was set at -500 HU. Color-coded maps displaying a progression from blue, sky blue, green, and yellow to red in relation to HU ranging from 1 to 50 were applied to the WI and WO images.

Our proposed CAC system for regional ventilation at two-phase xenon CT consists of lung segmentation, deformable registration, determination of xenon attenuation types, pattern classification of regional ventilation, and generation of a color-coded map (Fig. 1).

To restrict the registration to the lung parenchyma, the left and right lungs were separately segmented using the three-dimensional (3D) region growing method with a threshold of -400 HU for the lung and airway, and then -950 HU for the airway with subsequent subtraction technique (20). To correct positional differences and local deformation of the lungs, the lungs on WO images were globally aligned to the WI image using affine registration based on narrow-band distance propagation, and affine-transformed lungs were nonlinearly deformed by a demon algorithm using a combined gradient force and active cells (21). Then, each pixel in the WI and WO images was assigned to one of the three attenuation types through Equation 1 (below). Lower and upper threshold values were estimated by calculating the mean and standard deviation of xenon attenuation histograms through Equations 2 and 3. Weight w was automatically estimated by calculating the ratio of the normal area to the whole lung. The normal-attenuating lung was defined as the HU of pixels -950 HU or greater in the original CT images (Fig. 2). where, t(x) is the attenuation type of pixel x, A(x) is the xenon attenuation of a pixel x, µ is the mean value of the xenon attenuation histogram, σ is the standard deviation of the xenon attenuation histogram, w is weight, and α, β are coefficients for weight.

The assigned attenuation types were used to classify the corresponding pixel of WI and WO images into the following four ventilation patterns (Fig. 3): pattern A (iso-attenuation or high attenuation in the WI image and iso-attenuation in the WO image), pattern B (iso-attenuation or high attenuation in the WI image and high attenuation in the WO image), pattern C (low attenuation in both the WI and WO images), and pattern D (low attenuation in the WI image and iso-attenuation or high attenuation in the WO image). For the visual pattern recognition of CAC maps; blue, sky blue, red, and yellow were each designated to patterns A, B, C, and D in color-coded maps.

After reviewing all of the structural abnormalities in the pre-xenon dual source CT images, a predominant xenon ventilation pattern for each structural abnormality on WI/WO images was assessed in consensus by two radiologists with experience in two-phase xenon ventilation CT (Fig. 4). For assessment of the predominant xenon ventilation pattern for each structural abnormality, a color overlay of xenon attenuation in the corresponding area on WI/WO images was compared with that in the adjacent normal-attenuating lung showing normal xenon enhancement. The assessed ventilation pattern in consensus served as the reference standard.

Two series of image datasets of structural abnormalities were randomly extracted for optimization and validation. Both the optimization and validation datasets consisted of a total of eighty image datasets including twenty structural abnormalities per pattern. Each image dataset contained axial CT, WI, and WO images, and a CAC map. A single freehand region of interest (ROI) covering the structural abnormality on CT was drawn and pasted into the WI and WO images, and the CAC map, so that the location of the structural abnormality could be easily identified by one of the authors. The average area of ROI was maintained at approximately 0.5 cm² (range, 0.26-0.69 cm²).

The proposed CAC system was optimized by qualitative and quantitative comparisons of CAC candidate models with the reference. A series of CAC candidate models were generated by applying a series of candidate coefficients (α, β) of weight used in an equation for two thresholds (T_low, T_high) with one digit after the decimal point. To narrow down the number of candidate coefficients to less than five, CAC maps for eight image datasets were preliminarily generated using coefficients ranging from 0.1 to 2.0. The predominant ventilation pattern of the ROI on CAC maps was visually compared to the reference by a radiologist. The top five candidate coefficients, for which the CAC map was in good agreement with the reference, were chosen. If CAC maps generated by candidate coefficients did not definitively work out when compared to the reference, those candidate coefficients were discarded even if the number of remnant candidate coefficients was less than five. After narrowing down the number of candidate coefficients, the same review processes were performed for the full optimization set with a series of CAC maps based on those coefficients. The agreement between the CAC map and the reference was calculated on a per-lesion basis for qualitative comparison. Quantitative comparison was performed on a per-pixel basis. The number of pixels on the CAC map within the ROI in agreement with the reference was calculated.

Six radiologists with 20, 9, 8, 7, 6, and 6 years of experience in interpreting CT scans, respectively reviewed the validation image datasets. The validation image datasets were presented in a predetermined randomized order, which was the same for each reader. Each reader was blinded to the clinical information and independently classified the ventilation pattern of ROIs twice sequentially without and with a CAC map. Prior to reviewing the validation image dataset, each reader was provided with instructions including the definition and representative images of each ventilation pattern in WI and WO images, and the CAC map. The interpretation time was not limited.

The CAC model with the highest proportion of agreement for all patterns was considered to have been qualitatively optimized. For quantitative comparison, an area under the receiver operating characteristic curve was assessed with regard to the number of pixels on the CAC map within the ROI in agreement with the reference for each pattern. The CAC system having the largest average area under the curve was considered to have been quantitatively optimized.

The average percentage of agreement between reader pairs on pattern classification per image set was evaluated for validating the optimized CAC system. The number of image sets with complete agreement across all readers was calculated. Interobserver agreement for the pattern classification was determined using multirater Cohen's kappa value (22). A κ value of less than 0.20 indicated poor agreement; a κ value of 0.21-0.40, fair agreement; a κ value of 0.41-0.60, moderate agreement; a κ value of 0.61-0.80, good agreement; and a κ value of more than 0.81, excellent agreement. Interobserver agreement between reader pairs was also calculated. All analyses were compared between assessments without a CAC map and with a CAC map. Statistical analyses were performed using the SPSS package (SPSS 21.0, SPSS Inc., Chicago, IL, USA). A two-sided significance level of 5% was considered to indicate statistical significance.

Computer-aided classification maps were successfully generated in 31 of 38 patients (81.5%). Segmentation of the right and left lungs failed in seven patients. Optimized coefficient of weight for T_high was directly judged to be 1.0 as CAC maps with other candidate coefficients clearly misclassified the ventilation pattern of ROIs. Candidate coefficients of weight for T_low were narrowed from 0.3 to 0.7. Qualitative and quantitative analyses revealed that the optimized coefficient for T_low was 0.5, with the proportion of agreement and average area under the curve of optimized CAC maps being 94% (75/80) and 0.994, respectively (Tables 1, 2) (Figs. 5, 6).

The number of cases in agreement between reader pairs regarding the classification of ventilation patterns ranged from 47 to 68 cases (mean, 55.5; standard deviation, 5.7) without a CAC map, and from 63 to 75 cases (mean, 70.2; standard deviation, 3.3) with a CAC map. The mean percentage of agreement between reader pair results increased from 55.4% without a CAC map to 69.2% with a CAC map. The number of cases reaching complete agreement among all readers doubled from 26 cases (32.5%) to 54 cases (67.5%). Multirater kappa value showed a significant increase from 0.59 (95% confidence interval [CI]: 0.56, 0.62) to 0.82 (95% CI: 0.79, 0.85). Among the patterns, interobserver agreement of pattern B was the highest regardless of the CAC map (without CAC; kappa, 0.72, 95% CI, 0.57-0.87; with CAC; kappa, 0.91, 95% CI, 0.75-1.00) (Table 3). Kappa values among all reader pairs ranged from 0.45 to 0.80 at the initial assessment and from 0.72 to 0.92 with a CAC map (Table 4).