Color radiography in lung nodule detection and characterization: comparison with conventional gray scale radiography

We reviewed images from 272 patients (age > 18 years) who underwent chest CT and DER studies for the evaluation of known or suspected pulmonary nodules between Mar 2005 and Dec 2007. Patients were excluded if there were any abnormal findings suspicious for 1) active inflammation/infection (consolidation, clustered centrilobular nodules, ground-glass opacity, pleural effusion), 2) post-infectious scar or stable tuberculosis (fibrotic bands, traction bronchiectasis, architectural distortion), 3) pulmonary fibrosis (subpleural honeycombing, reticulation), 4) any other structural distortion (severe bronchiectasis, lobar atelectasis, segmentectomy/lobectomy state) on CT scan. Patients with pulmonary nodule that measured < 4 mm or > 40 mm on CT were also excluded. As a result, 65 patients (14 patients with normal scans and 51 patients with pulmonary nodules) were included.

DER was performed using a flat panel detector system (Definium 8000, GE Healthcare, Waukesha, WI). The detector has an image size of 41 × 41 cm with 200 mm pixels. DER were acquired with respiration suspended in deep inspiration using the standard dual shot method (120 kVp and 60 kVp) with a 200-millisecond delay between the high- and low-energy exposures. The dual-energy examination consists of the standard digital posteroanterior gray scale radiograph (GSR) which is the equivalent of an 80-kVp examination as well as the soft tissue image (bone subtracted) and the bone image (soft tissue subtracted). The entrance surface doses for the DER were recorded by dose area product meter and the average of dose was 0.11 ± 0.04 mGy. Various CT scanners were used; 40-detector scanner (Brilliance-40, Philips Medical System, Eindhoven, The Netherlands), 16-detector scanners (LightSpeed 16; LightSpeed QX/I, both GE Healthcare, Waukesha, WI, USA), and an eight-detector scanner (LightSpeed Ultra, GE Healthcare). Scanning was performed from the thoracic inlet to the middle portion of the kidneys in supine position. Images were reconstructed with a high-spatial-frequency algorithm. CT image parameters were as follows: detector collimation, 1.25 or 0.625 mm; field of view, 36 cm; beam pitch, 1.35 or 1.375; gantry speed, 0.5 or 0.6 s/rotation; 120 kVp; 150–200 mA; and reconstruction interval, 1–2.5 mm. Image reformation with a 2.5-mm section thickness for transverse images and a 2.0-mm section thickness for coronal images was performed.

Our proposed DCR visualization method is based on the overlay of color on original DER image. The DER unit produces a material-decomposed data. A color space, derived from compounding existing manipulated data, was added on material-decomposed data according to the attenuation value (blue for high attenuation and red for low attenuation). The color image was then overlaid to unprocessed gray scale DER.

Three independent, board-certified radiologists with 3, 5, and 6 years of experience in chest imaging, respectively, evaluated the radiographs for the presence of pulmonary nodules using the free-response paradigm [13, 14]. A total of 130 images (65 pairs of GSR and DCR from each patient) were reviewed. Images were blinded and randomly sorted, and divided into two subgroups. Each subgroup included one random image within a pair and vice versa. Each observer reviewed subgroup 1 or 2 and vice versa with 1 month interval. The observers were informed about possible presence of multiple nodules per case. Images were evaluated using a standard picture archiving and communication system (Centricity RA 1000; GE Healthcare, Barrington, IL, USA). The observers were able to adjust the window level and the window width at their discretion. They were asked to identify (mark) suspected nodules on the radiographs and to grade the confidence of presence for each marked nodule on a scale from score one to nine, where score nine represented the highest degree of confidence (definitely a nodule) and score one represented the lowest degree of confidence (probably not a nodule). They also graded the confidence of nodule calcification for each marked nodule using scale from score one to nine, where score nine represented the highest degree of confidence (definitely calcified nodule) and score one represented the lowest degree of confidence (probably not calcified nodule).

Before the completion of the detection study, the reference standard was created in consensus by two experienced thoracic radiologists (M. J. C. and I. S.), who had 12 and 4 years of chest CT interpretation experience, respectively. First, chest CT examinations were used for nodule detection. The nodules were identified using axial and coronal reformations in all cases. The size, number and calcification of pulmonary nodules were assessed. A nodule was considered calcified if calcification was visually detected in mediastinal window settings (level + 40, width 400) [15]. Based on the nodule localization on the multidetector CT images, the nodules were marked on the radiographs to verify the true locations. Some of the nodules could not be seen but were marked at the locations where they should have been according to the multidetector CT images.

The accuracy of prediction of nodule calcification was determined by multi-reader, multi-case receiver operating characteristic (MRMCROC) methodology [16, 17]. Dorfman-Berbaum-Metz (DBM) MRMC beta 2.3 (20) was used for calculating the areas under curve (AUC) value.

The observers’ performance for the nodule detection in DCR and GSR images was shown by the average free-response ROC curves in Fig. 3. The reader-averaged values of the JAFROC FOM for the detection of pulmonary nodules were not significantly different between DCR and GSR (p = 0.93). When GSR images were given to the readers, the reader-averaged FOM was 0.811 while the reader-averaged FOM was 0.807 when DCR images were added. FOM values for both interpretation sessions and radiologists are shown in Table 1.

The MRMCROC analysis for the accuracy of prediction of nodule calcification revealed that the DCR improved the AUC values for all readers. Average AUCs were 0.828 with GSR and 0.944 with DCR. The differences were statistically significant between GSR and DCR (p < 0.044). AUC values for both interpretation sessions and radiologists are shown in Table 2.