This study is a side study of the Dutch-Belgian multi-centre randomised low-dose CT lung cancer screening trial NELSON . The NELSON study has been approved by the Minister of Health of the Netherlands and the ethics committees of all four participating hospitals. Informed consent was obtained from all participants. The original approval and informed consent for the screening study included the ability to use data for future research, including the current side study.
This side study concerned an analysis of category II and III solid nodules (15–500 mm3, effective diameter 3.1–9.8 mm) according to the NELSON classification . Participants screened in the Groningen NELSON study centre between September 2007 and February 2008 (n = 562, 556 males, 6 females), were consecutively enrolled in this side study when they were diagnosed with non-calcified, solid nodules with a volume smaller than 500 mm3. For every enrolled participant, all qualified nodules were included. Finally, 79 male and 3 female participants with 200 nodules were included in the analysis, having a mean ± SD age of 64 ±5 years.
Low-dose chest CT examinations were performed by using 64 detector-row helical CT (Somatom-64, Siemens Medical systems, Forchheim, Germany) with the following parameters: 0.5 s tube rotation, 32 × 0.6 mm collimation with z-sample system and 28.8 mm table feed per rotation (pitch = 1.5). A caudal-cranial direction without contrast was used. Images were obtained from the level of the lung bases (posterior recesses) to the lung apex with the help of a scout view. Exposure settings were 20–30 mAs and 100–140 kVp depending on the weight of the participant. This corresponds to a dose-length-product <107 mGy·cm and an effective radiation dose <1.6 mSv. The fields-of-view of all three settings were the same and large enough to cover the complete lung cross-section.
Each low-dose chest CT examination was reconstructed using three different settings immediately after imaging with the following three combinations of section thickness/increment and kernel: (A) 1.0 mm/0.7 mm and a soft kernel (Siemens B30 filter), (B) 2.0 mm/1.4 mm and a soft kernel (Siemens B30) and (C) 2.0 mm/1.4 mm and a sharp kernel (Siemens B60). Setting A was the standard setting for the screening project , settings B and C were used for this particular study. Since a thicker section is less effective for nodule detection, only a 2 mm section thickness was chosen in our study . The Siemens B30 kernel is the standard soft-tissue reconstruction kernel, and B60 is the bone reconstruction kernel, widely used in high resolution chest CT at normal dose levels.
To evaluate the effect of reconstruction settings on volumetric measurement variability, each nodule was measured three consecutive times on every reconstructed CT image by one radiologist with 8 years of experience in chest CT. Digital workstations (Leonardo, Siemens Medical Solutions, Erlangen, Germany) were used for nodule volumetric analysis, with U.S. Food and Drug Administration (FDA) approved, commercially available software for semi-automated volume measurements (LungCare, Somaris/5, VB28B-W, Siemens Medical Solutions).
After the nodule was marked with a mouse click, the program automatically defined the volume of interest (VOI) around the nodule. A 3D template was generated, which represented the nodule in its optimal way, if possible without inclusion of surrounding structures. A second mouse click initiated the automated volume measurement, which included the following steps : first, a fixed attenuation threshold of −400 HU was applied to extract a 3D-connected “structure of interest”. This “structure of interest” can also include surrounding structures such as vessels or parts of the chest wall. Apart from this process, a small spherical 3D template originating from the click point was gradually expanded and its cross-correlation with the segmented nodule was computed for each step. The peak value of the cross-correlation curve was determined and an empirical cut-off value close to the peak value was used to separate the nodule from its potential adjacent structures. Finally, segmentation was completed by fusing the optimal 3D template and the “structure of interest”. The segmented nodule was then shown in yellow on the volume-rendered display as the VOI. If the reader was not satisfied with the fit of the VOI, a manual modification of the segmentation by increasing or decreasing the volume could be performed. Manual adjustment of the VOI was not performed in this side study in order to avoid reader-induced volume measurement variability.
The nodule characteristics (location and morphology) were determined by consensus double reading based on visual analysis of 1.0 mm thick axial images. Based on location, the nodules were classified into two subgroups: (1) purely intra-parenchymal, with little or no contact with the pleura, vessels or fissures or (2) attached, including pleural-based, fissure-attached and juxtavascular, with a pleura, fissure or vessel contact line larger than 50% of the nodule diameter.
Based on their morphology, nodules were classified into two groups: smooth-round and non-smooth-round (including polylobulated, spiculated and irregular nodules). Although attached and non-smooth-round nodules were reported to have high odds ratios of measurement variability, we decided to include them in order to be able to evaluate the effect of reconstruction settings on volume measurements for all nodule types encountered by CT screening [15, 16].
The measured volumes were logarithmically transformed (lnV) because of their skewed distribution. As the anti-log of differences on a log scale provides dimensionless ratios, the differences on the log scale were back-transformed to provide ratios among actual measurements. The ratios were further changed into relative volume difference (RVD, %) minus 100%.
Repeatability of each setting
As three repeated volume measurements of pulmonary nodules were performed by the same reader and the same software for each reconstruction setting, the mean difference between replicates is expected to be zero, implying that the backward transformed mean difference has a ratio of 100%. Thus, we calculated the repeatability coefficients according to the Bland and Altman method, which defined the value below which 95% of the differences will lie and thus a larger coefficient corresponds to a lower repeatability . Repeatability coefficients were calculated with one way analysis of variance, defined as 2.77 times the within-subject standard deviation.
Effect of setting and repeat measures on the measured volume
A multi-level model approach was performed to get unbiased effect estimates of reconstruction settings on the measured volume. We used nested data in the sense that some persons had more than one nodule and within each nodule we had several measurements. The multi-level model approach is particularly appropriate for these types of data where correlations between different measurements within participants could be present . In detail, we used a three-level model accounting for correlations within participants, between nodules and repeat measures, and accounting for correlations within nodules between repeated measures. Consequently, the highest level consists of the participants, the second level of the nodules and the lowest level of the repeated measures.
The reconstruction settings, nodule location and morphology were used as independent variables. Both random and fixed effects were considered. In the modelling process, variables were included in the model sequentially. After each step the goodness of fit was determined by the difference in deviance (-2 × log likelihood) between the present and the previous model. The modelling started with an empty model, in which only a fixed intercept and three random terms were modelled. Thereafter, reconstruction setting, nodule characteristics and their interaction were subsequently added. For this multi-level analysis, the MLwiN software package for fitting multi-level models was used (version 2.02; Centre for Multilevel Modelling, Institute of Education, University of London, UK).