Table 1.
Patient demographics, smoking status, and tumor stage.
Fig 1.
CT scans from a cohort of biopsy-confirmed ADCs from VUMC/TVHS (n = 50) and Mayo (n = 45) were analyzed by three independent software users from two institutions using CANARY segmentation. All ADCs were <3 cm, according to previous CANARY studies. A pulmonologist and radiologist confirmed nodule location prior to segmentation. A. Once the observer had identified the nodule on the CT scan, placing the pointer over the nodule established a volume of interest around the nodule. B. Next, CANARY defined the nodule border (red area) on each CT slice. C. If the border appeared inconsistent with the perceived nodule edge (i.e. extension into vasculature or chest wall), the observer adjusted the nodule borders by using an eraser tool (dotted yellow circle).
Fig 2.
ADC characterization based upon the CANARY class composition.
CANARY detected nine unique voxel characteristics within CT image data, which were color coded as CANARY classes: Violet (V), Indigo (I), Blue (B), Green (G), Yellow (Y), Orange (O), Red (R), Cyan (C), and Pink (P). Voxels of class V, I, R, and O (VIRO) were associated with invasion, while the classes B, C, and G represented lepidic growth. P and Y class voxels were between lepidic and frankly invasive growth, such those found in MIA or AIS. The composition of each class within the total ADC voxels was used to define overall ADC risk characterization as Good (G), Intermediate (I), or Poor (P). Above are three ADCs at the completion of CANARY analysis and characterization. A, B, and C are examples of G, I, and P nodule characterizations respectively.
Fig 3.
Observer segmentation differences contribute minimally to the variance of CANARY classes.
The variance between observer segmentations was calculated for each CANARY class and the V, I, R, O (VIRO) classes collectively. CANARY classification variance was compared between the individual ADCs (red), inter-observer segmentations (green), or residual intrinsic variability (blue).
Table 2.
Intra-class correlation coefficient (ICC) is highest amongst the CANARY classes representing most invasive ADC features.
Intra-class correlation coefficient (ICC), a measure of agreement of quantitative assessments made by different observers evaluating the same quantity, was calculated for each CANARY class, the VIRO group, and the average (Avg.) of all classes. An ICC of 0.8–1 reflects high agreement between users. These calculations were performed for the VUMC/TVHS and the Mayo cohorts, as well as for each subgroup of patients scanned by a particular CT scanner (GE Medical Systems, Philips, and Siemens). ICCs and 95% CI for each voxel class are shown below. The number of patients within each cohort is listed in parentheses.
Fig 4.
CANARY class patterns between observers.
Two ADC cases are shown with the segmentations as completed by the three observers. The segmentations numbered 1 and 2 for each case were performed by a VUMC/TVH observer, while segmentation numbered 3 was performed by the Mayo observer. A. The VUMC/TVHS observers and the Mayo observer segmented different portions of this nodule while following the standard operating procedure (SOP), yet this had minimal impact upon the nodule class composition as shown by the high percentage of the R class in the nodule. B. A representative case showing low segmentation variability between observers.
Table 3.
Intra-class correlation coefficient (ICC) amongst CANARY voxel subtypes from nodules less than 1cm in diameter.
Avg. is the average of all voxel classes. 95% CI is shown in parentheses below the ICC.