Quantitative image variables reflect the intratumoral pathologic heterogeneity of lung adenocarcinoma

We aimed to compare quantitative radiomic parameters from dual-energy computed tomography (DECT) of lung adenocarcinoma and pathologic complexity. A total 89 tumors with clinical stage I/II lung adenocarcinoma were prospectively included. Fifty one radiomic features were assessed both from iodine images and non-contrast images of DECT datasets. Comprehensive histologic subtyping was evaluated with all surgically resected tumors. The degree of pathologic heterogeneity was assessed using pathologic index and the number of mixture histologic subtypes in a tumor. Radiomic parameters were correlated with pathologic index. Tumors were classified as three groups according to the number of mixture histologic subtypes and radiomic parameters were compared between the three groups. Tumor density and 50th through 97.5th percentile Hounsfield units (HU) of histogram on non-contrast images showed strong correlation with the pathologic heterogeneity. Radiomic parameters including 75th and 97.5th percentile HU of histogram, entropy, and inertia on 1-, 2- and 3 voxel distance on non-contrast images showed incremental changes while homogeneity showed detrimental change according to the number of mixture histologic subtypes (all Ps < 0.05). Radiomic variables from DECT of lung adenocarcinoma reflect pathologic intratumoral heterogeneity, which may help in the prediction of intratumoral heterogeneity of the whole tumor.


Imaging and analysis
All patients underwent CT examination using a dual-source CT system (Somatom Definition Flash; Siemens Healthcare, Forchheim, Germany) with the dual-energy technique (Supplementary Figure S2). The DECT system was composed of two X-ray tubes and two corresponding 128-row detectors mounted in a perpendicular arrangement. DECT scanning was obtained 90 seconds after administration of contrast material (100 mL of iopamidol: Iomeron 300; Bracco, Milan, Italy) at a rate of 1.5 mL/sec using a power injector. This was followed by a 20-cc saline flush at a rate of 1.5 mL/sec. Imaging parameters were as follows: 105 mAs (effective) at 140 kV, 248 mAs (effective) at 80 kV, 32×0.6-mm collimation, pitch of 0.7, rotation time of 0.5 seconds, and a 512×512-pixel matrix and an in-plain resolution of 0.69mm. Imaging was performed from the thoracic inlet to the middle portion of the kidneys. Three different data sets were generated from the DECT imaging: 80-kV, 140-kV, and enhanced weighted-average images. The weightedaverage images were generated by combining the 140-kV and 80-kV data sets with a weighting factor of 0.6 (60% of information derived from the 80-kV image and 40% derived from the 140-kV image), and thus these were approximately 120-kV images.

Data post-processing and image reconstruction
The virtual non-enhanced images and iodineenhanced images were generated using the liver virtual non-contrast (VNC) application mode of dedicated dualenergy post-processing software (Syngo Dual Energy; Siemens Medical Solutions) (Supplementary Figure S1). To obtain the iodine value of both solid and ground-glass opacity (GGO) components in each tumor, post-processing was performed with two different software applications. For the solid component, the parameters for the material decomposition method were as follows: -110 Hounsfield units (HU) for fat at 80 kV, -87 HU for fat at 140 kV, 52 HU for soft tissue at 80 kV, and 51 HU for soft tissue at 140 kV. For the GGO component, since the lesion was composed of a mixture of air and soft tissue, the HU value of fat in the liver VNC application mode should be replaced with that of air, which is a HU value located at the interconnecting line between air and soft tissue [1,2].
Thus, the material parameters were -110 HU for fat at 80 kV, -115 HU for fat at 140 kV, 60 HU for soft tissue at 80 kV, and 54 HU for soft tissue at 140 kV. Image data were reconstructed with a section thickness of 1 mm using a D30f (medium smooth) kernel for the iodine-enhanced image and a D45f (medium sharp) kernel for the virtual non-enhanced image.

Quantitative CT image parameters
We adopted radiomic features to extract multidimensional information of a given ROI (Supplementary  Table S1) [3,4]. For physical features, the volume of the ROI was computed by multiplying the number of voxels by the unit volume of a voxel. Physical density (g/cm 3 ) was estimated by extrapolating from the mean CT scan attenuation [5]. Next, histogram features were derived from the intensity (HU) distribution of a given region of interest (ROI). From the histogram, we computed skewness, kurtosis, uniformity, entropy, and CT numbers or HU at the 2.5th, 25th, 50th, 75th and 97.5th percentiles. For regional features, the ROI was computed based on the gray level size zone matrix. The value of the matrix (m, n) was defined by the number of homogenous regions given the homogeneous tumor size (n) to their intensity (m). The intensity variability and size-zone variability were also computed from the gray level size zone matrix [5][6][7][8][9]. For local features, a gray level co-occurrence matrix (GLCM) [10], which describes the frequency of various combinations of grey values within an ROI, was created. For each ROI, GLCMs were created with 13 directions according to 1-voxel, 2-voxel, and 3-voxel distances (3 GLCMs for each ROI). From each GLCM, 12 local features (the energy, entropy, correlation, contrast, variance, sum mean, inertia, cluster shade, cluster tendency, homogeneity, maximum probability, and inverse variance) were computed. In summary, we extracted four physical, nine histogram-based, two regional, and 36 local features from the manually derived ROI.

Pathologic index
The individual prognostic impact of each subtype of lung adenocarcinoma was assessed from the development cohort using the disease-free survival (DFS) curve of a previous large-scale study [11]. Based on the  [12]. A higher entropy score represents increased heterogeneity of the tumor. The second pathologic index (pathologic index 2) was defined as the sum of the proportions of each subtype multiplied by their HRs, with the addition of the Gini coefficient. The Gini coefficient was calculated using the following equation: . The third pathologic index (pathologic index 3) was calculated as the sum of all subtype percentages multiplied by their HRs. the fourth pathologic index (pathologic index 4) was defined as the simple arithmetic sum of the scores of the subtypes multiplied by their HRs.
These four pathologic indices were used with a validation group of 148 patients with comprehensive histologic subtyping for completely resected lung adenocarcinomas. DFS curves were plotted and the predictive ability of each pathologic index was evaluated. Among the four pathologic indices, only pathologic index 3 enabled significant patient stratification in the validation cohort according to DFS (P = 0.005). Furthermore, pathologic index 3 showed the highest concordance probability estimate of 0.691 (P = 0.049; 95% CI: 0.633-0.749) of all four pathologic indices. Thus, we used index 3 as our pathologic heterogeneity index.