Value renal CT volumetric texture analysis with machine learning radiomics in assessment of pathological grade of clear cell renal cell carcinoma

Objective: To investigate the value of renal CT volumetric texture analysis with machine learning radiomics in assessment of pathological grade of clear cell renal cell carcinoma (ccRCC). Methods: Thirty-four biopsy-confirmed ccRCC subjects who had four-phase CT scanning (NC: non-contrast, CM: Corticomedullary, N: Nephrographic, E: Excretory) were collected retrospectively from June 2013 to October 2017 for the study. Non-rigid registration was performed on multi-phase CT images in reference to CM-phase. Each lesion was segmented on CM-phase CT images using our in-house volumetric image analysis platform, '3DQI'. A set of fifty-nine volumetric textures, including histogram, gradient, gray level co-occurrence matrix (GLCM), run-length (RL), moments, and shape, was calculated for each segment lesion in each phase as parameters for the training/testing of Random Forest (RF) classifier. Four groups according to pathological Fuhrman grade on a scale I to IV, these tumors were then divided into low (I + II) and high grade (III + IV) groups. Feature selection was performed by Boruta algorithm. A 10-fold cross-validation method was applied to validate the RF performance by receiver operating characteristic (ROC) curves analysis to determine the diagnostic accuracy of the model. Results: Subjects were divided into four groups by Fuhrman grade on a scale I to IV : 3 cases grade I, 19 cases grade II, 8 cases grade III and 4 cases grade IV. In CM-phase, kurtosis and long-run-emphasis (RLE) were selected the most important textures for ccRCC staging among 59 features. The area under curve (AUC) of ROC was 0.88 (79% sensitivity and 82% specificity) by using kurtosis and RLE textures. The mean values of kurtosis and RLE were (-20.00 ± 22.00) × 10^-2 and (3.00 ± 0.40) × 10^-2 for low group, whereas (31.00 ± 32.00) × 10^-2 and (5.00 ± 0.02) × 10^-2 for high group. Within the mean ± SD range of statistics, radiomics can distinguish between low and high grade tumors. In multi-phase analysis, three most important features were selected among 236 (59 × 4) textures: kurtosis (CM-phase), GLCM homogeneity I (HOMO 1) (E-phase), and GLCM homogeneity 2 (HOMO2) (E-phase). The mean values of HOMO 1 (E-phase) and HOMO 2 (E-phase) were (19.00 ± 0.03) × 10^-2 and (11.00 ± 0.02) × 10^-2 for low group, whereas (22.00 ± 0.03) × 10^-2 and (14.00 ± 0.02) × 10^-2 for high group. The AUC was 0.92 (93% sensitivity and 87% specificity) by using these three textures. Conclusion: This study has demonstrated that renal CT volumetric texture analysis with machine learning radiomics could preoperative accurately perform cancer staging for ccRCC. (authors)