In Abdominal radiology (New York)
OBJECTIVE : To compare machine learning (ML) of texture analysis (TA) features for classification of solid renal masses on non-contrast-enhanced CT (NCCT), corticomedullary (CM) and nephrographic (NG) phase contrast-enhanced (CE) CT.
MATERIALS AND METHODS : With IRB approval, we retrospectively identified 177 consecutive solid renal masses (116 renal cell carcinoma [RCC]; 51 clear cell [cc], 40 papillary, 25 chromophobe and 61 benign tumors; 49 oncocytomas and 12 fat-poor angiomyolipomas) with renal protocol CT between 2012 and 2017. Tumors were independently segmented by two blinded radiologists. Twenty-five 2-dimensional TA features were extracted from each phase. Diagnostic accuracy for 1) RCC versus benign tumor and 2) cc-RCC versus other tumor was assessed using XGBoost.
RESULTS : ML of texture analysis features on different phases achieved mean area under the ROC curve (AUC [SD]), sensitivity/specificity for 1) RCC vs benign = 0.70(0.19), 96%/32% on CM-CECT and 0.71(0.14), 83%/58% on NG-CECT and; 2) cc-RCC vs other = 0.77(0.12), 49%/90% on CM-CECT and 0.71(0.16), 22%/94% on NG-CECT. There was no difference in AUC comparing CECT to NCCT (p = 0.058-0.54) and no improvement when combining data across all three phases compared single-phase assessment (p = 0.39-0.68) for either outcome. AUCs decreased when ML models were trained with one phase and tested on a different phase for both outcomes (RCC;p = 0.045-0.106, cc-RCC; < 0.001).
CONCLUSION : Accuracy of machine learning classification of renal masses using texture analysis features did not depend on phase; however, models trained using one phase performed worse when tested on another phase particularly when associating NCCT and CECT. These findings have implications for large registries which use varying CT protocols to study renal masses.
Schieda Nicola, Nguyen Kathleen, Thornhill Rebecca E, McInnes Matthew D F, Wu Mark, James Nick
Computed tomography, Machine learning, Renal cell carcinoma, Texture analysis