Differentiation combined hepatocellular and cholangiocarcinoma from intrahepatic cholangiocarcinoma based on radiomics machine learning

Abstract

Background: Combined hepatocellular and cholangiocarcinoma (CHC) and intrahepatic cholangiocarcinoma (ICC) are hard to identify in clinical practice preoperatively. This study looked to develop and confirm a radiomics-based model for preoperative differentiation CHC from ICC.

Methods: The model was developed in 86 patients with ICC and 46 CHC, confirmed in 37 ICC and 20 CHC, and data were collected from January 2014 to December 2018. The radiomics scores (Radscores) were built from radiomics features of contrast-enhanced computed tomography in 12 regions of interest (ROI). The Radscore and clinical-radiologic factors were integrated into the combined model using multivariable logistic regression. The best-combined model constructed the radiomics-based nomogram, and the performance was assessed concerning its calibration, discrimination, and clinical usefulness.

Results: The radiomics features extracted from tumor ROI in the arterial phase (AP) with preprocessing were selected to build Radscore and yielded an area under the curve (AUC) of 0.800 and 0.789 in training and validation cohorts, respectively. The radiomics-based model contained Radscore and 4 clinical-radiologic factors showed the best performance (training cohort, AUC =0.942; validation cohort, AUC =0.942) and good calibration (training cohort, AUC =0.935; validation cohort, AUC =0.931).

Conclusions: The proposed radiomics-based model may be used conveniently to the preoperatively differentiate CHC from ICC.

Keywords: Combined hepatocellular and cholangiocarcinoma (CHC); intrahepatic cholangiocarcinoma (ICC); machine learning; radiomics.

Figure 1

Flowchart of ICC and CHC patients’ enrollment. ICC, intrahepatic cholangiocarcinoma; CHC, combined hepatocellular and cholangiocarcinoma.

Figure 2

Workflow of construction and validation of the models.

Figure 3

Receiver operating characteristic analysis for 4 combined models and 3 single-factor models. (A) Training cohort; (B) validation cohort. R, radiomics model. C. Clinical model. L, liver imaging model; RC, radiomics-clinical model; LC, liver imaging-clinical model; RL, radiomics-liver imaging model; RCL, radiomics-clinical-liver imaging model.

Figure 4

Decision curve analysis of combined models for predicting ICC and CHC. The y-axis measures the net benefits, wherein the red and blue line represents the RCL and RC model, respectively. The gray line represents the assumption that all patients have ICC. The black line represents the assumption that all patients have CHC. It shows that combined models are better than the treat-all-patients scheme, while if LI-scores are added, it does not result in more benefits for patients’ discrimination (chi-squared =0.1, P=0.7). RCL, radiomics-clinical-liver imaging; RC, radiomics-clinical; CHC, combined hepatocellular and cholangiocarcinoma; ICC, intrahepatic cholangiocarcinoma.

Figure 5

Nomogram of indipendent predictors of ICC/CHC and calibration curve of nomogram in training/validation cohorts. (A) The model presented with a nomogram scaled by the proportional regression coefficient of each predictor, the probability in nomogram describes the probability of a patient with CHC; (B) calibration curve for training and validation set. The Hosmer-Lemeshow test gave a chi-square of 2.87 (P=0.94) for the training cohort and validation cohort, indicating that the RCL model was appropriate in both data sets. RCL, radiomic-clinical-liver image model; CHC, combined hepatocellular and cholangiocarcinoma.

Figure S1

ROC analysis for differentiating between ICC and CHC patients in 12 R models. (A) Training cohort; (B) validation cohort. AP, arterial phase. PVP, portal vein phase; ROI, region of interest; Pre, imaging preprocessing; ICC, intrahepatic cholangiocarcinoma; CHC, combined hepatocellular and cholangiocarcinoma.