Prediction of Post-hepatectomy Liver Failure in Patients With Hepatocellular Carcinoma Based on Radiomics Using Gd-EOB-DTPA-Enhanced MRI: The Liver Failure Model

Abstract

Objectives: Preoperative prediction of post-hepatectomy liver failure (PHLF) in patients with hepatocellular carcinoma (HCC) is significant for developing appropriate treatment strategies. We aimed to establish a radiomics-based clinical model for preoperative prediction of PHLF in HCC patients using gadolinium-ethoxybenzyl-diethylenetriamine (Gd-EOB-DTPA)-enhanced magnetic resonance imaging (MRI). Methods: A total of 144 HCC patients from two medical centers were included, with 111 patients as the training cohort and 33 patients as the test cohort, respectively. Radiomics features and clinical variables were selected to construct a radiomics model and a clinical model, respectively. A combined logistic regression model, the liver failure (LF) model that incorporated the developed radiomics signature and clinical risk factors was then constructed. The performance of these models was evaluated and compared by plotting the receiver operating characteristic (ROC) curve and calculating the area under the curve (AUC) with 95% confidence interval (CI). Results: The radiomics model showed a higher AUC than the clinical model in the training cohort and the test cohort for predicting PHLF in HCC patients. Moreover, the LF model had the highest AUCs in both cohorts [0.956 (95% CI: 0.955-0.962) and 0.844 (95% CI: 0.833-0.886), respectively], compared with the radiomics model and the clinical model. Conclusions: We evaluated quantitative radiomics features from MRI images and presented an externally validated radiomics-based clinical model, the LF model for the prediction of PHLF in HCC patients, which could assist clinicians in making treatment strategies before surgery.

Keywords: hepatocellular carcinoma; magnetic resonance imaging; post-hepatectomy liver failure; prediction model; radiomics.

Figure 1

The radiomics workflow. (A) Regions of interests (ROIs) were delineated in the non-tumor area; (B) Radiomics features were extracted from ROIs; (C) The selection of radiomics features; (D) The evaluation of the performance of prediction models.

Figure 2

Heatmaps of correlations among radiomics features. (A) Heatmap depicting correlation coefficients matrix of 1,044 radiomics features in the training cohort. (B) Heatmap depicting correlation coefficients matrix of 24 selected radiomics features in the training cohort.

Figure 3

Radiomics features selection. (A) The determination of 24 radiomics features; (B) The ranks of the 24 selected features.

Figure 4

Comparison of ROC curves between the clinical model, radiomics model, and the LF model in the training (A) and test (B) cohorts.