The value of using deep learning to predict proliferative hepatocellular carcinoma based on multiphasic magnetic
Fengxi Chen1, Yuxin He2, Lizhen Liu1
17T Magnetic Resonance Translational Medicine Research Center, Department of Radiology, Southwest Hospital, Army Medical University (Third Military Medical University), Chongqing, China.
Background:
Proliferative hepatocellular carcinoma (PHCC) is an aggressive subtype of hepatocellular carcinoma (HCC) characterized by high recurrence rates and poor prognosis. Accurate preoperative identification of PHCC is essential for prognostic assessment and individualized treatment planning. However, conventional imaging methods often fail to reliably diagnose PHCC. This study aimed to develop an effective deep learning (DL)-based model using multiphasic gadoxetate disodium (Gd-EOB-DTPA)-enhanced magnetic resonance imaging (MRI) to preoperatively predict PHCC.
Methods:
This retrospective multicenter study included 1,111 HCC patients who underwent Gd-EOB-DTPA-enhanced MRI followed by curative resection between January 2015 and June 2021. Patients were allocated to training (n=818), internal (n=150), and external (n=143) validation cohorts. A hybrid model combining DL and clinical-radiological features was developed using the AutoGluon framework. The clinical-radiological model was constructed using multivariate logistic regression, whereas the DL model was developed utilizing DL confidence scores derived from arterial and venous phase images. Model performance was assessed by the area under the curve (AUC) and 95% confidence intervals (CIs), and recurrence-free survival (RFS) was analyzed using the log-rank test.
Results:
The hybrid model demonstrated superior performance in both the training and internal validation cohorts, achieving AUCs of 0.922 (95% CI: 0.900-0.939) and 0.894 (95% CI: 0.834-0.934), respectively. In the external validation cohort, the hybrid and DL models demonstrated comparable performance (AUC: 0.771 vs. 0.788; P=0.065). Notably, the DL model outperformed the clinical-radiological model in predicting PHCC, with AUCs of 0.857 (95% CI: 0.781-0.912) vs. 0.747 (95% CI: 0.652-0.825) in the internal validation cohort and 0.788 (95% CI: 0.686-0.865) vs. 0.625 (95% CI: 0.505-0.735) in the external validation cohort (all P<0.05). Furthermore, patients classified as high-risk by the hybrid model had significantly shorter RFS compared to those in the low-risk group (P<0.05).
Conclusions:
The hybrid model showed potential for predicting PHCC, which may assist clinicians in making personalized treatment decisions.
