Deep–Learning Radiomics and Hand-Crafted Radiomics Utilizing Contrast-Enhanced MRI to …

ORIGINAL RESEARCH article

Front. Oncol.

Sec. Cancer Imaging and Image-directed Interventions

• ¹Department of Radiology, Second Affiliated Hospital of Chongqing Medical University, Chongqing, China
• ²GE Healthcare, Advanced analytics team, Shanghai, China
• ³Department of Radiology, The First Hospital Affiliated to the Army Medical University, Chongqing, China

Purpose: To investigate early peritumoral recurrence (EPR) after drug-eluting bead transarterial chemoembolization (DEB-TACE) in a multicenter cohort of patients with hepatocellular carcinoma (HCC) using deep learning radiomics (DLR) based on preoperative multiphase magnetic resonance imaging (MRI). Patients and methods: A total of 157 patients with HCC from two institutions who received DEB-TACE were retrospectively enrolled and divided into a training cohort (n=114) and an external validation cohort (n=43). A total of 960 radiomics features were extracted from five different phases: arterial phase (AP), delayed phase (DP), portal venous phase (PVP), peritumoral 3 mm portal venous phase (PVP_Pri3mm), and tumoral plus peritumoral portal venous phase (PVP_Plus3mm). A total of 512 deep learning features were extracted from PVP using ResNet34 (PVP_DLR). The features selected through the minimum Redundancy and Maximum Relevance (mRMR) and Least Absolute Shrinkage and Selection Operator (LASSO) methods were utilized for model construction. The performance of the model was evaluated using area under the curve (AUC), calibration curves, net reclassification (NRI), and decision curve analysis (DCA). Results: PVP_Pri3mm and PVP_Plus3mm showed comparable performance to the PVP model (P>0.05). The final deep learning radiomics and radiomics nomogram (DLRRN) included three predictors: PVP-signature, PVP_ DLR signature, and AFP, which showed effectively discrimination of between EPR to DEB-TACE, with AUCs of 0.802 (95% CI, 0.718-0.887) in the training cohort and 0.770 (95% CI, 0.623-0.916) in the external validation cohort, demonstrating good calibration (P>0.05). Additionally, the DLRRN model performed significantly better than the clinical model (P0.05). DCA confirmed that DLRRN was clinically useful. Conclusion: DLRRN has good efficacy in predicting EPR after DEB-TACE, which can provide value for preoperative treatment selection and postoperative prognostic assessment of patients with HCC.

Keywords: Hepatocellular Carcinoma, DEB-TACE, contrast-enhanced MRI, deep learning, Radiomics

Received: 07 Jun 2025; Accepted: 28 Oct 2025.

Copyright: © 2025 Liu, Wang, Liu, Li, Ma, Chen, Luo and Zhou. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

* Correspondence: Xi Liu,