MRI-based 2.5D deep learning and radiomics effectively predicted microvascular invasion and Ki-67 expression in hepatocellular carcinoma

TL;DR

This study developed a model using MRI and clinical data to predict aggressive features in liver cancer, improving pre-surgery risk assessment.

Contribution

The novel integration of 2.5D deep learning with radiomics and clinical features improves prediction of MVI and Ki-67 dual positivity in HCC.

Findings

The integrated model achieved high accuracy (AUROC 0.939) in predicting MVI/Ki-67 dual positivity.

The dual-positive group had significantly higher early recurrence rates after surgery.

2.5D DL, radiomics, and clinical features together outperformed single-modality models.

Abstract

To develop and validate an integrated 2.5D deep learning (DL) and Radiomics model using gadoxetic acid-enhanced MRI hepatobiliary phase (HBP) images combined with clinical features for preoperative prediction of microvascular invasion (MVI) and high Ki-67 expression (>20%) dual positivity in hepatocellular carcinoma (HCC).

This retrospective study included 235 pathologically confirmed HCC patients categorized as MVI/Ki-67 double-positive (n = 129) or non-double-positive (n = 106). Clinical data (tumor diameter, AFP, GGT, differentiation grade, etc.) and HBP MRI images were collected. Tumor ROIs were segmented on HBP images. A 2.5D DL approach utilized axial, sagittal, and coronal planes of the largest tumor cross-section. LASSO regression selected key features from clinical, radiomic, and DL feature sets. Multivariate logistic regression identified independent predictors, and a nomogram was built. Model performance was evaluated via ROC curves, calibration plots, DCA, confusion matrices, and waterfall plots. Assessment of early recurrence within 2 years after HCC surgery was performed using alpha-fetoprotein (AFP) levels and imaging examinations.

Significant intergroup differences existed in tumor diameter, AFP, GGT, and differentiation grade (P < 0.05). LASSO selected 38 key features (7 clinical, 23 DL, 8 radiomic). Multivariate analysis confirmed the derived clinical feature score, DL_Radscore, and radiomics Radscore as independent predictors of dual positivity. The integrated nomogram model (combining 2.5D DL, radiomics, and clinical features) achieved optimal prediction performance: AUROC, sensitivity, specificity, precision, accuracy, and F1-score values of 0.939, 0.793, 0.940, 0.942, 0.859, and 0.861, respectively.Calibration curves demonstrated good agreement, and DCA indicated clinical utility. Furthermore, postoperative follow-up confirmed that the MVI/Ki-67 dual-positive group exhibited a significantly higher early recurrence rate compared to the non-dual-positive group (P < 0.05).

The integrated MRI 2.5D DL model synergizing radiomics and clinical features surpasses single-modality models for preoperative prediction of MVI/Ki-67 dual positivity in HCC. This tool shows strong potential for enhancing HCC risk stratification and guiding personalized treatment planning.

Figures