Cross-modal interactive fusion framework for preoperative histological grading of hepatocellular carcinoma

Objective.To develop and validate a cross-modal interactive fusion (CMIF) framework that addresses the limitations of single-modality imaging by integrating complementary CT and MRI features for noninvasive preoperative prediction of high-grade hepatocellular carcinoma (HCC).Approach.To address the clinical need for noninvasive preoperative identification of high-grade HCC, we retrospectively analyzed paired CT and MRI data from 270 pathologically confirmed HCC patients across four centers. Leveraging the complementary anatomical and tumor-heterogeneity information provided by CT and MRI, we developed a CMIF framework. The framework used a frozen DINOv2 backbone for CT and MRI feature extraction, a lightweight multilayer perceptron for task-specific adaptation, and a cross-modal feature alignment and fusion (CMAF) module to model bidirectional CT-MRI interactions and generate fused representations for histological grading. The model was evaluated in internal and external validation cohorts and compared with unimodal models and representative multimodal fusion methods.Main results.The proposed CMIF model achieved an area under the receiver operating characteristic curve (AUC) of 0.760 in the internal validation cohort and AUCs of 0.754 and 0.790 in two external validation cohorts. Compared with unimodal models and alternative fusion methods, CMIF showed more stable performance across validation cohorts. Ablation experiments further indicated that both DINOv2-based feature extraction and the cross-modal fusion module contributed to the improvement in predictive performance.Significance.The proposed framework provides a noninvasive and quantitative approach for preoperative HCC histological grading. By explicitly modeling complementary CT and MRI information, CMIF may assist radiologists in identification of high-grade tumors and provide valuable support for individualized preoperative treatment planning.