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

Purpose ： Deep learning has shown great promise in medical image analysis, yet challenges remain in histological grading of hepatocellular carcinoma (HCC), particularly in effectively integrating multimodal features and enhancing model generalizability. This study aimed to develop a Cross-Modal Interactive Fusion 2 (CMIF) framework to leverage complementary information from CT and MRI for improved noninvasive HCC grading. Methods ： We proposed CMIF, which utilizes the self-supervised representational capacity of DINOv2 to combine the high-resolution anatomical details from CT with the functional and physiological features of MRI. We retrospectively enrolled 270 HCC patients with concurrent CT and MRI from four centers, and allocated them into training, internal validation, and two independent external validation cohorts. Performance of CMIF was compared against unimodal models and existing fusion approaches. Results：CMIF consistently outperformed baseline methods. In the internal validation cohort, CMIF achieved an AUC of 0.760, surpassing comparison methods (0.556-0.751). In external validation cohorts, CMIF obtained AUCs of 0.754 and 0.790, while competing methods ranged from 0.53 to 0.745. These results highlight the superior discriminative capability and robust generalizability of the proposed framework across multicenter datasets.Conclusion：The CMIF framework demonstrates significant potential for improving HCC histological grading by effectively integrating CT and MRI features. Its robust performance across independent cohorts suggests that CMIF may assist radiologists in accurate tumor differentiation and provide valuable support for individualized preoperative treatment planning.