Hepatocellular carcinoma (HCC) is a common malignancy in China with a high mortality rate. Early recurrence and long-term prognosis are closely linked to tumor aggressiveness. Microvascular invasion (MVI), defined as the presence of tumor cells within small branches of the portal or hepatic veins, is a critical marker of malignant biological behavior. Clinically, MVI is strongly associated with early postoperative recurrence and serves as an important reference for determining surgical margin extension, adjuvant treatment, and postoperative management strategies. At present, definitive diagnosis of MVI still relies on postoperative pathological examination, and reliable preoperative assessment methods are lacking. Although prior studies have attempted to predict MVI using preoperative imaging, their clinical application remains limited by poor generalizability, weak interpretability, and insufficient cross-center adaptability.
This study aims to develop a high-performance, generalizable, and interpretable computer-aided diagnostic (CAD) system for preoperative prediction of HCC-MVI using multiphase CT imaging, artificial intelligence techniques, and clinical prior knowledge. The system will be evaluated prospectively in an observational, multicenter clinical study to assess its diagnostic value and clinical applicability.
The CAD system integrates three categories of imaging features: (1) high-level representations automatically extracted by deep neural networks; (2) predefined radiomics features such as tumor morphology, texture, and intensity distributions; and (3) structured prior features derived from radiological expertise, including tumor margin blurriness and spatial relationships with adjacent portal veins. Sparse constraints and redundancy suppression mechanisms will be applied to identify stable and efficient MVI-related representations. In addition, the system adopts a spatial domain strategy covering tumor, peritumoral, and distant regions, in order to capture invasion patterns from both local morphology and microenvironmental context, thereby constructing reproducible and clinically interpretable imaging biomarkers.
To overcome the limitations of single-domain models, the system employs a multi-source heterogeneous fusion strategy that integrates morphological-textural features, dynamic enhancement patterns, and spatial graph structures. The model architecture combines convolutional neural networks (CNNs) to capture fine-grained textures, Transformer modules to model long-range dependencies, and graph neural networks (GNNs) to represent tumor-vascular topological relationships. This hybrid approach enables comprehensive understanding of both local details and global structures. Furthermore, the model incorporates uncertainty quantification and attention-like mechanisms to dynamically adjust prediction confidence and generate saliency heatmaps. These outputs are designed to enhance clinicians' interpretability and trust in the system. An interactive visualization interface will also be developed to support risk interpretation and surgical planning.
The study will conduct a prospective observational validation across multiple clinical centers, with unified inclusion/exclusion criteria and standardized data collection protocols. Model predictions will be blindly compared against postoperative pathological results. In addition to conventional metrics (accuracy, sensitivity, specificity, and AUC), the study will observationally evaluate the impact of model-based predictions on preoperative risk stratification and surgical decision-making. By testing the system across diverse patient populations, the study aims to confirm its generalizability, clinical utility, and potential for real-world translation of intelligent diagnostic technologies.
