Multi-tissue deep fusion network for prediction of pulmonary metastasis in hepatocellular carcinoma

Abstract

Pulmonary metastasis is a critical adverse prognostic factor in patients with hepatocellular carcinoma (HCC), underscoring the need for accurate prediction to guide prognoses and treatment decisions. However, current prediction methods are hindered by two major challenges: (1) inter-class similarity and intra-class variation in computed tomography (CT) images, and (2) the predominant methods focus on extracting tumor-associated features, despite evidence that metastasis may often be related to the degree of hepatic cirrhosis and deformation of hepatic vessels. To address these limitations, we propose a multi-tissue deep fusion network (MDFNet) for predicting pulmonary metastasis from CT images. The network employs MeshNet as the backbone to extract spatial structural features and capture tumor heterogeneity, cirrhosis severity, and vascular deformation. A dual-level contrastive learning module highlights feature disparities across tissues to enhance the network’s feature representational ability, while a triple attention mechanism-based feature fusion module integrates multi-tissue features to identify essential predictive information. MDFNet was validated on a multi-center dataset including seven clinical centers. The experimental results demonstrate that, compared to existing methods, MDFNet exhibits the highest area under the receiver operating characteristic curve of 0.7948 and accuracy of 0.7622 on an independent testing set. Despite its effectiveness, the model currently uses only single time-point venous-phase CT images; future work will incorporate multi-phase CT sequences and dynamic follow-up scans to further improve prediction performance.