GICAF-Net: A cross-attentional graph-image fusion network for hyperspectral pathological diagnosis of FNH and HCC.

[BACKGROUND AND OBJECTIVE] Accurate intraoperative differentiation between focal nodular hyperplasia (FNH) and hepatocellular carcinoma (HCC) remains a major clinical challenge, especially in atypical cases where conventional imaging and histopathology are constrained by turnaround time, cost, or spectral resolution. This study aims to develop a novel deep learning framework to improve the precision and efficiency of hyperspectral pathological diagnosis for liver tumors.

[METHODS] We propose GICAF-Net, a Graph-Image Cross-Attentional Fusion Network, designed to leverage hyperspectral imaging (HSI) for capturing fine-grained spatial-spectral information. The network employs a dual-branch architecture: (1) a residual convolutional branch for extracting pseudo-color image features, and (2) a residual graph convolutional branch for modeling topological spatial-spectral features. A Topology-Aware Cross-Attention Fusion (TACA) module enables bidirectional information exchange between the two modalities, while a multi-constraint fusion loss-combining cross-entropy, prediction confidence, and cross-modal attention consistency-enhances classification stability. A balanced hyperspectral liver tumor dataset consisting of 60 HCC and 60 FNH cases was constructed and evaluated using ten-fold cross-validation.

[RESULTS] GICAF-Net achieved an AUC of 0.9571 ± 0.0068, accuracy of 88.34 % ± 1.10 %, and F1-score of 88.32 % ± 1.11 %, outperforming state-of-the-art baseline models. Ablation experiments further validated the contributions of both the TACA module and the multi-constraint loss function in enhancing cross-modal fusion and improving classification performance.

[CONCLUSION] The integration of graph-based spectral-structural modeling with deep visual features through cross-attention provides a powerful approach for hyperspectral pathological diagnosis. The proposed GICAF-Net demonstrates strong potential for rapid, accurate, and minimally invasive intraoperative differentiation of FNH and HCC, offering valuable clinical support in liver tumor surgery.