• A tensor-based discriminative multi-graph method for MHSI dimensionality reduction. • Enhanced intra-/inter-class and local graphs improve discriminative learning. • Superior OA, AA, and Kappa on three real medical hyperspectral datasets. Medical hyperspectral imaging (MHSI) integrates microscopic spatial information with spectral data, providing rich physicochemical characterization for medical applications. However, high dimensionality introduces redundant information and computational challenges, limiting the efficiency and accuracy of subsequent analysis. Existing tensor-based graph embedding methods often rely on single or fixed adjacency strategies, which cannot fully capture the complex geometric structures of medical tissues. To address this, we propose a Tensor-based Discriminative Multi-Graph Embedding (TDMGE) method for dimensionality reduction. TDMGE constructs three types of adjacency graphs that capture intra-class similarity, inter-class separability, and local structural relationships. During the graph construction process, bidirectional edge weights and adaptive modulation factors are introduced to characterize local geometric relationships and class-discriminative structures. Bidirectional weights enhance intra-class compactness and inter-class separability by considering mutual contributions of node pairs, while adaptive modulation factors dynamically adjust the influence of individual samples according to local geometry. A unified tensor-based graph embedding objective jointly optimizes multi-graph information to obtain effective low-dimensional representations. Experimental results on three MHSI datasets demonstrate the effectiveness and robustness of the proposed method in medical image analysis tasks.
Zhu et al. (Fri,) studied this question.