Graph Neural Network with LCMV-Constrained Optimization for Lesion-Level Tumor Classification from MR Images

Lesion-level tumor classification from magnetic resonance (MR) images remains challenging due to the difficulty in effectively aggregating multi-slice instance features under weakly supervised conditions. Traditional multiple instance learning (MIL) methods often fail to capture the underlying relational structures and spatial dependencies among instances. To address this challenge, we propose DBF-GNN, a graph neural network (GNN) framework that incorporates a Linearly Constrained Minimum Variance (LCMV)-inspired multi-constraint loss function for lesion-level tumor classification. In the proposed framework, each lesion is modeled as a graph, where nodes correspond to patch-level features extracted from MR images, and edges are constructed via a k-nearest neighbors (k-NN) strategy based on inter-node feature similarity. A Graph Isomorphism Network with Edge features (GINE)-based GNN, equipped with an edge-feature-aware readout module, is employed to aggregate the patch-level instance features into lesion-level representation. To guide the aggregation process, we introduce the LCMV-inspired multi-constraint loss function to promote the discriminative feature aggregation by minimizing total output power and enforcing class-directional alignment constraints. Experimental results on renal and ovarian tumor MR image datasets demonstrate that the proposed method outperforms the baseline methods. Experiments across multiple GNN backbones validate the general applicability of the proposed constraint mechanism to various GNN architectures.