MAPI-GNN: Multi-Activation Plane Interaction Graph Neural Network for Multimodal Medical Diagnosis

TL;DR

This work tackles multimodal medical diagnosis by moving beyond a single static graph to a dynamic, patient-specific multi-activation graph profile. It introduces the Multi-Dimensional Feature Discriminator (MDFD) to identify salient features across semantic dimensions and a Multi-Activation Graph Construction Strategy (MAGCS) to build multiple activation graphs, followed by a Hierarchical Feature Dynamic Association Network (HFDAN) that fuses intra-sample graphs and models inter-sample relations. The approach achieves state-of-the-art results on PI-CAI csPCa and CHD datasets, with ablations confirming that each component (MDFD, MAGCS, HFDAN) contributes to performance gains. The method offers a practical, efficient pathway toward more accurate, reliable AI-assisted diagnosis in complex clinical settings.

Abstract

Graph neural networks are increasingly applied to multimodal medical diagnosis for their inherent relational modeling capabilities. However, their efficacy is often compromised by the prevailing reliance on a single, static graph built from indiscriminate features, hindering the ability to model patient-specific pathological relationships. To this end, the proposed Multi-Activation Plane Interaction Graph Neural Network (MAPI-GNN) reconstructs this single-graph paradigm by learning a multifaceted graph profile from semantically disentangled feature subspaces. The framework first uncovers latent graph-aware patterns via a multi-dimensional discriminator; these patterns then guide the dynamic construction of a stack of activation graphs; and this multifaceted profile is finally aggregated and contextualized by a relational fusion engine for a robust diagnosis. Extensive experiments on two diverse tasks, comprising over 1300 patient samples, demonstrate that MAPI-GNN significantly outperforms state-of-the-art methods.

Figures (10)

• Figure 1: Conceptual fusion strategies. (a) CNN-based methods use fixed fusion points (e.g., early/late), which limits the modeling of complex inter-modal dependencies. (b) GNNs offer a flexible paradigm, using a graph topology for explicit relationship modeling and hierarchical aggregation.
• Figure 2: Overview of the MAPI-GNN architecture. Stage I (detailed in Fig. \ref{['fig:fig4']}) generates multiple, semantically-aware activation graphs from patient-specific multimodal data. Stage II (detailed in Fig. \ref{['fig:fig5']}) then performs a hierarchical, two-level fusion on these graphs, first modeling intra-sample relationships and then inter-sample dependencies for the final diagnosis.
• Figure 3: Workflow of Stage I: Multi-Activation Graph Construction. From compressed modality features, a Multi-Dimensional Feature Discriminator identifies salient activated features for multiple semantic dimensions, each guiding the construction of a corresponding activation graph.
• Figure 4: Workflow of the Hierarchical Feature Dynamic Association Network (Stage II): 1) Intra-sample encoding of multiple activation graphs into representations ($\mathbf{g}_m$); and 2) Inter-sample classification on a global graph of fused patient vectors ($\mathbf{F}_p$) processed by a GCN.
• Figure 5: Visual comparison of MAPI-GNN against key baselines on the PI-CAI dataset. This figure visualizes the primary metrics presented in Table \ref{['tab:comparison']}, highlighting our model's competitive performance.