MAPI-GNN: Multi-Activation Plane Interaction GNN

• The paper presents a novel framework that decomposes the feature space into multiple semantic dimensions to build patient-specific activation graphs.
• MAPI-GNN employs a Multi-Dimensional Feature Discriminator with Graph Attention Networks for dynamic graph construction and hierarchical fusion.
• Experimental results show that it outperforms traditional CNN, transformer, and static GNN methods, significantly improving diagnostic metrics.

The Multi-Activation Plane Interaction Graph Neural Network (MAPI-GNN) is a graph-based framework designed to address limitations of conventional fusion methods in multimodal medical diagnosis. Unlike prior approaches relying on a single, static graph constructed from pre-extracted features, MAPI-GNN decomposes the feature space into multiple semantic dimensions, dynamically constructs a stack of activation graphs per example, and fuses these representations in a hierarchical, context-aware manner. This architecture achieves patient-specific adaptive graph modeling and hierarchical intra- and inter-sample information integration, leading to state-of-the-art diagnostic performance across disparate modalities and disease domains (Qin et al., 23 Dec 2025).

1. Motivation and Conceptual Foundations

Traditional deep learning methods for multimodal diagnosis—spanning CNNs, transformers, and static GNNs—struggle to robustly model complex, non-Euclidean relationships between heterogeneous information sources (e.g., MRI, CT, structured clinical data). The standard paradigm indiscriminately fuses modalities and constructs a single, static topology, resulting in:

• Mixing of relevant and irrelevant (redundant/noisy) features.
• Non-adaptive topologies unable to capture patient-specific pathological interactions.
• Limited capacity to propagate information across semantically distant features.

MAPI-GNN addresses these deficiencies via:

• Semantic disentanglement: Partitioning feature space into multiple "activation planes" to expose latent, clinically-relevant subspaces.
• Dynamic, multi-plane graph construction: Building a tailored activation graph per semantic dimension, yielding a patient-specific, multifaceted graph profile.
• Hierarchical relational fusion: Sequentially aggregating intra-sample feature relationships and global inter-sample dependencies (Qin et al., 23 Dec 2025).

2. Architecture and Workflow

The MAPI-GNN framework comprises two main stages, each with tightly integrated components:

Stage I: Multi-Activation Graph Construction

A Multi-Dimensional Feature Discriminator (MDFD) receives a compressed multimodal feature vector $x \in \mathbb{R}^C$ (e.g., generated by an autoencoder handling multiple imaging or clinical modalities). The MDFD projects this vector onto $M$ orthogonal semantic dimensions, each reflecting a different "activation plane." For each dimension $m$, the top-$P$ fraction of features—those most influential for that semantic subspace—are selected to form the node set $A_m$ of the activation graph $G_m$.

Graph construction proceeds as follows:

• For each node $i \in A_m$, identify its $k$ nearest neighbors (in Euclidean feature space), creating the edge set $E_m$.
• Edge weights $w_{ij}^{(m)}$ are defined as the mean influence of the connected nodes, where the influence score $C_m(i)$ computes the sensitivity of the $m$-th semantic dimension to perturbing feature $i$.
• Each $G_m$ is an undirected, weighted graph on the common feature node set $V = \{1, ..., C\}$, with adjacency $A_m$.

Stage II: Hierarchical Feature Dynamic Association Network (HFDAN)

This module encodes and integrates the multifaceted graph profile:

• Intra-sample encoding: Each $G_m$ is processed with a planar graph encoder implemented as a single-layer Graph Attention Network (GAT), leveraging both learned attention and the pre-defined influence-based edge weights.
• Aggregation: The resulting $M$ graph embeddings $\{ g_m \}$ are concatenated with the original feature vector $x_p$ for patient $p$, yielding an extended feature profile $F_p$.
• Representation regularization ensures that $g_m$ retains information about the original node features via a reconstruction penalty.
• Inter-sample fusion: A global patient graph $\mathcal{G}_P$ is constructed, typically via a $k$-NN topology on the $F_p$. This graph is encoded by a Graph Convolutional Network (GCN) and outputs predictions through an MLP classifier.

The complete system is optimized end-to-end, leveraging classification, representation, and semantic disentanglement losses.

3. Detailed Algorithmic Components

Multi-Dimensional Feature Discriminator (MDFD)

The MDFD employs a shallow feed-forward network with orthogonality regularization to achieve disentangled semantic projections. Key operations include:

• Zero-out perturbation: Measures each feature's influence on each activation plane by individually nullifying entries of $x$ and observing the effect on the MDFD output.
• Discriminator loss:

$$L_\text{sd} = L_\text{AE}(x, \hat{x}) + L_\text{reg}(\Theta_\text{sd})$$

where $L_\text{AE}$ is mean squared reconstruction error, and $L_\text{reg}$ enforces sparsity, weight decay, and orthogonality.

Dynamic Activation Graph Construction

• Graph building: For each $m$, edges are restricted to high-importance features, with similarity computed in the original feature space.
• Edge weighting: Incorporates influence-based scores to guide subsequent message passing.
• No further normalization is performed before GAT encoding, deferring to the attention mechanism to absorb the weighting.

Relational Fusion Engine

• Planar Graph Encoder (GAT): Node update rule combines learned attention with explicit edge weights, followed by feature aggregation (readout) over nodes.
• Global Patient Graph: Allows relational reasoning across patients, modeling cohort-level structure and supporting end-to-end learning.

4. Experimental Protocol and Results

Experiments evaluate MAPI-GNN against strong CNN-based, transformer, GNN, and fusion method baselines on two multimodal medical datasets:

• PI-CAI (Prostate csPCa): 440 balanced samples, modalities include T2w, ADC, HBV MRI.
• CHD (Coronary Heart Disease): 974 cases with CCTA scans and structured clinical data.

Key experimental details:

• Five-fold cross-validation, fixed random seed.
• Metrics: ACC, AUC, PRE, REC, F1, SPE.

Performance on PI-CAI:

Method	ACC	AUC	PRE	REC	F1	SPE
HGM2R	0.9242	0.9798	0.9246	0.9242	0.9242	0.9394
ViT (Transformer)	0.9053	0.9728	0.8587	0.9491	0.9145	0.8069
MAPI-GNN	0.9432	0.9838	0.9361	0.9545	0.9438	0.9318

Ablation studies on both PI-CAI and CHD demonstrate that omission of any key component—MDFD, the dynamic multi-activation graph construction stack (MAGCS), or HFDAN—substantially degrades performance (up to −12.3% ACC).

5. Analysis, Limitations, and Advantages

MAPI-GNN exhibits the following properties:

• Adaptive graph topology: Each patient receives a personalized, feature-driven graph stack, overcoming the rigidity of static graph schemes.
• Semantic disentanglement: The MDFD extracts multiple clinically-relevant perspectives from noisy or redundant multimodal features.
• Hierarchical fusion: Sequential intra/inter-sample operations ensure robust and balanced performance across tasks, metrics, and cohort heterogeneity.
• Efficiency: Lightweight design (12.3M parameters, 1.93 GFLOPs), with inference latency (~45 ms/case) compatible with clinical workflows (Qin et al., 23 Dec 2025).

Observed limitations include:

• Dependence on full modality availability; the framework is not directly robust to missing data scenarios.
• Abstract semantic planes do not directly map to known pathological or radiomic markers, presenting challenges for interpretability.

6. Future Directions and Open Problems

Potential research avenues include:

• Extending the architecture to operate under partial modality missingness, integrating strategies like modality-dropout or imputation.
• Aligning learned semantic planes with clinically-understood features to improve interpretability and facilitate human-in-the-loop diagnostics.
• Data-driven adaptation of architecture hyperparameters (number of activation planes $M$, neighborhood size $k$) based on population or patient-level characteristics.

A plausible implication is that principled, patient-specific graph construction and semantic disentanglement may generalize to broader applications in heterogeneous biomedical data fusion, contingent on future advances in handling incomplete or ambiguous modality composition (Qin et al., 23 Dec 2025).