ARMA-C3: A Contrastive ARMA Convolutional Framework for Unsupervised and Semi-supervised Classification

Technical Summary of ARMA-C3: A Contrastive ARMA Convolutional Framework

Problem Statement

Accurate and early identification of neurodegenerative and medical disorders remains challenging due to the scarcity of labeled data, the complexity of imaging patterns, and high-dimensional, noisy modalities (e.g., diffusion MRI, ultrasound, X-ray). Traditional machine learning approaches often treat subjects or images as independent instances, ignoring crucial relational information and subject-level structures that may indicate disease progression. Furthermore, existing Graph Neural Network (GNN) methods in biomedical settings predominantly rely on supervised learning, requiring large amounts of expensive labeled data, or fail to fully exploit underlying data structures in unsupervised and semi-supervised regimes. Current graph-based methods often struggle to simultaneously preserve community-level cohort structures and enforce augmentation-invariant subject representations under noisy graph construction.

Methodology

The authors propose ARMA-C3, a unified unsupervised and semi-supervised graph learning framework designed for node classification. The framework integrates three core components: Contrastive learning, ARMA (Auto-Regressive Moving Average) Convolution, and graph-Cut regularization.

1. Graph Construction and Preprocessing:

   • Feature Extraction: For diffusion MRI (dMRI) datasets (ADNI, NIFD), the method extracts Fractional Anisotropy (FA) maps, registers them to a white-matter atlas, and constructs histogram-based descriptors from predefined neuroanatomically relevant Regions of Interest (ROIs). For image-based datasets (BreastMNIST, PneumoniaMNIST, Liver Ultrasound), features are extracted using frozen pre-trained backbones (ViT-DINO for ultrasound, ResNet-18 for X-ray) to avoid overfitting in low-data regimes.
   • Graph Formation: Subjects are modeled as nodes. A subject-level graph is constructed where edges represent inter-subject similarity, computed via cosine similarity of feature vectors. A threshold $\alpha$ is applied to sparsify the adjacency matrix, acting as an implicit denoising mechanism.

2. Architectural Components:

   • ARMA Convolution: The framework employs ARMA-based graph convolutional networks. Unlike standard GCNs, ARMA layers use recursive filter stacks to approximate spectral graph convolutions. This allows for flexible frequency responses, improved control over oversquashing through iterative message refinement, and the capture of both local and long-range graph dependencies.
   • Structural Regularization (Graph-Cut/Modularity): To prevent feature mixing and oversmoothing (where embeddings from different clusters merge), the model incorporates modularity maximization. This imposes explicit constraints on soft cluster assignments, encouraging strong connections within communities and weak connections between them. A balancing regularizer prevents the collapse of all nodes into a single cluster.
   • Contrastive Learning (MERIT): The framework adopts the Multi-scale Contrastive Siamese Network (MERIT) approach. It enforces consistency across multiple augmented views of the same graph using a teacher-student paradigm (with an Exponential Moving Average teacher). This ensures that learned representations are invariant to local feature noise and global structural perturbations.

3. Optimization Objective:
   The model is trained by jointly optimizing a structural regularization loss ($L_{struct}$) and a contrastive loss ($L_{con}$). In semi-supervised settings, a supervised cross-entropy loss ($L_{sup}$) is added for labeled nodes. The total objective is:
   $$L = L_{struct} + \lambda_{con} L_{con} \quad (\text{unsupervised})$$
   $$L_{total} = \lambda_{con} L_{con} + L_{sup} + \lambda_{struct} L_{struct} \quad (\text{semi-supervised})$$

Key Contributions

1. Unified Framework: The paper introduces a novel joint formulation that simultaneously optimizes modularity-based structural regularization and multi-view contrastive consistency within a single objective. This coupling is claimed to be unique in the context of biomedical population graph learning, addressing the gap where previous methods treated these components independently.
2. Robust Unsupervised Clustering: The framework performs fully unsupervised node-level clustering without relying on diagnostic labels. It uncovers stable subject groups, enabling robust disease classification in real-world settings characterized by label scarcity and severe class imbalance.
3. Cross-Modal Generalization: The method demonstrates strong adaptability across fundamentally different biomedical modalities, including diffusion MRI (neuroimaging), breast ultrasound, liver ultrasound, and chest X-ray.
4. Clinical Feature Engineering: For dMRI data, the authors construct a compact representation using clinically informed ROIs (e.g., corpus callosum, fornix) and histogram-based FA distributions, preserving disease-sensitive microstructural information.

Experimental Results

The framework was evaluated on five datasets: ADNI (CN vs. MCI, CN vs. AD), NIFD (Frontotemporal Dementia), BreastMNIST, PneumoniaMNIST, and a Liver Ultrasound dataset.

• Unsupervised Performance: ARMA-C3 achieved competitive or superior performance compared to classical clustering techniques (K-Means, SpectralNet), state-of-the-art graph clustering (MAGI, TokenCut), and standard GNN backbones (GCN, GAT). Notably, it achieved statistically significant improvements over MAGI across all datasets (e.g., F1-score of 0.786 vs. 0.778 on CN vs. MCI; 0.841 vs. 0.790 on BreastMNIST).
• Semi-Supervised Performance: Using only 10% labeled data, ARMA-C3 consistently outperformed traditional classifiers (SVC, Random Forest) and other graph-based semi-supervised methods (APPNP, JacobiConv). It maintained balanced precision-recall trade-offs, which is critical for disease screening where false negatives are costly.
• Ablation Studies: Experiments confirmed that:
  • Finer histogram bin resolution (20 bins) is crucial for distinguishing subtle microstructural differences in early-stage disease.
  • The choice of feature backbone matters (ViT-DINO outperformed ResNet-18 on ultrasound, while ResNet-18 was slightly better on X-ray).
  • Moderate graph sparsification thresholds ($\alpha$) and contrastive loss weights ($\lambda_{con}$) yield the best balance between structural preservation and noise suppression.
  • The learned embeddings preserve intrinsic data geometry even under aggressive dimensionality reduction (PCA), unlike raw features.

Significance and Claims

The paper claims that ARMA-C3 offers a unified and clinically relevant framework for biomedical data analysis, particularly in scenarios where annotation is expensive or data is imbalanced. By modeling inter-subject relationships and leveraging spectral filtering with contrastive consistency, the method learns stable, well-separated, and robust representations.

The authors position ARMA-C3 as a label-efficient pre-screening tool for clinical pipelines (e.g., dMRI cohort studies, ultrasound triage). It aims to reduce the burden of manual annotation while maintaining diagnostically meaningful subgroup separation. The framework's ability to generalize across heterogeneous imaging modalities suggests its potential as a versatile tool for diverse biomedical imaging tasks, moving beyond the limitations of modality-specific or purely supervised approaches.

Limitations Acknowledged: The authors note that the framework is sensitive to graph construction hyperparameters (specifically the sparsification threshold $\alpha$), requires dataset-specific tuning, and currently employs a static transductive graph construction strategy that necessitates reconstruction when new subjects are introduced. Future work is planned to address multi-class modeling and inductive graph learning.