Ada-FCN: Adaptive Frequency-Coupled Network for fMRI-Based Brain Disorder Classification

AI summary

Overview

• Research area: Neuroimaging machine learning — graph neural networks applied to resting-state fMRI functional connectivity for neurological disorder diagnosis.
• Technical level: Intermediate to Advanced. Readers should be comfortable with graph neural networks, attention mechanisms, correlation-based connectivity matrices, and basic signal decomposition concepts.
• Scope: The paper proposes a single end-to-end framework (Ada-FCN) that learns subject- and region-specific frequency sub-bands from fMRI time series and models both within-band and across-band functional connectivity for brain disorder classification.

What This Paper Is About

Most fMRI-based diagnostic models treat each brain region's BOLD time series as one undifferentiated signal, even though neurological disorders typically disrupt neural oscillations in specific frequency ranges. Existing attempts to use frequency information rely on fixed, hand-picked bands and then simply merge the resulting features, which can miss disease-relevant rhythms and blur the interactions between different bands. Ada-FCN addresses both weaknesses by learning the frequency decomposition itself and by explicitly modeling how different frequency bands talk to each other, all inside one trainable classification pipeline.

Key Contributions

1. Adaptive Cascade Decomposer. A learnable, cascaded decomposition that splits each region's fMRI time series into paired low-frequency approximations and high-frequency residuals, replacing handcrafted band definitions with task-driven ones.
2. Frequency-Coupled Connectivity Learning. A construction of a unified functional connectivity network that contains sparse intra-band connections (Pearson correlation plus adaptive dynamic thresholding) and cross-band connections learned through a Dual-Projection Bilinear Attention mechanism.
3. Unified-GCN message passing. A graph convolution operating on the combined intra- and cross-band adjacency matrix, with a learnable scale factor balancing the two, producing refined per-region embeddings that are pooled per band and concatenated for classification.
4. Auxiliary objectives and strong empirical validation. A band diversity loss and a cross-band sparsity loss guide training, and the method outperforms eleven baselines on ADNI (Alzheimer's disease) and ABIDE (autism spectrum disorder).

Main Findings

• Improved classification accuracy on both datasets. Ada-FCN reaches 79.68% accuracy and 75.30% AUROC on ADNI, and 77.89% accuracy and 77.62% AUROC on ABIDE, exceeding all eleven baselines including BrainGNN, BrainNetCNN, Contrasformer, and frequency-domain methods such as Hu et al., MFHC, and Tewarie et al.
• The accuracy gain is substantial. The next-best ADNI accuracy is 74.31% (BrainGNN) and 74.18% (Tewarie et al.); the next-best ABIDE accuracy is 74.92% (BrainGNN), so the improvement is several percentage points rather than marginal.
• Every architectural component contributes. Ablating the dynamic threshold causes the largest drop (ADNI falls from 79.68% to 77.35% in one configuration and 76.14% in another), while removing either the diversity loss or the sparsity loss also consistently lowers accuracy on both datasets.
• Two decomposition levels are optimal. Setting the cascade depth K = 2, yielding four sub-bands, worked best on both ADNI and ABIDE.
• Learned connectivity patterns separate diagnostic groups. Group-averaged unified connectivity matrices show the Alzheimer's group with stronger intra-band connectivity in the first high-frequency band (H1) than cognitively normal and mild cognitive impairment groups, and generally weaker cross-frequency interactions.
• MCI looks like an intermediate stage. Mild cognitive impairment subjects display connectivity patterns between those of Alzheimer's and cognitively normal subjects, with a unique enhancement of intra-band connectivity in the H1 band — consistent with MCI being a prodromal Alzheimer's phase.
• Frequency band pairs behave differently. The L1–L2 cross-band connection is uniformly weak across all groups, while L1–H1 and L2–H2 connections are similar across groups, indicating that not all cross-band coupling is diagnostically informative.

Methodology in Plain English

The input is a matrix of BOLD signal time series, one row per brain region of interest (116 regions from the AAL atlas), over a fixed number of time points.

First, instead of applying a fixed filter or wavelet, the model learns to separate each region's signal into frequency components. It does this in stages: at each stage, a dilated one-dimensional convolution extracts a smooth low-frequency part, and the leftover signal is passed through a second, smaller convolution to produce a high-frequency residual. Repeating this process produces a cascade of paired low- and high-frequency signals per region.

Second, the model builds a functional connectivity network from these components. Within each frequency band, it computes Pearson correlations between regions and keeps only connections above an adaptively chosen threshold, which removes noise-driven links. Across bands, it uses a bilinear attention scheme: each band's correlation matrix is projected into a source space or a target space with its own learnable weights, and the interaction between a source band and a target band is computed as a product of these projections. This allows directed, band-specific coupling rather than a symmetric average.

Third, the within-band blocks and the cross-band blocks are combined into one large adjacency matrix, with a learnable weight controlling how much cross-band structure matters. A standard graph convolution — normalized adjacency multiplied by node features and a weight matrix — propagates information over this unified graph, so a region in one band can influence a region in another band.

Finally, the per-band node embeddings are averaged into graph-level vectors, concatenated, and passed through a multi-layer perceptron that outputs the diagnosis. Training uses cross-entropy, plus a loss that discourages different bands from producing similar embeddings (encouraging each band to capture something distinct) and an L1 penalty that keeps cross-band connections sparse.

Why This Matters

Impact on research. The paper challenges a widely held simplification in fMRI-based diagnosis — that BOLD signals can be treated as single-band series — and shows that where the frequency decomposition comes from (learned versus fixed) measurably changes diagnostic performance. It also provides a reusable mechanism for modeling interactions between frequency bands rather than merging them, which applies beyond fMRI to any multi-band graph-structured signal.

Real-world applications:

• Earlier Alzheimer's detection. Group-specific connectivity signatures, particularly the H1 band patterns distinguishing MCI from CN and AD, could support screening tools that flag prodromal patients before severe impairment.
• Autism spectrum disorder triage. The ABIDE results suggest a non-invasive, scan-based classifier could serve as a supporting signal in diagnostic workflows where behavioral assessment is currently the primary route.
• Clinical trial enrichment. Better subject stratification could help select participants most likely to respond in trials of disease-modifying therapies.
• Biomarker discovery. The learned unified connectivity matrices provide candidate frequency-specific targets for further neuroscience validation rather than only a black-box label.

Industry relevance. Computational neuroimaging and digital-health companies building diagnostic decision-support products, and medical imaging platform vendors, can adopt the adaptive decomposition and unified graph construction as drop-in components. The modest computational footprint (a single RTX 4090 for training) keeps it accessible to small research and startup teams.

Future Directions

• Improving interpretability of the learned bands. The authors explicitly name this as future work: mapping the learned sub-bands back to physiological frequency ranges would let clinicians reason about what the model found.
• Scaling and validating across datasets. Evaluation is limited to ADNI and ABIDE with the fixed AAL116 atlas; testing on Parkinson's, schizophrenia, and other cohorts, and with alternative parcellation schemes, would test generalization.
• Handling class imbalance and small cohorts. ADNI's AD class has only 65 samples against 914 CN samples, and ABIDE has roughly 61 ASD samples; strategies for imbalance and for federated or multi-site learning remain open.
• Making cross-band coupling more interpretable and efficient. The unified adjacency matrix has dimension 2KN by 2KN, which grows with the number of bands and regions; reducing this cost and quantifying which cross-band edges actually drive predictions are natural extensions.

Target Audience

Graduate students and researchers in neuroimaging, biomedical engineering, and graph machine learning who work on brain connectome analysis; practitioners building fMRI-based diagnostic models who want a concrete alternative to fixed-band connectivity pipelines; and method-oriented machine learning researchers interested in learnable signal decomposition and multi-band graph learning, particularly those studying multivariate time series where frequency structure matters.

Authors’ abstract

Resting-state fMRI has become a valuable tool for classifying brain disorders and constructing brain functional connectivity networks by tracking BOLD signals across brain regions. However, existing mod els largely neglect the multi-frequency nature of neuronal oscillations, treating BOLD signals as monolithic time series. This overlooks the cru cial fact that neurological disorders often manifest as disruptions within specific frequency bands, limiting diagnostic sensitivity and specificity. While some methods have attempted to incorporate frequency informa tion, they often rely on predefined frequency bands, which may not be optimal for capturing individual variability or disease-specific alterations. To address this, we propose a novel framework featuring Adaptive Cas cade Decomposition to learn task-relevant frequency sub-bands for each brain region and Frequency-Coupled Connectivity Learning to capture both intra- and nuanced cross-band interactions in a unified functional network. This unified network informs a novel message-passing mecha nism within our Unified-GCN, generating refined node representations for diagnostic prediction. Experimental results on the ADNI and ABIDE datasets demonstrate superior performance over existing methods. The code is available at https://github.com/XXYY20221234/Ada-FCN.