VarCoNet: Variability-Aware fMRI Connectome Pipeline

• VarCoNet is a variability-aware self-supervised framework designed to extract robust functional connectomes from parcellated rs-fMRI data.
• It employs a hybrid 1D-CNN plus Transformer architecture with contrastive learning on augmented sub-sequences to preserve intra- and inter-subject variability.
• Empirical results show improved subject fingerprinting and ASD classification over standard PCC methods, emphasizing its applicability in research and clinical contexts.

Searching arXiv for "VarCoNet" and closely related papers to ground the article. {"query":"VarCoNet arXiv", "max_results": 10} I’m checking arXiv for the exact term and neighboring literature to verify the topic’s publication context. VarCoNet is a variability-aware self-supervised framework for extracting robust functional connectomes from resting-state fMRI by treating functional inter-individual variability as meaningful signal rather than nuisance variation. It is formulated as both a representation-learning framework for rs-fMRI and a learned functional connectome estimator: it consumes parcellated rs-fMRI time series, learns ROI-wise temporal representations, forms a functional-connectivity matrix through cosine similarity, and exports a vectorized FC embedding for downstream tasks such as subject fingerprinting and autism spectrum disorder classification (Lamprou et al., 2 Oct 2025).

1. Definition and scope

VarCoNet was introduced in the paper "VarCoNet: A variability-aware self-supervised framework for functional connectome extraction from resting-state fMRI" (Lamprou et al., 2 Oct 2025). The framework is centered on resting-state functional connectivity, with the explicit claim that a useful representation of brain function should keep within-subject variation low across sessions or segments while preserving or enhancing between-subject variation. That design target is directly tied to two downstream settings emphasized in the study: subject fingerprinting and brain-disorder classification.

The input is a parcellated rs-fMRI signal

$x \in \mathbb{R}^{R \times T},$

where $R$ is the number of ROIs and $T$ is the number of time points. Two atlases are used: AAL3 with 166 ROIs and AICHA with 384 ROIs. All datasets were resampled to a common $TR = 1.5$ s, and inputs shorter than the maximum training length were zero-padded to 320 samples (Lamprou et al., 2 Oct 2025).

The output has two forms. First, VarCoNet produces a learned FC matrix of size $R \times R$, computed via cosine similarity between learned ROI representations. Second, it produces a vectorized FC embedding of size

$1 \times \frac{R(R-1)}{2},$

obtained by discarding the lower triangle of the symmetric FC matrix. This vector is the principal representation used for contrastive learning and downstream evaluation (Lamprou et al., 2 Oct 2025).

A recurring misconception is to read VarCoNet as a generic latent encoder. The paper instead defines it as an explicit connectome-construction pipeline operating on raw parcellated time series rather than on precomputed PCC matrices. Another possible confusion is terminological: in the power-systems literature included here, “VarCoNet” appears informally as a comparative label for Volt/Var-control architectures, but the formal arXiv paper titled "VarCoNet" concerns rs-fMRI and functional connectome extraction (Shi et al., 2018).

2. Scientific premise and objective

The framework is motivated by the claim that each brain has distinctive anatomical and functional traits shaped by genetics and environment, and that these differences manifest in resting-state functional connectivity. VarCoNet therefore rejects the standard tendency to suppress subject-specific variation as noise; instead, it treats functional inter-individual variability as part of the target signal (Lamprou et al., 2 Oct 2025).

This premise is operationalized through the ratio of within-subject to between-subject variability. In the paper’s framing, two segments from the same subject should remain close in representation space despite differing temporal windows, differing lengths, and dynamic FC fluctuations, whereas segments from different subjects should remain distinguishable. That representation geometry is particularly relevant for subject fingerprinting, where cross-session identity should be retained, and for ASD classification, where clinically relevant heterogeneity should not be averaged away (Lamprou et al., 2 Oct 2025).

VarCoNet is also presented as a response to several limitations of prior approaches. The paper identifies sensitivity to scan duration, weak temporal modeling, over-reliance on labels, and indirect or conventional FC estimation as central deficiencies. PCC-based FC is described as discarding temporal dynamics; some time-series methods are said to crop all data to a shortest common duration; earlier contrastive approaches using fixed crops are said to reflect heterogeneous multi-site acquisitions poorly; 1D-CNNs or GRUs alone may miss long-range dependencies; and Transformers alone are described as inefficient on raw rs-fMRI time points because individual time points are not semantically meaningful tokens (Lamprou et al., 2 Oct 2025).

This suggests that VarCoNet is best understood not merely as a new classifier but as a duration-aware, label-free FC-learning strategy whose main invariance target is temporal segmentation and whose main discriminative target is subject identity. The paper explicitly evaluates whether that premise yields stable short-duration fingerprinting and clinically useful cross-cohort generalization (Lamprou et al., 2 Oct 2025).

3. Self-supervised contrastive learning formulation

VarCoNet is trained with SimCLR-style contrastive self-supervised learning. For a batch of $N$ subjects, two augmented views are generated for each subject, yielding $2N$ examples; the two views from the same subject form a positive pair, and all views from different subjects in the batch form negative pairs (Lamprou et al., 2 Oct 2025).

Its central augmentation is segmentation of each subject’s rs-fMRI signal into two random sub-sequences. At each epoch, segment lengths $L_1$ and $L_2$ and their positions are randomly selected, with

$R$0

Because the data are standardized to $R$1 s, these correspond to 2 minutes and 8 minutes. The paper distinguishes this strategy from earlier fixed-crop methods by emphasizing a continuous length range and randomized position as well as duration (Lamprou et al., 2 Oct 2025).

The contrastive loss is given as

$R$2

with cosine similarity

$R$3

The optimized temperature was approximately $R$4, which the authors interpret as emphasizing hard negatives and forcing positive pairs to be very close (Lamprou et al., 2 Oct 2025).

The training semantics are subject-level but the training instances are segments extracted from subject recordings. During self-supervised training, no labels are used. For ASD classification, the encoder remains self-supervised and only the final linear classification layer is trained using labels. Repeated recordings from the same subject are not allowed in the same contrastive batch, and subjects with multiple recordings are used exclusively for training in the classification setting to prevent leakage across train, validation, and test partitions (Lamprou et al., 2 Oct 2025).

A second misconception addressed by the paper is that the framework is supervised end-to-end for diagnosis. It is not: the core encoder is trained without labels, and the supervised component is limited to a linear head in the ASD experiments. The reported ablation that contrastive SSL outperformed supervised training of the same encoder reinforces that distinction (Lamprou et al., 2 Oct 2025).

4. Encoder and connectome construction

The core encoder is a hybrid 1D-CNN plus Transformer. The rationale is explicitly wav2vec-like: the 1D-CNN extracts local temporal features and converts raw time series into more meaningful tokens, after which the Transformer models longer-range temporal dependencies among those tokens (Lamprou et al., 2 Oct 2025).

The encoder pipeline is described as follows. Instance normalization is first applied per sample. A 1D convolution is then applied independently to each ROI time series, producing an output of shape

$R$5

where $R$6 is the number of kernels and $R$7 is the reduced temporal length after convolution. Global average pooling over the kernel dimension reduces this to

$R$8

The $R$9 columns are treated as Transformer tokens, trainable positional encodings are added, and a Transformer encoder outputs another $T$0 representation. Cosine similarity across ROI-wise learned representations then forms the $T$1 FC matrix, after which lower-triangular entries are removed to produce the final embedding (Lamprou et al., 2 Oct 2025).

The CNN hyperparameters are specified explicitly: a single convolutional layer, 16 kernels, kernel size 8, and stride 4. The kernel size is justified by the observation that, at $T$2 s, 8 samples correspond to about 12 seconds, approximately the duration of a typical hemodynamic response. The stride of 4 is described as reducing overlap and token redundancy, thereby making attention more efficient (Lamprou et al., 2 Oct 2025).

Transformer hyperparameters were selected separately for each atlas by Bayesian optimization. For AAL3, the best configuration used $T$3, $T$4, and $T$5. For AICHA, it used $T$6, $T$7, and $T$8. The paper emphasizes that the best architecture consistently used only one Transformer layer, which it interprets as suggesting that deeper attention was unnecessary or harmful in this setting (Lamprou et al., 2 Oct 2025).

Bayesian optimization is performed with the Tree-structured Parzen Estimator over 125 trials. The tuned hyperparameters are the number of Transformer layers, the number of attention heads, the feedforward dimension, batch size, temperature, and learning rate. Model selection is tied to duration robustness: on a validation set of 200 HCP subjects, each with two sessions, 30 segments are extracted from each validation recording—10 of length $T$9, 10 of length $TR = 1.5$0, and 10 of length $TR = 1.5$1—and fingerprinting is evaluated under the six length combinations $TR = 1.5$2, $TR = 1.5$3, $TR = 1.5$4, $TR = 1.5$5, $TR = 1.5$6, and $TR = 1.5$7. The intended selection objective is stated to be the harmonic mean of the average and minimum of these six fingerprinting rates (Lamprou et al., 2 Oct 2025).

5. Empirical evaluation and reported performance

The empirical study spans HCP, ABIDE I, and ABIDE II. HCP is used for subject fingerprinting and hyperparameter optimization. ABIDE I is used for ASD classification, and ABIDE II is used as an independent external generalization test. The reported HCP cohort contains 1,092 subjects and 2,117 recordings, with runs of 14.33 minutes originally acquired at $TR = 1.5$8 s and resampled to 1.5 s. ABIDE I contains 946 subjects and 995 recordings, with 500 NC and 446 ASD. ABIDE II contains 730 subjects; the text notes that the table and narrative totals are not perfectly aligned because of exclusions (Lamprou et al., 2 Oct 2025).

Training uses PyTorch, Adam, and a linear warmup plus cosine annealing schedule with 10 warmup epochs on a single NVIDIA RTX 6000 Ada with 48 GB VRAM under Python 3.11.8. For both AAL3 and AICHA, batch size is 64. The learning rate is $TR = 1.5$9 for AAL3 and $R \times R$0 for AICHA, with $R \times R$1 for both. The total number of training epochs is not explicitly stated (Lamprou et al., 2 Oct 2025).

For subject fingerprinting on HCP, the test protocol extracts FC embeddings from session 1 and session 2, forms

$R \times R$2

computes Pearson correlation between rows of $R \times R$3 and rows of $R \times R$4, and declares identification correct when the maximal correlation maps to the same subject. The paper reports that VarCoNet consistently outperforms PCC and both deep-learning baselines across all duration combinations and atlases, especially for short recordings; that accuracy increases with duration for all methods; that AICHA generally performs better than AAL3; and that VarCoNet has the lowest variability across duration combinations. Exact fingerprinting accuracies are not tabulated in the provided text, but the paper attributes the gains to both increasing intra-subject similarity and decreasing inter-subject similarity (Lamprou et al., 2 Oct 2025).

For ASD classification on ABIDE I, the main evaluation uses 10 independent 10-fold cross-validation runs, yielding 100 performance samples. On AAL3, VarCoNet reports $R \times R$5, $R \times R$6, and $R \times R$7 in 10-fold CV, and $R \times R$8, $R \times R$9, and $1 \times \frac{R(R-1)}{2},$0 on the external Caltech test. The strongest baseline, BolT, reports $1 \times \frac{R(R-1)}{2},$1, $1 \times \frac{R(R-1)}{2},$2, and $1 \times \frac{R(R-1)}{2},$3 in CV, and $1 \times \frac{R(R-1)}{2},$4, $1 \times \frac{R(R-1)}{2},$5, and $1 \times \frac{R(R-1)}{2},$6 on Caltech (Lamprou et al., 2 Oct 2025).

On AICHA, VarCoNet reports $1 \times \frac{R(R-1)}{2},$7, $1 \times \frac{R(R-1)}{2},$8, and $1 \times \frac{R(R-1)}{2},$9 in 10-fold CV, and $N$0, $N$1, and $N$2 on Caltech. The best baseline, again BolT, reports $N$3, $N$4, and $N$5 in CV, and $N$6, $N$7, and $N$8 on Caltech (Lamprou et al., 2 Oct 2025).

The external ABIDE II test is presented as evidence of out-of-cohort generalization without fine-tuning. On AAL3, VarCoNet reports $N$9, $2N$0, $2N$1, and prediction change $2N$2, compared with BolT’s $2N$3, $2N$4, $2N$5, and prediction change $2N$6. On AICHA, VarCoNet reports $2N$7, $2N$8, $2N$9, and prediction change $L_1$0, compared with BolT’s $L_1$1, $L_1$2, $L_1$3, and prediction change $L_1$4 (Lamprou et al., 2 Oct 2025).

The ablation study reports that all major design choices contribute: the proposed augmentation outperforms UCGL-style augmentation, the 1D-CNN plus Transformer outperforms both 1D-CNN-only and Transformer-only variants, and contrastive SSL outperforms supervised training of the same encoder across both atlases and almost all reported metrics. Exact ablation values are not included in the provided text, but significance levels shown in the figures include $L_1$5, $L_1$6, $L_1$7, and $L_1$8 (Lamprou et al., 2 Oct 2025).

6. Interpretation, limitations, and terminological scope

VarCoNet includes an explicit interpretability layer, though the paper also acknowledges that interpretability remains limited. In fingerprinting analyses, the ReX toolbox is used to compare VarCoNet with PCC and other methods using intra-class correlation, individual variation fields, and gradient flow maps. Using the AICHA atlas and 2-minute segments, VarCoNet improved ICC relative to PCC, Cai et al., and Lu et al. The authors interpret this as evidence that the framework reduces intra-subject variation while increasing inter-subject variation, thereby preserving meaningful person-specific structure (Lamprou et al., 2 Oct 2025).

For ASD classification, feature importance is derived from the final linear classifier. If $L_1$9 and $L_2$0 denote the NC and ASD class weights, connection importance is defined as

$L_2$1

Positive $L_2$2 favors ASD, negative $L_2$3 favors NC, and $L_2$4 measures magnitude of influence. These scores are averaged over 110 trained linear layers. The most informative connections are reported to involve temporo-parietal regions associated with social cognition and frontal cortices implicated in prior ASD neuroimaging work, although the exact ROI pairs are not enumerated in the text (Lamprou et al., 2 Oct 2025).

The principal limitations acknowledged by the authors are interpretability and fixed-TR implementation. The model was implemented for data resampled to $L_2$5 s so that tokens would have consistent temporal meaning for the Transformer, but resampling can introduce interpolation errors or alter temporal properties. The authors also note that resting-state signals lack direct semantic meaning and that Transformer attention further reduces intuitive interpretability. Proposed future directions include integrating VarCoNet with GNNs on its graph-like FC outputs, learning a binary mask to discard noisy or irrelevant connections, building an end-to-end contrastive framework combining FC extraction and graph reasoning, and removing the need for fixed-TR resampling by feeding TR as input or using multiple CNNs with different kernel sizes (Lamprou et al., 2 Oct 2025).

In a broader methodological landscape, VarCoNet belongs to the growing class of representation-learning systems that embed domain structure directly into the learning architecture. A plausible implication, by analogy with physics-aware graph learning in Volt-VAR control (Wu et al., 2022) or stability-constrained local policy learning in distribution networks (Yuan et al., 2022), is that its main contribution is architectural rather than purely predictive: it defines which variability should be invariant, which variability should remain discriminative, and how that inductive bias should be encoded in the training objective and representation geometry. Within the present literature set, however, only (Lamprou et al., 2 Oct 2025) uses the name VarCoNet as the formal title of a method; the appearances of the term in Volt/Var-control discussions are comparative and informal rather than nominal (Byeon et al., 2023).