MAC-Net: A Multi-Task Deep Learning Framework for Modeling Cognitive Function From Task-Based fMRI

Md. Tanvir Rahman; Nabil Anan Orka; Asaduzzaman Khan; Mohammad Ali Moni

cs.AISep 27, 2026

MAC-Net: A Multi-Task Deep Learning Framework for Modeling Cognitive Function From Task-Based fMRI

Authors: Md. Tanvir Rahman, Nabil Anan Orka, Asaduzzaman Khan, Mohammad Ali Moni

Organizations: School of Health and Rehabilitation Sciences, The University of Queensland, QLD 4072, Australia · Department of Information and Communication Technology, Mawlana Bhashani Science and Technology University, Tangail 1902, Bangladesh

Abstract

Objective cognitive assessment from neural signals supports neurorehabilitation, but individual-level prediction from task-based fMRI (tfMRI) remains difficult because neural features coexist with substantial demographic and scanner-related variation. We present the Multi-task Activation and Contrast Network (MAC-Net), a covariate-aware deep learning framework for modeling individual cognitive function from regional tfMRI. By isolating tfMRI features into a dedicated neural pathway and restricting participant variables to a terminal late-fusion pathway, MAC-Net prevents dominant covariates from suppressing high-dimensional clinical representations during feature learning. Evaluating baseline data from 6,500 Adolescent Brain Cognitive Development Study participants under family-aware cross-validation, MAC-Net was benchmarked against linear models, random forests, and alternative deep architectures. The N-back plus Monetary Incentive Delay configuration achieved $R^{2}$ values of 0.174, 0.238, and 0.277 for fluid, crystallized, and total cognition, outperforming covariate-only baselines (0.178) and alternative deep models (0.217). N-back was the most informative paradigm, whereas incorporating the Stop Signal Task marginally degraded performance. Feature attributions via Integrated Gradients, DeepLIFT, and Input Gradient were highly concordant, localizing working-memory-related frontal, parietal, and cingulate regions. These findings demonstrate that covariate-aware multi-task modeling yields reproducible cognitive-function estimations, establishing a robust neural engineering framework for clinical translation.

Figures & tables

Fold	Split	N	Males	Females	Age (months)	$G_{f}$	$G_{c}$	$G_{t}$
1	Train	3891	1947	1944	119.63 $\pm$ 7.53	94.05 $\pm$ 9.44	87.82 $\pm$ 6.50	88.57 $\pm$ 7.97
	Val	1303	682	621	119.48 $\pm$ 7.54	93.84 $\pm$ 9.70	87.72 $\pm$ 6.51	88.39 $\pm$ 8.20
	Test	1306	638	668	119.84 $\pm$ 7.59	93.64 $\pm$ 9.56	87.84 $\pm$ 6.47	88.36 $\pm$ 8.03
2	Train	3891	1930	1961	119.78 $\pm$ 7.53	93.99 $\pm$ 9.43	87.87 $\pm$ 6.49	88.58 $\pm$ 7.99
	Val	1306	655	651	119.39 $\pm$ 7.59	93.82 $\pm$ 9.57	87.69 $\pm$ 6.47	88.33 $\pm$ 7.99
	Test	1303	682	621	119.48 $\pm$ 7.54	93.84 $\pm$ 9.70	87.72 $\pm$ 6.51	88.39 $\pm$ 8.20

TABLE I : Participant characteristics and cognitive-score distributions across the training, validation, and test partitions of the five family-aware cross-validation folds.

Fig. 1 : Schematic of the MAC-Net architecture. Regional tfMRI beta estimates ( $C$ conditions $\times$ 94 regions) are projected into a 64-dimensional embedding, flattened, and processed through a feed-forward backbone (256 $\rightarrow$ 128 $\rightarrow$ 64 units) with batch normalization, SiLU, and dropout. Participant and acquisition covariates bypass the tfMRI extractor via terminal late fusion. The primary N-back+MID configuration utilizes $C=19$ inputs; the N-back+MID+SST sensitivity ablation utilizes $C=26$ .

Fig. 2 : Distributions of continuous covariates and uncorrected NIH Toolbox cognitive outcomes. Histograms and kernel density estimates denote interview age, estimated total intracranial volume (eTIV), task-specific mean framewise displacement, and fluid, crystallized, and total intelligence scores. Categorical covariates (sex, MRI device) are omitted from visualization.

Target	Model / Modality Configuration	$R^{2}$ ( $\uparrow$ )	Pearson $r$ ( $\uparrow$ )	MAE ( $\downarrow$ )	MSE ( $\downarrow$ )
$G_{f}$	Covariate-Only Baselines
	– Ridge	0.113 $\pm$ 0.006	0.338 $\pm$ 0.008	7.091 $\pm$ 0.073	80.257 $\pm$ 1.652
	– Lasso	0.114 $\pm$ 0.006	0.340 $\pm$ 0.008	7.083 $\pm$ 0.075	80.112 $\pm$ 1.686
	– ElasticNet	0.113 $\pm$ 0.005	0.338 $\pm$ 0.008	7.089 $\pm$ 0.091	80.207 $\pm$ 1.867
	– Random Forest	0.087 $\pm$ 0.013	0.300 $\pm$ 0.018	7.207 $\pm$ 0.135	82.637 $\pm$ 2.738
	Single-Task MAC-Net with Covariates

TABLE II : Baseline modeling performance. The top panel for each target displays covariate-only machine learning baselines (using age, sex, eTIV, scanner, and motion). The bottom panel displays MAC-Net performance using isolated single-task tfMRI inputs along with these common covariates. Values reflect mean

\pm

standard deviation across five held-out test folds.

Target	Model Configuration	$R^{2}$ ( $\uparrow$ )	Pearson $r$ ( $\uparrow$ )	MAE ( $\downarrow$ )	MSE ( $\downarrow$ )
$G_{f}$	MAC-Net	0.174 $\pm$ 0.009	0.420 $\pm$ 0.009	6.855 $\pm$ 0.046	74.748 $\pm$ 1.373
	Flattened Machine Learning
	– Ridge	0.105 $\pm$ 0.020	0.357 $\pm$ 0.018	7.132 $\pm$ 0.100	80.939 $\pm$ 3.230
	– Lasso	0.150 $\pm$ 0.017	0.392 $\pm$ 0.019	6.947 $\pm$ 0.082	76.892 $\pm$ 2.321
	– ElasticNet	0.151 $\pm$ 0.015	0.395 $\pm$ 0.018	6.933 $\pm$ 0.105	76.814 $\pm$ 2.577
	– Random Forest	0.119 $\pm$ 0.018	0.348 $\pm$ 0.030	7.070 $\pm$ 0.163	79.718 $\pm$ 3.194

TABLE III : Performance comparison of MAC-Net against conventional machine-learning and advanced deep-learning baselines. All models utilize N-back+MID tfMRI inputs and identical covariates under a common family-aware cross-validation framework.

Task	Condition / Contrast	Psychological Process	Total Features
N-Back	0-back (Baseline)	Baseline attention and visual processing.	846
	2-back (Load)	High working memory load.
	2-back vs. 0-back	Pure working memory updating signature.
	Places	Visual processing of place stimuli.
	Emotional Faces	Visual processing of emotional stimuli.
	Face vs. Place	Category-specific visual processing.

TABLE S1 : Summary of task-fMRI conditions and functional contrasts included in the analysis. Beta estimates were extracted for each condition or contrast across 94 anatomical regions (comprising 68 Desikan–Killiany cortical parcels and 26 bilateral ASEG anatomical segments). This extraction yielded 846 features for N-back, 940 for the Monetary Incentive Delay (MID) task, and 658 for the Stop Signal Task (SST).

Target	Model	$R^{2}$ ( $\uparrow$ )	Pearson $r$ ( $\uparrow$ )	MAE ( $\downarrow$ )	MSE ( $\downarrow$ )
$G_{f}$	Ridge	0.045 $\pm$ 0.022	0.284 $\pm$ 0.022	7.382 $\pm$ 0.133	86.447 $\pm$ 3.617
	Lasso	0.087 $\pm$ 0.017	0.306 $\pm$ 0.020	7.200 $\pm$ 0.114	82.599 $\pm$ 2.753
	ElasticNet	0.090 $\pm$ 0.019	0.310 $\pm$ 0.024	7.183 $\pm$ 0.120	82.366 $\pm$ 3.027
	Random Forest	0.084 $\pm$ 0.008	0.293 $\pm$ 0.015	7.217 $\pm$ 0.137	82.904 $\pm$ 2.529
	MAC-Net	0.110 $\pm$ 0.010	0.341 $\pm$ 0.011	7.107 $\pm$ 0.125	80.532 $\pm$ 2.436
$G_{c}$	Ridge	0.093 $\pm$ 0.032	0.345 $\pm$ 0.029	4.819 $\pm$ 0.094	38.239 $\pm$ 1.439

TABLE S2 : Brain-only ablation performance using N-back+MID tfMRI inputs without participant or acquisition covariates. Values reflect mean

\pm

standard deviation across five held-out test folds.

Explore similar work

Oct 15, 2025cs.CV

Scaling Vision Transformers for Functional MRI with Flat Maps

We study the problem of training self-supervised foundation models for functional MRI. Our main contributions are: (1) we introduce a new model family (CortexMAE) trained using the masked autoencoder framework on 2.1K hours of open fMRI data, and (2) we release the first open evaluation suite (Brainmarks) for fMRI foundation models. Our core innovation is simple: we adapt the Vision Transformer to fMRI by first converting each 3D fMRI volume to a 2D map using a cortical flat map projection. We directly compare flat maps to both parcellation and volume-based representations. While each has its advantages, flat maps generally perform best. We perform the first systematic scaling analysis for fMRI and observe strict power law scaling, albeit with limits. Finally, we use Brainmarks to do controlled benchmark comparisons. On subject-level trait prediction, we report a challenging null result: no single model achieves clear state-of-the-art performance. Moreover, all models struggle to outperform a simple functional connectivity baseline. On cognitive state decoding, we observe more robust performance, and in this setting our CortexMAE family outperforms prior models by a large margin. Code, models, and datasets are available at https://github.com/MedARC-AI/CortexMAE and https://github.com/MedARC-AI/Brainmarks.

Connor Lane, Mihir Tripathy, Leema Krishna Murali +15

MedARC · Sophont · Baylor College of Medicine +3

Functional Magnetic Resonance Imaging Vision Transformer Transformer Architectures

Ask PDF

May 29, 2026q-bio.NC

The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail

Brain foundation models (BFMs) are self-supervised Transformers pretrained on fMRI data. We posit that these models should capture each subject's cognitive performance from their fMRI signal. Yet across three state-of-the-art BFMs and every readout we test, they predict cognition worse than a linear regression from the

\sim

80K parameters of the functional connectivity matrix (FC). The gap widens with scale: BrainLM's 650M model predicts cognition worse than its 111M. We attribute this to a \textbf{variance allocation problem}: BFM pretraining captures the variance components that dominate fMRI but not the higher-order structure that predicts cognition. Our per-cumulant analysis of the reconstructed signal shows that the second-order covariance is partially preserved, while the third-order co-skewness tensor is largely destroyed. To recover what BFMs lose, we design a linear pipeline that projects the fMRI signal into the subspace that best preserves its co-skewness and computes FC there. This \textbf{exceeds raw FC and every pretrained BFM} on every dataset and parcellation we test, outperforming prior state-of-the-art under controlled evaluation \textbf{with no pretraining and no GPU}. We \textbf{recover the raw-FC ceiling on BrainLM's forward pass} by finetuning with a loss targeted at this same subspace. This shows that the bottleneck is the pretraining objective, not the architecture or the model size.

Giovanni Marraffini, Gabriel Mahuas, Trinidad Borrell +2

Inria Saclay Île-de-France, CEA, Université Paris-Saclay, Palaiseau, France · Sigma Nova · Sorbonne Université, Institut du Cerveau - Paris Brain Institute - ICM +1

Functional Magnetic Resonance Imaging Wireless Foundation Models Cognitive Science

Ask PDF

Aug 12, 2026cs.CV

GenFAR: A generalized representation of brain structure, derived from 49,246 multi-cohort MRIs via deep learning

Deep learning models for neuroimaging have largely been developed for individual tasks, limiting knowledge transfer across applications. Here we introduce GenFAR, a modular deep learning framework that learns general, clinically informed features from brain MRIs. We trained this modular architecture on 49,246 individuals across 11 cohorts, using 17 diverse classification and regression tasks spanning cognition, clinical, diagnosis, demographics, and biomarkers. This yields aggregated, focused feature sets that capture rich, clinically- and biologically-relevant brain representations. We developed a sequential learning approach where tasks progressively build on previously learned representations. Through an analysis of 5,000 task sequences, we identified an optimal sequence length of six tasks and introduced a Donor Score metric to quantify each task's contribution to downstream performance. This analysis revealed five consistently strong donor tasks (Age, AD/MCI, MMSE, Hypertension, Hyperlipidemia) that formed the base of our sequential model. We demonstrated the utility of our learned representation, in various tasks beyond those included in the training set, to serve as the foundation for specialized secondary predictors. We further showed that using the learned feature representation can substantially increase the sample efficiency of secondary deep learning training tasks and models, as well as improve their accuracy.

Vishnu M. Bashyam, Guray Erus, Junhao Wen +29

Artificial Intelligence in Biomedical Imaging Lab, University of Pennsylvania · Department of Electrical and Systems Engineering, University of Pennsylvania · Florey Institute of Neuroscience and Mental Health, University of Melbourne +16

Functional Magnetic Resonance Imaging Deep Learning

Ask PDF