Attention-Gated Convolutional Networks for Scanner-Agnostic Quality Assessment

Abstract: Motion artifacts present a significant challenge in structural MRI (sMRI), often compromising clinical diagnostics and large-scale automated analysis. While manual quality control (QC) remains the gold standard, it is increasingly unscalable for massive longitudinal studies. To address this, we propose a hybrid CNN-Attention framework designed for robust, site-invariant MRI quality assessment. Our architecture integrates a hierarchical 2D CNN encoder for local spatial feature extraction with a multi-head cross-attention mechanism to model global dependencies. This synergy enables the model to prioritize motion relevant artifact signatures, such as ringing and blurring, while dynamically filtering out site-specific intensity variations and background noise. The framework was trained end-to-end on the MR-ART dataset using a balanced cohort of 200 subjects. Performance was evaluated across two tiers: Seen Site Evaluation on a held-out MR-ART partition and Unseen Site Evaluation using 200 subjects from 17 heterogeneous sites in the ABIDE archive. On seen sites, the model achieved a scan-level accuracy of 0.9920 and an F1-score of 0.9919. Crucially, it maintained strong generalization across unseen ABIDE sites (Acc = 0.755) without any retraining or fine-tuning, demonstrating high resilience to domain shift. These results indicate that attention-based feature re-weighting successfully captures universal artifact descriptors, bridging the performance gap between diverse imaging environments and scanner manufacturers.

• The paper introduces an end-to-end CNN with channel-wise spatial self-attention that prioritizes motion-artifact signatures while reducing scanner-specific variation through minimal preprocessing.
• The model achieves 99.2% scan-level accuracy on MR-ART and transfers without retraining to 17 ABIDE sites, reaching 75.5% accuracy and 84.6% sensitivity.
• The findings suggest attention improves cross-scanner sensitivity, but larger cohorts, graded artifact labels, and controlled comparisons with prior attention-based MRI QC methods remain necessary.

Motivation and problem statement

Motion artifacts in structural MRI (sMRI) degrade both clinical diagnostics and downstream automated analyses, and manual quality control (QC) does not scale to large longitudinal efforts such as ABCD, ADNI, and UK Biobank. The authors identify a specific failure mode in the existing literature: both classical machine learning pipelines built on image quality metrics (IQMs) and deep CNN classifiers perform poorly on "unseen" sites—scanners absent from the training set (Kaur et al., 17 Mar 2026). Prior remedies either rely on multi-stage preprocessing (skull-stripping, tissue segmentation), nested cross-validation, or handcrafted histogram-based features. This paper addresses the generalization gap purely architecturally: a hybrid CNN–attention model trained end-to-end on MR-ART, evaluated without any retraining on 17 heterogeneous ABIDE sites.

Architecture

The framework is a three-stage pipeline operating on 2D axial slices:

1. CNN encoder: six convolutional layers arranged in three residual-style blocks with batch normalization and ReLU, max-pooling after the first two blocks, adaptive average pooling at the end, and channel expansion to 256. Pooled features are projected into 128-dimensional embeddings.
2. Attention module: a two-layer network (hidden dimension 256) applying channel-wise spatial self-attention independently per feature channel over flattened spatial dimensions—analogous to depthwise convolution but with content-dependent weights. Softmax-normalized weights redistribute importance across features, prioritizing artifact signatures such as ringing and blurring while suppressing low-variance site-specific background.
3. Classification head: an MLP ($256 \rightarrow 128 \rightarrow 64 \rightarrow 1$) with dropout regularization and sigmoid output producing a continuous motion-corruption probability.

Training uses BCE loss, Adam (learning rate $10^{-3}$, batch size 16), plateau-based LR decay, gradient clipping (max norm 1.0), and early stopping over 50 epochs. Preprocessing is deliberately minimal: the middle 50 axial slices are extracted, near-empty slices discarded, and each slice min-max normalized independently—a choice the authors argue forces attention to key on texture rather than global intensity, which varies across sites. Scan-level decisions use majority voting (>50% corrupted slices).

Results on seen sites

On a held-out MR-ART partition of 100 balanced subjects, the model reaches scan-level accuracy of 0.9920 with precision of 1.0000 and recall of 0.9840, and slice-level accuracy of 0.9859 with AUC-ROC of 0.9994. Against prior work on MR-ART, this outperforms MIA's 3D-CNN (84.17% accuracy despite training on four datasets), Naveetha et al.'s lightweight DHoGM+MLP pipeline (93%), and IJNS's slice-stack CNN (95% on 1000BRAINS). The comparison is favorable but should be read carefully: several baselines were evaluated on different test sets or trained on different data volumes, so the comparison is indicative rather than strictly controlled.

Generalization to unseen sites

The central claim concerns zero-shot transfer: trained only on MR-ART, the model achieves 75.5% scan-level accuracy and 84.6% sensitivity across 200 subjects from 17 ABIDE sites spanning multiple manufacturers and field strengths. Notably, the sensitivity exceeds the 77% reported by JMRI [jimeno2024automated], which was both trained and tested on ABIDE. The precision–recall asymmetry on ABIDE (precision 0.7249 vs. recall 0.8460) indicates the model errs toward flagging scans as corrupted, which is arguably the safer failure mode for QC triage.

The ablation study substantiates the architectural hypothesis. On ABIDE, a CNN plus classification head alone drops recall to 0.69, suggesting it latches onto MR-ART-specific patterns; adding attention recovers recall to 0.8460, though at a modest cost in accuracy relative to the CNN+Attention configuration (0.78). On MR-ART all configurations perform comparably, so the benefit of attention is specifically cross-domain sensitivity retention rather than in-distribution gains. Per-slice normalization is presented as critical to this behavior, though no ablation isolates its contribution separately from the attention module.

Limitations and open questions

The paper is candid about scope constraints. Training used only 100 MR-ART subjects (50 clean score-1, 50 corrupted score-3), discarding intermediate-quality score-2 scans and collapsing a graded artifact severity scale into binary labels; how the model behaves on moderate artifacts is not characterized. The unseen-site evaluation covers 200 subjects across 17 sites—a small sample given site heterogeneity—and the full ABIDE archive (900+ scans) remains untested. The claimed novelty ("first application of attention mechanisms to MRI QA") sits uneasily with the cited BrainQCNet work [garcia2024brainqcnet], which also employs attention for artifact detection; the distinction appears to lie in the specific cross-attention formulation and the unseen-site evaluation protocol, but the claim warrants qualification. Finally, the ablation compares only three configurations, and the interaction between per-slice normalization, majority-vote thresholding, and attention weighting is not disentangled.

Conclusion

This paper demonstrates that coupling a compact CNN encoder with channel-wise spatial self-attention yields near-ceiling performance on seen sites (0.9920 scan-level accuracy) and usable zero-shot transfer to 17 unseen scanner environments (0.755 accuracy, 0.846 sensitivity), with the ablation attributing the cross-site sensitivity retention specifically to the attention module. The results support attention-based feature re-weighting as a mechanism for capturing site-invariant artifact descriptors, though validation on larger cohorts, graded severity labels, and controlled comparisons against attention-based predecessors remain open.