Interpretable Motion Artificat Detection in structural Brain MRI
A lightweight, interpretable model achieves 89% accuracy on unseen data while avoiding false positives.
A team of researchers has introduced a breakthrough framework for automated quality assessment of structural brain MRI scans, specifically targeting motion artifacts that can compromise neuroimaging analysis. The method, detailed in the arXiv preprint "Interpretable Motion Artifact Detection in structural Brain MRI," extends the Discriminative Histogram of Gradient Magnitude (DHoGM) technique into three dimensions. It uniquely integrates complementary 2D slice-level and 3D volume-level DHoGM features through a parallel decision strategy, capturing both localized and global motion-induced degradation. This hybrid approach allows the model to perform volumetric analysis using overlapping 3D cuboids, achieving comprehensive spatial coverage while maintaining high computational efficiency.
Remarkably, the researchers employed a simple threshold-based classifier and a low-parameter multilayer perceptron, resulting in an entire model with only 209 trainable parameters. This extreme lightweight design stands in stark contrast to typical deep learning models that require millions of parameters and extensive preprocessing. The framework was rigorously evaluated on the MR-ART and ABIDE datasets under both seen-site and unseen-site conditions, demonstrating strong generalization. It achieved up to 94.34% accuracy in-domain and maintained 89% accuracy on completely unseen acquisition sites, while almost completely avoiding the critical error of falsely accepting poor-quality scans. Ablation studies confirmed the complementary benefits of combining the 2D and 3D features, validating the core architectural insight.
The proposed solution directly addresses major challenges in neuroimaging: the high computational cost of existing methods and their poor generalization across different MRI machines and clinical sites. By being both interpretable—relying on handcrafted gradient features—and highly efficient, it offers a practical path for integration into large-scale clinical and research pipelines. This could significantly improve the reliability of downstream analyses in studies of neurological disorders, brain development, and aging by ensuring only high-quality, artifact-free scans are used.
- Achieves 94.34% in-domain and 89% cross-site accuracy on MRI artifact detection, demonstrating strong generalization.
- Uses a novel 3D extension of DHoGM features and a parallel decision strategy, resulting in a model with only 209 trainable parameters.
- Effectively avoids false acceptance of poor-quality scans, a critical requirement for reliable clinical and research analysis.
Why It Matters
Enables fast, reliable automated quality control for brain MRI across clinical sites, improving the integrity of neuroimaging research and diagnostics.