Synthetic Motion-Corrupted MRI Volumes for Training a 3D CNN Motion Estimator

A method and associated training data for estimating head motion in structural MRI scans. The approach uses a 3D convolutional neural network trained on a large dataset of synthetically motion-corrupted volumes, validated across 14 independent datasets. The final model is available as the open-source tool 'Agitation'.

Use Cases

• Training motion estimation models based on synthetically corrupted MRI volumes described in the paper.
• Assessing potential biases in cortical thickness and volume analyses based on predicted motion scores.
• Validating automated quality control tools across different scanner brands and protocols as demonstrated in the validation.
• Studying the correlation between subject age and motion artifacts in neuroimaging data.

Strengths

• Model validation achieved a Spearman Rank correlation of 0.71 versus manual motion labels.
• Method was tested and shown to generalize across 14 fully independent datasets.
• The final model and tool ('Agitation') are made openly and freely available under a CC-BY-4.0 license.

Limitations

• The dataset's specific row count and column-level documentation are unknown.
• The primary artifact is a 9.6 MB PDF describing the method; the underlying training data files are not explicitly listed.
• Data may reflect bias inherent to the specific research scans used for training and validation.

Provenance

Source

Charles Bricout via figshare.

Collection Method

Leverages a large training dataset of synthetically motion-corrupted MRI volumes.

Freshness

Last updated 2026-05-08 05:55:48.