A reviewed MRI workflow finds duplicate identities across datasets
In plain words
When brain scans from different studies are gathered together, the same person can appear under different ID numbers. HAPPEN is a workflow that helps repository curators find these cases before they confuse analyses.
First it matches scans with exactly the same image data. Then it finds similar scans under different records and asks a person to review them; accepted links are grouped. In one audit, it recovered all 54 duplicate pairs identified from genetic data, a result limited to that reference set.
The research
When separate MRI datasets are combined, can repository curators find people whose scans now appear under different identifiers? If they miss those links, one person’s scans may end up in both the training and test data used to assess a model.
HAPPEN combines exact scan matching, retrieval of similar cross-record scans, and human review to build repository-level groups of repeated identities.
After grouping exact image copies, HAPPEN retrieved non-identical cross-record scan pairs for human review; only accepted links joined repository subject groups. In a 21-dataset audit, reviewers supported 1,275 near-duplicate groups, and 82.4% connected records across datasets.
That shows the workflow found reviewed links at repository scale. It also placed both records together for all 54 pairs independently identified from genetic data.
The genetic comparison covered only 54 reference pairs, and neither that reference nor image review reliably distinguishes identical twins from one person recorded twice. The workflow was evaluated only on T1-weighted brain MRI and does not address multiple people sharing one identifier.
Analysis
Deep
- License
- arXiv non-exclusive