QC your neuroimaging cohort.

Cortex has seen 1,545 brain scans during training. For each of your 40 uploaded scans, it tried to recreate the scan from its learned understanding. If a scan matches what Cortex expects, it reconstructs accurately. If the scan is unusual, Cortex struggles — and that reconstruction error is your outlier signal.

Think of it like a music teacher who has heard thousands of students play a piece. When a new student plays, the teacher can tell how "typical" the performance is. A wildly unusual performance doesn't mean bad — it might be brilliant, or it might be a technical mistake. But it warrants a closer listen.

23 clean 10 borderline 7 outlier — that's 43% flagged for review.

18% outliers is moderate. Worth reviewing — flagged scans likely have motion artifacts, registration issues, or novel phenotypes. Spot-check 2-3 outliers first before assuming a pipeline bug.

For each outlier, check in order: (1) head motion logs — framewise displacement > 0.3 mm is a red flag. (2) preprocessing — did registration complete? are there masking errors? (3) scan duration — too short (<5 min) produces unreliable connectivity. (4) phenotype — if technical checks pass, the subject may have a genuinely unusual neural pattern worth separate analysis.

Atypical does not equal "bad" — Cortex's training data was ABIDE (mixed autism/typical). A scan from a completely different population (e.g., stroke patients, children under 6) will look atypical because Cortex has never seen that phenotype. Outlier flags are not quality judgments; they flag differences from training distribution.