Validate before training
A dataset can load successfully and still be invalid. Repeated examinations from one patient can leak across splits; controls can contain unreported disease; derived files can be paired with the wrong source; and a dominant scanner can create a shortcut signal.
Validation layers
Validation combines automated checks with clinical and privacy review. Findings are recorded in a versioned report rather than corrected silently.
- Patient, study and series uniqueness with leakage prevention
- Required sequence or reconstruction completeness
- Image geometry, orientation, artefact and corruption checks
- Ground-truth source, missingness and reviewer agreement
- Hospital, scanner, protocol and demographic distributions
- De-identification exceptions and residual-risk review
Acceptance criteria and change control
The buyer and supplier agree on thresholds before the final cohort is assembled. Replacements, exclusions and new versions update the manifest and dataset card. This prevents a dataset from changing after evaluation without an audit trail.
Questions, answered directly.
Is model performance part of dataset validation?+
It can be a downstream acceptance test, but dataset validation first checks whether the cohort itself matches the agreed specification.
Why separate by patient rather than scan?+
The same patient can have repeated examinations. Splitting by scan can leak patient-specific signals into validation data.
What is an external validation cohort?+
A cohort kept independent from model development, ideally differing by site, scanner, time or population in a controlled way.