A diffusion tensor imaging report is full of numbers. Fractional anisotropy values, percentiles, z-scores. Every one of them is a comparison, and a comparison is only as trustworthy as the reference it is measured against. That reference is the normative database, and it is the quiet foundation the entire study stands on.
A DTI value is a comparison, not a picture
Diffusion tensor imaging is not an image you simply read. It is a quantitative measurement. When a report states that a white matter tract sits at the 4th percentile, it is not describing how the tract looks. It is stating where its fractional anisotropy falls against what that value should be. Remove the reference and the number loses its meaning. An FA of 0.42 is neither good nor bad until you know what 0.42 is supposed to be for this tract, in this kind of patient, on this scanner.
Normal is not universal. It belongs to the scanner.
Here is the part that surprises people. Normal is not a fixed, published value you can look up once and apply everywhere. It is specific to the scanner, the field strength, the acquisition protocol, and the site. Two magnets can image the same healthy brain and return different numbers, because the hardware, the coils, the gradients, and the sequence all shape the measurement. A value that is completely unremarkable on one 1.5T machine can land at the 3rd percentile on another, for reasons that have nothing to do with the patient's brain and everything to do with the equipment.
What borrowed normals quietly cost
When a scanner has no normative database of its own, the only option is to borrow another site's reference set. On the surface the report looks the same. The percentiles still populate, the z-scores still calculate, the impression still reads cleanly. Underneath, every one of those numbers is being measured against the wrong population.
That is what makes the problem dangerous. It is invisible. A perfectly healthy tract can be flagged as abnormal, or a genuine injury can be smoothed into a normal-looking range, purely because the yardstick came from a different machine. Nothing in the report announces the error. It simply produces a confident number built on the wrong foundation.
A measurement is only as good as what you compare it to. Borrow the comparison and you inherit its uncertainty along with it.
What a site-specific normative database actually takes
The reassuring part is that building one is straightforward, and far less work than most sites expect. A qualified normative database for a scanner needs three things:
- A small set of healthy volunteers scanned on that exact scanner and protocol. A handful is enough to begin, with steady growth over time to keep the reference current.
- A screening questionnaire for every volunteer, covering prior head trauma or loss of consciousness, neurological and psychiatric history, medications, substance use, and MRI safety and data quality. Any scan that shows a clinical finding is set aside, no matter how the volunteer answered.
- A short acquisition. A 3D T1 and a 30-direction DTI sequence, roughly five to eight minutes of table time. Several volunteers can often be completed in a single hour, and they can be recruited from patients already on site for unrelated studies, as long as they have never sustained a traumatic brain injury.
Once those first cases are in hand, the database is live, and every report that follows is measured against the scanner's own normals rather than a stranger's.
The payoff
When a scanner has its own reference set, the numbers finally mean what they say. A percentile stops being an educated guess borrowed from elsewhere and becomes a real statement: this measurement, on this machine, for a person like this, sits here. That is the difference between a number you can stand behind and a number you are hoping is right.
Trustworthy DTI does not start at the read. It starts at the reference. Get the normative database right and everything downstream, the percentiles, the comparisons, the conclusions, inherits that same solid ground.