The Header Fields That Break Quantitative Analysis Across Sites

When a quantitative imaging study fails to replicate across sites, the post-mortem usually reaches for field strength, vendor or slice thickness. Those are visible, they are easy to tabulate, and they are rarely the whole story. The differences that survive matching on the obvious parameters are generally encoded in fields nobody stratified on.

Reconstruction kernel, the largest uncontrolled CT variable

Convolution Kernel (0018,1210) records the reconstruction filter. It is a short vendor-specific string — B30f, STANDARD, FC18 — with no cross-vendor mapping and no ordering defined anywhere in the standard.

Kernel choice changes the noise power spectrum and the edge response of the reconstructed image directly. Any texture feature, any measure of heterogeneity, and any small-structure segmentation is sensitive to it, often more sensitive than to slice thickness. Two sites both scanning at 1 mm with the same tube current can produce systematically different texture distributions purely through kernel selection.

Because the strings do not map across vendors, the honest options are to restrict a cohort to a single kernel family per vendor, or to record kernel as a covariate and accept that the model is fitted, not eliminated. Harmonising after the fact treats a deterministic instrument setting as though it were random noise.

Rescale attributes, and the assumption that they were applied

Rescale Slope (0028,1053) and Rescale Intercept (0028,1052) convert stored pixel values into the modality’s output units. The conversion is elementary; the failures are procedural.

Two recur. The first is a pipeline that applies rescale in one branch and not another — typically because one path reads DICOM directly and another reads a NIfTI conversion in which the transform was already baked in. Applying it twice is as wrong as never applying it, and both produce plausible-looking images.

The second is Value of Interest LUT handling. Window Center (0028,1050) and Window Width (0028,1051) are display parameters and must not enter a quantitative path at all, but conversion tools sometimes apply them, silently clipping the dynamic range before any measurement is taken. A cohort assembled from mixed conversion tooling can contain both clipped and unclipped versions of the same distribution.

The check is the one in what DICOM metadata guarantees: transform, then confirm the result lands in a physically plausible range before anything downstream sees it.

PET units, which are a declaration and not a guarantee

Units (0054,1001) declares what a PET pixel represents — BQML, CNTS, GML and so on. SUV computation depends on it absolutely, and on a cluster of other attributes: Radionuclide Total Dose (0018,1074), Radiopharmaceutical Start DateTime (0018,1078), Radionuclide Half Life (0018,1075), and Patient Weight (0010,1030).

Every one of these is a place where a site convention can quietly diverge. Decay correction reference time differs between vendors. Patient weight is sometimes the weight at a previous visit. Injected dose may be recorded as assayed or as residual-corrected. None of this is visible in the image, and all of it moves SUV by amounts comparable to the effects a study is trying to detect.

A cohort intended for SUV analysis should carry an explicit per-series record of which of these attributes were present, which were imputed, and from what.

Diffusion b-values in classic objects

Enhanced MR objects have a standard attribute for the diffusion b-value, (0018,9087). A great deal of clinical diffusion data is still classic single-frame, where the b-value lives in manufacturer-private tags with different locations and different encodings per vendor.

This makes b-value extraction vendor-conditional by necessity. The risk is not that the code is hard to write; it is that a pipeline handling three vendors correctly and a fourth incorrectly will still produce ADC maps for all four. A wrong b-value yields a plausible map with a systematically wrong scale, and nothing downstream will flag it.

Any multi-vendor diffusion cohort needs a per-vendor extraction test with a known-answer case, not merely an assertion that extraction succeeded.

Field strength as a string

Magnetic Field Strength (0018,0087) is a decimal string. Grouping on it without numeric coercion splits 3 from 3.0, and the actual value on a nominal 3 T system is often 2.89. Cohorts stratified naively by this field routinely contain a “3 T” group and a “2.89 T” group that are the same scanner population.

The fix is trivial. It is listed because it survives review so often: the code looks correct, the strings are exactly what the header contains, and the bug is entirely in the assumption that two representations of the same number compare equal.

What to record at ingestion

The through-line is that none of these are recoverable later. Once a cohort is converted, harmonised and analysed, the evidence of what was in the original headers is gone unless it was captured deliberately.

A defensible ingestion step records, per series: the raw values of every attribute above; whether each was present, absent or empty; the tool and version that performed any conversion; and the result of the plausibility checks. That record is what lets a reviewer ask whether a cross-site difference is biology or instrumentation — which is the same argument made for AI validation cohorts in metadata provenance in imaging AI validation sets.