The CT archive is smaller than its file count suggests.
A CT_ filename does not mean a chest CT, and a chest CT is not automatically usable for quantitative work. Reading the real DICOM headers puts the analysis-ready cohort at roughly half the archive.
Source: anonymized UNOS DICOM archive · all 16,404 CT archives swept via DICOMDIR, plus a random 399 read at full header level.
~9,291of 16,404 archives
Only 56.6% (±4.9 pp) of CT archives carry a thin-slice chest series — the combination radiomics and 3-D CNN methods actually need. The headline count of 16,404 CT studies overstates the analysable set by roughly a factor of two.
01 · Scale
What the archive contains
The delivery manifest lists ~48,000 studies, but that counts every modality. Restricted to CT and resolved to STAR donor IDs, this is the set.
16,404
CT archives, over 8,811 donors
120,747
imaging series
16.4M
image instances
2,631
archives with no DICOMDIR index
02 · Body part
Not every CT is a chest CT
Donor imaging includes the head CT used in brain-death evaluation, plus abdominal and trauma studies. Study-level composition from the DICOMDIR sweep across all 13,761 indexed archives.
Share of archives by study type (%)
Two methods, two answers — use the deeper one. The cheap DICOMDIR sweep puts chest-containing archives at 62%, but full header reads find 83.0%. DICOMDIR descriptions are sparser than per-series headers, so the sweep misses chest series inside studies it labels otherwise. BodyPartExamined is missing outright in 38% of series, so keyword inference over study, series and protocol descriptions is mandatory rather than optional.
03 · Resolution
Most series are too thick for quantitative CT
Reconstruction slice thickness, over the 64% of series that record one. Thin-slice (≤1.5 mm) is the practical threshold for radiomics and 3-D convolutional models.
Series by slice thickness (mm, % of series)
thin-slicetoo thick
Just 29.4% of series clear the threshold; the median reconstruction is 3.0 mm. Combined with the body-part filter, that is what reduces 16,404 archives to roughly 9,291.
04 · What to do with it
Define the imaging cohort by header, not by filename
Three consequences for how VitalMatch selects and reports its training data.
Quote the usable count, not the file count. "16,404 donor CTs" is true of the archive and false of the analysable cohort. The defensible figure is ~9,291 archives carrying a thin-slice chest series, and it should be stated with its sampling interval.
Filter on headers at ingest. Body part and slice thickness both have to be read from the DICOM, since neither is recoverable from the filename and 38% of series omit BodyPartExamined entirely.
Close the 2,631-archive gap. Archives without a DICOMDIR are unlabelled by the cheap sweep. They are covered proportionally in the random sample, so the estimates above are unbiased — but there is no per-archive label for them yet.