Skip to content
Imaging Evidence Path

CT and DICOM — reference

Reference · Dense lookup material — return here later, do not learn from it cold. The teaching versions live in the lessons.

The identifiers (lookup)

Tag Name Scope Why it matters
(0010,0020) PatientID Patient De-identification/site key; grouping. Only meaningful within its assigning authority — pair with IssuerOfPatientID (0010,0021) across sites.
(0020,000D) StudyInstanceUID Study “This exam”
(0020,000E) SeriesInstanceUID Series “This acquisition/reconstruction” — the unit of work for a volume
(0008,0018) SOPInstanceUID Instance “This slice” — the join key a SEG uses to reference its source images
(0020,0052) FrameOfReferenceUID Spatial frame Names the patient coordinate frame. A match lets you compare geometry; it does not prove identical grids.
(0008,0016) SOPClassUID Object type What kind of object this is before you parse it
(0008,0060) Modality Object type CT, SEG, MR, OT, … — the one-tag filter

Intensity / pixel tags

Tag Name Purpose
(0028,1052) RescaleIntercept Modality LUT offset; HU = stored × slope + intercept
(0028,1053) RescaleSlope Modality LUT slope (often 1)
(0028,0100) BitsAllocated typically 16
(0028,0103) PixelRepresentation 0 = unsigned, 1 = signed
(7FE0,0010) PixelData the raw pixel buffer (bulk OW/OB)
(0018,1210) ConvolutionKernel the reconstruction kernel (texture variable)

Geometry tags (full treatment in Chapter 2)

Tag Name Purpose
(0020,0032) ImagePositionPatient (x,y,z) of the top-left voxel centre, LPS mm
(0020,0037) ImageOrientationPatient row and column direction cosines
(0028,0030) PixelSpacing [row spacing, column spacing]
(0018,0050) SliceThickness collimated/imaged slab
(0018,0088) SpacingBetweenSlices distance between slice centres

SOP Classes you will meet

  • 1.2.840.10008.5.1.4.1.1.2 — CT Image Storage (single CT slice).
  • 1.2.840.10008.5.1.4.1.1.66.4 — Segmentation Storage (DICOM SEG).
  • 1.2.840.10008.5.1.4.1.1.481.3 — RT Structure Set Storage (RTSTRUCT).

Value Representation (VR) — useful level

Each data element has a two-letter type code. You do not need the full table; you need three facts:

  • Numeric values are often stored as text (DS, decimal string) — so "0.78125\0.78125" is a normal way to store pixel spacing.
  • UIDs (UI) are dot-separated strings; compare them as strings.
  • Pixel Data is bulk (OW/OB) in (7FE0,0010).

Common VRs: PN (person name), DA (date), UI (UID), US (unsigned short), DS (decimal string), OB/OW (other byte/word).

Transfer Syntax

TransferSyntaxUID (0002,0010) is the byte-level encoding of Pixel Data: uncompressed (e.g. Explicit VR Little Endian 1.2.840.10008.1.2.1), JPEG-lossy, JPEG-2000, etc. pydicom decodes most of these for you; if you ever read raw PixelData bytes you must know the transfer syntax first.

Classic vs Enhanced CT

The “one instance = one slice” model is classic single-frame CT Image Storage. Enhanced CT is a multi-frame object where one SOP Instance contains many frames, each described by per-frame functional group macros (the same mechanism DICOM SEG uses). Geometry concepts are identical; only packaging differs.

What usually goes wrong

  • Treating stored pixels as HU without applying slope/intercept — the most common silent error.
  • Assuming uint16 everywhere — read PixelRepresentation and BitsAllocated.
  • Trusting InstanceNumber or filename order for anatomical z-order. Order by geometry (Chapter 2).
  • Pooling reconstructions — a study contains several series (thin/thick axial, coronal reformats, scout). A localizer/scout is a 2D planning image, not an axial slice; stacking it corrupts the volume. Filter on ImageType / ConvolutionKernel / SliceThickness.
  • Ignoring transfer syntax when reading raw bytes — you get garbage or a decompression error. Let pydicom handle it; check for a missing package on compressed syntaxes.

A practical reading sequence (curated external)

Do not read the whole DICOM standard. Read these in order and stop when the model above becomes concrete:

  1. “DICOM is Easy” (Ronnie) — dicomiseasy.blogspot.com. The canonical gentle introduction: the Patient→Study→Series→Instance intuition from a real CT instance, tag by tag. Skip the networking posts (that is Chapter 5).
  2. pydicom docs — pydicom.github.io. “Dataset basics”/“elements” + the Pixel Data tutorial cover ~90% of what you need (dcmread, ds.pixel_array, RescaleSlope/Intercept). Skip creating new SOP instances.
  3. Innolitics DICOM Standard Browser — dicom.innolitics.com. Use as a reference, not a textbook; keep it open permanently.
  4. The DICOM standard (normative) — dicom.nema.org/standard.html. Read on demand: PS3.3 Image Plane Module / Image Pixel Module / Modality LUT. Skip PS3.4/5/7/8 (services/encoding/networking) until imaging informatics.

TRACE-CT connection

Every concept in this chapter is exercised by real code in TRACE-CT (read-only — do not copy it here):

  • HU conversion — scripts/seg_decode_resample.py load_ct_volume_explicit() applies volume[k] = pixel × slope + intercept while building the volume.
  • Series grouping & metadata — scripts/rider_preflight.py classifies each object by SOPClassUID/Modality, then builds per-series metadata (kernel, kVp, SliceThickness, ImageType, PixelSpacing, IOP/IPP).
  • Reconstruction selection — rider_preflight.py selects both a 1.25 mm LUNG and a 5.0 mm STANDARD reconstruction per acquisition — the two-reconstructions-of-one-acquisition effect, made real.
  • HU clipping before features — rider_radiomics_extraction.py clip_intensities() clips to [-1000, +400] HU after resampling and before PyRadiomics (it operates in physical HU, not stored pixels).
  • Localizer filtering & geometry validation — load_ct_volume_explicit() drops localizers and validates orientation/spacing regularity before declaring a stack a usable volume.

Go deeper

  • Innolitics entries for Rescale Slope/Intercept, Image Plane Module, Image Pixel Module.
  • DICOM PS3.3 Image Plane / Image Pixel / Modality LUT modules — the normative source for geometry and slope/intercept semantics.
  • PyRadiomics docs on image type and binWidth — why working in HU (not stored pixels) is mandatory before feature extraction (Chapter 3).
  • Bushberg, Essential Physics of Medical Imaging — for CT reconstruction and dose depth when reconstruction/dose becomes an active TRACE-CT variable.

What to retain (chapter summary)

  1. CT measures projections, reconstructs slices; the stored pixel is an algorithmic output, not raw physics.
  2. HU = stored_pixel × RescaleSlope + RescaleIntercept. Never use stored pixels as HU.
  3. Patient → Study → Series → Instance; geometry and reconstruction live at the Series level.
  4. The hierarchy organises data but does not define volume geometry — that needs IOP/IPP/spacing per instance (Chapter 2).
  5. Reconstruction kernel and slice thickness are quantitative variables: two reconstructions of one scan are different images and can give different features.