Skip to content
Imaging Evidence Path

What does a CT scanner actually measure?

Tonight · ~20 min · read · energy: low · setup: none

You open a CT series and see a clean axial slice of a chest: lung, ribs, soft tissue. It looks like a photograph of a thin cross-section. It is not. The single most important idea in this chapter is that no detector in a CT scanner ever measured that slice.

The anchor: what the detectors actually see

A CT gantry rotates an x-ray source around the patient. Opposite the source is a ring of detectors. At each of hundreds of angles, the source fires and the detectors record how much x-ray intensity made it through the patient along each line. What you collect is a set of projections — one-dimensional attenuation profiles from many directions. This raw collection is called the sinogram.

A detector element answers one question: how much radiation reached me after passing through this path through the body? It answers nothing about any particular point inside the body. A single projection cannot tell air from bone at a given location, because it integrates attenuation along the entire path.

flowchart LR
    S["X-ray source"] -->|beam through patient| P["Patient"]
    P --> D["Detector row<br/>(one projection)"]
    D --> R["Sinogram<br/>(all angles)"]
    R --> REC["Reconstruction<br/>algorithm"]
    REC --> IMG["Axial slice<br/>(computed pixels)"]

The slice you read is the third box’s output. The detectors produced the first.

Reconstruction is where the pixels come from

To get from projections to a 2D image of the slice, the scanner runs a reconstruction algorithm. The classical method is filtered back-projection (FBP): each projection is smeared back across the image plane, the contributions from all angles are summed, and a mathematical filter sharpens the result. Modern scanners increasingly use iterative or model-based reconstruction, which build the image by repeatedly comparing a candidate image against the measured projections under a noise/physics model.

The consequence that matters for everything downstream:

The pixel is an algorithmic output, and the algorithm makes choices. Change the choices and you get a different image from the same measured projections.

Those choices include the reconstruction kernel (a frequency filter), the slice thickness, the field of view, the matrix size, and the reconstruction algorithm itself. Two reconstructions of the same raw data are different images and can give different quantitative answers. This is not a footnote to skip — it is the single biggest source of radiomics non-reproducibility, and we devote a whole lesson to it.

So is CT “quantitative” or not?

Both, and the tension is the point. CT is quantitative in a way few other imaging modalities are: its reconstructed values map onto the Hounsfield scale, a defined scale where water is 0 HU and air is about −1000 HU. That mapping is why you can threshold bone, measure a tumour’s mean attenuation, or compare a value across patients. Chapter 8 exists largely because MRI has no such scale.

But “quantitative” does not mean “unambiguous”. The HU value at a voxel is the output of a reconstruction algorithm applied to noisy projections, filtered by a kernel, averaged over a slice thickness, at a dose and contrast phase. Each of those is a variable. CT is quantitative under a stated protocol; it is not a pure physical measurement of a single point.

A concrete feel for the numbers

Hounsfield units, the scale the reconstructed values are mapped onto:

Substance Approximate HU
Air −1000
Lung parenchyma −900 to −500
Fat −100
Water 0 (by definition)
Soft tissue −100 to +60
Acute blood / muscle +40 to +60
Bone +400 to +3000

Water ≡ 0 HU is a definition (ESTABLISHED EVIDENCE — the scale is conventional). The approximate tissue values are COMMON PRACTICE ranges, not exact — they shift with kVp, contrast, and partial volume. Use them as sanity checks, not as diagnostic thresholds.

Stop and think — then reveal

A CT study contains two series reconstructed from the same raw projections with different kernels. Are they “the same image” for a radiomics texture feature? Why does this single fact matter for reproducibility?

No — they are different images. The kernel is a reconstruction choice, so the same projections yield different spatial-frequency content and noise, and a texture feature (GLCM/GLRLM statistics) can differ substantially between them. This matters because it means a radiomic value is never a property of “the tumour” alone — it is a property of the tumour under a stated reconstruction. Pool series reconstructed differently and you are mixing different images as if they were one. (Shape features are largely kernel-independent; texture features are not — that asymmetry returns in Chapter 3.)

What to retain

  1. A CT scanner measures x-ray projections; the slice you read is reconstructed from them by an algorithm.
  2. A pixel is an algorithmic output that makes choices (kernel, thickness, algorithm). Same data, different choices → different image.
  3. CT is quantitative because reconstructed values map onto the Hounsfield scale — but “quantitative” means under a stated protocol, not “a pure measurement”.
  4. Because reconstruction is a choice, it is a variable you must record, not ignore — which is the whole subject of lesson 3.

Next: those HU numbers sound clean, but the value stored in the file is usually not HU at all — and that gap is where the first, most common quantitative error hides.