30. Quantitative Imaging

Volumetric measurement is the most direct expression of CT as a metrological instrument: because the reconstructed dataset is a regular three-dimensional lattice of voxels, each with known physical dimensions, the volume of any segmented structure is in principle simply the voxel count multiplied by the per-voxel volume. Formally, if $S$ is the set of voxels belonging to the structure and each voxel has in-plane dimensions $\Delta x,\Delta y$ and slice spacing $\Delta z$, then

$V = \sum_{(i,j,k)\in S}\Delta x\,\Delta y\,\Delta z = N_S\,\Delta x\,\Delta y\,\Delta z,$

the discrete approximation to the continuous integral $V=\iiint_{\Omega}\mathbf{1}_{\text{structure}}\,dV$. The accuracy of this estimate is governed almost entirely by how the boundary set $S$ is defined, and here the partial-volume effect dominates. Voxels straddling the interface between the structure and its surroundings contain a mixture of tissues, and the reconstructed attenuation of such a voxel is the volume-weighted average of its contents, $\mathrm{HU}_{\text{voxel}}=\sum_m f_m\,\mathrm{HU}_m$ with $\sum_m f_m=1$. A binary threshold therefore systematically includes or excludes boundary voxels depending on where the cut is placed, and the resulting bias scales with the surface-area-to-volume ratio; small or spiculated lesions, where this ratio is large, suffer the greatest fractional error. This is why near-isotropic acquisition ($\Delta x\approx\Delta y\approx\Delta z$) and consistent reconstruction kernels are prerequisites for reproducible volumetry, and why a fixed-threshold segmentation repeated on the same lesion with different slice thickness can disagree by tens of percent for sub-centimetre nodules.

The clinical payoff of volume over linear measurement is fundamentally geometric. Tumour-burden assessment by RECIST 1.1 deliberately uses the sum of longest diameters of target lesions, declaring progressive disease at a 20% increase and partial response at a 30% decrease in that sum, with the additional absolute requirement of a 5 mm rise to prevent small measurements from triggering false progression. For an idealized sphere, however, volume varies as the cube of diameter, $V=\tfrac{\pi}{6}d^{3}$, so a 20% diameter increase corresponds to $1.2^{3}\approx1.73$, a 73% volume increase, and the 30% diameter decrease of partial response corresponds to $0.7^{3}\approx0.34$, a 66% volume reduction. Diameter is thus a deliberately conservative, reproducibility-driven surrogate that lags true volumetric change. Volumetric tracking detects growth earlier because exponential tumour kinetics are linear in the logarithm of volume. The volume doubling time, the standard growth metric for pulmonary nodules under the Fleischner framework, follows from assuming exponential growth $V(t)=V_0\,2^{t/\mathrm{VDT}}$, which inverts to

$\mathrm{VDT}=\frac{(t_2-t_1)\,\ln 2}{\ln\!\big(V_2/V_1\big)}.$

A solid nodule whose volume rises from $V_1=180\,\mathrm{mm^3}$ to $V_2=320\,\mathrm{mm^3}$ over 90 days has $\mathrm{VDT}=90\ln2/\ln(320/180)\approx108$ days, well inside the malignant range (roughly 30–400 days), whereas the same lesion measured by diameter would have grown only from about 7.0 to 8.5 mm — a change near the limit of reader reproducibility. The price of this sensitivity is that volumetric software is itself a measurement device with its own repeatability limits, so an apparent volume change must exceed the technique's repeatability coefficient before it can be called real growth, a point developed quantitatively in the biomarker-development section. Modern practice extends voxel integration well beyond tumours: automated organ volumetry (liver remnant before hepatectomy, splenic and renal volumes), body-composition analysis (skeletal-muscle and visceral-fat cross-sectional areas predicting sarcopenia and surgical risk), and air-trapping or low-attenuation lung volumes all rest on the same summation of calibrated voxels, and all inherit the same partial-volume and segmentation dependencies.

Densitometry is the use of the reconstructed CT number as a quantitative physical measurement of tissue composition, and it is the oldest quantitative application of CT because it was built into Hounsfield's original calibration. Each voxel value is a linear rescaling of the local linear attenuation coefficient relative to water,

$\mathrm{HU}=1000\times\frac{\mu_{\text{tissue}}-\mu_{\text{water}}}{\mu_{\text{water}}-\mu_{\text{air}}},$

so that water is $0$ and air is $-1000$ by construction. The power of this scale is that, within its calibration, the number reports something about the tissue itself: fat is strongly negative (roughly $-50$ to $-120$ HU), most soft tissues cluster near $20$–$70$ HU, acute clotted blood is hyperdense ($50$–$80$ HU) because of the high electron density of haemoglobin protein, and calcium and cortical bone rise into the hundreds and thousands. The single most clinically entrenched densitometric threshold exploits intracytoplasmic lipid: a homogeneous, unenhanced adrenal nodule measuring $\le 10$ HU is diagnostic of a lipid-rich adenoma with high specificity, because abundant lipid drives the volume-averaged attenuation toward fat values; this one calibrated number resolves the majority of incidental adrenal masses without further work-up. The essential caveat — the recurring theme of all CT quantitation — is that HU is not an intrinsic tissue property but a polychromatic, beam-hardening-sensitive, energy-dependent estimate of $\mu$. Because the photoelectric cross-section of high-$Z$ elements such as iodine and calcium varies steeply with photon energy (roughly as $Z^{3\text{–}4}/E^{3}$), the measured HU of a calcified plaque or an opacified vessel changes substantially with tube potential, patient size, and reconstruction kernel. A density threshold validated at 120 kVp with a specific kernel cannot be transplanted unexamined to an 80 kVp or virtual-monoenergetic image, a fact that spectral and photon-counting CT both exploit (deliberate energy selection) and must control for (cross-protocol comparability).

Beyond single-voxel thresholds, densitometry becomes statistical when applied to whole organs. Pulmonary emphysema is quantified from the histogram of lung attenuation: the low-attenuation-area index reports the fraction of lung voxels below a threshold (classically $\mathrm{LAA}_{-950}$, the percentage below $-950$ HU on inspiration), and the complementary 15th-percentile point, $\mathrm{Perc}15$, reports the HU value below which 15% of lung voxels fall. These indices were validated against macroscopic and microscopic morphometry — Gevenois and colleagues showed that CT density measurements correlate with the pathological extent of emphysema — establishing densitometry as a genuine in-vivo surrogate for tissue destruction rather than a mere image statistic. Their interpretation is acutely sensitive to inspiratory level, because lung density is the mass of tissue divided by the air-inflated volume; identical parenchyma scanned at different lung volumes yields different attenuation, so volume normalization or spirometric gating is required for longitudinal comparison. The most consequential densitometric biomarker, however, is the coronary artery calcium score. The Agatston method, defined on the original ECG-gated calcium-scoring acquisition, identifies calcified plaque as voxels exceeding $130$ HU, then weights each lesion's area by a density factor derived from its peak attenuation (1 for 130–199 HU, 2 for 200–299, 3 for 300–399, 4 for $\ge 400$), summing across all lesions:

$\text{Agatston} = \sum_{\text{lesions}} A_{\text{lesion}}\times w(\mathrm{HU}_{\max}).$

This density-weighted score is among the best-validated prognostic biomarkers in all of imaging: in the Multi-Ethnic Study of Atherosclerosis, the calcium score independently predicted incident coronary events across ethnic groups and substantially reclassified risk beyond conventional factors, with a score of zero conferring a very low near-term event rate and rising scores tracking steeply increasing hazard. Its existence depends entirely on the calibrated, reproducible HU scale, and its standardization — fixed $130$ HU threshold, $2.5$–$3$ mm slices, prospective gating — is precisely what makes scores comparable across machines and studies, a discipline that the unstandardized densitometric measurements above still struggle to achieve.

CT perfusion converts the modality from a map of static anatomy into a map of tissue physiology by tracking the first pass of an iodinated contrast bolus through a volume with rapid, repeated scanning. Because the CT number rises approximately linearly with local iodine concentration over the diagnostic range, each voxel yields a time–attenuation curve that is a surrogate for the tissue contrast concentration over time, $C_{\text{tissue}}(t)$, while a curve drawn over a feeding artery provides the arterial input function $C_a(t)$. The haemodynamic parameters follow from indicator-dilution theory. The central-volume principle relates the three primary quantities — cerebral (or tissue) blood flow $\mathrm{CBF}$, blood volume $\mathrm{CBV}$, and mean transit time $\mathrm{MTT}$ — by

$\mathrm{CBF}=\frac{\mathrm{CBV}}{\mathrm{MTT}},$

so that any two determine the third. Blood volume is recovered from the ratio of the areas under the tissue and arterial curves, $\mathrm{CBV}\propto \int C_{\text{tissue}}\,dt\,/\!\int C_a\,dt$, while flow and transit time require disentangling the tissue response from the shape of the input bolus. This is a deconvolution problem: the tissue concentration is the convolution of the arterial input with a flow-scaled residue function $R(t)$, the fraction of injected tracer still present in the voxel at time $t$,

$C_{\text{tissue}}(t)=\mathrm{CBF}\,\big(C_a * R\big)(t)=\mathrm{CBF}\int_0^{t}C_a(\tau)\,R(t-\tau)\,d\tau.$

Estimating $R(t)$ by inverting this relation — most commonly with singular-value decomposition, in either a standard or delay-insensitive (block-circulant) form, or with Bayesian estimators — yields CBF as the peak height of $\mathrm{CBF}\cdot R(t)$ and MTT from its first moment. A fourth, widely used parameter is $T_{\max}$, the time to the peak of the deconvolved residue function, which captures bolus delay and dispersion and is robust to be computed without a precise absolute flow calibration. Because deconvolution is mathematically ill-posed, the smoothing (regularization) imposed to stabilize it directly shapes the parameter values, which is the technical reason perfusion outputs differ between software packages and algorithms.

The clinical reasoning that this machinery supports is the operational distinction between irreversibly infarcted core and salvageable penumbra in acute ischaemic stroke, and the mathematics maps cleanly onto pathophysiology. Autoregulation initially defends flow as perfusion pressure falls by dilating the vasculature, which raises CBV and prolongs MTT while CBF is maintained; once this reserve is exhausted, CBF and CBV both fall and tissue moves toward infarction. The core is therefore characterized by severely reduced CBF and CBV, whereas the penumbra shows prolonged MTT and $T_{\max}$ with relatively preserved CBV — the perfusion signature of tissue that is hypoperfused but not yet dead. Quantitative thresholds operationalize this: the ischaemic core is commonly defined as relative CBF below roughly 30% of the contralateral normal tissue ($\mathrm{rCBF}<30\%$), and critically hypoperfused tissue as $T_{\max}>6\,\mathrm{s}$, with the penumbra (and hence the target mismatch) estimated as the $T_{\max}>6\,\mathrm{s}$ volume minus the core volume. These exact thresholds are not arbitrary; they are the selection criteria validated in the late-window thrombectomy trials, in which automated perfusion software (using these rCBF and $T_{\max}$ cut-points and a mismatch ratio) identified patients with small cores and large penumbras who benefited from reperfusion 6–16 hours after onset. The decisive expert caveat is that these numbers are software- and algorithm-specific: because the underlying deconvolution and its regularization differ between vendors, a given absolute threshold does not transfer between platforms, and reported core volumes can diverge for the same patient depending on the package used. CT perfusion is thus a paradigm of the chapter's central tension — a genuinely physiological, decision-changing measurement whose validity is inseparable from the standardization of how it is computed.

A quantitative imaging biomarker is an objectively measured characteristic, derived from an image, that serves as an indicator of a normal biological process, a pathological process, or a response to therapy — and the discipline of biomarker development is precisely what separates a defensible biomarker from a number that merely happens to be computable. The governing framework is metrological, articulated for imaging by the Quantitative Imaging Biomarkers Alliance (QIBA) and codified in its technical-performance statistical methods. Three properties must be established before a measurement can be trusted to track biology. The first is technical accuracy, decomposed into bias and linearity: bias is the systematic deviation of the measured value from a known truth, and linearity is the consistency of that relationship across the measurand's range, jointly assessed against reference phantoms or ground truth. The second is repeatability — the variability of repeated measurements under identical conditions (same scanner, protocol, operator, and subject in the test–retest interval) — quantified by the within-subject standard deviation $w$ or the within-subject coefficient of variation $\mathrm{wCV}=w/\mu$. The single most clinically actionable derived quantity is the repeatability coefficient,

$\mathrm{RC}=1.96\sqrt{2}\,w\approx 2.77\,w,$

the magnitude of change between two measurements of an unchanged subject that will be exceeded by chance only 5% of the time. This is the quantitative threshold that makes longitudinal imaging interpretable: an observed change in tumour volume, lung density, or coronary calcium must exceed the RC of its measurement technique before it can be attributed to real biological change rather than measurement noise — the rigorous justification for the conservative thresholds embedded in RECIST and nodule-growth criteria. The third property is reproducibility, the variability when conditions deliberately change (different scanners, vendors, sites, or reconstruction settings); the intraclass correlation coefficient, $\mathrm{ICC}=\sigma^{2}_{\text{between}}/(\sigma^{2}_{\text{between}}+\sigma^{2}_{\text{within}})$, summarizes the fraction of total variance attributable to true between-subject differences rather than measurement error, and a biomarker with poor reproducibility cannot support multi-centre trials or guideline thresholds however good its single-site repeatability.

The contemporary expansion of this field is radiomics — the high-throughput extraction of large numbers of quantitative descriptors (first-order histogram statistics, shape, and texture features such as those from grey-level co-occurrence and run-length matrices) from segmented regions, on the hypothesis that these features encode sub-visual phenotype correlated with genomics, prognosis, or treatment response. The canonical articulation, that images are more than pictures and are in fact mineable data, frames the standard pipeline: reproducible segmentation, feature extraction, feature selection, model building, and — the step that most often fails — independent validation. The threats to validity are severe and specific. The number of candidate features routinely exceeds the number of patients, so models overfit, discovering spurious associations that do not generalize; rigorous practice therefore demands correction for multiple comparisons, internal cross-validation, and ideally external validation on data from other institutions and scanners. Feature values are exquisitely sensitive to acquisition and reconstruction — voxel size, kernel, dose, and vendor — producing batch (scanner) effects that can dominate biological signal unless controlled by acquisition standardization, voxel resampling, intensity discretization, and post hoc harmonization methods. The discipline's hard-won lesson, and the unifying message of quantitative imaging, is that computability is not validity: a feature or index becomes a biomarker only after its bias, repeatability, and reproducibility are characterized, its association with the biological or clinical endpoint is demonstrated to be robust, and its performance is shown to generalize beyond the data that generated it. Absent that evidentiary chain, a quantitative output is a hypothesis, not a measurement on which a patient's care may be staked.