Diffusion: DWI, ADC & DTI

Imaging Motion That Is Invisible to the Eye

Every water molecule in tissue is in ceaseless thermal motion, jostled by collisions with its neighbors. Over a few tens of milliseconds a single proton wanders only a handful of microns, yet that microscopic random walk turns out to be one of the most clinically powerful contrasts in all of MRI. Diffusion-weighted imaging (DWI) made cytotoxic edema visible within minutes of an ischemic stroke, decades before any other modality could reliably do so. The same physics, generalized to three dimensions, lets us map the orientation of white-matter tracts in the living brain.

The central trick is that diffusion is incoherent motion. Unlike flowing blood, where spins move together, diffusing spins scatter in all directions. A gradient pair that perfectly rephases stationary spins cannot rephase spins that have moved randomly, and the residual phase dispersion produces signal loss. By controlling how strongly we sensitize the signal to that loss, we convert an unmeasurable micron-scale displacement into image contrast that radiologists read every day.

The Physics of the Random Walk

Free self-diffusion is described by the Einstein relation, which connects the mean squared displacement of a molecule to a single material constant, the diffusion coefficient $D$. In one dimension over a time $t$ the root-mean-square displacement grows as the square root of time, not linearly as it would for coherent flow.

For free water at body temperature $D \approx 3 \times 10^{-3}$ mm squared per second. Plugging a typical diffusion time of $t \approx 40$ ms into Eq. 9.1 gives a displacement of about $\sqrt{2 \cdot 3\times10^{-3} \cdot 0.04} \approx 15$ micrometers. That is roughly the scale of a cell, which is precisely why DWI is exquisitely sensitive to cellular architecture: membranes, organelles, and macromolecules are the obstacles that hinder the walk.

Sensitizing the Signal: the Stejskal-Tanner Gradients

Consider a spin echo with two identical gradient lobes of amplitude $G$ and duration $\delta$, one applied before the 180 degree refocusing pulse and one after, separated by an interval $\Delta$. A stationary spin accumulates phase $+\gamma G \delta x$ from the first lobe; the 180 degree pulse flips its sign, and the second lobe adds $+\gamma G \delta x$ again so the net phase is zero. The spin is perfectly rephased. A spin that moves between the two lobes experiences different field at each position, so the two phases no longer cancel. Averaged over the random displacement distribution, this residual phase scatter attenuates the echo.

Carrying out the average over a Gaussian displacement distribution yields the signature exponential attenuation. The single parameter that bundles together all of the gradient timing is the b-value, with units of seconds per mm squared.

The Apparent Diffusion Coefficient and ADC Maps

In tissue the diffusion is not free: membranes, fibers, and macromolecules restrict and hinder the molecules. We therefore cannot measure the true intrinsic $D$, and we call the effective value we observe the apparent diffusion coefficient (ADC). The word apparent is a confession: ADC lumps together genuine restriction, tortuosity, exchange, and even microcirculation, and it depends on the diffusion time and direction we chose to probe.

Measuring ADC requires at least two b-values. Acquiring images at $b_1$ and $b_2$ (commonly $b_1 = 0$ and $b_2 \approx 1000$ s/mm squared) and taking the ratio removes $S_0$, so a simple logarithm gives ADC voxel by voxel.

The resulting ADC map is a quantitative parametric image: each pixel value is a diffusion coefficient, conventionally displayed so that high ADC (free water, like cerebrospinal fluid) is bright and low ADC (restricted, like an acute infarct) is dark. Normal adult brain white matter has ADC roughly $0.7$ to $0.8 \times 10^{-3}$ mm squared per second; cortical gray matter is slightly higher; CSF approaches the free-water value near $3 \times 10^{-3}$. An acute infarct typically falls to about $0.3$ to $0.5 \times 10^{-3}$.

| Tissue / fluid | Approx. ADC (1e-3 mm2/s) | Appearance on ADC map | | --- | --- | --- | | CSF / free water | about 3.0 | very bright | | Normal gray matter | 0.8 to 0.9 | intermediate | | Normal white matter | 0.7 to 0.8 | intermediate to dark | | Acute infarct core | 0.3 to 0.5 | dark (restricted) | | Cellular tumor (e.g. lymphoma) | 0.5 to 0.7 | dark (restricted) | | Vasogenic edema | 1.2 to 1.8 | bright (facilitated) |

Restricted Diffusion versus T2 Shine-Through

The most common interpretive trap in DWI is that the diffusion-weighted image is never purely diffusion-weighted. Because the underlying sequence is a spin echo (or a long-TE echo planar readout), the $b = 0$ image and therefore $S_0$ in Eq. 9.3 carry heavy T2 weighting. A lesion with a very long T2, such as subacute infarct or vasogenic edema, can be bright on the high-b DWI simply because it was bright to begin with, even though its water diffuses freely.

This artifact is called T2 shine-through, and the ADC map is the antidote. True restricted diffusion is bright on DWI and dark on ADC. T2 shine-through is bright on DWI but bright (or at least not dark) on ADC, because Eq. 9.4 has already divided out the $S_0$ term that carried the T2 signal. Always read DWI and ADC together.

Explore the Trade-offs

Diffusion contrast is a tug-of-war: raising the b-value sharpens the distinction between restricted and free water but, per Eq. 9.3, costs raw signal everywhere, degrading the signal-to-noise ratio. Use the playground below to vary b-value and tissue diffusivity and watch how the DWI signal and the apparent contrast respond before reading the interpretation that follows.

Notice three things. First, at $b = 0$ all tissues sit at their T2-weighted baseline and diffusion contrast is absent. Second, as $b$ climbs toward 1000 the freely diffusing compartments lose signal fastest while restricted tissue holds on, opening up the contrast we want. Third, pushing $b$ much higher (2000 to 3000) keeps improving lesion conspicuity in principle but the falling SNR eventually swamps the benefit, which is why $b = 1000$ s/mm squared is the clinical workhorse for the brain.

Clinical Anchor: Acute Ischemic Stroke

When an artery occludes, the ischemic neurons can no longer power their sodium-potassium ATPase pumps. Ions and water shift from the extracellular space into the cells, swelling them: this is cytotoxic edema. The extracellular space, normally the highway for water diffusion, becomes narrow and tortuous, and net diffusion drops sharply. ADC can fall by 30 to 50 percent within minutes, long before any change appears on CT or conventional T2 imaging.

Pairing DWI with perfusion imaging gives the diffusion-perfusion mismatch. The DWI lesion approximates the irreversibly infarcted core; the larger perfusion deficit marks hypoperfused but still-viable tissue. The mismatch, perfusion abnormality minus diffusion core, estimates the ischemic penumbra, the tissue that thrombectomy or thrombolysis might salvage. Mismatch-based selection underpinned the extended-window thrombectomy trials and is now embedded in stroke triage.

From a Scalar to a Tensor: DTI

Equation 9.3 treats ADC as a single number, which assumes diffusion is the same in every direction (isotropic). In organized white matter it is emphatically not: axonal membranes and myelin let water diffuse readily along a tract but hinder it across the fibers. Capturing this directionality requires replacing the scalar ADC with a $3 \times 3$ symmetric matrix, the diffusion tensor $\mathbf{D}$.

Because $\mathbf{D}$ is symmetric it has six independent entries, so DTI needs at least six non-collinear diffusion directions plus one $b = 0$ image to solve the system; in practice 20 to 64 directions are acquired for robustness. Diagonalizing the tensor yields three orthogonal eigenvectors (the principal diffusion axes) with three eigenvalues $\lambda_1 \geq \lambda_2 \geq \lambda_3$ giving the diffusivity along each. The leading eigenvector points along the dominant fiber direction, the seed of tractography.

Two rotationally invariant scalars summarize the tensor. Mean diffusivity (MD) is the average of the eigenvalues, the directionless analog of ADC. Fractional anisotropy (FA) measures how elongated the diffusion ellipsoid is, ranging from 0 (a perfect sphere, fully isotropic) to 1 (a needle, all diffusion along one axis).

Dense, coherent tracts such as the corpus callosum reach FA near 0.7 to 0.9; cortical gray matter sits near 0.1; CSF is essentially 0. Tractography then follows the leading eigenvector field from voxel to voxel to reconstruct three-dimensional fiber bundles, a tool used in presurgical mapping of the corticospinal tract and arcuate fasciculus, and in detecting microstructural injury in trauma, demyelination, and aging.

Acquisition Reality and Beyond the Tensor

Diffusion contrast is fragile: any bulk motion during the strong sensitizing gradients, even a pulse beat or a swallow, produces gross phase errors and ghosting. The solution is to freeze motion by reading out an entire image in one shot. Clinical DWI therefore almost universally uses single-shot echo planar imaging (EPI), filling all of k-space after a single excitation in tens of milliseconds.

Single-shot EPI buys speed at the price of artifacts. Its long echo train accrues phase along the slow phase-encode direction, so any off-resonance, principally at air-tissue interfaces such as the skull base, sinuses, and temporal bones, produces geometric distortion and signal pile-up. The huge diffusion gradients also induce eddy currents that linger into the readout, shearing and scaling the image differently for each direction, which must be corrected before tensor fitting. Parallel imaging, readout-segmented EPI, and field-map or reversed-phase-encode (blip-up blip-down) correction all mitigate these problems.

| Model | What it adds | Key parameter(s) | | --- | --- | --- | | DWI / ADC | Single scalar diffusivity | ADC | | DTI | Directionality (anisotropy) | MD, FA, eigenvectors | | IVIM | Separates microcirculation from diffusion | perfusion fraction f, pseudo-diffusion D* | | DKI | Non-Gaussian (restricted) diffusion | mean kurtosis K |

Two refinements push past the mono-exponential, single-tensor picture. Intravoxel incoherent motion (IVIM) notes that at very low b-values (below about 200) blood flowing through the randomly oriented capillary bed mimics fast diffusion; fitting a bi-exponential separates a perfusion fraction and pseudo-diffusion coefficient from true tissue diffusion, giving perfusion information without contrast. Diffusion kurtosis imaging (DKI) acknowledges that real tissue diffusion is non-Gaussian because of all those barriers; adding a kurtosis term quantifies how much the decay deviates from Eq. 9.3, a more sensitive marker of microstructural complexity in tumor grading and subtle injury.