Image Quality, SNR/CNR & QA

Why we quantify quality at all

Every clinical MRI decision is a contest between signal, resolution, and scan time, all corrupted by noise. A radiologist's confidence in calling a 4 mm metastasis or a subtle cartilage defect is not a matter of taste; it is governed by measurable quantities. The two that matter most are the signal-to-noise ratio (SNR) and the contrast-to-noise ratio (CNR). SNR tells you how cleanly the scanner has recorded any signal at all; CNR tells you whether two tissues are distinguishable from one another against that noise floor. A pristine image with high SNR but no contrast between lesion and background is diagnostically useless, which is why CNR, not SNR, ultimately drives detectability.

This lesson develops the quantitative machinery: how SNR is defined and correctly measured, how it scales with the parameters you control at the console, how the resolution-SNR-time triangle forces explicit trade-offs, and how the American College of Radiology (ACR) phantom QA program verifies that the machine is performing within tolerance. We close with the routine calibrations that keep a scanner honest day to day.

Defining signal, noise, and SNR

Noise in MRI originates dominantly from thermal (Johnson-Nyquist) fluctuations in the patient and the receive chain. In the complex raw data, noise is additive, white, and Gaussian with zero mean and standard deviation $\sigma_0$ in each of the real and imaginary channels. Signal is the coherent transverse magnetization the coil detects. The fundamental figure of merit is the ratio of the mean signal to the noise standard deviation.

Measuring SNR in practice

Three ROI-based methods dominate. Each estimates the same SNR but handles the noise term differently, and only the last is robust to modern reconstruction.

1. Background-noise method. Place a signal ROI in the tissue and a noise ROI in air outside the patient. Because background magnitude noise is Rayleigh-distributed, the measured background standard deviation must be divided by 0.655 to recover the true $\sigma_0$. The working formula is $\mathrm{SNR} = 0.655 \cdot \bar{S} / \sigma_{bg}$. Simple, but invalid wherever the background is non-uniform (ghosting, parallel-imaging artifacts, or filtered air).
2. Difference-image (two-acquisition) method. Acquire two identical images, subtract them, and measure the standard deviation in a tissue ROI on the difference image. Subtraction cancels the signal but adds the noise in quadrature, so $\sigma_{diff} = \sqrt{2}\,\sigma_n$. Then $\mathrm{SNR} = \sqrt{2}\cdot \bar{S}/\sigma_{diff}$. This is the NEMA-recommended reference standard because it samples noise in the same tissue where the signal lives.
3. Single-ROI (statistical) method. Within a uniform phantom region, the pixel standard deviation across a large ROI estimates $\sigma_n$ directly. Convenient for phantoms, but it conflates true noise with low-spatial-frequency non-uniformity, so the ROI must be small relative to any shading.

Contrast-to-noise ratio: the real currency of detectability

Lesion conspicuity depends not on absolute signal but on the difference between lesion and background relative to the noise. CNR formalizes this.

By the Rose criterion, a feature is reliably detected by a human observer when its CNR is roughly 3 to 5 for objects spanning several pixels. This explains why a sequence can be redesigned to lower raw SNR yet improve diagnosis: a fluid-attenuated or fat-suppressed acquisition sacrifices overall signal to widen the lesion-to-background separation, raising CNR. Gadolinium works the same way, boosting $\bar{S}_A$ of enhancing tissue so $|\bar{S}_A - \bar{S}_B|$ grows. The takeaway is that maximizing SNR is a means, not the end; maximizing CNR for the clinical question is the end.

How SNR scales with the parameters you control

The acquired signal is averaged over the voxel and the readout, while noise integrates over the sampled bandwidth. Combining the voxel signal with the noise from $N_{avg}$ averages sampled at receiver bandwidth $BW$ gives the canonical proportionality. For a Cartesian acquisition the per-voxel SNR is:

Reading Eq. 15.3 term by term gives the rules of thumb every technologist uses:

• Voxel volume — linear. Double any one dimension (e.g. slice thickness from 3 to 6 mm) and SNR doubles. This is the most powerful single lever, which is why thick slices look so clean.
• Averages — square root. SNR grows as $\sqrt{N_{avg}}$, so quadrupling NEX from 1 to 4 only doubles SNR while quadrupling scan time. Diminishing returns make averaging an expensive way to buy signal.
• Receiver bandwidth — inverse square root. Halving $BW$ raises SNR by $\sqrt{2}\approx 1.41$, at the cost of a longer readout, increased chemical-shift, and greater susceptibility to off-resonance.
• Matrix and FOV. Holding FOV fixed, a finer matrix shrinks the voxel (lowering SNR) but the $\sqrt{N}$ term partially compensates, so net SNR scales with voxel volume.
• Field strength — roughly linear. Equilibrium magnetization scales with $B_0$ and coil noise considerations make in-vivo SNR scale close to linearly with field, the chief motivation for 3T over 1.5T (about a factor of two before accounting for T1 lengthening).

The resolution-SNR-time triangle

These three goals are mutually antagonistic. Improving any one degrades at least one of the others, and acceleration techniques (parallel imaging, partial Fourier, compressed sensing) buy time by spending SNR.

| Change requested | Direct effect | Cost | | --- | --- | --- | | Higher spatial resolution (smaller voxel) | Better delineation of small structures | SNR falls with voxel volume; longer scan to refill matrix | | Higher SNR (averaging) | Cleaner image, $\propto\sqrt{N_{avg}}$ | Scan time rises linearly with NEX | | Shorter scan (parallel imaging, factor R) | Time falls by ~R | SNR falls by $\sqrt{R}$ plus the g-factor penalty | | Lower receiver bandwidth | SNR rises as $1/\sqrt{BW}$ | Longer echo train, more chemical shift and susceptibility |

The interactive below lets you feel the trade-off directly. Adjust the tissue and sequence parameters and watch how pushing for more contrast or finer detail interacts with the noise floor and the resulting CNR.

The ACR phantom QA program

Protecting quality requires more than good acquisition choices; the hardware must be verified objectively. The ACR MRI Accreditation Program uses a fluid-filled cylindrical phantom and a defined set of measurements, performed at accreditation and on a routine schedule by the technologist and annually by a medical physicist. The seven core tests map directly onto the image-quality concepts above.

| Test | What it checks | Representative ACR limit | | --- | --- | --- | | Geometric accuracy | Length measurements (gradient calibration) | Within ±2 mm of true 148 mm and 190 mm | | High-contrast spatial resolution | Smallest resolvable hole array | Resolve 1.0 mm (or better) hole pairs | | Slice-thickness accuracy | Ramp-pair signal profile | 5.0 mm slice within ±0.7 mm | | Slice-position accuracy | Bar length difference at ends | Bar difference within ±5 mm | | Low-contrast object detectability | Spoke patterns of subtle disks (CNR-limited) | Minimum number of complete spokes (e.g. ≥9 at 1.5T) | | Percent image uniformity (PIU) | Signal flatness across a large ROI | PIU ≥ 87.5 percent at 1.5T (≥ 82 percent at 3T) | | Percent-signal ghosting (PSG) | N/2 ghost from periodic instability | PSG ≤ 2.5 percent (≤ 3.0 percent some setups) |

Two of these deserve elaboration. Percent image uniformity is computed from the maximum and minimum mean signal in small ROIs swept across a large uniform ROI: $\mathrm{PIU} = 100\,(1 - (S_{max}-S_{min})/(S_{max}+S_{min}))$. It is the bench test for coil and B1 homogeneity. Percent-signal ghosting quantifies the ghost artifact along the phase-encode direction from periodic data inconsistency: $\mathrm{PSG} = |(\,(top+bottom) - (left+right)\,)/(2\,S_{phantom})|$, using background ROIs placed inside and outside the ghost path. The low-contrast object detectability test is the program's explicit CNR check, because the visible spoke count rises and falls with system CNR exactly as the Rose criterion predicts.

Routine calibrations that keep SNR and contrast honest

Before nearly every clinical sequence the scanner runs an automatic prescan that performs three calibrations. Each protects a different part of the quality chain:

• Center frequency. The Larmor frequency $f_0 = \gamma B_0 / 2\pi$ is measured from the patient's water peak so that excitation and fat suppression are correctly tuned. A mis-set $f_0$ causes failed fat saturation, spatial shift, and slice mis-registration.
• Transmit gain (reference voltage / flip-angle calibration). The RF amplitude needed for an exact 90° (or 180°) is calibrated per patient because loading changes the coil. An incorrect transmit gain delivers the wrong flip angle, directly corrupting both signal level and the intended T1/T2 weighting.
• Gradient calibration. Gradient amplitude scaling is calibrated so that 1 mm of prescribed FOV equals 1 mm in the patient; this is exactly what the ACR geometric-accuracy test audits. Gradient miscalibration produces geometric distortion and incorrect distance measurements.