m48 · Imaging across the EP workflow: which modality answers which question, resolution, contrast, noise and dose, accuracy of imaging tests
Education only. Not for diagnosis, treatment or any clinical decision.
In 60 seconds
Every imaging test used around an EP procedure trades the same few things: how fine a detail it resolves, how different two tissues look, how much noise sits on top, how much radiation or time it costs, and whether it can be used in the lab at all. This module gives one vocabulary for all of them. Blur is a point-spread function with a full width at half maximum; contrast and noise combine into a contrast-to-noise ratio; smoothing trades one for the other. Published resolutions span more than an order of magnitude, from a 0.34-mm CT section (Rajendran et al. 2022) to 4–7 mm in clinical PET (Kinahan and Fletcher 2010), and the left atrial wall is about 1.9 mm thick on average (Beinart et al. 2011), so the same wall looks different in each. An imaging test’s accuracy also depends on who is tested: CT for appendage thrombus had 96% sensitivity and 92% specificity against TEE, but a pooled positive predictive value of 41% at 8.9% prevalence (Romero et al. 2013). The EACVI/EHRA consensus reviews echo, MRI, CT and nuclear imaging in AF (Donal et al. 2016); modules m49–m55 take each modality in turn.
From the field
- Start from the question, not the scanner. Where is the esophagus, is there thrombus, how thick is the wall, where is the scar, is there inflammation: each question has a modality that answers it best, and the table below gives one published example for each.
- A resolution number is a blur, not a detection limit. At routine LGE resolution the thinnest published atrial wall keeps about a third of its contrast (lab part B); whether it is “seen” then depends on noise and thresholds.
- Contrast is chosen. Blood and heart wall differ by about 5 HU in CT without contrast agent; iodine, gadolinium and tracers are how imaging separates them (lab part C).
- Read the prevalence before the sensitivity. The same CT protocol gives very different predictive values in a population where thrombus is rare and in one where it is common (Equation 5).
Derivations
Which modality answers which question
One published example per question. “Evidence” says what kind of study it is; each module goes further.
| Question | Modality | What one study found | Evidence | Module |
|---|---|---|---|---|
| How much radiation does a procedure give? | Fluoroscopy | Reference levels from 4.10 Gy·cm² (accessory pathway ablation) to 12.65 Gy·cm² (persistent AF ablation) across 33 centers (Greffier et al. 2026) | National survey | m49 |
| Can AF ablation be done without fluoroscopy? | Mapping system, ICE | Zero fluoroscopy feasible in 95.1% of 1,593 patients, with no significant difference in success or complications (Debreceni et al. 2023) | Meta-analysis of observational studies | m49, m50 |
| ICE or TEE during AF ablation? | ICE, TEE | No significant difference in complications, shorter fluoroscopy with ICE; certainty low to very low (Kanniyan et al. 2026) | Meta-analysis, 6 studies | m50 |
| Is there thrombus in the appendage? | CT, TEE | Sensitivity 96%, specificity 92%, pooled PPV 41% at 8.9% prevalence; with delayed imaging 100%, 99% and 92% (Romero et al. 2013) | Meta-analysis, TEE as reference | m48, m51 |
| Where is the esophagus? | CT | It touched the mid posterior left atrial wall in every patient (Tsao et al. 2005) | Cross-sectional CT study | m51 |
| How thick is the atrial wall? | CT | 1.89 mm on average, 0.5–3.5 mm (Beinart et al. 2011) | Observational CT study | m51 |
| How accurate is CT merged with the map? | CT and mapping | Reported registration error 2.3 mm with a surface fit against 6.4 mm with landmarks (Kistler et al. 2006) | Clinical series | m51, m55 |
| Is “atrial fibrosis” a stable number? | MRI | Average burden from 5.0% to 37.4% of the wall depending on the threshold (Boyle et al. 2022) | Methods study | m52 |
| Where are the VT channels? | MRI | Conducting channels on MRI colocalized with functional mapping targets (Guichard et al. 2024) | Single-center cohort | m52 |
| Did the ablation leave gaps? | MRI | Scar gaps predicted the extent of vein reconnection but not its location (Al Jefairi et al. 2019) | Prospective cohort | m53 |
| Is there active inflammation? | FDG PET | In cardiac sarcoidosis with VT, no improvement on serial PET was associated with events (Muser et al. 2018) | Single-center cohort | m54 |
| Is a device infected? | FDG PET/CT, SPECT | Multisociety consensus on uptake at device pockets and leads (Bourque et al. 2024) | Expert consensus | m54 |
Key takeaways
- Each question has its own modality and its own evidence, from randomized trials to consensus.
- A finding belongs to its study’s population, protocol and reference standard.
Blur and resolution
Model a scanner’s blur as a Gaussian point-spread function (PSF) \(h\) of standard deviation \(\sigma\). Its full width at half maximum (FWHM) \(\Gamma\) is the distance between the two points where it falls to half its peak. Two equal point sources a distance \(d\) apart, each blurred, show a dip between them only when the curvature of their sum at the midpoint is positive:
\[ h(x) = \frac{1}{\sqrt{2\pi}\,\sigma}\,e^{-x^2/(2\sigma^2)}, \qquad \Gamma = 2\sqrt{2\ln 2}\;\sigma \approx 2.355\,\sigma, \qquad \text{dip between two points} \iff d > 2\sigma \approx 0.85\,\Gamma \tag{1}\]
Two points one FWHM apart show a dip of 7%; three standard deviations apart, 36% (lab part A). Published resolutions, read as Gaussian FWHMs, span a wide range. A photon-counting CT system measured a thinnest section (section-sensitivity FWHM) of 0.34 mm in its high-resolution mode and 0.64 mm in its standard mode, and reported finer in-plane resolution still (Rajendran et al. 2022). LGE MRI resolves about 1.5–2 mm routinely and about 1 mm when thin walls are the target (Kellman and Arai 2012). Clinical PET resolves about 4–7 mm FWHM (Kinahan and Fletcher 2010). ICE resolves about 0.2 mm along the beam at 7.5 MHz with a short pulse, and less across it (eq-m50-axial). Reading each as a Gaussian is a simplification: real PSFs have other shapes, and an MRI voxel size is not exactly a FWHM.
Key takeaways
- FWHM ≈ 2.355 σ; two structures closer than about 0.85 FWHM merge.
- Across modalities, resolution differs by more than a factor of ten.
Thin walls
A wall of thickness \(w\) and unit contrast, blurred, peaks at \(\operatorname{erf}(w/(2\sqrt2\,\sigma))\) of its contrast (eq-m51-wall, derived in m51). For the thinnest published left atrial wall, 0.5 mm (Beinart et al. 2011), that is 92% at a 0.34-mm CT section, 44% at 1 mm, 31% at 1.5 mm and 12% at 4 mm. The mean wall, 1.89 mm, keeps 86% at 1.5 mm and 42% at 4 mm (lab part B). Read with a fixed threshold at half the contrast, the thin wall disappears once its peak falls below one half; read as the width at half of its own peak, it never reads thinner than the blur.
Exact
Assumptions
- Model: eq-m51-wall, a uniform wall between two equal backgrounds, blurred by a 1-D Gaussian; readings found by bisection. Tested against
epml.overviewandepml.ct. - Presets: CT section FWHM (Rajendran et al. 2022), LGE resolution (Kellman and Arai 2012), PET resolution (Kinahan and Fletcher 2010), each read as a Gaussian FWHM. Wall thickness 0.5 and 1.89 mm marked (Beinart et al. 2011).
- What it ignores: noise, unequal backgrounds on the two sides (blood and fat), the true shape of each scanner’s PSF, voxel sampling and motion.
Key takeaways
- A wall thinner than the blur loses contrast before it loses width.
- Every thickness read from an image belongs with its resolution and its reading method.
Contrast, noise and dose
Contrast is the difference between two tissues’ signals; noise is the random scatter on each. Their ratio decides what can be told apart:
\[ \mathrm{CNR} = \frac{|S_A - S_B|}{\sigma_n}, \qquad \mathrm{CNR}_n = \sqrt{n}\;\mathrm{CNR}_1 \tag{2}\]
where \(S_A\) and \(S_B\) are mean signals, \(\sigma_n\) is the noise standard deviation and \(\mathrm{CNR}_n\) the value after averaging \(n\) independent samples. For photon-limited images the noise falls as one over the square root of the counts, so halving it takes four times the dose (eq-m49-cnr). Contrast comes from physics and agents. With the 60-keV attenuation data of m49 and m51, wall and blood are 54 and 59 HU, and 2 mm of wall against blood changes an X-ray projection by 0.019%; 10 mg/mL of iodine (illustrative) lifts blood to 427 HU. With the T1 values of Bauner et al. (Bauner et al. 2012), an infarct keeps 34% of the full signal when normal myocardium is nulled (lab part C, m52).
Smoothing averages neighboring pixels. For white noise on a grid of spacing \(\Delta\), a Gaussian smoothing kernel of standard deviation \(\sigma_s\) (much larger than \(\Delta\)) leaves
\[ \sigma_n' = \frac{\sigma_n\,\Delta}{2\sqrt{\pi}\,\sigma_s}, \qquad \sigma_{\text{tot}}^2 = \sigma^2 + \sigma_s^2 \tag{3}\]
in two dimensions, while the blur grows to \(\sigma_{\text{tot}}\). Putting blur and noise together, a wall of contrast \(C\) between equal backgrounds has, at its center,
\[ \mathrm{CNR}_{\text{wall}} = \frac{C}{\sigma_n}\operatorname{erf}\!\left(\frac{w}{2\sqrt2\,\sigma}\right), \qquad w_{\min} = 2\sqrt2\,\sigma\,\operatorname{erf}^{-1}\!\left(\frac{k\,\sigma_n}{C}\right) \tag{4}\]
where \(w_{\min}\) is the thinnest wall that reaches a chosen CNR \(k\) (it exists only if \(k\sigma_n < C\)). With \(C/\sigma_n = 5\), 1 mm of blur and an illustrative \(k = 3\), \(w_{\min}\) = 0.71 mm. In lab part E, smoothing the mean wall from 0.25 to 2 mm raises its CNR from 4.2 to 12.5 (illustrative noise), but from 1.5 mm on the wall merges with an esophageal wall 2 mm away, and the esophagus spills 8% of its signal into the wall’s center.
Key takeaways
- CNR, not contrast alone, decides visibility; averaging gains \(\sqrt n\).
- Smoothing trades noise for resolution; there is no free lunch across a structure.
Accuracy of an imaging test
Sensitivity (Se) and specificity (Sp) describe the test; predictive values describe what a result means in a population with prevalence \(\pi\) (eq-m35-ppv). Likelihood ratios separate the two:
\[ \mathrm{LR}^+ = \frac{\mathrm{Se}}{1 - \mathrm{Sp}}, \qquad \mathrm{LR}^- = \frac{1 - \mathrm{Se}}{\mathrm{Sp}}, \qquad \frac{p_{\text{post}}}{1 - p_{\text{post}}} = \frac{\pi}{1 - \pi}\,\mathrm{LR} \tag{5}\]
where \(p_{\text{post}}\) is the probability of disease after the result; with \(\mathrm{LR}^+\) it equals the PPV, and with \(\mathrm{LR}^-\) it equals one minus the NPV. For CT and appendage thrombus (Romero et al. 2013), \(\mathrm{LR}^+\) = 12 and \(\mathrm{LR}^-\) = 0.043. At 8.9% prevalence the formula gives a PPV of 54% and an NPV of 99.6%; the meta-analysis reported pooled values of 41% and 99%. They need not agree, because each measure was pooled across studies separately and PPV depends on each study’s prevalence. PPV is concave in prevalence, so the average of per-study PPVs lies below the PPV at the average prevalence (lab part G shows this on illustrative studies); that is one possible reason, not a full account. With delayed imaging (Se 100%, Sp 99%) the formula gives 91%, against 92% reported.
Exact
Assumptions
- Model: eq-m35-ppv and Equation 5, with sensitivity and specificity that do not change with prevalence. Tested against
epml.overviewandepml.ml. - Presets and reported pooled values: CT against TEE for appendage thrombus, 19 studies and 2,955 patients (Romero et al. 2013).
- What it ignores: uncertainty in the pooled estimates, differences between studies, and the imperfection of TEE as the reference standard.
Key takeaways
- Sensitivity and specificity belong to the test; predictive values belong to the test in a population.
- Pooled predictive values from a meta-analysis are not the formula applied to pooled sensitivity and specificity.
Clinical correlate
This section describes published findings only. It is not guidance for patient care, and it has not yet been reviewed by an EP physician.
Where this shows up in EP. Patients with AF may have imaging before, during and after ablation. The EACVI/EHRA expert consensus reviews echocardiography (left atrial volume and strain), MRI (left atrial fibrosis), CT (with emphasis on planning interventions) and nuclear imaging in these patients, and reaches fourteen consensus statements meant to help imagers and electrophysiologists choose and interpret the technique (Donal et al. 2016).
What the evidence is. A 2016 expert consensus (Donal et al. 2016), and, for the worked example, one meta-analysis of diagnostic accuracy with TEE as the reference (Romero et al. 2013). The table of questions above mixes randomized trials, registries, cohorts and consensus; each module m49–m55 discusses its own evidence. Newer studies may have changed what the consensus describes; it is read here for its structure, not as current advice.
What a reader should ask:
- Which question is the image meant to answer? The modality follows from it.
- At what resolution and contrast? A thin structure or a small difference may be below what the image can show (Equation 1, Equation 4).
- In which population? Predictive values move with prevalence (Equation 5).
Visual summary
Schematic
Assumptions
- Schematic: not to scale, no quantitative claim. Labels follow Equation 1 to Equation 5.
How experts think
Before choosing or trusting an image, ask:
- What is the question? Anatomy, wall, scar, thrombus, inflammation, catheter position or dose.
- Can this modality resolve it? Compare the structure’s size with the FWHM (Equation 1).
- Is there contrast, and enough CNR? Which agent or sequence creates the difference, and how noisy is it (Equation 2, Equation 4)?
- What did processing do? Smoothing and thresholds change what is seen and measured (Equation 3).
- What does a result mean here? Use the prevalence in this population, not the one in the paper (Equation 5).
One sentence to keep: an image answers a question at a resolution, a contrast and a prevalence; change any one and the answer changes.
Cheat sheet
| Quantity | Result |
|---|---|
| PSF and FWHM (Equation 1) | \(\Gamma = 2\sqrt{2\ln 2}\,\sigma \approx 2.355\,\sigma\); two points show a dip only if \(d > 2\sigma\) |
| Thin wall (eq-m51-wall) | Peak \(\operatorname{erf}(w/(2\sqrt2\,\sigma))\); FWHM reading never below \(\Gamma\) |
| CNR (Equation 2) | \(\lvert S_A - S_B\rvert/\sigma_n\); averaging \(n\) samples gains \(\sqrt n\); dose ×4 halves photon noise |
| Smoothing (Equation 3) | 2-D white noise falls to \(\sigma_n\Delta/(2\sqrt\pi\,\sigma_s)\); blur adds in quadrature |
| Wall visibility (Equation 4) | \(\mathrm{CNR}_{\text{wall}} = (C/\sigma_n)\operatorname{erf}(w/(2\sqrt2\,\sigma))\) |
| Likelihood ratios (Equation 5) | \(\mathrm{LR}^+ = \mathrm{Se}/(1-\mathrm{Sp})\); post-test odds = pre-test odds × LR |
Values used in this module
| Value | Source |
|---|---|
| CT section FWHM 0.34 mm (high-resolution) and 0.64 mm (standard), photon-counting system | Rajendran et al. (2022) |
| LGE resolution 1.5–2 mm routinely, about 1 mm for thin walls | Kellman and Arai (2012) |
| Clinical PET resolution 4–7 mm FWHM | Kinahan and Fletcher (2010) |
| Left atrial wall 1.89 mm on average, 0.5–3.5 mm | Beinart et al. (2011) |
| CT for appendage thrombus: Se 96%, Sp 92%, PPV 41%, NPV 99%, prevalence 8.9%; delayed imaging 100%, 99%, 92%, 100% | Romero et al. (2013) |
| Myocardium and infarct T1 after contrast | Bauner et al. (2012) |
Studies in this module
| Study | Question | Design |
|---|---|---|
| Donal et al. (2016) | Multi-modality imaging in AF | Expert consensus, 14 statements |
| Romero et al. (2013) | CT for appendage thrombus | Meta-analysis, 19 studies, 2,955 patients |
| Rajendran et al. (2022) | Technical performance of a photon-counting CT system | Phantom measurements and 4 participants |
| Beinart et al. (2011) | Left atrial wall thickness | Observational CT study |
Interview questions
Try each question before you open its model answer. In a real interview, say which assumptions you are making.
Q1. What is the FWHM of a Gaussian blur, and why use it?
Model answer
The distance between the points where the PSF falls to half its peak, \(2\sqrt{2\ln 2}\,\sigma \approx 2.355\,\sigma\) (Equation 1). It is easy to measure on a line or point phantom and roughly the separation at which two structures start to merge.
Q2. When do two blurred points show a dip between them?
Model answer
When their separation exceeds \(2\sigma\), about 0.85 FWHM: the curvature of the sum at the midpoint is proportional to \(d^2/4 - \sigma^2\) (Equation 1). At one FWHM the dip is only about 7%, so with noise they may still look like one.
Q3. A 0.5-mm wall imaged at 1.5 mm FWHM. What do you expect?
Model answer
Its peak reaches about 31% of its contrast (eq-m51-wall); a half-contrast threshold misses it, and a half-peak reading reports about 1.5 mm, the blur, not the wall.
Q4. Why do blood and myocardium look almost the same on fluoroscopy and non-contrast CT?
Model answer
Their X-ray attenuation is nearly equal: at 60 keV they differ by about 5 HU, and 2 mm of wall against blood changes a projection by about 0.02% (lab part C). Iodine in the blood creates the difference.
Q5. Define CNR. How does it change with averaging and with dose?
Model answer
\(|S_A - S_B|/\sigma_n\) (Equation 2). Averaging \(n\) independent samples multiplies it by \(\sqrt n\). For photon-limited noise, CNR grows with the square root of the dose, so doubling CNR costs four times the dose (eq-m49-cnr).
Q6. What does smoothing do to noise and resolution?
Model answer
In 2-D it divides white noise by \(2\sqrt\pi\,\sigma_s/\Delta\) and widens the blur to \(\sqrt{\sigma^2 + \sigma_s^2}\) (Equation 3). A wall’s CNR rises, but neighbors merge and spill into it (lab part E).
Q7. How thin a wall can you see?
Model answer
Pick a CNR you need, \(k\). Then \(w_{\min} = 2\sqrt2\,\sigma\,\operatorname{erf}^{-1}(k\sigma_n/C)\) (Equation 4), and no wall is visible if \(k\sigma_n \ge C\). Better contrast or less noise lowers it, as does a smaller \(\sigma\).
Q8. Sensitivity 96%, specificity 92%, prevalence 8.9%. What is the PPV?
Model answer
About 54% (eq-m35-ppv). Equivalently, LR+ = 12, pre-test odds 0.098, post-test odds 1.17 (Equation 5).
Q9. A meta-analysis reports pooled sensitivity, specificity and PPV that do not satisfy Bayes’ rule. Is it wrong?
Model answer
Not necessarily. Each measure is pooled across studies separately, and PPV depends on each study’s prevalence; since PPV is concave in prevalence, averaging per-study PPVs gives less than the PPV at the average prevalence. For CT and appendage thrombus the formula gives 54% against a reported 41% (Romero et al. 2013).
Q10. Why are likelihood ratios useful for imaging tests?
Model answer
They summarize the test independently of prevalence and update any pre-test probability by multiplying odds (Equation 5). A high LR+ rules in, a low LR− rules out.
Q11. How would you compare two modalities for measuring atrial wall thickness?
Model answer
By resolution against the thickness range (0.5–3.5 mm (Beinart et al. 2011)), wall contrast against blood and fat, noise and reading method (threshold or half-peak), against a reference such as histology or a phantom of known thickness. Report bias as a function of thickness, not one mean error.
Q12. Why is PET poor at showing wall thickness but still useful in EP?
Model answer
Its 4–7 mm resolution (Kinahan and Fletcher 2010) is coarser than the wall, so thickness is lost and uptake is underestimated by partial volume (m54). It answers a different question: where a tracer accumulates, for example inflammation or infection (Muser et al. 2018; Bourque et al. 2024).
Code lab
The lab describes several modalities with one set of numbers and ends with the accuracy of a published test.
| Part | What you do | Data |
|---|---|---|
| A | Blur and resolution (Equation 1) | published resolutions |
| B | Thin walls at published resolutions (eq-m51-wall) | published resolutions and wall thickness |
| C | Where contrast comes from | our own code (m49, m51, m52) |
| D | Contrast-to-noise ratio and averaging (Equation 2) | our own simulation |
| E | Smoothing: less noise, less resolution (Equation 3, Equation 4) | our own phantom |
| F | One cross-section imaged four ways | our own phantom, published resolutions |
| G | Accuracy of an imaging test (Equation 5) | published meta-analysis (Romero et al. 2013) |
- Python notebook: labs/python/m48_lab.ipynb. Run it from the course repository after
pip install -r requirements.txt, or open it in Google Colab. - Data: none to download. Published values are listed in the notebook; contrasts, noise levels, the esophageal geometry and the simulated meta-analysis are illustrative.
Each part ends with checks: lines that pass with the code as given. If you change something and a check fails, that is information, not an error to hide.
Self-check
Short checks you can do in your head. Open each answer after you have tried it.
S1. A PSF has σ = 1 mm. FWHM?
Answer
About 2.4 mm (Equation 1).
S2. With that PSF, how close can two points be and still show a dip?
Answer
Just over 2 mm (Equation 1).
S3. CNR is 2 per pixel. Average 9 independent pixels. CNR?
Answer
6 (Equation 2).
S4. Smoothing σ_s doubles. What happens to 2-D white noise?
Answer
It halves (Equation 3).
S5. Se 100%, Sp 99%. LR+?
Answer
100 (Equation 5).
S6. True or false: a test’s PPV is a property of the scanner.
Answer
False: it depends on prevalence (eq-m35-ppv).
Literature figures
These figures are worth looking at alongside this module. We link to them instead of copying them: some licenses do not allow reuse here, and others we could not confirm. Each entry says what to look for and where this module explains it.
- Measured resolution of a CT system. Rajendran et al. show modulation transfer functions and section sensitivity profiles for standard and high-resolution modes [Rajendran et al. (2022); free full text]. This module: blur and resolution.
- Resolution and SUV in PET. Kinahan and Fletcher discuss how resolution and partial volume affect measured uptake [Kinahan and Fletcher (2010); free full text]. This module: thin walls.
Further reading
- Imaging in AF. EACVI/EHRA expert consensus (Donal et al. 2016).
- Diagnostic accuracy. CT for appendage thrombus (Romero et al. 2013); predictive values and Bayes’ rule (Altman and Bland 1994).
- Resolution in practice. Photon-counting CT (Rajendran et al. 2022); LGE (Kellman and Arai 2012); PET (Kinahan and Fletcher 2010).
- Related modules. m35 for diagnostic accuracy, m49 for fluoroscopy, m50 for ICE, m51 for CT, m52 and m53 for MRI, m54 for nuclear imaging, and m55 for putting them together.