m52 · Cardiac MRI for EP: relaxation and inversion recovery, late gadolinium enhancement, atrial fibrosis, ventricular scar, MRI with implanted devices
Education only. Not for diagnosis, treatment or any clinical decision.
In 60 seconds
Magnetic resonance imaging (MRI) images hydrogen nuclei, which precess in a strong magnetic field at a frequency proportional to the field. Contrast comes from how fast their magnetization recovers after it is disturbed, the time constant \(T_1\). Gadolinium contrast agents shorten \(T_1\) in proportion to their concentration, and they stay outside cells, so tissue with more extracellular space, such as scar, ends up with a shorter \(T_1\). Late gadolinium enhancement (LGE) inverts the magnetization and reads the image when healthy myocardium has no signal: scar is then bright (Kellman and Arai 2012). In the left atrium, fibrosis on LGE was associated with recurrence after AF ablation (Marrouche et al. 2014), but adding ablation of that fibrosis did not reduce recurrence in a randomized trial (Marrouche et al. 2022). How much “fibrosis” a scan shows depends on the method and the threshold (Boyle et al. 2022; Kamali et al. 2020). In the ventricle, conducting channels found on LGE colocalized with functional mapping targets (Guichard et al. 2024). In registries of patients whose pacemakers and defibrillators were not labelled for MRI, scans under a protocol at 1.5 T reported no long-term clinically significant adverse events (Nazarian et al. 2017).
From the field
- From association to trial. In the DECAAF cohort, each 1% more atrial fibrosis on LGE was associated with more recurrence after AF ablation (hazard ratio 1.06), and 57 of 329 enrolled patients were excluded for poor image quality (Marrouche et al. 2014). DECAAF II then randomized 843 patients with persistent AF to pulmonary vein isolation (PVI) with or without ablation of MRI-defined fibrosis: recurrence was 43.0% and 46.1%, not significantly different, with more safety events in the fibrosis-guided group (Marrouche et al. 2022). Its lesions did not uniformly become scar on follow-up MRI (Akoum et al. 2024).
- One scan, many burdens. In 102 patients, average IIR fibrosis went from 5.0% to 37.4% of the wall as the threshold moved from 1.2 to 0.97; a second method agreed only modestly in space (Dice 0.49), and only that method predicted recurrence (Boyle et al. 2022).
- Repeat scans. In 45 patients scanned twice, Otsu thresholding reproduced better (Dice 71) than 3.3 standard deviations above the blood pool (58) or IIR above 1.61 (46); only voxels above about the 70th percentile of the wall histogram agreed better than chance (Kamali et al. 2020).
- Where the threshold comes from. In the right atrium at 3 T, the upper normal limit (mean + 2 SD in nine healthy volunteers) was an IIR of 1.21, and a dense-scar threshold was 1.29; agreement with voltage mapping was 56% (Gunturiz-Beltrán et al. 2022).
- Ventricular scar and channels. In 40 patients, conducting channels found automatically on LGE were unmasked by deceleration zones in 78% with a double-extrastimulus mapping protocol, against 48% with baseline pacing (Guichard et al. 2024). In 37 patients with nondilated left ventricular cardiomyopathy, a ring-like LGE pattern came with more and longer channels, and channel number and length were associated with VT recurrence (Pittorru et al. 2026).
- Devices in the scanner. At 1.5 T, 1,500 nonthoracic scans in patients with devices not labelled for MRI had no deaths, lead failures or losses of capture; one defibrillator not programmed per protocol had to be replaced (Russo et al. 2017). In 2,103 thoracic and nonthoracic scans, nine devices (0.4%) reset to a backup mode (Nazarian et al. 2017). In pigs, an LGE sequence designed for use with a defibrillator present found more heterogeneous tissue corridors (candidate channels), with lower specificity (Sheagren et al. 2025).
Derivations
Spins and the Larmor relation
A hydrogen nucleus has a magnetic moment. In a static field \(B_0\) the net magnetization of many nuclei, \(M_0\), points along the field, and any part tipped away from it precesses at
\[ f = \frac{\gamma}{2\pi}\,B_0 \tag{1}\]
with \(\gamma/2\pi\) = 42.577 MHz/T for the proton (National Institute of Standards and Technology 2024): 63.87 MHz at 1.5 T and 127.73 MHz at 3 T. An RF pulse at this frequency tips the magnetization, and a field gradient makes the frequency depend on position, which is how MRI encodes space (lab part A).
Key takeaways
- The scanner talks to protons at the Larmor frequency.
- Gradients turn frequency into position.
Inversion recovery
After a 180° pulse the longitudinal magnetization starts at \(-M_0\) and relaxes back toward \(M_0\): \(dM_z/dt = (M_0 - M_z)/T_1\). Solving,
\[ M_z(t) = M_0\left(1 - 2e^{-t/T_1}\right), \qquad M_z = 0 \ \text{at}\ t = T_1 \ln 2 \tag{2}\]
so the signal of a tissue is zero about \(0.69\,T_1\) after the inversion (Kellman and Arai 2012). Lab part B integrates the full Bloch equations through a finite RF pulse and matches Equation 2 within 1.1%, the loss to \(T_2\) decay during the pulse.
Key takeaways
- Each tissue crosses zero at its own time, \(T_1 \ln 2\).
- Real readouts shorten the apparent \(T_1\) (Kellman and Arai 2012).
Gadolinium, T1 and extracellular volume
A gadolinium agent raises the relaxation rate \(R_1 = 1/T_1\) in proportion to its concentration:
\[ \frac{1}{T_1} = \frac{1}{T_{1,0}} + r_1\,[\mathrm{Gd}] \tag{3}\]
with relaxivity \(r_1\) = 4.5 L/(mmol·s) (Kellman and Arai 2012). In Kellman and Arai’s worked example, 0.15 mmol/kg in blood that is 8% of body weight gives about 2 mmol/L in blood and 0.5 mmol/L in tissue with 25% extracellular volume, which shortens a native \(T_1\) of 950 ms to about 300 ms (lab part C). The agent stays outside cells, so a tissue’s concentration is its extracellular volume fraction (ECV) times the plasma’s, and blood’s is \(1 - \mathrm{Hct}\) times the plasma’s, where Hct is the hematocrit. Dividing the two changes in \(R_1\):
\[ \mathrm{ECV} = (1 - \mathrm{Hct})\,\frac{\Delta R_{1,\text{myo}}}{\Delta R_{1,\text{blood}}} \tag{4}\]
In 26 patients with chronic infarction at 1.5 T, \(T_1\) ten minutes after contrast was 379 ms in remote myocardium and 238 ms in the infarct, with ECV 28% and 53% (Bauner et al. 2012). The infarct’s change in \(R_1\) is 2.04 times the myocardium’s, close to their ECV ratio of 1.89. Scar has more room for gadolinium, so its \(T_1\) falls further.
Key takeaways
- Gadolinium marks extracellular space, not scar as such.
- ECV turns two pairs of \(T_1\) values and a hematocrit into a tissue fraction.
Choosing the inversion time
Ignoring the readout, a magnitude LGE image shows \(|1 - 2e^{-\mathrm{TI}/T_1}|\) at inversion time TI. With TI set to null healthy myocardium (\(T_{1m}\ln 2\)), scar with \(T_{1s} < T_{1m}\) shows \(|1 - 2^{\,1 - T_{1m}/T_{1s}}|\). The signed difference between scar and myocardium is \(2(e^{-\mathrm{TI}/T_{1m}} - e^{-\mathrm{TI}/T_{1s}})\), which peaks at
\[ \mathrm{TI}^* = \frac{\ln(T_{1m}/T_{1s})}{1/T_{1s} - 1/T_{1m}} \tag{5}\]
With the published means, the null is at 263 ms, scar reads 0.34 of full signal there, and \(\mathrm{TI}^*\) is 298 ms (lab part D). Set TI too short and, in a magnitude image, myocardium can look brighter than scar; a phase-sensitive reconstruction keeps the sign and is much less sensitive to TI (Kellman and Arai 2012). Blood often has an intermediate concentration and signal, and the contrast between scar and the blood pool is frequently poor (Kellman and Arai 2012). In our model, with an illustrative native blood \(T_1\) and hematocrit, blood recovers almost as fast as scar.
Exact
Assumptions
- Model: Equation 2 for an ideal inversion and an instantaneous readout, with Equation 5; tested against
epml.mri. - \(T_1\) of myocardium (379 ms) and infarct (238 ms) are the published means 10 minutes after contrast at 1.5 T (Bauner et al. 2012). Blood’s \(T_1\) (249 ms) follows from Equation 4 with an illustrative native blood \(T_1\) (1600 ms) and hematocrit (0.42).
- What would change the result: the readout (it shortens the apparent \(T_1\)), time since injection, dose, field strength, heart rate and incomplete recovery between inversions (Kellman and Arai 2012).
Key takeaways
- The null of myocardium makes scar conspicuous; the largest difference comes a little later.
- Scar next to the blood pool can be the hardest scar to see.
Image intensity ratio and thresholds
LGE intensities have no absolute scale; they depend on coil sensitivity, timing and dose. Atrial studies therefore divide each wall intensity by the mean blood-pool intensity and threshold the ratio:
\[ \mathrm{IIR}_i = \frac{I_{\text{wall},i}}{\bar I_{\text{blood}}}, \qquad B(\theta) = \frac{1}{N}\sum_{i=1}^{N} \mathbf{1}\!\left[\mathrm{IIR}_i > \theta\right] \tag{6}\]
where \(B(\theta)\) is the burden: the fraction of the \(N\) wall samples above the threshold \(\theta\). A common gain cancels in the ratio; the threshold does not. On our synthetic wall with 15% true fibrosis, the published thresholds give burdens from 4.1% (IIR 1.61) to 46.2% (IIR 0.97), and no method reproduces itself perfectly on a repeat scan (lab part E). Overlap is measured with the Dice coefficient (m40).
Simulation
Assumptions
- Model: a synthetic, unwrapped atrial wall (60 × 120 samples) with smooth fibrotic patches covering 15% of it, scanned twice with independent noise, normalized by Equation 6; made by
epml.mri.synthetic_wallwith a fixed seed and shipped asdemos/m52-d2-wall.json. Burden, Otsu’s threshold and Dice are tested against Python. - Normal wall has mean IIR 0.9 and SD 0.155, so that mean + 2 SD equals the published upper normal limit of 1.21, measured in the right atrium at 3 T (Gunturiz-Beltrán et al. 2022). Patch shape, fibrotic intensity (1.5), noise and burden are illustrative. Thresholds are the published ones (Boyle et al. 2022; Gunturiz-Beltrán et al. 2022; Kamali et al. 2020).
- What it ignores: wall segmentation errors, partial volume with blood and fat, coil shading, the wall’s thickness, and how real fibrosis is distributed.
Key takeaways
- The reported burden is a property of the method as much as of the atrium.
- Normalization removes gain; it does not remove noise or segmentation error.
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 MRI shows up in EP. The cited studies use LGE to quantify atrial fibrosis before AF ablation (Marrouche et al. 2014), to guide added ablation of fibrosis in a trial (Marrouche et al. 2022), to look at lesions after ablation (Akoum et al. 2024), to find conducting channels in ventricular scar before VT ablation (Guichard et al. 2024; Pittorru et al. 2026), and to scan patients with implanted devices under a protocol (Russo et al. 2017; Nazarian et al. 2017).
What the evidence is. The fibrosis-outcome link comes from a cohort (Marrouche et al. 2014); the one randomized trial of acting on it found no difference in recurrence (Marrouche et al. 2022). Methods studies show that the number depends on the threshold and method (Boyle et al. 2022; Kamali et al. 2020; Gunturiz-Beltrán et al. 2022). The VT studies are single-center cohorts (Guichard et al. 2024; Pittorru et al. 2026). The device data come from prospective registries scanned at 1.5 T under protocols (Russo et al. 2017; Nazarian et al. 2017), and the device-artifact study is in pigs (Sheagren et al. 2025).
What a reader should ask:
- Which threshold and which normalization? A burden without its method is not comparable across studies.
- Association or intervention? A marker that predicts recurrence is not yet a target that reduces it.
- Which scanner, timing and protocol? \(T_1\) after contrast depends on dose, time and field strength.
Visual summary
Schematic
Assumptions
- Schematic: not to scale, no quantitative claim. Labels follow Equation 1 to Equation 6.
How experts think
Before trusting an LGE finding or a fibrosis number, ask:
- Was myocardium nulled? A wrong TI in a magnitude image can hide scar or invent it (see Equation 2 and Equation 5).
- Where is the scar relative to the blood pool? Scar and blood can have similar \(T_1\) after contrast (Kellman and Arai 2012).
- What does the number depend on? Normalization, threshold, wall segmentation and resolution (Boyle et al. 2022; Kamali et al. 2020).
- Is it reproducible? A repeat scan is the simplest test (Kamali et al. 2020).
- What did the trial show? Association is not benefit (Marrouche et al. 2014, 2022).
One sentence to keep: LGE shows where gadolinium has room to go, read at one moment, through a threshold someone chose.
Cheat sheet
| Quantity | Result |
|---|---|
| Larmor relation (Equation 1) | \(f = (\gamma/2\pi)B_0\); 42.577 MHz/T; 63.87 MHz at 1.5 T |
| Inversion recovery (Equation 2) | \(M_z = M_0(1 - 2e^{-t/T_1})\); null at \(T_1\ln 2 \approx 0.69\,T_1\) |
| Relaxivity (Equation 3) | \(1/T_1 = 1/T_{1,0} + r_1[\mathrm{Gd}]\); \(r_1\) = 4.5 L/(mmol·s) |
| Extracellular volume (Equation 4) | \((1 - \mathrm{Hct})\,\Delta R_{1,\text{myo}}/\Delta R_{1,\text{blood}}\) |
| Inversion time (Equation 5) | null \(T_{1m}\ln 2\); scar there \(|1 - 2^{1 - T_{1m}/T_{1s}}|\); peak difference at \(\ln(T_{1m}/T_{1s})/(1/T_{1s} - 1/T_{1m})\) |
| Image intensity ratio (Equation 6) | \(\mathrm{IIR} = I_{\text{wall}}/\bar I_{\text{blood}}\); burden = share above \(\theta\) |
Values used in this module
| Value | Source |
|---|---|
| Proton \(\gamma/2\pi\) = 42.577478461 MHz/T | National Institute of Standards and Technology (2024) |
| Relaxivity 4.5 L/(mmol·s); dose example; native myocardial \(T_1\) 950 ms | Kellman and Arai (2012) |
| \(T_1\) after 10 minutes: myocardium 379 ms, infarct 238 ms; ECV 28% and 53% (1.5 T) | Bauner et al. (2012) |
| IIR thresholds 0.97 and 1.2 | Boyle et al. (2022) |
| IIR 1.21 upper normal and 1.29 dense scar (right atrium, 3 T) | Gunturiz-Beltrán et al. (2022) |
| IIR 1.61; blood pool + 3.3 SD; Otsu | Kamali et al. (2020) |
Studies in this module
| Study | Question | Design |
|---|---|---|
| Marrouche et al. (2014) | Fibrosis and recurrence after AF ablation | Multicenter cohort |
| Marrouche et al. (2022) | Fibrosis-guided ablation added to PVI | Randomized trial, 843 patients |
| Akoum et al. (2024) | Lesions and scar after that ablation | Secondary analysis of the trial |
| Boyle et al. (2022) | Two quantification methods | Single-center cohort, 102 patients |
| Kamali et al. (2020) | Reproducibility of thresholds | Cohort with repeat scans, 45 patients |
| Gunturiz-Beltrán et al. (2022) | Thresholds for the right atrium | Cohort with healthy volunteers, 53 people |
| Guichard et al. (2024) | LGE channels and functional mapping | Single-center cohort, 40 patients |
| Pittorru et al. (2026) | Ring-like scar and VT recurrence | Single-center cohort, 37 patients |
| Russo et al. (2017); Nazarian et al. (2017) | MRI with devices not labelled for MRI | Prospective registries |
| Sheagren et al. (2025) | LGE with a defibrillator present | Animal study |
Interview questions
Try each question before you open its model answer. In a real interview, say which assumptions you are making.
Q1. What frequency does a 3 T scanner transmit at, and why?
Model answer
About 127.7 MHz: the proton Larmor frequency, \((\gamma/2\pi)B_0\) with 42.577 MHz/T (Equation 1). Only spins precessing at that frequency respond to the RF pulse.
Q2. Derive the null point of inversion recovery.
Model answer
\(dM_z/dt = (M_0 - M_z)/T_1\) with \(M_z(0) = -M_0\) gives \(M_0(1 - 2e^{-t/T_1})\), which is zero at \(t = T_1\ln 2\) (Equation 2).
Q3. Why does scar look bright on LGE?
Model answer
Gadolinium stays outside cells and scar has a larger extracellular volume, so it holds more agent, its \(T_1\) falls further (Equation 3), and it recovers faster after the inversion. At the TI that nulls healthy myocardium it still has signal (Kellman and Arai 2012; Bauner et al. 2012).
Q4. Why is subendocardial scar sometimes hard to see?
Model answer
Blood also holds gadolinium and can recover about as fast as scar, so the two have similar signal at the chosen TI (Kellman and Arai 2012). In our model with illustrative blood values, blood’s \(T_1\) after contrast (249 ms) is close to scar’s (238 ms).
Q5. How would you compute ECV, and what do you need?
Model answer
\(T_1\) of myocardium and blood before and after contrast, and the hematocrit: \(\mathrm{ECV} = (1 - \mathrm{Hct})\,\Delta R_{1,\text{myo}}/\Delta R_{1,\text{blood}}\) (Equation 4). It assumes the agent has reached equilibrium between blood and tissue.
Q6. You set TI too short on a magnitude LGE image. What happens?
Model answer
Myocardium is still negative and appears with some signal; scar may be nearer zero, so contrast drops or inverts (lab part D). Phase-sensitive reconstruction keeps the sign and is much less sensitive to TI (Kellman and Arai 2012).
Q7. Why divide atrial wall intensity by the blood pool?
Model answer
LGE intensities have no absolute scale; dividing by the blood-pool mean cancels a common gain (Equation 6). It does not cancel noise, coil shading across the wall, or differences in blood signal from timing and dose.
Q8. Two papers report 5% and 37% atrial fibrosis. Could it be the same patients?
Model answer
Yes. In one cohort, the average IIR burden went from 5.0% to 37.4% as the threshold moved from 1.2 to 0.97 (Boyle et al. 2022). Burden depends on the threshold and the method, so it is only comparable within one method.
Q9. How would you test whether a fibrosis threshold is reliable?
Model answer
Scan the same patients twice and measure the spatial overlap of what each scan calls fibrosis (Dice, m40); compare methods on the same scans; and check against an independent reference such as voltage mapping. Repeat scans favored Otsu thresholding in one study (Kamali et al. 2020); agreement with voltage mapping was 56% in another (Gunturiz-Beltrán et al. 2022).
Q10. DECAAF found fibrosis predicted recurrence. Why did DECAAF II not show benefit from ablating it?
Model answer
A predictor is not necessarily a cause one can remove. In the trial, recurrence was 43.0% against 46.1% with no significant difference (Marrouche et al. 2022), and the fibrosis-guided lesions did not uniformly become scar (Akoum et al. 2024). Measurement variability (lab part E) and lesion delivery both stand between the marker and the outcome.
Q11. What can LGE add before VT ablation?
Model answer
The extent of scar and candidate conducting channels through it. In one cohort, channels found automatically on LGE were often unmasked by functional mapping (Guichard et al. 2024); in another, channel number and length were associated with recurrence (Pittorru et al. 2026). Both are single-center studies.
Q12. A patient has a defibrillator. What do the registries say about MRI, and what limits them?
Model answer
Under protocols at 1.5 T, scans in patients with devices not labelled for MRI had no long-term clinically significant adverse events in 2,103 scans (Nazarian et al. 2017) and no lead failures in 1,500 nonthoracic scans, with one replacement when the protocol was not followed (Russo et al. 2017). They are registries at one field strength under strict protocols; the device also degrades the image, which special sequences try to address (Sheagren et al. 2025).
Code lab
The lab goes from spins to a burden number and ends with the evidence chain.
| Part | What you do | Data |
|---|---|---|
| A | The Larmor frequency (Equation 1) | CODATA 2022 |
| B | Inversion recovery from the Bloch equations (Equation 2) | our own simulation |
| C | Gadolinium, T1 and extracellular volume (Equation 3, Equation 4) | published T1 values |
| D | Choosing the inversion time (Equation 5) | published T1 values |
| E | Thresholds and burden on a synthetic wall (Equation 6) | our own phantom |
| F | Blood-pool normalization on real heart MRI | Medical Segmentation Decathlon (Antonelli et al. 2021) |
| G | Reading the evidence | published results |
- Python notebook: labs/python/m52_lab.ipynb. Run it from the course repository after
pip install -r requirements.txt, or open it in Google Colab. - Data: part F downloads the Decathlon heart task once (several hundred MB, CC BY-SA 4.0), as in m40; it is MRI but not LGE, so it shows normalization, not fibrosis. Set
EPML_SKIP_REAL_DATA=1to skip it. Everything else is published values or our own code.
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. What is the Larmor frequency at 1.5 T, roughly?
Answer
\(42.6 \times 1.5 \approx 64\) MHz (Equation 1).
S2. A tissue has \(T_1\) = 400 ms. When is it nulled after an inversion?
Answer
\(400 \times 0.69 \approx 277\) ms (Equation 2).
S3. Native \(T_1\) is 1000 ms and \([\mathrm{Gd}]\) = 0.4 mmol/L with \(r_1\) = 4.5 L/(mmol·s). What is \(T_1\)?
Answer
\(1/(1 + 1.8)\) s ≈ 357 ms (Equation 3).
S4. \(\Delta R_1\) is 1.5/s in myocardium and 5/s in blood, Hct 0.4. What is the ECV?
Answer
\(0.6 \times 1.5/5 = 0.18\) (Equation 4).
S5. A scanner’s gain doubles between two scans. Does the IIR change?
Answer
No; wall and blood both double and the ratio is unchanged (Equation 6). Noise and shading are another matter.
S6. True or false: a higher IIR threshold always gives a lower burden.
Answer
True for one scan: burden is the share above the threshold, which cannot grow as the threshold rises (Equation 6). Across scans or methods, anything can happen.
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.
- How late enhancement is made. Kellman and Arai explain the choice of inversion time, magnitude and phase-sensitive reconstruction, and common artifacts, with example images and protocol tables [Kellman and Arai (2012); free full text]. This module: choosing the inversion time.
- Repeat scans and thresholds. Kamali and colleagues compare four thresholding methods on scans repeated three months apart [Kamali et al. (2020); free full text]. This module: image intensity ratio and thresholds.
- Two methods, one cohort. Boyle and colleagues compare fibrosis maps from two methods on the same atria [Boyle et al. (2022); free full text]. This module: from the field.
Further reading
- Physics. Late enhancement techniques (Kellman and Arai 2012); \(T_1\) and ECV in infarcts (Bauner et al. 2012); constants (National Institute of Standards and Technology 2024).
- Atrial fibrosis. From association to trial (Marrouche et al. 2014, 2022; Akoum et al. 2024); methods and thresholds (Boyle et al. 2022; Kamali et al. 2020; Gunturiz-Beltrán et al. 2022).
- Ventricular scar. Channels and mapping (Guichard et al. 2024); scar pattern and outcome (Pittorru et al. 2026).
- Devices. Registries (Russo et al. 2017; Nazarian et al. 2017); imaging with a defibrillator present (Sheagren et al. 2025).
- Related modules. m40 for AI segmentation and Dice, m51 for CT anatomy, m53 for lesions and real-time MRI, m35 for diagnostic accuracy and m45 for levels of evidence.