m55 · Capstone: a multimodal image-guided procedure on a digital phantom, from segmentation and registration to overlay and an error and dose budget
Education only. Not for diagnosis, treatment or any clinical decision.
In 60 seconds
An image-guided procedure chains several steps, and each adds error. In this capstone one synthetic left atrium goes through all of them: a CT-like volume (m51), segmentation, a surface, registration to simulated map points, prediction of the error at targets, an overlay on a simulated fluoroscopy view (m49), the same registration on real anatomy, and an error and dose budget. The central lesson comes from registration theory: the error a system reports at its own landmarks (fiducial registration error) does not tell you the error at a target, and the expected target error grows with distance from the landmarks (Fitzpatrick et al. 1998; Danilchenko and Fitzpatrick 2011). Early clinical and animal studies of CT merged with the map show both kinds of number (Kistler et al. 2006; Dong et al. 2006). Independent errors add in quadrature, so the largest one dominates.
From the field
- Reported is not true. In 30 patients, merging CT with the map reported 6.4 ± 2.8 mm with landmarks and 2.3 ± 0.4 mm with a surface fit (Kistler et al. 2006); those are distances between fitted points. In nine dogs with markers sewn to the heart, ablation guided by the registered CT landed 1.8–2.7 mm from the markers, depending on the chamber (Dong et al. 2006).
- A small fit error can hide a poor registration. The expected fiducial error depends only on the number of landmarks and their localization error, not on their layout, so a poor layout can look as good as a good one (Fitzpatrick et al. 1998).
- No goodness-of-fit number predicts a given case’s target error. In simulations, the correlation between fiducial and target errors was negligible (below 0.1) for the weightings used in commercial guidance systems (Danilchenko and Fitzpatrick 2011).
- One view hides depth. A projection shows error across the beam only; the part along the beam needs a second view (lab part E).
Derivations
C-arm projection
A C-arm has a point source a distance SOD (source to object) before the isocenter and a flat detector a distance SID (source to image) from the source. By similar triangles, a point at lateral position \(x\) and depth \(y\) beyond the isocenter lands at
\[ u = \frac{\mathrm{SID}}{\mathrm{SOD} + y}\,x, \qquad v = \frac{\mathrm{SID}}{\mathrm{SOD} + y}\,z, \qquad M(y) = \frac{\mathrm{SID}}{\mathrm{SOD} + y} \tag{1}\]
where \(z\) is the head-foot position and \(M\) the magnification. With an illustrative SOD of 750 mm and SID of 1100 mm, \(M\) = 1.47 at the isocenter, 1.59 for a point 60 mm nearer the source and 1.36 for one 60 mm beyond it (lab part A). Rotating the C-arm turns the beam. An error along the beam changes only the magnification, by about 1% per 6 mm here, so one view barely shows it.
Exact
Assumptions
- Model: Equation 1, a point source and a flat detector, with the C-arm rotated about the head-foot axis; tested against
epml.capstone. - The atrium is an ellipsoid with semi-axes 25, 20 and 22 mm, drawn as latitude and longitude rings. SOD, SID, the error and the angles are illustrative.
- What it ignores: rotational errors, deformation, cardiac and respiratory motion, image distortion, and how overlays are aligned in real systems.
Key takeaways
- Magnification depends on depth.
- One projection shows two of three error directions.
Fiducial error and target error
Register \(N\) landmarks (fiducials), each localized with isotropic error of RMS value FLE. To first order (Fitzpatrick et al. 1998; Danilchenko and Fitzpatrick 2011),
\[ \langle \mathrm{FRE}^2 \rangle = \left(1 - \frac{2}{N}\right)\langle \mathrm{FLE}^2\rangle, \qquad \langle \mathrm{TRE}^2(\mathbf r) \rangle \approx \frac{\langle \mathrm{FLE}^2\rangle}{N}\left(1 + \frac13\sum_{k=1}^{3}\frac{d_k^2}{f_k^2}\right) \tag{2}\]
where FRE is the fiducial registration error the system can report, TRE is the target registration error at a point \(\mathbf r\), \(d_k\) is the distance of \(\mathbf r\) from principal axis \(k\) of the landmarks and \(f_k^2\) is the mean squared distance of the landmarks from that axis. FRE does not depend on the layout; TRE grows away from the landmarks, fastest along directions in which they are bunched. In lab part D the formula matches simulation within about 10% at 2 mm FLE; with larger errors relative to the layout the simulated TRE exceeds it by about 12%, because the formula is first order.
Approximate model
Assumptions
- Model: the first-order expression Equation 2 for rigid point registration with independent, isotropic localization error of the same size at every landmark (Fitzpatrick et al. 1998; Danilchenko and Fitzpatrick 2011); tested against
epml.capstone, whose Monte Carlo agrees within about 10% at 2 mm FLE. - Landmark layouts, their number and the localization error are illustrative; the map is a plane through the landmarks.
- What it ignores: anisotropic or unequal localization error, deformation between image and map, and surface refinement after the landmarks.
Key takeaways
- A small FRE is not evidence of a small TRE.
- Spread the landmarks around the region where accuracy matters.
The error budget
If the steps contribute independent, zero-mean errors with RMS values \(\sigma_i\),
\[ \sigma_{\text{total}} = \sqrt{\sum_i \sigma_i^2} \tag{3}\]
so the largest term dominates. In lab part G, segmentation (0.5 mm), registration at targets (1.5 mm) and an illustrative catheter localization error (1.0 mm) combine to 1.8 mm, not the 3.0 mm a simple sum would give, and the registration term carries 63% of the variance. A shared bias, such as a segmentation that always sits half a voxel inside the wall, adds directly instead. The dose side of the budget adds the kerma-area product of each step (m49) and can be compared with published reference levels for EP procedures (Greffier et al. 2026).
Key takeaways
- Improve the largest term first.
- Biases do not average out; budget them separately.
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 chain shows up in EP. Merging pre-procedural CT or MRI with the electroanatomic map, and overlaying models on fluoroscopy, are described in the cited studies and in modules m21 and m49–m52. The validation studies measured registration either as a reported fit error in patients (Kistler et al. 2006) or as the error at marked targets in animals (Dong et al. 2006).
What the evidence is. One clinical series and one animal study (Kistler et al. 2006; Dong et al. 2006), and a body of registration theory and simulation (Fitzpatrick et al. 1998; Danilchenko and Fitzpatrick 2011). The capstone’s own numbers come from a phantom with illustrative noise levels and are teaching values, not estimates for any system.
What a reader should ask:
- Which error is quoted? Fit error at landmarks or surface points, or error at a target (Equation 2).
- Where are the targets relative to the landmarks? Error grows away from them.
- How many views? One projection hides the error along the beam (Equation 1).
Visual summary
Schematic
Assumptions
- Schematic: not to scale, no quantitative claim. Labels follow Equation 1 to Equation 3.
How experts think
Before trusting an overlay or a merged image, ask:
- What was registered to what, with which points? Landmarks, a surface, or both.
- Which error is reported? FRE and surface distance are not TRE (Equation 2).
- Where are my targets? Far from the landmarks, expect more error.
- Did I check from a second angle? One view hides the error along the beam (Equation 1).
- What dominates the budget? Fix the largest term first (Equation 3).
One sentence to keep: the number a system reports about its fit is not the error where you will act.
Cheat sheet
| Quantity | Result |
|---|---|
| Projection (Equation 1) | \(u = \mathrm{SID}\,x/(\mathrm{SOD} + y)\); magnification \(\mathrm{SID}/(\mathrm{SOD} + y)\) |
| Fiducial error (Equation 2) | \(\langle\mathrm{FRE}^2\rangle = (1 - 2/N)\langle\mathrm{FLE}^2\rangle\), layout-independent |
| Target error (Equation 2) | \(\langle\mathrm{TRE}^2\rangle \approx \frac{\langle\mathrm{FLE}^2\rangle}{N}(1 + \frac13\sum d_k^2/f_k^2)\); at the centroid \(\mathrm{FLE}/\sqrt N\) |
| Budget (Equation 3) | \(\sqrt{\sum\sigma_i^2}\) for independent, zero-mean errors |
Values used in this module
| Value | Source |
|---|---|
| Landmark 6.4 ± 2.8 mm and surface 2.3 ± 0.4 mm reported registration errors | Kistler et al. (2006) |
| 1.8–2.7 mm error at marked targets in dogs | Dong et al. (2006) |
| Reference levels for KAP: accessory pathway 4.10, persistent AF 12.65 Gy·cm² | Greffier et al. (2026) |
| Fiducial-target error correlation below 0.1 in simulation | Danilchenko and Fitzpatrick (2011) |
Studies in this module
| Study | Question | Design |
|---|---|---|
| Fitzpatrick et al. (1998) | Predicting registration error | Theory and simulation |
| Danilchenko and Fitzpatrick (2011) | First-order error prediction, FRE-TRE correlation | Theory and simulation |
| Kistler et al. (2006) | CT merged with the map in AF ablation | Clinical series, 30 patients |
| Dong et al. (2006) | Error at targets after CT registration | Animal study, 9 dogs |
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 difference between FRE and TRE?
Model answer
FRE is the residual at the landmarks used for the registration; TRE is the error at other points, the ones that matter (Equation 2) (Fitzpatrick et al. 1998).
Q2. Why can a small FRE coexist with a large TRE?
Model answer
Expected FRE depends only on \(N\) and FLE, not on the layout, while TRE depends strongly on the layout and the target’s position (Equation 2). A bunched or nearly collinear layout can fit well and still rotate badly about its own axis.
Q3. How would you place four landmarks to register a left atrium?
Model answer
Spread them around the chamber, not in a line or on one side, so that their principal-axis spreads \(f_k\) are large and the targets lie near their centroid (Equation 2).
Q4. Doubling the number of landmarks does what to TRE?
Model answer
For the same layout shape and FLE, it divides \(\langle\mathrm{TRE}^2\rangle\) by two, so TRE by about \(\sqrt2\) (Equation 2).
Q5. A vendor reports a 2-mm surface registration error. What else do you want to know?
Model answer
Whether any error at targets was measured, and how; surface distance can be small even when the model is shifted along the surface (Kistler et al. 2006). Target-level validation needs an independent reference, as in the animal study (Dong et al. 2006).
Q6. An overlay matches the fluoroscopy perfectly in one view. Is the registration right?
Model answer
Not necessarily: an error along the beam changes only the magnification slightly (Equation 1). Check a second view at an angle (lab part E).
Q7. What is the magnification of a point at the isocenter with SOD 750 mm and SID 1100 mm?
Model answer
\(1100/750 \approx 1.47\) (Equation 1).
Q8. Errors of 0.5, 1.5 and 1.0 mm: what is the total?
Model answer
\(\sqrt{0.25 + 2.25 + 1} \approx 1.9\) mm if independent and unbiased (Equation 3); the 1.5-mm term carries most of the variance.
Q9. When does the TRE formula fail?
Model answer
It is first order and assumes independent, isotropic, equal localization errors and a rigid object (Fitzpatrick et al. 1998; Danilchenko and Fitzpatrick 2011). With large errors relative to the layout, deformation, or anisotropic error, it underestimates or misdescribes TRE (lab part D).
Q10. Why does surface registration usually beat landmark registration?
Model answer
It uses many points, so random localization error averages out, and it does not depend on identifying the same anatomical point in both spaces. It can still slide along a symmetric surface, so a landmark start and target-level checks matter (lab part C).
Q11. How do you make the budget useful for a design decision?
Model answer
Estimate each term, add in quadrature, and list each term’s share of the variance (Equation 3); spend effort on the largest share. Keep biases separate, because they add linearly.
Q12. What does the capstone’s dose budget compare against, and what does it leave out?
Model answer
Fluoroscopy kerma-area product against published reference levels (Greffier et al. 2026). It leaves out CT dose, staff dose and any dose that real image-guided steps add or save; those need their own sources (m49, m51).
Code lab
The lab runs the whole chain on one phantom and ends with a budget.
| Part | What you do | Data |
|---|---|---|
| A | C-arm projection and magnification (Equation 1) | our own geometry |
| B | Segment the atrium and measure the segmentation error | m51 phantom |
| C | Register map points: landmarks, then surface | our own simulation |
| D | Predict target error from landmark error (Equation 2) | published formula |
| E | Overlay the model on a fluoroscopy view | m49 phantom |
| F | The same registration on real anatomy | Medical Segmentation Decathlon (Antonelli et al. 2021) |
| G | An error and dose budget (Equation 3) | parts B–E and published reference levels |
- Python notebook: labs/python/m55_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 and m52; set
EPML_SKIP_REAL_DATA=1to skip it. Everything else is our own code and published values.
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. Six landmarks, FLE 3 mm. Expected FRE?
Answer
\(3\sqrt{1 - 2/6} \approx 2.4\) mm (Equation 2).
S2. Same landmarks: expected TRE at their centroid?
Answer
\(3/\sqrt6 \approx 1.2\) mm (Equation 2).
S3. A point 100 mm beyond the isocenter (SOD 750, SID 1100): magnification?
Answer
\(1100/850 \approx 1.29\) (Equation 1).
S4. Errors of 3 and 4 mm, independent. Total?
Answer
5 mm (Equation 3).
S5. True or false: if FRE is small, TRE is small.
Answer
False (Fitzpatrick et al. 1998; Danilchenko and Fitzpatrick 2011).
S6. Which error direction can a single anterior-posterior view not show?
Answer
Front-back, along the beam (Equation 1).
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.
- FRE and TRE. Danilchenko and Fitzpatrick compare predicted and simulated errors and show how weakly FRE and TRE are correlated [Danilchenko and Fitzpatrick (2011); free full text]. This module: fiducial error and target error.
Further reading
- Registration error. Prediction of TRE (Fitzpatrick et al. 1998); first-order analysis with general weighting (Danilchenko and Fitzpatrick 2011).
- Image integration in EP. Clinical series (Kistler et al. 2006); animal validation at targets (Dong et al. 2006).
- Dose. Reference levels for EP procedures (Greffier et al. 2026).
- Related modules. m21 for registration methods, m40 for segmentation, m49 for fluoroscopy and dose, m51 and m52 for the images, and m48 for comparing modalities.