m20 · Geometry from catheter positions: point clouds to surfaces
Education only. Not for diagnosis, treatment or any clinical decision.
In 60 seconds
Before there is an activation map there is a shape. A mapping system records where the catheter has been, mostly against the wall, and turns that cloud of positions into a closed surface. The simplest wrap, the convex hull, cannot follow a concavity, so its error at a notch equals the notch’s depth however many points you collect. A volumetric approach inflates each position by a radius \(r\), fills the enclosed cavity and shrinks back by \(r\), a morphological closing: it follows concavities wider than about \(2r\), but leaves gaps if positions are farther apart than that, and rounds off detail if \(r\) is large. Surfaces are then smoothed, and plain Laplacian smoothing shrinks them: a regular \(n\)-gon loses the fraction \(\lambda(1 - \cos 2\pi/n)\) of its radius every step, which Taubin’s alternating steps cancel. Volumes follow from the divergence theorem, and two surfaces are compared by their Hausdorff and mean distances.
From the field
- Geometry comes from sampled sites. An early electroanatomic system reconstructed a chamber’s 3D geometry by sampling many endocardial sites with a located catheter, and found geometry and activation repeatable in pigs (Gepstein et al. 1997).
- Acquisition conditions change the shape. Left atrial maps collected regardless of respiratory phase overestimated pulmonary vein ostia compared with maps collected at end expiration (Beinart et al. 2011).
- Anatomy is checked against imaging. One study compared pulmonary vein diameters from a field-based mapping system with angiography, reporting the mean difference and limits of agreement (Rottner et al. 2021).
Derivations
Symbols used throughout. \(P\) the set of catheter positions; \(H(P)\) its convex hull; \(B_r\) a ball (disk in 2D) of radius \(r\); \(\oplus\), \(\ominus\) dilation and erosion; \(\mathbf x_i\) surface vertices and \(N(i)\) their neighbors; \(\lambda\), \(\mu\) smoothing steps; \(n\) polygon vertices; \(a = 1 - \cos(2\pi/n)\); \(\mathbf v_a\), \(\mathbf v_b\), \(\mathbf v_c\) triangle vertices; \(V\) volume; \(A\), \(B\) two surfaces; \(d\) Euclidean distance.
The convex hull
The convex hull is the smallest convex set that contains every position:
\[ H(P) = \Bigl\{\textstyle\sum_i w_i\mathbf p_i : w_i \ge 0,\ \sum_i w_i = 1\Bigr\}, \qquad d(\mathbf w, \partial H) = \text{depth of the concavity at a wall point } \mathbf w \tag{1}\]
For a convex chamber the hull of dense wall samples converges to the wall. For the lab’s chamber, with a shallow notch and a vein-like stub, the hull’s largest distance to the true wall stayed above 7 mm whether 40 or 640 positions were used, mostly where it bridges between the stub and the body (lab part B).
Inflate, fill, shrink
Treat each position as a ball of radius \(r\), fill the cavity they enclose, and shrink back:
\[ S_r(P) = \operatorname{fill}\bigl(P \oplus B_r\bigr) \ominus B_r \tag{2}\]
Dilating a disk of radius \(R\) by \(B_r\) gives radius \(R + r\), and eroding returns \(R\), so the shrink step removes the inflation. Concavities wider than about \(2r\) survive; gaps between positions wider than about \(2r\) leave the band open, so nothing is enclosed. In the lab, 40 or 80 positions left gaps at \(r\) = 3 and 5 mm; 320 positions with \(r\) = 5 mm gave a mean distance to the true wall of 0.50 mm, while \(r\) = 8 mm rounded the notch (largest error 3.0 mm). Without the shrink step the wall moved out by about \(r\) (lab part C). Real systems add catheter size, contact and filtering to this idea; this is a generic version.
Simulation
Assumptions
- Model: a 2D slice of an illustrative chamber on a 1-mm grid; catheter positions within 2 mm of the wall from a fixed random sequence; reconstruction by Equation 1 and Equation 2 with disk structuring elements and 4-connected hole filling; distances by Equation 5. Tested against
epml.geometry. - Parameters: the number of positions and the radius are adjustable.
- What it ignores: the third dimension, catheter size and contact, motion (m19), and localization error (m16–m18).
Key takeaways
- A convex hull cannot see concavities.
- Inflate–fill–shrink needs positions closer than about \(2r\) along the wall; \(r\) trades gaps against detail.
Smoothing without shrinking
The umbrella operator moves each vertex a fraction \(\lambda\) toward the mean of its neighbors. On a regular \(n\)-gon of radius \(\rho\):
\[ \mathbf x_i \leftarrow \mathbf x_i + \lambda\Bigl(\tfrac{1}{\lvert N(i)\rvert}\textstyle\sum_{j\in N(i)}\mathbf x_j - \mathbf x_i\Bigr), \qquad \rho \leftarrow (1 - \lambda a)\rho, \qquad \text{Taubin: } \rho \leftarrow (1 - \lambda a)(1 - \mu a)\rho \tag{3}\]
Repeated Laplacian steps shrink every surface. Taubin’s method follows each shrinking step \(\lambda\) with an inflating step \(\mu < -\lambda\) (Taubin 1995). In the lab, 200 Laplacian steps shrank a noisy 20-mm outline to 12.3 mm, while Taubin smoothing kept it at 20.5 mm; on a noisy sphere, 30 Laplacian steps cut the volume by 40% and Taubin steps changed it by 2.5% (lab part D).
Simulation
Assumptions
- Model: umbrella-operator smoothing of a closed 64-vertex outline with fixed random radial noise of ±1.5 mm, by Equation 3 with \(\mu = -\lambda - 0.03\). Tested against
epml.geometry. - Parameters: the number of iterations and the step size are adjustable.
- What it ignores: mesh irregularity, cotangent weights (m06) and feature-preserving smoothing.
Volume and surface distance
For a closed triangle mesh with outward-facing triangles, the divergence theorem gives the volume, and two surfaces are compared by their worst-case and mean distances:
\[ V = \frac16\sum_{\text{triangles}}\mathbf v_a\cdot(\mathbf v_b\times\mathbf v_c), \qquad d_H(A, B) = \max\Bigl\{\sup_{\mathbf a\in A} d(\mathbf a, B),\ \sup_{\mathbf b\in B} d(\mathbf b, A)\Bigr\} \tag{4}\]
\[ \bar d(A, B) = \frac{\sum_{\mathbf a\in A} d(\mathbf a, B) + \sum_{\mathbf b\in B} d(\mathbf b, A)}{\lvert A\rvert + \lvert B\rvert} \tag{5}\]
Icospheres of radius 10 mm approach the sphere’s volume from below as they are refined: 4152.7 mm³ with 1280 faces and 4179.7 mm³ with 5120, against 4188.8 mm³ (lab part E). Isosurfaces of volumetric data are commonly extracted as triangle meshes by marching cubes (Lorensen and Cline 1987).
Key takeaways
- Laplacian smoothing shrinks; Taubin smoothing mostly does not.
- Report both the worst-case and the mean surface distance.
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. Every activation or voltage map is drawn on a geometry built from catheter positions (Gepstein et al. 1997). How positions are collected changes that geometry: respiratory phase changed pulmonary vein ostial dimensions in one study (Beinart et al. 2011). Anatomy from mapping systems is validated against imaging (Rottner et al. 2021), and registered to CT or MRI in m21.
What the evidence is. An early validation in animals (Gepstein et al. 1997), a clinical comparison of acquisition methods (Beinart et al. 2011) and a comparison with angiography (Rottner et al. 2021).
What a reader should ask:
- How was the surface built, and with what radius or smoothing? (Equation 2, Equation 3)
- Are there concavities the method cannot follow? (Equation 1)
- How was geometry accuracy measured: worst case or mean? (Equation 5)
Visual summary
Schematic
Assumptions
- Schematic: not to scale, no quantitative claim. Labels follow Equation 1 and Equation 2.
How experts think
When looking at a reconstructed chamber, ask:
- Where were positions dense, and where sparse? Gaps become flat patches or holes (Equation 2).
- Which concavities could the method miss? (Equation 1)
- How much smoothing, and does it shrink? (Equation 3)
- Against what was the geometry checked, and by which distance? (Equation 5)
- Were positions corrected for motion and localization error first? (m16–m19)
One sentence to keep: a reconstructed chamber is the shape of where the catheter went, filtered by the algorithm.
Cheat sheet
| Quantity | Result |
|---|---|
| Convex hull (Equation 1) | smallest convex set; error at a notch equals its depth |
| Closing (Equation 2) | \(\operatorname{fill}(P\oplus B_r)\ominus B_r\); needs wall spacing below about \(2r\) |
| Smoothing (Equation 3) | n-gon factor \(1 - \lambda a\); Taubin \((1 - \lambda a)(1 - \mu a)\) |
| Volume (Equation 4) | \(\frac16\sum\mathbf v_a\cdot(\mathbf v_b\times\mathbf v_c)\); Hausdorff distance |
| Mean distance (Equation 5) | symmetric mean of point-to-surface distances |
Values used in this module
| Value | Source |
|---|---|
| Chamber shape, sampling depth, radii, noise, step sizes | illustrative |
| Hull and reconstruction errors, shrinkage, volumes | our own code |
Interview questions
Try each question before you open its model answer. In a real interview, say which assumptions you are making.
Q1. Why does a convex hull overestimate a chamber with concavities?
Model answer
It is convex by definition, so it bridges any notch; the error equals the notch’s depth regardless of sample count (Equation 1).
Q2. Describe a volumetric way to build a surface from catheter positions.
Model answer
Inflate each position by a ball of radius \(r\), fill the enclosed cavity, shrink by \(r\), and extract an isosurface (Equation 2) (Lorensen and Cline 1987).
Q3. What goes wrong if \(r\) is too small or too large?
Model answer
Too small: gaps in the inflated band, so nothing is enclosed. Too large: concavities and fine detail are rounded off (Equation 2).
Q4. Why shrink after inflating?
Model answer
Inflation moves the wall out by \(r\); erosion by the same ball removes that bias (Equation 2).
Q5. Derive the shrink factor of umbrella smoothing on a regular polygon.
Model answer
The neighbor mean of a vertex at radius \(\rho\) lies at radius \(\rho\cos(2\pi/n)\) along the same direction; moving by \(\lambda\) toward it gives \(\rho(1 - \lambda(1 - \cos 2\pi/n))\) (Equation 3).
Q6. How does Taubin smoothing avoid shrinkage?
Model answer
It alternates a shrinking step \(\lambda\) with an inflating step \(\mu < -\lambda\), so smooth shapes keep their size while high-frequency noise is still damped (Equation 3) (Taubin 1995).
Q7. How do you compute a closed mesh’s volume?
Model answer
Sum the signed tetrahedron volumes \(\frac16\mathbf v_a\cdot(\mathbf v_b\times\mathbf v_c)\) over outward-oriented triangles (Equation 4).
Q8. Hausdorff or mean surface distance: which would you report?
Model answer
Both: the Hausdorff distance shows the worst local error, the mean shows overall agreement (Equation 4, Equation 5).
Q9. Why are wall positions more useful than interior ones?
Model answer
The wall is what is being reconstructed; interior positions only confirm the cavity is open, and enclosure depends on the spacing along the wall (Equation 2).
Q10. A map’s ostium looks larger than on CT. List causes.
Model answer
Respiratory motion during collection, a large inflation radius, smoothing, localization scaling (m17) and registration (m21) (Beinart et al. 2011).
Q11. How would you validate a reconstruction algorithm?
Model answer
On phantoms or synthetic shapes with known truth, varying sample density and noise, reporting volume error and surface distances (Equation 4, Equation 5).
Q12. What did the early electroanatomic validation say about geometry?
Model answer
Sampling many endocardial sites reconstructed the chamber’s 3D geometry, and geometry and activation were repeatable in pigs (Gepstein et al. 1997).
Code lab
The lab reconstructs a chamber from catheter positions.
| Part | What you do | Data |
|---|---|---|
| A | A chamber and catheter positions | synthetic |
| B | The convex hull (Equation 1) | synthetic |
| C | Inflate, fill, shrink (Equation 2) | synthetic |
| D | Smoothing without shrinking (Equation 3) | synthetic |
| E | Volume and surface distance (Equation 4, Equation 5) | synthetic |
- Python notebook: labs/python/m20_lab.ipynb. Run it from the course repository after
pip install -r requirements.txt, or open it in Google Colab. - Data: none to download.
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 disk of radius 20 mm dilated by a 5 mm disk. New radius?
Answer
25 mm (Equation 2).
S2. A notch 6 mm deep. Convex hull error there?
Answer
6 mm (Equation 1).
S3. Wall positions 12 mm apart. Smallest inflation radius that closes the band, roughly?
Answer
About 6 mm (Equation 2).
S4. Regular hexagon (\(n\) = 6), \(\lambda\) = 0.5. Radius factor per Laplacian step?
Answer
\(1 - 0.5(1 - 0.5) = 0.75\) (Equation 3).
S5. A tetrahedron with legs 3, 4 and 5 mm along the axes. Volume?
Answer
\(3 \times 4 \times 5/6 = 10\) mm³ (Equation 4).
S6. Concentric spheres of radius 20 and 22 mm. Hausdorff distance?
Answer
2 mm (Equation 4).
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.
- Smoothing without shrinkage. Taubin’s signal-processing view of surface smoothing and its two-step filter (Taubin 1995). This module: smoothing without shrinking.
- Marching cubes. Lorensen and Cline’s case table and medical surface examples (Lorensen and Cline 1987). This module: volume and surface distance.
- Geometry from sampled sites. Gepstein and colleagues’ reconstructed chambers (Gepstein et al. 1997). This module: the whole of it.
Further reading
- Surface algorithms. Marching cubes (Lorensen and Cline 1987); fair surface design (Taubin 1995).
- Mapping geometry. Early electroanatomic reconstruction (Gepstein et al. 1997); respiration and geometry (Beinart et al. 2011); anatomy against angiography (Rottner et al. 2021).
- Related modules. m06 (meshes), m13 (catheters), m19 (motion), m21 (image integration), m24 (activation maps), m30 (capstone mapping system).