Bench characterization toolkit for the drive mechanics of intravascular OCT (Optical Coherence Tomography) catheters — rotational stability, NURD (Non-Uniform Rotational Distortion) measurement and correction, and pullback accuracy — with ground-truth-validated metrics and an automated pass/fail characterization report.
Engineering verification & validation of drive mechanics only. This is a simulation / signal-analysis bench and makes no clinical or diagnostic claims.
Same acquired A-lines, two angular grids: the uncorrected panel paints them at
assumed-uniform angles (the vessel wall warps away from the true outline); the
corrected panel rebuilds the grid from the encoder via correct_nurd() and the
wall lands back on the phantom. Regenerate with python scripts/make_demo_gif.py.
An OCT catheter images by spinning a fiber-optic core (typically ~100 rev/s) at the distal tip while pulling it back along the vessel. Image quality is governed by how uniformly that core rotates and how accurately it is pulled back:
- NURD — torque-cable friction/binding makes the distal optics speed up and slow down within a single revolution, smearing features azimuthally. It is the #1 image-quality killer and the headline metric here.
- Rotational jitter / wow & flutter — revolution-to-revolution speed instability.
- Pullback accuracy & uniformity — sets the longitudinal (pullback-axis) image scale.
A synthetic (or measured) drive signal — the rotation and pullback encoder
channels — flows through the metric analyzers, each result is compared against
spec.yaml, and a PASS/FAIL characterization report is stamped.
flowchart LR
A["Drive signal<br/>rotation + pullback<br/>encoders"] --> B["Metrics<br/>• NURD index<br/>• spectral NURD<br/>• rotational stability<br/>• pullback accuracy<br/>• angular error"]
B --> C["spec.yaml<br/>acceptance limits"]
C --> D["PASS / FAIL<br/>characterization report"]
| Metric | Definition |
|---|---|
| NURD index | std(ω) / mean(ω) over a revolution (dimensionless, ×100 → %) |
| Spectral NURD | Hann-windowed FFT of ω(t); harmonic content at k·f_rot relative to DC, with the dominant harmonic |
| Per-rotation NURD map | NURD index for each completed revolution |
| Rotational stability | period-to-period std & peak-to-peak; wow (slow) vs flutter (fast) split; RPM mean/std |
| Pullback accuracy | measured vs commanded mm/s, % error, speed CV, max deviation, position-vs-time linearity R² |
| Angular position error | cumulative commanded-vs-measured θ per revolution |
A synthetic drive-signal generator injects known NURD harmonics, jitter and pullback error, so every metric is validated against ground truth in the test suite (inject a known distortion → assert the analyzer recovers it).
Generated from the real API by scripts/make_figures.py (fixed seeds, headless),
contrasting a clean in-spec drive against a degraded one (binding torque cable:
strong 1× + 3× NURD, jitter, 8% pullback overspeed). Engineering characterization
of drive mechanics only — no clinical or diagnostic interpretation.
Instantaneous rotation rate across four revolutions. The clean drive (blue)
holds ~6000 rpm; the degraded drive (red) speeds up and slows down within each
revolution — this within-rev modulation is NURD. The NURD index std(ω)/mean(ω)
rises from 0.7% to 13.3%.
Harmonic content of ω(t) at multiples of the rotation frequency (f_rot =
100 Hz), normalized to DC. The dominant once-per-revolution (1×) component
flags the bind; the recovered ratios match the injected harmonic amplitudes
(H_k ≈ a_k/2).
Pullback position vs time with a constant-velocity fit overlaid. The measured speed (21.6 mm/s) exceeds the commanded 20.0 mm/s by +8.0% — a calibration bias — while linearity stays high (R² = 0.99998).
NURD index computed for each completed revolution, against the spec.yaml
acceptance limit (5%). The clean drive stays near 0.7%; the degraded drive sits
~13% across every revolution, in the fail zone.
Measuring NURD is only half the job — the toolkit can also correct it. NURD arises because A-lines are acquired uniformly in time while the fiber core, due to torque-cable friction, rotates non-uniformly in angle; consecutive A-lines are therefore separated by a varying azimuthal angle, smearing the frame. Real OCT systems fix this by resampling the A-line stream from its non-uniform angular spacing back onto a uniform angular grid, using the measured instantaneous rotation (rotary encoder, or here the generator ground truth).
oct_drive_test.correction implements exactly that re-gridding:
- Build a uniform angular grid (
n_per_revpoints per revolution, an exact integer per 2π so every revolution samples identically). - Invert the measured cumulative angle
θ(t)— for each target grid angle, interpolate the time at which the core actually passed through it (t = θ⁻¹(angle)). - Resample the per-A-line payload at those corrected times → an equi-angular A-line stream.
from oct_drive_test import correction
# measured (NURD-distorted) drive signal from the generator/encoder
corr = correction.correct_drive_signal(sig, n_per_rev=512)
corr.nurd_index_before # std(ω)/mean(ω) of the raw, uniform-time stream
corr.nurd_index_after # residual on the corrected uniform-angle grid (~0)
corr.improvement_factor # before / after
corr.theta_grid, corr.t_grid # uniform angle grid + corrected A-line timesEncoder-limited residual. Correcting with the generator's own angle makes
the residual collapse to numerical zero — a tautology. oct_drive_test/encoder.py
therefore models the real measurement chain (edge-timestamp decode: count-boundary
crossings, electrical timing jitter, monotone decode), and
encoder.true_residual_nurd() evaluates the true angle at the corrected
A-line times. Under a 2048-count encoder with 0.5 µs edge jitter at 6000 rpm,
a 0.16 NURD index corrects to a genuine ~0.026 residual — the physically
meaningful limit, validated in tests/test_encoder_limited.py.
The residual is the same std/mean NURD index applied to the corrected
stream's angular increments (residual_nurd), not a relaxed metric — on a
uniform-angle grid those increments are constant, so the index collapses to the
numerical floor. The validation suite injects a known ~13% NURD, confirms the
index is high, then asserts the residual drops by many orders of magnitude after
correction, while a clean (NURD-free) signal is left essentially untouched (no
distortion is injected).
A-line azimuthal sampling before vs after angular resampling, for the degraded (binding-cable) drive. Top: the spacing between consecutive A-lines wobbles 0.75°–2.1° when sampled uniformly in time (red), but is pinned to the ideal 1.406°/A-line after re-gridding (blue). Bottom: the cumulative angular position error vs the ideal uniform grid sweeps ±12° before correction and collapses to ~0 after. The NURD index falls from 13.9% to the numerical floor. Engineering characterization of drive-correction signal processing only — no clinical or diagnostic interpretation.
pip install -e ".[dev]" # numpy, scipy, pyyaml (+ pytest, matplotlib)# Two-scenario example: a clean in-spec drive (PASS) + a degraded drive (FAIL).
python examples/run_characterization.py
# Characterize a synthetic signal against a spec and stamp PASS/FAIL:
oct-drive-report --spec spec.yamlfrom oct_drive_test import nurd
from oct_drive_test.generator import generate_drive_signal, NurdHarmonic
sig = generate_drive_signal(...) # inject known NURD/jitter/pullback (see example)
summary = nurd.nurd_summary(sig)
print(summary["nurd_index_pct_overall"], summary["dominant_harmonic_k"])spec.yaml declares acceptance limits (nominal RPM and tolerance, max NURD index,
pullback accuracy, jitter ceiling). The report runner measures a dataset and
stamps PASS/FAIL per limit — the structure of a real V&V characterization record.
oct_drive_test/
├── generator.py # synthetic drive-signal generator (ground-truth injection)
├── nurd.py # NURD index + spectral / per-rotation NURD
├── correction.py # NURD correction: encoder-based angular resampling
├── rotation.py # rotational stability: jitter, wow/flutter, RPM stats
├── pullback.py # pullback speed accuracy & uniformity
├── angular.py # angular position error
├── spec.py # spec.yaml loading + pass/fail characterization report
├── backlash.py # torque-cable wind-up / backlash (stretch)
└── aline.py # A-line trigger-interval timing (stretch)
pytest -q # ground-truth validation of every metricSample reports: clean drive · degraded drive — generated by oct-drive-report, no setup needed to read them.
MIT © Eric Rosenfeld — see LICENSE.