TrueMargin

The question

Image registration aligns two scans. This study asks whether variation between registrations identifies local spatial error. Error ranking, blind spots and calibrated bounds are evaluated separately.

Method

A fixed nine-setting ensemble produces displacement fields. Their local spread is compared with known error from synthetic deformations, inverse-consistency error, image residual and Jacobian deviation.

1. 01Align images with fixed parameter settings
2. 02Measure displacement disagreement
3. 03Compare with known error and baselines

The controlled study contains 30 cases across 10 anatomies and 1,500 sampled locations. Anatomy-level summaries avoid counting spatial samples as independent subjects.

Experiments

The first study measured associations and failure cases. A separate calibration study used 30 calibration and 30 evaluation anatomies across two sources. The interactive plot shows retained comparator data, not simulated results.

Results

All ten anatomy-level spread/error associations were positive. The median correlation confidence interval was [0.3048, 0.8284]. No advantage over inverse-consistency error was established: the paired difference interval crossed zero.

Controlled study10 anatomies

Registration spread and alignment error

Each anatomy has three synthetic deformations with known alignment errors.

Anatomy-level correlation

Each dot shows an anatomy’s median Spearman correlation with known error.

Registration spread
Median 0.6841Inverse-consistency error
Median 0.7203

Spread did not show an advantage over inverse-consistency error. The comparator’s valid-case counts vary by anatomy.

Case-level correlations

Spread/error correlations for all 30 deformation cases.

Six correlations were negative.

39 of 1,500 sampled locations had high error despite low spread. Locations are not independent patients.

Exact retained values and sampling units

Ten anatomies × three deformations × 50 locations. The source tables preserve exact values and case ordering.

Anatomy-level results and valid-case counts · Case-level results

Original technical figure. Calibration results below use a separate cohort.

Failures and bounds

Six of 30 case-level correlations were negative. Thirty-nine of 1,500 observations had high error despite low spread; these failures remain in the analysis.

At nominal 90% coverage, the separate evaluation reached 97.73% and 99.73% mean anatomy-level coverage, with median anatomy radii of 3.93 mm and 4.36 mm. One anatomy had a 70.69 mm median radius. The 95% thresholds were infinite with fifteen calibration groups per source. Reverse-registration failures prevented the planned full-cohort calibrated baseline comparison.

Limitations

These studies do not establish clinical validity, external generalization or uniformly useful bounds. A later calibration-only analysis found source-dependent effects, not a consistent improvement. The work has a technical report, not an accepted peer-reviewed publication.

Reproducibility

From a checkout with Python 3.12, check the saved calibration and evaluation records without imaging data:

python scripts/verify_m6_phase_a_result.py results/m6_phase_a
python scripts/verify_m6_phase_b_result.py results/m6_phase_b

These checks do not rerun registrations. Full experiments require the source data and documented imaging dependencies.

Reproduction guide · Protocols and results

Deferred external study

An independent lung-CT protocol and initial image checks were prepared. Two retained Case7 fields were almost identical after a tolerance change; both reached the iteration cap. This establishes neither convergence nor useful ensemble diversity. Numerical landmarks remain sealed and external landmark validation has not run.

External protocol · Historical execution record