longitudinal/REGISTRATION_BENCHMARK_AFFINE.md

Registration Pipeline Benchmark — CT → T1c (M6) — Affine

Companion to REGISTRATION_BENCHMARK_RIGID.md and REGISTRATION_BENCHMARK_DEFORMABLE.md: the affine (12-DOF) variant of the same benchmark. The rigid benchmark asked "which toolkit aligns the head best with 6 DOF?"; this one asks "how much does the full 12-DOF affine buy on top of rigid — and do the toolkits' affine workflows still rank the same way?" — on the same 10 M6 CT→T1c test cases, the same unified contract, and the same metrics.

Run date: 2026-10-03, 13:50–14:40 local (single run, 6 workers, 24 threads, 2 workers per GPU; ~50 min wall, dominated by impact's ~6 min phase). 10 M6 test cases, 3× NVIDIA A100 80GB PCIe, 96-core host, conda env longitudinal-313.

Unified contract

Identical to the rigid benchmark: the returned transform T must satisfy sitk.Resample(ct, t1c, T, ...) = CT on the T1c grid, i.e. T maps T1c → CT coordinates. fixed = T1c, moving = CT, full volumes (no head box, no mask, no crop, no pad), each toolkit's native automatic transform initialization, CT clipped to [0, 80] HU only for the pipelines where the rigid A/B showed clipping helps (niftyreg, synthmorph).

One change vs the rigid benchmark: every pipeline estimates a full affine (12-DOF) transform instead of a rigid (6-DOF) one — each toolkit's native affine workflow (or, for the antspyx rows, a stock linear-only preset).

The nine pipelines (affine configurations)

pipeline	affine configuration	metric
antspyx	antspyx stock Affine preset (the toolkit's native affine): Affine[0.25], 2100×1200×1200×10 iters, shrink 6×4×2×1, smoothing 3×2×1×0, center-of-mass auto init	Mattes MI (32 bins, 20 % regular sampling)
ants_trsaa	antspyx stock TRSAA preset: Translation[1] → Rigid[1] → Similarity[1] (each 2000×2000×0 iters, 4×2×1, 2×1×0) → Affine[1] × 2 (each 40×20×0, 4×2×1, 2×1×0), COM init — the full TRSA cascade	Mattes MI (32 bins), all stages
ants_reproa	antsRegistrationSyNRepro[a]: Rigid[0.1] + Affine[0.1] (each 1000×500×250×100, 8×4×2×1, 3×2×1×0vox), COM init — the linear-only half of the Repro workflow	GC (1/1/Regular/0.25), both stages
ants_quickreproa	antsRegistrationSyNQuickRepro[a]: Rigid[0.1] + Affine[0.1] (each 1000×500×250×0 — no full-res level —, 8×4×2×1, 3×2×1×0vox), COM init	GC (1/1/Regular/0.25), both stages
elastix	AffineTransform (12-DOF, the rigid benchmark's map with Euler→Affine), ASGD 2000 iters × 3 resolutions, CenterOfGravity auto init	Mattes MI (32 bins), 5000 samples/level
impact	the ImpactReg MR_CT_TS reference map with AffineTransform in place of EulerTransform (same MIND + TotalSegmentator static features, 4 scales 400/300/200/200 iters, 2000 samples, same CoG-init / 0.1 sample-gate adaptations)	IMPACT semantic (L1 + Dice) + Mattes MI
niftyreg	reg_aladin full affine (no -rigOnly), 3-level pyramid, 15 TLS iters/level	NCC block matching (init) + TLS
simpleitk	two-stage (the toolkit's canonical affine workflow): Euler3D stage (2000 iters/level, pyramid [4,2,1], identity start) → AffineTransform stage (same budget) initialized to reproduce the rigid stage exactly — a single-stage affine from identity collapses its scale on these pairs	Mattes MI (32 bins), both stages
synthmorph	affine matrix model (the toolkit's native 12-DOF deep model)	anatomy-aware contrast-invariant network (pretrained)

The ants_trsaa / ants_reproa / ants_quickreproa rows are the linear-only ([a]) antspyx presets — the same stage configurations the antsRegistrationSyNRepro*.sh / antsRegistrationSyNQuickRepro*.sh workflows use for their Rigid + Affine head.

Test cases

Same 10 as the rigid / deformable benchmarks (first 10 M6 labelled cases of m6.discover(), all with a GTV label): m6_22GD5VL2_20230724, m6_22IRRKU6_20240115, m6_23XIMO4M_20230222, m6_24V62QQM_20230223, m6_24V62QQM_20231017, m6_24V62QQM_20240820, m6_26TYQI3C_20221027, m6_2BBN5XA4_20250411, m6_2C4TKI6H_20240506, m6_2CDNN7BH_20240102.

(m6_22GD5VL2_20230724 has an incomplete CT head — a known data issue, not a pipeline failure; it is scored as ct_incomplete in every pipeline.)

Metrics (identical for every pipeline)

• Mattes MI (nats) between the warped CT and the T1c over the MR head (largest connected component of the T1c mask) — achieved alignment, higher = better.
• IMPACT loss (vboussot/ImpactLoss; pretrained TotalSegmentator TS/M730 deep features — weighted L1 over the two deepest layers, full volumes; the same scorer as the other benchmarks) between the warped CT and the T1c — semantic-anatomy similarity, lower = better.
• Registration wall time (seconds, excludes I/O + scoring).

Per-case JSONs additionally retain the production skull-ring QA (gap + concentricity, and the acceptance verdict) for reference.

Result

Alignment is scored by the achieved Mattes MI (nats, warped-CT ~ T1c over the MR head; higher = better) and the ImpactLoss semantic score (warped-CT ~ T1c over the full volumes; lower = better), on the same unclipped scoring volumes as the other benchmarks, so all rows are comparable on the same scales and against the rigid / deformable numbers. niftyreg and synthmorph clip the CT (the moving image) to [0, 80] HU before registration (per the rigid benchmark's clipping A/B); the others register the CT as-is. All nine transforms are affine (12-DOF).

All nine pipelines registered 9/9 cases (the 10th is the known incomplete-CT data issue, unscored in every pipeline). Bold = best for the column (MI higher is better, ImpactLoss lower is better).

pipeline	toolkit	registered	s/reg (med)	ImpactLoss (med/mean)	MI (med/mean, nats)
antspyx	ANTsPy 0.6.3 (Affine preset)	9	12.6	0.156 / 0.155	0.660 / 0.660
ants_trsaa	ANTsPy 0.6.3 (TRSAA preset)	9	30.4	0.156 / 0.155	0.658 / 0.659
ants_reproa	ANTsPy 0.6.3 (Repro[a])	9	63.1	0.160 / 0.159	0.525 / 0.514
ants_quickreproa	ANTsPy 0.6.3 (QuickRepro[a])	9	28.4	0.161 / 0.159	0.523 / 0.514
elastix	ITK-Elastix 0.25.4	9	12.3	0.156 / 0.156	0.651 / 0.645
impact	Docker vboussot/elastix_impact (Elastix 5.2.0 + IMPACT, affine MIND+TS reference map)	9	215.5	0.159 / 0.158	0.623 / 0.567
niftyreg	NiftyReg 2.1.1 (reg_aladin full affine, Docker + CUDA)	9	19.8	0.167 / 0.170	0.278 / 0.330
simpleitk	SimpleITK 2.5.6	9	23.3	0.156 / 0.156	0.654 / 0.641
synthmorph	Docker freesurfer/synthmorph (TensorFlow, pretrained, affine model)	9	27.1	0.167 / 0.171	0.282 / 0.320

Per-case MI (nats, unclipped score; bold = best for the case):

case	antspyx	ants_trsaa	ants_reproa	ants_quickreproa	elastix	impact	niftyreg	simpleitk	synthmorph
m6_22GD5VL2_20230724	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT
m6_22IRRKU6_20240115	0.655	0.647	0.547	0.546	0.638	0.391	0.278	0.654	0.282
m6_23XIMO4M_20230222	0.661	0.660	0.519	0.521	0.656	0.657	0.611	0.660	0.520
m6_24V62QQM_20230223	0.710	0.708	0.568	0.569	0.684	0.708	0.254	0.710	0.252
m6_24V62QQM_20231017	0.628	0.635	0.425	0.423	0.622	0.623	0.473	0.619	0.479
m6_24V62QQM_20240820	0.670	0.673	0.459	0.464	0.670	0.659	0.121	0.545	0.122
m6_26TYQI3C_20221027	0.719	0.717	0.525	0.523	0.675	0.676	0.593	0.699	0.570
m6_2BBN5XA4_20250411	0.590	0.588	0.505	0.507	0.568	0.360	0.069	0.590	0.071
m6_2C4TKI6H_20240506	0.646	0.645	0.546	0.542	0.636	0.451	0.241	0.635	0.256
m6_2CDNN7BH_20240102	0.660	0.658	0.535	0.534	0.651	0.575	0.327	0.656	0.328

Per-case ImpactLoss (semantic, unclipped score; bold = best (lowest) for the case):

case	antspyx	ants_trsaa	ants_reproa	ants_quickreproa	elastix	impact	niftyreg	simpleitk	synthmorph
m6_22GD5VL2_20230724	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT	incomplete-CT
m6_22IRRKU6_20240115	0.1514	0.1516	0.1527	0.1528	0.1509	0.1607	0.1667	0.1513	0.1666
m6_23XIMO4M_20230222	0.1565	0.1564	0.1598	0.1599	0.1565	0.1564	0.1585	0.1563	0.1616
m6_24V62QQM_20230223	0.1466	0.1468	0.1527	0.1529	0.1470	0.1468	0.1670	0.1466	0.1672
m6_24V62QQM_20231017	0.1519	0.1517	0.1599	0.1599	0.1527	0.1533	0.1587	0.1520	0.1581
m6_24V62QQM_20240820	0.1560	0.1559	0.1618	0.1617	0.1564	0.1570	0.1840	0.1599	0.1843
m6_26TYQI3C_20221027	0.1563	0.1570	0.1604	0.1615	0.1615	0.1657	0.1621	0.1585	0.1624
m6_2BBN5XA4_20250411	0.1617	0.1617	0.1622	0.1622	0.1593	0.1662	0.1845	0.1617	0.1849
m6_2C4TKI6H_20240506	0.1559	0.1561	0.1612	0.1613	0.1561	0.1594	0.1759	0.1559	0.1766
m6_2CDNN7BH_20240102	0.1591	0.1590	0.1609	0.1617	0.1587	0.1587	0.1740	0.1587	0.1741

Interpretation

• The four intensity tools form a statistical tie, and affine barely moves them. antspyx (0.660), ants_trsaa (0.658), simpleitk (0.654) and elastix (0.651) all sit within 0.009 nats of each other on the median (ImpactLoss medians 0.156), and per-case MI bests split antspyx 5/9, trsaa 2/9, simpleitk 2/9 (ties). Against the rigid benchmark's per-pipeline medians, affine buys only antspyx 0.652 → 0.660 (+0.008), simpleitk 0.647 → 0.654 (+0.007), elastix 0.646 → 0.651 (+0.005) — the M6 head is essentially rigid between sessions, so the extra 6 DOF mostly absorb inter-session anisotropic scaling (≤ a few %), not misalignment.
• ants_trsaa is the full TRSA cascade for free quality — and it is not free. Translation → Rigid → Similarity → Affine → Affine (5 linear stages) reaches 0.658, i.e. the same as the single-stage Affine preset (0.660) at 2.4× the wall time (30.4 s vs 12.6 s): the cascaded 6-DOF and 9-DOF stages just re-solve what the one 12-DOF stage finds.
• ants_quickreproa ≈ ants_reproa quality at 45 % of the time (0.523 vs 0.525 MI; 28.4 s vs 63.1 s): dropping the full-resolution rigid level costs nothing. But both are clearly the weakest of the top-tier eligible antspyx presets — the GC-based Rigid+Affine head (the linear half of the Repro T1w-T1w workflow) locks a cross-modality-weak pose, the same story as the Repro family in the deformable benchmark (0.530–0.549 MI there).
• impact's affine run degrades vs its own rigid. Median 0.623 (mean 0.567) vs 0.652 for rigid: the semantic 12-DOF landscape has more local minima, and the single CenterOfGravity start falls into two of them (m6_22IRRKU6 0.391, m6_2BBN5XA4 0.360 — both 0.645 / 0.572 in the rigid run). It also roughly doubles in wall time (215.5 s vs 111.2 s) for the extra 6 DOF.
• niftyreg full affine does not fix its gross-pose failures (0.278 vs 0.292 rigid; the two repositioned cases still at 0.069 / 0.121): reg_ aladin's block-matching + TLS is a local optimizer either way, and the output affine is still flipped by the orientation gate on 8/9 cases.
• synthmorph's affine model confirms the rigid-benchmark investigation (0.282 vs 0.277; within ±0.03 per case): the linear part of its output is a near-pure rotation — the extra 6 DOF are essentially unused, and the pretrained matrix head's gross-pose estimation weakness stands for 12 DOF as for 6.
• The deformable stage is still worth it. The best affine median here is 0.660 (antspyx); the best deformable in the companion benchmark is 0.753 (ants_rigid_synonly; 0.723 for the stock SyN preset) — the deformable stage adds +0.06 to +0.09 MI on top of the best affine, which is fine-tissue alignment, not gross pose.
• Production skull-ring QA (reference only, kept in the per-case JSONs): acceptance 4/9 antspyx / ants_reproa / ants_quickreproa / elastix / simpleitk, 3/9 ants_trsaa, 2/9 impact, 1/9 niftyreg / synthmorph. Concentricity: ants_reproa tightest mid-gap (5.6 mm), niftyreg / impact / synthmorph worst on the rotational term (23.5–25.3 mm std) — the same weak correlation with perceived alignment noted in the other benchmarks, so it is reported for reference only.

Transform conventions (affine-specific)

• antspyx linear-only presets write one merged affine per consecutive linear run. The TRSAA cascade (5 linear stages) comes back as a single binary 0GenericAffine.mat that SimpleITK cannot parse; the pipeline recovers it by point-probing the transform (exact for affine maps, verified to NCC 1.0 against ANTs' own warpedmovout on the pilot cases). Repro[a] / QuickRepro[a] (Rigid + Affine) merge the same way.
• niftyreg's -aff file is documented as Affine * Reference = Floating; the fix_orientation gate resolves the direction empirically and inverts it on 8/9 M6 cases (as in the rigid benchmark).
• elastix / impact return the full combination transform (auto-init folded in) in the contract direction (auto_inverted: False); the impact affine read-back is TransformParameters.0.itk.txt (parsed by SimpleITK, with a manual 12-parameter fallback).
• simpleitk returns the two-stage composite (rigid-init affine o optimized affine) condensed to a single AffineTransform by point mapping.
• synthmorph's -T LTA (LINEAR_RAS_TO_RAS 4×4) is the contract transform directly, direction-checked by the usual gate.

Reproducing

source /opt/conda/etc/profile.d/conda.sh && conda activate longitudinal-313
cd /mnt/b4/xfr/git26/longitudinal

# all nine affine pipelines (results go to results/bench_registration_affine/;
# 2026-10-03 run, 6 workers, ~50 min wall dominated by impact's ~6 min phase)
python scripts/bench_registration.py --affine --pipelines "antspyx ants_trsaa ants_reproa ants_quickreproa elastix impact niftyreg simpleitk synthmorph" --workers 6 --threads 24 --qa-images

# single pipeline / case subset
python scripts/bench_registration.py --affine --pipelines antspyx simpleitk --n-cases 3

# combine the per-pipeline JSONs into summary.json / summary.md / plot
python scripts/bench_registration.py --affine --merge

Visual QA images

--qa-images writes a checkreg-style PNG per (pipeline, case) to results/bench_registration_affine/qa/<pipeline>__<case>.png (9 per pipeline; the incomplete-CT case is skipped) — same format as the other benchmarks: a 3×2 grid of the mid-axial / mid-coronal / mid-sagittal slices through the head center, left column checkerboard blend of the warped CT vs the T1c (clean registration = seamless mosaic), right column additive color overlay (CT red, T1c green; agreement = yellow, misregistration = red/green fringing). CT is brain-windowed 0–80 HU, T1c 2–98 % head window; the title carries the pipeline, case key and the achieved Mattes MI.