Spine benchmark results

[Luugaaa]

Results of the combination of our method (v19_c, outdated!) + Nathan and Hendrik's results:

The combination out performs synthseg, our method and the their method alone.
Should now compare using v26_6_2.

[Luugaaa]

v26_6_2 results on healthy-spine-tum (TUM external import, 2026-08-05) — supersedes the v19_c comparison above.

method	spider T1w	spider T2w	spinegan ct	spinegan fat	spinegan inph.	spinegan water	all
baseline	43.5	44.0	50.6	55.7	53.2	43.8	48.5
synthseg_noEM	41.6	46.8	57.2	54.7	51.7	50.6	50.4
synthseg_EM	80.1	79.4	83.7	78.7	81.1	80.8	80.6
auglab_default	79.9	80.0	84.5	80.6	83.1	77.1	80.9
v26_6_2 (ours)	80.7	80.7	84.8	81.9	84.0	81.4	82.3

Ours wins every column, beats everything at p ≤ 8.5e-16 (Holm-corrected). Margin over auglab_default is thin at the pooled level (+1.4) but consistent. Per-contrast: baseline massively overfits in-domain (CT→spinegan_ct 90.0) and collapses OOD (CT→spider_T2w 4.8); ours doesn't chase the in-domain peak but is by far the most consistent. Caveats: externally computed pipeline, no HD95 yet, no SRCSM arm yet — answers "compare using v26_6_2" for Dice, not yet a drop-in replacement for the internal 6-method suite. Paper text update is queued for later.