Liver HCC

[Luugaaa]

Liver-HCC (atlas-liver-hcc) -- onboarding, cross-dataset eval, elevation to 5th headline paper task

What it is

HCC tumour segmentation on ATLAS-Liver-HCC (contrast-enhanced T1w, single modality -- no in-house second training contrast). 48 train / 12 test subjects, 3-fold (0/1/2), 2000 epochs (same budget as ON-Harmony/Open-MS). 6-method suite trained 2026-08-04, plus the DualVal val100 mirror (train050_val000 + train050_val100), 7 rows total.

Because it has no in-house cross-contrast axis, its generalization read is cross-dataset instead: same checkpoints evaluated on two independent external cohorts --

• lld-mmri-hcc -- HCC subset of LLD-MMRI, T2WI + DWI (genuinely cross-contrast), masks from MedSAM2
• liverhccseg -- same contrast family (CE-T1w), independent cohort/scanner, 4 phases (pre/arterial/portal-venous/delayed)

Headline numbers (task-level; t1w column pools the in-domain atlas_t1w test set together with liverhccseg's 4 CE-T1w phases -- see the note above on that trade-off)

Dice (up) / HD95mm (down), source: 8_results_atlas-liver-hcc/02_metrics/atlas_liver_hcc_model/combined_contrasts/01_results_summary.md (regenerated 2026-08-31)

method	t1w	lldmmrihcc_t2wi	lldmmrihcc_dwi	all	sig. vs ref
baseline	62.1	28.2	9.5	33.3	3.5e-12
synthseg_noEM	22.4	10.8	15.5	16.2	1.5e-44
synthseg_EM	53.9	33.5	42.4	43.3	0.029
auglab_default	60.1	35.9	26.6	40.9	1.6e-07
srcsm	57.6	25.0	18.3	33.6	1.3e-22
ours (train050_val000)	59.0	36.7	40.2	45.3	--
ours (train050_val100)	56.5	37.4	41.8	45.2	0.412

HD95 mm (down):

method	t1w	lldmmrihcc_t2wi	lldmmrihcc_dwi	all	sig. vs ref
baseline	45.3	108.5	111.3	88.4	1.000
synthseg_noEM	171.3	179.5	176.0	175.6	2.2e-47
synthseg_EM	62.0	85.2	84.3	77.2	1.000
auglab_default	68.6	94.1	111.5	91.4	0.366
srcsm	64.7	106.6	119.7	97.0	0.001
ours (train050_val000)	65.3	99.9	103.6	89.6	--
ours (train050_val100)	60.3	83.3	86.3	76.7	1.000