Liver-lesion nnU-Net fine-tunes: an LR-schedule ablation
Two nnU-Net v2 checkpoints that differ only in the fine-tuning learning-rate schedule.
They exist to make one teaching claim reproducible: that a short linear warm-up to 1e-3
beats jumping straight to a constant 1e-3 when fine-tuning a pretrained segmentation
encoder. Companion to the mini-course at
https://github.com/Kappapapa123/internal-mini-course (notebook 07).
The two variants
Variant
Trainer
Schedule
Liver-lesion Dice
plain1e3
nnUNetTrainer_plain1e3_wandb
LR 1e-3 with default PolyLR (PANTHER-style default)
0.7889520
warmup1e3
nnUNetTrainer_warmup1e3_wandb
50-epoch linear warm-up (2e-5 → 1e-3 over 50 epochs), then offset PolyLR (TBI-style)
0.8026699
Difference: +0.0137 Dice (+1.37 points) for warm-up.
This is the liver-lesion class Dice from nnU-Net's own end-of-training validation
(fold_0/validation/summary.json), computed on the 169 fold-0 validation cases of the
842 — cases neither checkpoint trained on, but not an independent test set held out from
the whole 5-fold design. It is not the mean-foreground pseudo-Dice logged during training;
see Limitations.
Base model
MultiTalentV2 Challenge Edition, Constantin Ulrich (DKFZ) — a multi-dataset CT
segmentation ResEnc-L in the MultiTalent (arXiv:2303.14444) / AIMS-TBI
(arXiv:2504.06741) line of work from MIC-DKFZ.
License: CC BY 4.0 — attribution only, no further restrictions, so redistributing a
fine-tuned derivative is permitted provided the base model is credited
Related publication: 10.1007/978-3-031-43898-1_6
Specific weights used: the native-CT Z-score variant,
Dataset617_nativect/...nnUNetResEncUNetL1x1x1_Plans_znorm_bs24.../fold_all/checkpoint_final.pth.
checkpoint_final.pth rather than checkpoint_best.pth, following PANTHER's finding that
final works better for this foundation model.
The fine-tuning plans (..._znorm_bs24_mig_bs1) are the base model's own plans with only
batch_size changed 24 → 1 to fit a MIG slice. This is what makes the preprocessing match
— the point NB01 makes.
Transfer check: 946 / 946 transferable non-segmentation keys loaded.
The whole chain is CC BY 4.0: base checkpoint (Zenodo 13753413) and fine-tuning data
(Zenodo 20272572).
Fine-tuning dataset
Dataset591_liver_lesions — "Training dataset for TotalSegmentator task liver_lesions",
Jakob Wasserthal, University Hospital Basel. CC BY 4.0.
The license stated here is the one on the Zenodo deposit, which is the authoritative record.
(The dataset.json inside the archive carries a leftover "licence": "Apache 2.0" from the
TotalSegmentator template; both licenses permit redistribution with attribution, so nothing
about the mirroring depends on which one applies.)
A small number of fold-0 validation cases (image + ground-truth label) are mirrored in
this repo under cases/ in nnU-Net raw layout — cases neither checkpoint trained on. This
is because a Colab session cannot download 38.9 GB to obtain two cases. Please cite
Wasserthal if you use them.
Training configuration
Configuration: 3d_fullres, fold 0, 1000 epochs
Plans: nnUNetResEncUNetL1x1x1_Plans_znorm_bs24_mig_bs1 — 1 mm isotropic spacing,
Z-score normalization, patch size 192×192×192, batch size 1, batch_dice: True, matched
to the pretrained checkpoint's preprocessing
Peak LR 1e-3 for both variants (the peak is identical; only the ramp differs). Verified in
the training logs: plain starts at Current learning rate: 0.001 at epoch 0, warm-up
starts at 2e-05.
nnU-Net version: 2.6.4 (/home/kcuoft/nnunet_env_fir, Python 3.11). The two LR
variants are installed as
nnunetv2/training/nnUNetTrainer/variants/lr_schedule/{plain1e3,warmup1e3}_wandb.py, so a
reader loading these checkpoints needs the trainer files too, not just a matching pip
version.
Provenance: Phase-1 / S5 array job 42949211, tasks _0 (plain) and _1 (warm-up); both
COMPLETED, exit 0:0, per sacct. This array resumed a partially OOM-killed earlier
array (42820974). These runs predate the every-100-epoch snapshot mixin, so there are no
checkpoint_ep*.pth files — only best, final, and latest.
Files
Laid out exactly as $nnUNet_results expects, so nnUNetv2_predict finds them by path:
Dataset591_liver_lesions/
nnUNetTrainer_plain1e3_wandb__nnUNetResEncUNetL1x1x1_Plans_znorm_bs24_mig_bs1__3d_fullres/
plans.json
dataset.json
fold_0/checkpoint_final.pth
nnUNetTrainer_warmup1e3_wandb__.../fold_0/checkpoint_final.pth
...__3d_fullres/fold_0/validation_summary.json # nnU-Net's own scores, as provenance
cases/images/<CASE>_0000.nii.gz # 2 mirrored fold-0 validation cases
cases/labels/<CASE>.nii.gz
README.md # this card
The two mirrored cases are QTbgDDsctkRUgSTt (207×199×141, 10,195 lesion voxels) and
eZKVPnX7guzPW2Ph (342×316×101, 19,942 lesion voxels) — both with substantial lesions, so
Dice on them is meaningful rather than dominated by a handful of voxels.
These are the unmodified nnU-Net checkpoints, byte for byte as trained — the sha256 sums
below match the files on Fir, so you can verify what you downloaded. That means each file
also carries the optimizer state nnU-Net saves for resuming training, which is most of the
819 MB and is not needed for inference. Stripping it would cut the download by roughly two
thirds; it is not done here, in favour of hashes that match the originals.
The training facts in this card were read off the original run directories on the Alliance
Canada Fir cluster on 2026-07-30, and both checkpoints' sha256 sums were re-verified
against the values above at upload time.
Intended use
Teaching and methods research. Specifically: letting a reader of the mini-course verify
the plain-versus-warm-up comparison themselves instead of trusting a precomputed CSV.
Not for clinical use. These are single-fold research checkpoints trained on one public
dataset from one institution, with no prospective validation, no regulatory clearance, and
no evaluation of subgroup performance. Do not use them to inform patient care.
Limitations
The gain is task dependent. +1.37 Dice points on liver lesions here, about +0.4 on
TBI segmentation, and no benefit on the PANTHER pancreas-MRI task. Warm-up is not
universally better; run the ablation for your own task.
The training curves nearly overlap. For these same two runs, nnU-Net's
mean-foreground pseudo-Dice peaks at 0.899 (plain) versus 0.902 (warm-up) and ends at
0.882 versus 0.881 — i.e. the training curve does not predict the final gap. The
separation appears in per-class evaluation. Same runs, different metric.
Fold 0 only, single seed, and the reported Dice is that fold's validation split
rather than a set held out from the whole 5-fold design. No cross-fold variance estimate,
so the +0.0137 gap has no error bar.
The same ablation on a second dataset agreed but by a different margin: on
Dataset315_pleural_pericard_effusion, warm-up won by +0.0205 on the decision metric
(nanmean of pleural and pericardial effusion Dice). Two datasets, both favouring warm-up,
is still two datasets.
Domain: portal-venous-ish abdominal CT at 1 mm isotropic, one institution. Expect
degradation on other scanners, protocols, and contrast phases.
Untested on a 16 GB GPU. These were trained and validated on a 20 GB H100 MIG slice.
Inference needs less memory than training, so a 16 GB Colab T4 should be enough, but that
has not been measured. Notebook 07 carries a fallback ladder if you hit out-of-memory.
Citation
Please cite the dataset and the underlying methods, not this teaching artifact:
