rhpg/liver-us-nnunet-baselines

Liver ultrasound segmentation baselines (nnU-Net)

Status: Pre-publication. Paper under review, results and weights are final but documentation may change.

Model weights from "Resource efficiency of nnU-Net for malignant liver mass segmentation on B-mode ultrasound" (Price-Gauger, 2026).

Checkpoints

File	Training images	Config	Mass Dice
checkpoint_best_625.pth	625	PlainConv, 150 epochs	0.652
checkpoint_best_500.pth	500	PlainConv, 150 epochs	0.639
checkpoint_best_400.pth	400	PlainConv, 150 epochs	0.587
checkpoint_best_300.pth	300	PlainConv, 150 epochs	0.559
checkpoint_best_200.pth	200	PlainConv, 150 epochs	0.546
checkpoint_best_100.pth	100	PlainConv, 150 epochs	0.506
checkpoint_best_50.pth	50	PlainConv, 150 epochs	0.358
checkpoint_best_25.pth	25	PlainConv, 150 epochs	0.292
checkpoint_best_200_300epochs.pth	200	PlainConv, 300 epochs	0.525
checkpoint_best_200_resenc.pth	200	ResEnc M, 300 epochs	0.556

Usage

These are nnU-Net v2 checkpoints. To run inference:

nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_ID -c 2d -f 0 -tr nnUNetTrainer150 -chk checkpoint_best.pth

Links

• Paper: [forthcoming]
• Code: github.com/roprice/liver-us-nnunet-baselines
• Dataset: AUL on Zenodo