OpenHeart segmentation model (HCM)

nnU-Net v2 weights that segment short-axis cine cardiac MRI into 1 = LV blood pool, 2 = LV myocardium, 3 = RV blood pool. This is the model used by the Segment step of OpenHeart.

Framework	nnU-Net v2, nnUNetTrainer, 2d configuration
Folds	all five folds of a five-fold cross-validation (hcm/fold_0 … hcm/fold_4), each trained on 1,039–1,040 volumes and validated on the other 259–260
Input	one short-axis cine volume (NIfTI), one channel
Training data	1,299 labelled short-axis cine volumes (end-diastole and end-systole) from patients with hypertrophic cardiomyopathy at 5 centres

Use with OpenHeart

from huggingface_hub import snapshot_download
import openheart as oh

model = snapshot_download("ImperialCollegeLondon/OpenHeart") + "/hcm"
case = oh.Segment(model)(oh.Case.load("cases/subject1", layout="legacy"))

MODEL=$(python -c 'from huggingface_hub import snapshot_download as d; print(d("ImperialCollegeLondon/OpenHeart"))')/hcm
openheart run cases/ -o out/ --model "$MODEL"

Without folds, all five folds are used and their predictions averaged. folds=[1] (--folds 1) runs a single fold, which is about five times faster. To download one fold only, pass allow_patterns=["hcm/*.json", "hcm/fold_1/*"] to snapshot_download.

hcm/ is a standard nnU-Net results folder, so it also loads directly with nnUNetPredictor.initialize_from_trained_model_folder(model, use_folds=(0, 1, 2, 3, 4)).

Limitations

Trained on hypertrophic cardiomyopathy cohorts; accuracy on other populations has not been evaluated here. For research use only, not for clinical decisions.

Credits

Computational Cardiac Imaging Group, MRC Laboratory of Medical Sciences and Imperial College London. Built with nnU-Net; please cite Isensee et al., nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation, Nature Methods 18, 203–211 (2021).