PULSE: A Unified Multi-Task Architecture for Cardiac Segmentation, Diagnosis, and Few-Shot Cross-Modality Clinical Adaptation

Published in the IEEE Journal of Biomedical and Health Informatics (JBHI), 2026.

Hania Ghouse, Maryam Alsharqi, Farhad Nezami, Muzammil Behzad.

• Code: https://github.com/BRAIN-Lab-AI/PULSE
• Project page: https://brain-lab-ai.github.io/PULSE/

This repository hosts the pretrained weights. The model, training, and inference code live in the GitHub repository above.

Model summary

PULSE reads one short-axis cardiac MRI study and returns three outputs in a single pass: ventricular segmentation (RV, myocardium, LV), a patient-level cardiomyopathy diagnosis, and a structured clinical report. The segmentation network is a DINOv2 ViT-B/14 encoder with a Dense Prediction Transformer (DPT) decoder and deep supervision. Diagnosis uses a 23-dimensional clinical biomarker vector (computed from the predicted masks) fed to a Random Forest, and the report is produced by a deterministic, rule-based template. The released model is a 5-fold ensemble.

Files

folds_vdino/
├── fold_0/best_model.pth
├── fold_1/best_model.pth
├── fold_2/best_model.pth
├── fold_3/best_model.pth
└── fold_4/best_model.pth

Results

Benchmark	Metric	Score
ACDC segmentation (5-fold + TTA)	Mean Dice	88.8%
ACDC diagnosis	Accuracy / macro-AUC	90.0% / 0.982
M&Ms-2 (360 subjects, zero-shot)	Mean Dice	85.3%
Sunnybrook (zero-shot)	LV Dice	88.1%
CAMUS (few-shot, N=20)	Mean Dice	73.2%

Usage

from huggingface_hub import hf_hub_download
import torch
import pulse_seg as ps   # from the GitHub repo: https://github.com/BRAIN-Lab-AI/PULSE

model = ps.PULSESeg()
ckpt = hf_hub_download(repo_id="hg-0403/PULSE", filename="folds_vdino/fold_0/best_model.pth")
model.load_state_dict(torch.load(ckpt, map_location="cpu")["model"])
model.eval()

For single-volume inference, use pulse/infer.py from the GitHub repository:

python pulse/infer.py --input scan.nii.gz --folds_dir checkpoints/folds_vdino --output pred.nii.gz

Citation

@article{ghouse2026pulse,
  title   = {PULSE: A Unified Multi-Task Architecture for Cardiac Segmentation,
             Diagnosis, and Few-Shot Cross-Modality Clinical Adaptation},
  author  = {Ghouse, Hania and Alsharqi, Maryam and Nezami, Farhad and Behzad, Muzammil},
  journal = {IEEE Journal of Biomedical and Health Informatics (JBHI)},
  year    = {2026}
}

License

MIT