PULSE Cardiac MRI

Abstract

Cardiac image analysis requires accurate ventricular segmentation, disease classification, and structured clinical reporting; these tasks are typically handled by separate models, limiting clinical deployment. To address this, we introduce PULSE, a unified three-task framework that performs: (1) ventricular segmentation using a DINOv2 Vision Transformer (ViT-B/14) backbone with a Dense Prediction Transformer (DPT) decoder and deep supervision; (2) cardiomyopathy diagnosis via a 23-dimensional clinical biomarker vector fed to a Random Forest classifier; and (3) a structured, template-based clinical reporting module that populates a predefined report with the measured indices and rule-based abnormality flags. The framework is entirely vision-based: all reported text is produced by deterministic, rule-driven templates. Additionally, the model takes 2.5D inputs (three adjacent short-axis slices) and is evaluated via a 5-fold stratified ensemble. With extensive experiments on the ACDC benchmark, PULSE achieves a mean Dice of 88.8% (RV: 90.3%, Myo: 84.7%, LV: 91.6%, HD95 ≤ 4.6 mm), 90.0% patient-level diagnostic accuracy (macro-AUC 0.982), and 92.7% clinical flag agreement in the generated reports (LVEF MAE 3.09%, within inter-observer tolerance). Without retraining, PULSE achieves 85.3% mean Dice on M&Ms-2 (360 subjects, RV: 87.9%, Myo: 80.4%, LV: 87.5%) and 88.1% LV Dice on Sunnybrook MRI. We further demonstrate that few-shot fine-tuning on CAMUS echocardiography samples yields a mean Dice of 73.2%, showing strong cross-modality transfer from cardiac MRI priors.

Pipeline

Segment, diagnose, report

1. Segment

DINOv2 + DPT delineate RV, myocardium, and LV from three adjacent short-axis slices.

2. Diagnose

A 23-dimensional biomarker vector feeds a Random Forest for patient-level cardiomyopathy diagnosis.

3. Report

Measured indices and rule-based flags populate a structured clinical report, with no language model.

Highlights

What makes PULSE different

3 in 1

Unified multi-task

Segmentation, diagnosis, and reporting from a single model, with no task-specific retraining.

ViT-B/14

Foundation backbone

Self-supervised DINOv2 features, fine-tuned with a two-stage curriculum and a DPT decoder.

23-D

Interpretable diagnosis

Every classifier input is a named clinical measurement: EF, volumes, mass, wall thickness.

5-shot

Cross-modality transfer

Adapts from MRI to echocardiography with only a handful of labels.

Results

Accurate, generalizable, and transferable

ACDC segmentation, 5-fold ensemble with test-time augmentation (50 test patients)
Structure	Dice (%)	IoU (%)	HD95 (mm)	ASSD (mm)
Left ventricle (LV)	91.6	84.5	3.87	0.95
Myocardium (Myo)	84.7	73.5	3.72	0.87
Right ventricle (RV)	90.3	82.3	4.57	1.05
Mean	88.8	80.1	4.05	0.96

Comparison with published methods on ACDC (Dice %)
Method	RV	Myo	LV	Mean	Diagnosis
U-Net	82.3	78.7	94.9	85.3	No
nnU-Net	90.1	88.4	95.7	91.4	No
TransUNet	84.5	77.7	94.1	85.4	No
SwinUNet	90.7	79.1	93.9	87.9	No
PULSE (ours)	90.3	84.7	91.6	88.8	90.0%

Zero-shot generalization (no retraining)
Dataset	Subjects	Mean Dice
M&Ms-2 (4 vendors)	360	85.3
Sunnybrook (LV)	45	88.1

Few-shot CAMUS echocardiography (2-class)
Labelled N	Myo	LV	Mean
5	65.2	72.1	68.7
10	67.8	73.3	70.6
20	70.8	75.6	73.2

Per-disease Dice across the five ACDC classes.

Confusion matrix — Diagnosis confusion matrix on the ACDC test set.

Diagnosis performance — Per-class accuracy, precision, recall, and F1 for cardiomyopathy classification.

DINOv2 attention maps — DINOv2 self-attention concentrates on the cardiac region across all disease classes.

Qualitative

Segmentation across vendors and modalities

ACDC segmentation across cardiomyopathy classes — End-diastole (ED) and end-systole (ES) segmentation across the five ACDC cardiomyopathy classes (DCM, HCM, MINF, NOR, RVA): RV (red), Myocardium (green), LV (blue).

M&Ms-2 generalization — Zero-shot generalization to multi-vendor M&Ms-2 cine MRI: RV (red), Myocardium (green), LV (blue).

Sunnybrook segmentation — Zero-shot transfer to the Sunnybrook Cardiac Data with well preserved LV structure.

CAMUS few-shot — Few-shot adaptation to CAMUS echocardiography from a handful of labelled scans.

Citation

Cite PULSE

@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}
}

PULSE

One examination in, three clinical outputs out