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

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 $\leq$ 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.