Research map / Vision-language and multimodal / Cross-modal learning

EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function

NeurIPS 2026 · Vision-language and multimodal › Cross-modal learning · approach: Unimodal-to-multimodal representation transfer · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces EchoXFlow, a 666-exam beamspace echocardiography dataset with separable B-mode, Doppler, ECG and dense annotations, plus tooling and three cross-modal/segmentation benchmarks using 2D and 3D U-Nets.

Problem

Public echocardiography datasets are scan-converted B-mode videos with Doppler absent or fused as RGB overlays, so cross-modal learning from native acquisition signals (beamspace, separate Doppler streams) is not possible.

Approach

Release routine-care echo recordings as separable modality-specific streams in probe-centered beamspace geometry, paired with ECG and dense/sparse annotations, and define benchmark tasks (Doppler estimation from B-mode, beamspace vs. Cartesian segmentation) that existing datasets cannot support.

Key results

• 3D U-Net on beamspace input reaches the lowest Task 1 tissue-velocity loss (5.07 cm/s vs 7.76 for temporal-mean baseline).
• 3D U-Net achieves 81.2% Dice on LV segmentation with scan-converted input, versus 45.8% for the temporal-mean baseline.
• For blood velocity (Task 2), 3D U-Net reduces the loss from 29.29 to 24.70 cm/s, a modest ~16% improvement over baseline.

Surprising findings

• Beamspace input won slightly on Doppler tasks while scan-converted input won on segmentation, suggesting input domain should match supervision domain.
• Temporal context was the main driver of cross-modal signal: 2D U-Net gains were much smaller than 3D U-Net gains on Tasks 1 and 2.

Limitations

• Single vendor (GE Vivid E95) at a single hospital; transfer untested.
• Annotations are routine clinical reports with inter-operator variability and non-random missingness.
• Doppler measures only beam-aligned velocity and the Doppler reference contains noise, so the irreducible error is unknown; no CMR validation; cross-sectional only, no diagnostic labels.

Experimental setup

Datasets: EchoXFlow

Benchmarks: EchoXFlow Task 1 tissue velocity EchoXFlow Task 2 blood velocity/power/turbulence EchoXFlow Task 3 LV segmentation

Metrics: Nyquist-wrapped L1 masked L1 Dice %

Baselines: Temporal mean 2D U-Net

Builds on: U-Net Muon

Closest papers by meaning

• Self-supervised Learning of Echocardiographic Video Representations via Online Cluster Distillation (NeurIPS 2025 · similarity 0.86)
• ECHOPulse: ECG Controlled Echocardio-gram Video Generation (ICLR 2025 · similarity 0.86)
• EchoWorld: Learning Motion-Aware World Models for Echocardiography Probe Guidance (CVPR 2025 · similarity 0.86)
• Bidirectional Recurrence for Cardiac Motion Tracking with Gaussian Process Latent Coding (NeurIPS 2024 · similarity 0.84)
• CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations (NeurIPS 2026 · similarity 0.84)
• Q-PART: Quasi-Periodic Adaptive Regression with Test-time Training for Pediatric Left Ventricular Ejection Fraction Regression (CVPR 2025 · similarity 0.83)