Research map / Detection and segmentation / CNN features

EchoONE: Segmenting Multiple Echocardiography Planes in One Model

CVPR 2025 · Detection and segmentation › CNN features · approach: Domain-specific segmentation and detection · 5 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.

Proposes EchoONE, a SAM-based model with a prior-composable mask learning dense-prompt module and a CNN local-feature fusion branch that segments LV, LA and myocardium across multiple echocardiography planes in a single model.

Problem

Multi-plane segmentation of echocardiography: existing models must be tailored per plane, and training one model on combined planes causes a clear performance drop; SAM fails zero-shot on echo images.

Approach

Generates semantic-aware dense mask prompts without plane labels by weighting cluster-center prior masks (from ResNet34 latent clustering) by cosine similarity and refining them with a light U-Net, and adapts SAM via a CNN branch whose multi-layer features are fused into the mask decoder by LFFA modules.

Key results

• Internal plane-wise evaluation: EchoONE reaches 89.67 mDice on 2CH, 87.27 on 3CH, 90.28 on 4CH and 88.26 on PSAX, outperforming all listed baselines including SAMUS (87.51/85.54/88.37/85.92).
• External evaluation: EchoONE achieves 73.94 mDice on HMC QU (vs. 71.49 for SAMUS) and 88.36 on EchoNet-Dynamic (vs. 87.56 for SAMUS).
• Ablation: LFFA alone improves PSAX mIoU from 74.86 to 77.34, and PC-Mask alone to 76.34; both together reach 79.68.

Measured effects reported in the paper

Surprising findings

• Vanilla SAM applied directly to echocardiography achieves only about 25-35 mDice across planes, and the authors note that SAM-derived methods were weaker than plain U-Net/H2Former on most planes.

Limitations

• Validated only on echocardiography ultrasound images, so generalization to other modalities is untested.
• Private datasets are not released, limiting reproducibility.
• Only three cardiac structures (LV, LA, MYO) across four plane groups (PSAX treated as one) are covered; the three PSAX sub-planes are merged.

Experimental setup

Datasets: CAMUS HMC QU EchoNet-Dynamic private datasets A-D

Benchmarks: CAMUS HMC QU EchoNet-Dynamic private datasets A-D

Metrics: mDice mIoU HD95

Baselines: U-Net DeepLabV3+ SwinUNet H2Former SAM MedSAM SAMed SAM-Med2D MSA SAMUS

Builds on: Segment Anything Model SAMUS SAM-LST Ladder Side-Tuning

Closest papers by meaning

• MemSAM: Taming Segment Anything Model for Echocardiography Video Segmentation (CVPR 2024 · similarity 0.85)
• Self-supervised Learning of Echocardiographic Video Representations via Online Cluster Distillation (NeurIPS 2025 · similarity 0.83)
• Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.83)
• EchoWorld: Learning Motion-Aware World Models for Echocardiography Probe Guidance (CVPR 2025 · similarity 0.82)
• MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling (CVPR 2024 · similarity 0.82)
• Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation (NeurIPS 2025 · similarity 0.82)