CATCH: Cross-View Attention for Consistent Multi-Chamber Heart Segmentation

Official implementation of "CATCH: Cross-View Attention for Consistent Multi-Chamber Heart Segmentation" (MICCAI ASMUS 2026).

CATCH is a segmentation framework for the left ventricle (LV), left atrium (LA), and right atrium (RA) in 2D echocardiography. It pairs a DINOv3 vision-transformer encoder (adapted for echo pretraining) with a cross-view modeling module that exchanges information across standard apical views, producing anatomically consistent chamber delineations rather than independent per-view predictions.

The model is trained on ~10,000 echocardiography studies from a private multi-chamber dataset and evaluated on ~4,000 held-out studies, plus zero-shot evaluation on the public CAMUS and EchoNet-Dynamic benchmarks.

---

Highlights

• DINOv3 encoder for echo. Self-supervised ViT backbone, modified for ultrasound-specific pretraining and dense feature extraction (see dinov3/).
• Cross-view modeling module. Attention across paired apical views enforces consistency of shared chambers, reducing view-dependent disagreement.
• Three-chamber segmentation. Joint LV / LA / RA prediction in a single forward pass.
• Scale. Trained on ~10k videos; validated on ~4k videos.
• Zero-shot transfer. Evaluated without fine-tuning on CAMUS and EchoNet-Dynamic.
• Streaming data pipeline. Unified loader over private and public corpora via MosaicML Streaming.

Repository structure

.
├── src/                     # Core implementation
│   ├── model.py             #   CATCH model, cross-view attention module, seg head
│   ├── datasets/            #   Private dataset + public datasets (CAMUS, EchoNet-Dynamic)
│   └── utils.py             #   Metrics, logging, checkpointing, visualization
│
├── scripts/                 # Entry points
│   ├── train_catch.py       #   Main training script
│   ├── train_ablations.sh   #   Ablation sweeps (encoder, cross-view module)
│   └── test_catch.sh        #   Evaluation on the private / zero-shot public datasets
│
├── dinov3/                  # DINOv3 (upstream code, modified for echo pretraining)
├── configs/                 # YAML configs for the echo-pretrained DINOv3
├── docs/
│   └── figures/             # Figures and result plots from the paper
└── requirements.txt

---

Installation

git clone https://github.com/DeepRCL/CATCH.
cd CATCH

conda create -n catch python=3.10 -y
conda activate catch

# Install PyTorch matching your CUDA version (see https://pytorch.org)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

pip install -r requirements.txt

Requirements: Python ≥ 3.10, PyTorch ≥ 2.1, CUDA ≥ 12.1. Training was performed on NVIDIA B200 and B300 GPUs.

DINOv3 backbone weights

The encoder is initialized from DINOv3 weights. Download the official checkpoint from the DINOv3 repository, then point the config at it:

# configs/model.yaml
model:
  config_file: ./configs/dinov3_vitb16_pretrain.yaml
  pretrained_weights: /path/to/dinov3_vitb16.pth

Data

Private dataset

The private multi-chamber echo dataset (~10k train / ~4k test videos, LV/LA/RA annotations) is not publicly available.

Public datasets (zero-shot evaluation)

Dataset	Chambers	Source
CAMUS	LV, LA	https://www.creatis.insa-lyon.fr/Challenge/camus/
EchoNet-Dynamic	LV	https://echonet.github.io/dynamic/

Download the required CAMUS files from their official source, then place them in data/camus_files:

Training

run_name="train_catch"

python src/train.py \
    --epochs 10 \
    --dataset-name private_echo \
    --n-classes 4 \
    --batch-size 8 \
    --log-dir "outputs/${run_name}" \
    --use-attention \
    --run-name "${run_name}"

Training logs, checkpoints, and validation overlays are written to --log-dir.

---

Evaluation

Private test set & zero-shot on public benchmarks

python src/eval.py \
    --dataset-name private_echo \
    --n-classes 4 \
    --log-dir "outputs/${run_name}" \
    --use-attention \
    --start-ckpt "outputs/train_catch/ckpt_ep_1.pth" \
    --calc-ef

--dataset-name choices: private_echo, camus, echonet.

Reports per-chamber Dice, IoU, Hausdorff distance (95th percentile), and mean absolute error for ejection fraction, plus the cross-view consistency metric described in the paper.

---

Ablations

Reproduce the ablation study from the paper:

python src/train.py \
    --epochs 10 \
    --dataset_name private_echo \
    --n_classes 4 \
    --batch_size 8 \
    --log-dir "outputs/${run_name}" \
    --use-attention \
    --run_name "${run_name}"

Covered axes:

Ablation	Description
Cross-view module	Concatenation / single-view self-attention / late fusion

To run a single variant, modify the command above:

• Concatenation — remove --use-attention
• Single-view self-attention — modify the dataset's required views
• Late fusion — add --use-late-fusion

---

Results

Full quantitative tables are in the paper; qualitative figures and plots are in docs/results/.

Private test set — per-chamber segmentation

Method	mDice LV ↑	mDice LA ↑	mDice RA ↑	mIoU LV ↑	mIoU LA ↑	mIoU RA ↑	HD95 LV ↓	HD95 LA ↓	HD95 RA ↓
U-Net	89.23	89.53	86.63	82.06	82.41	78.87	7.25	12.56	12.20
DeepLabv3	85.52	87.62	84.03	75.68	80.23	75.10	9.54	10.82	13.49
nnU-Net	91.38	90.42	89.50	84.80	84.10	82.58	5.60	8.83	8.18
AutoSAMUS	85.01	80.79	78.97	75.29	71.76	68.81	13.73	15.13	19.66
MedSAM	90.62	92.73	90.25	83.33	86.69	82.89	5.81	4.73	5.57
DINOv3	90.23	88.62	83.00	82.73	80.05	73.90	6.09	8.27	11.78
CATCH (ours)	91.49	91.04	89.20	84.69	84.19	81.57	5.11	5.79	6.80

---

Zero-shot / external validation

Method	CAMUS LV mDice ↑	CAMUS LV mIoU ↑	CAMUS LA mDice ↑	CAMUS LA mIoU ↑	EchoNet-Dynamic LV mDice ↑	EchoNet-Dynamic LV mIoU ↑
U-Net	89.28	80.94	65.22	57.57	88.30	79.80
DeepLabv3	91.36	84.33	79.02	69.63	90.41	82.99
nnU-Net	89.21	81.74	66.75	61.21	83.48	75.17
AutoSAMUS	86.41	76.97	79.05	69.70	84.59	75.00
MedSAM	89.79	81.74	75.33	66.30	90.91	83.63
DINOv3	91.98	85.31	86.85	77.52	89.41	81.37
CATCH (ours)	92.97	87.00	88.22	79.52	89.64	81.60

---

---

Acknowledgements

This work builds on excellent open-source projects:

• DINOv3 (Meta AI) — the encoder in dinov3/ is derived from the official implementation and modified for echocardiography pretraining. Original license and copyright are retained in that directory.
• rclstream — used for the unified streaming data pipeline over the private and public datasets.
• The CAMUS and EchoNet-Dynamic teams for releasing their datasets.

---

AI Assistance

Portions of this project were developed with AI assistance:

• Code refactoring
• README documentation

---

Citation

@inproceedings{catch2026,
  title     = {CATCH: Cross-View Attention for Consistent Multi-Chamber Heart Segmentation},
  author    = {Bassant Medhat, Nima Hashemi, Baraa Abdelsamad, Edward S. Chen, Samira Sojoudi, Christina L. Luong, Teresa S. M. Tsang, Purang Abolmaesumi},
  booktitle = {MICCAI Workshop on Advances in Simplifying Medical UltraSound (ASMUS)},
  year      = {2026}
}

---

Contact

Questions and issues: please open a GitHub issue or contact bassant@ece.ubc.ca.