Cardiac Structure Segmentation in 2D Echocardiography

Cardiac structure segmentation (LV, myocardium, left atrium) from 2D echo, trained on the CAMUS dataset with a U-Net + ResNet34 encoder.

0.910 mean foreground Dice with 5-fold patient-level cross-validation — outperforming nnU-Net (~0.893) on CAMUS through systematic experimentation across 20+ configurations.

Approach

U-Net with pretrained ResNet34 encoder, SCSE decoder attention, and deep supervision (24.5M parameters). Trained with DiceCE loss (CE=0.3, Dice=0.7), AdamW with differential learning rates (encoder 3e-5, decoder 3e-4), and CosineAnnealingLR. Evaluation follows nnU-Net conventions (1000 epochs, 250 batches/epoch, patience 100). Test-time augmentation is applied at inference.

All results use patient-level 5-fold cross-validation — images from the same patient never appear in both training and validation sets.

Results

Method	Mean FG Dice
UwU-Net (263K params)	~0.890
nnU-Net	~0.893
Ours (5-fold CV + TTA)	0.910
TC-SegNet (ASPP+SE)	0.928

Class	Dice	IoU	HD (px)	MAD (px)
LV Cavity	0.944 ± 0.002	0.895 ± 0.003	6.46 ± 0.17	2.00 ± 0.04
Myocardium	0.881 ± 0.002	0.791 ± 0.002	8.33 ± 0.30	2.19 ± 0.05
Left Atrium	0.906 ± 0.006	0.837 ± 0.009	7.69 ± 0.38	2.45 ± 0.11

What Worked (and What Didn't)

Helped: Pretrained encoder (+0.006 Dice), DiceCE loss (best across all experiments), deep supervision (+0.011 on myocardium), SCSE attention, patient-level splitting, test-time augmentation.

Didn't help: Attention gates, SE blocks, Focal Tversky loss, boundary loss, SWA, class weights, EfficientNet-B4 encoder, 384×384 resolution, OneCycleLR, label smoothing, 5-fold ensemble. All converged to the same ~0.91 ceiling — the bottleneck is data, not model capacity.

Worst predictions occur on images with poor acoustic windows, extreme cardiac geometries, or ambiguous myocardial boundaries.

Dataset

CAMUS: 500 patients, 2CH/4CH apical views, ED/ES phases + half-sequences. 3000 image-mask pairs, 4 classes, NIfTI format resized to 256×256. Must be downloaded separately.

Installation

git clone https://github.com/Slyche/cardiac-wall-segmentation.git
cd cardiac-wall-segmentation
pip install -r requirements.txt

Requires Python 3.8+, PyTorch 2.0+, and a GPU.

Usage

Training (5-fold CV): notebooks/training_cv.ipynb — designed for Google Colab with A100 GPU.

Evaluation: notebooks/evaluation.ipynb — full metrics, visualizations, and comparison figures.

from src.model import DeepSupSMPUNet

model = DeepSupSMPUNet(
    encoder_name='resnet34',
    encoder_weights='imagenet',
    in_channels=1,
    num_classes=4,
)

References

• Leclerc et al. (2019). Deep Learning for Segmentation Using an Open Large-Scale Dataset in 2D Echocardiography. IEEE TMI.
• Ronneberger et al. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI.
• Isensee et al. (2021). nnU-Net: A Self-Configuring Method for Deep Learning-Based Biomedical Image Segmentation. Nature Methods.
• Kervadec et al. (2019). Boundary Loss for Highly Unbalanced Segmentation. Medical Image Analysis.

Acknowledgments

The CAMUS dataset was provided by the CREATIS laboratory at INSA Lyon.