Tte Clutter Filtering Eval

tte-clutter-filtering-eval

Deep Spatiotemporal Clutter Filtering of Transthoracic Echocardiographic Images: Leveraging Contextual Attention and Residual Learning — Tabassian et al. (2024) (arXiv:2401.13147, 2024)

What this evaluates

This protocol evaluates a 3D convolutional auto-encoder for removing reverberation artifacts (clutter) from transthoracic echocardiographic (TTE) sequences. It measures how well the network preserves cardiac structures while suppressing simulated artifacts, using synthetic data with known ground truth for training and validation, and normal in-vivo sequences for testing.

Datasets

• Synthetic TTE sequences — total 28836; splits: train (28836), val (-1), test (-1); repo https://github.com/MahdiTabassian/Deep-Clutter-Filtering

Metrics

• reconstruction loss ($L_{rec}$) (primary) — range: other
  • Measures the pixel-wise difference between the filtered output and the artifact-free ground truth sequence. The paper also evaluates joint losses combining $L_{rec}$ with adversarial ($L_{rec&adv}$) and patch-based ($L_{rec&prc}$) terms.

Input / output format

Input: 3D TTE sequences consisting of 50 temporal frames with spatial dimensions processed through a 3D U-Net encoder/decoder using $(2\times 2\times 1)$ max-pooling windows.

Output: Filtered 3D TTE sequences of identical dimensions (50 frames) with reverberation artifacts removed.

Scoring recipe

def compute_metric(predictions, ground_truth):
    # L_rec typically refers to MSE or MAE over all voxels/frames
    loss = np.mean((predictions - ground_truth) ** 2)
    return loss

Common pitfalls

• The evaluation relies heavily on synthetic data for training and validation, which may not fully capture real-world in-vivo artifact distributions.
• Temporal shift augmentation is used during training to ensure cycle-independence, but the exact alignment strategy for test sequences is not specified in this section.
• The exact mathematical formulation of $L_{rec}$ and its variants is referenced in Section 2.3.2, which is not provided here.

Evidence (verbatim from paper)

The training set was composed of 28836 TTE sequences, which is the multiplication of the number of the generated clutter patterns (534, see Section 2.1), views (3), vendors (6) and ischemic groups (3). The data of the forth ischemic group were used as a validation set to tune the parameters of the network and determine the best set of network’s weights. The sequences of the normal group were used as the test set. The benchmark networks were trained using the reconstruction loss, $L_{rec}$.

Citation

@misc{tabassian2024clutterfiltering,
  title={Deep Spatiotemporal Clutter Filtering of Transthoracic Echocardiographic Images: Leveraging Contextual Attention and Residual Learning},
  author={Tabassian et al. (2024)},
  year={2024},
  note={arXiv:2401.13147}
}

• arXiv: 2401.13147