--- license: cc-by-nc-sa-4.0 datasets: [Ahus-AIM/EchoXFlow] library_name: pytorch tags: [echocardiography, ultrasound, masked-autoencoder] --- # 2D echocardiography masked autoencoder Convolutional masked autoencoder for single 2D echocardiography frames in native beamspace (256 x 128, depth by lateral direction, values in [0, 1]). It reconstructs the full frame from a frame in which 16 x 16 pixel blocks are hidden, and its encoder gives multi-scale feature maps (`model.encode`). ![Input, masked input and reconstruction](reconstruction.png) ## Usage Requires `torch` and `matplotlib`. Download the repository and run the code below from inside it (it is also saved as `example.py`). It loads the model and the sample, hides part of the input, and saves the figure above. ```python import matplotlib.pyplot as plt import numpy as np import torch from model import EchoMAE2D model = EchoMAE2D() model.load_state_dict(torch.load("model.pt", weights_only=True)) model.eval() frame = torch.from_numpy(np.load("sample.npy")).float()[None, None] / 255 # 1 x 1 x 256 x 128 torch.manual_seed(0) mask = (torch.rand(1, 1, 16, 8) < 0.5).float() # hide half of the 16 x 16 pixel blocks mask = mask.repeat_interleave(16, 2).repeat_interleave(16, 3) with torch.no_grad(): reconstruction = model(frame * (1 - mask)) panels = {"Input": frame, "Masked": frame * (1 - mask), "Reconstruction": reconstruction} figure, axes = plt.subplots(1, 3, figsize=(6, 4.4)) for axis, (title, image) in zip(axes, panels.items()): axis.imshow(image[0, 0], cmap="gray", vmin=0, vmax=1) axis.set_title(title) axis.axis("off") figure.tight_layout() figure.savefig("reconstruction.png", dpi=150) ``` ## Model - Residual encoder with block counts [3, 4, 6, 3] and base width 320 after 8 x 8 input patchification, batch normalization, and a bilinear U-Net decoder with skip connections; 671 million parameters. - Pretrained for 1,000 epochs with full-image L1 reconstruction and 50% block masking, then fine-tuned for 30 epochs together with auxiliary tasks (view, sector geometry, probe rotation, examination identity, ECG phase). The weights are the exponential moving average at the best validation step. - The auxiliary prediction heads are not included. ## Files | File | Content | |---|---| | `model.pt` | Model weights (state dict, 2.7 GB) | | `model.py` | Model definition | | `sample.npy` | One beamspace frame, `uint8`, 256 x 128 | | `example.py` | The code above | ## Training data and citation Trained on [EchoXFlow](https://huggingface.co/datasets/Ahus-AIM/EchoXFlow), a clinical echocardiography dataset in native beamspace. `sample.npy` is one apical four-chamber frame from its validation fold. - Paper: [EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function](https://arxiv.org/abs/2605.05447) - Dataset: https://huggingface.co/datasets/Ahus-AIM/EchoXFlow - Tooling: https://github.com/Ahus-AIM/EchoXFlow ```bibtex @misc{stenhede2026echoxflow, title = {EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function}, author = {Stenhede, Elias and Sulkowska, Joanna and Orstad, Eivind Bj{\o}rkan and Schirmer, Henrik and Ranjbar, Arian}, year = {2026}, eprint = {2605.05447}, archivePrefix = {arXiv}, primaryClass = {cs.CV} } ``` ## License CC BY-NC-SA 4.0, the same license as EchoXFlow, from which the weights and the sample are derived. The model definition is adapted from the MAE in [lambertae/drifting](https://github.com/lambertae/drifting) (MIT license).