#### NOTE
New to DeepInverse? Get started with the basics with the
[5 minute quickstart tutorial](https://deepinv.org/auto_examples/basics/demo_quickstart.html.md#sphx-glr-auto-examples-basics-demo-quickstart-py)..
# Ultrasound despeckling from B-mode images
This example despeckles real clinical ultrasound B-mode images with Speckle2Self
[1](#footcite-li2025speckle2self), a model pretrained without clean reference images.
We compare its performance against a classical BM3D [`deepinv.models.BM3D`](https://deepinv.org/api/stubs/deepinv.models.BM3D.html.md#deepinv.models.BM3D) baseline.
We use B-mode images from the carotid dataset acquired in Speckle2Self.
Ultrasound B-mode images are corrupted by speckle, a granular noise pattern that
arises from interference between backscattered echoes. Speckle lowers contrast and
hides fine tissue structure, which makes clinical interpretation harder.
Speckle arises from the beamforming inverse problem, but often in practice we only have access to
envelope-detected B-mode grayscale in the log domain. Therefore, here, we treat despeckling as a
standard image denoising problem.
This example requires the `ptwt` (for BM3D) and `zea` (for Speckle2Self) packages, which can be installed with
`pip install ptwt zea`.
```Python
import deepinv as dinv
import torch
device = dinv.utils.get_device()
```
```none
Selected GPU 0 with 9822.4375 MiB free memory
```
## Load the dataset
We download the Speckle2Self test set of 104 carotid B-mode images at 512x512 resolution,
acquired with a Clarius L7 portable scanner from two healthy volunteers, taken from the Clarius envelope collection API by Li *et al.*[1](#footcite-li2025speckle2self).
#### NOTE
The data is envelope-detected B-mode grayscale, already in the log domain, where the multiplicative speckle of the linear domain can now be approximated by additive noise.
#### NOTE
For this demo, we show despeckling on two slices on CPU or four on GPU.
```Python
url = "https://drive.usercontent.google.com/download?id=1rQxgyCzDuLao05tE8y5l8vw8YFrPjbzx&export=download"
y = dinv.io.load_np(dinv.io.load_url(url)).unsqueeze(1).to(device) # (B, 1, H, W)
y = y[: 2 if dinv.utils.devices_equal(device, "cpu") else 4]
y = y / y.max() # to allow a meaningful sigma
```
## Despeckle with BM3D
As a classical baseline, we use a custom efficient implementation of BM3D
[2](#footcite-dabov2007image), faster particularly on GPU. The `sigma` argument sets the assumed noise level.
#### NOTE
Fully-developed speckle is Gamma-distributed, so in log domain, the Gaussian is only an approximation for BM3D.
See the [denoisers user guide](https://deepinv.org/user_guide/reconstruction/denoisers.html.md#denoisers) for further classical and deep denoisers.
For example, you can train your own with [`deepinv.physics.FisherTippettNoise`](https://deepinv.org/api/stubs/deepinv.physics.FisherTippettNoise.html.md#deepinv.physics.FisherTippettNoise), which models the noise more accurately.
```Python
model = dinv.models.BM3D(use_legacy=False, device=device)
with torch.no_grad():
x_hat = model(y, sigma=0.3)
```
## Despeckle with Speckle2Self
We run the pretrained in-vivo Speckle2Self model, which is available in the `zea` toolbox [3](#footcite-stevens2026zea).
```Python
import os
os.environ["KERAS_BACKEND"] = "torch"
from zea.models.speckle2self import Speckle2Self
model = Speckle2Self.from_preset("hf://zeahub/speckle2self-invivo").to(device)
with torch.no_grad():
x_net = model(y.moveaxis(1, -1)).moveaxis(-1, 1) # zea expects channel last
```
```none
zea: Using backend 'torch'
```
## Visualise the results
We compare the input B-mode image with the BM3D and Speckle2Self reconstructions.
```Python
dinv.utils.plot(
{"Vendor bmode": y, "BM3D": x_hat, "Speckle2Self": x_net},
figsize=(7, 10),
)
```
* **References:**
* Xuesong Li, Nassir Navab, and Zhongliang Jiang. Speckle2self: self-supervised ultrasound speckle reduction without clean data. *Medical Image Analysis*, 106:103755, 2025.
* Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. *IEEE Transactions on image processing*, 16(8):2080–2095, 2007.
* Tristan S. W. Stevens, Wessel L. van Nierop, Ben Luijten, Vincent van de Schaft, Oisín Nolan, Beatrice Federici, Louis D. van Harten, Simon W. Penninga, Noortje I. P. Schueler, and Ruud J. G. van Sloun. Zea: a toolbox for cognitive ultrasound imaging. *Journal of Open Source Software*, 11(121):9881, 2026.
**Total running time of the script:** (0 minutes 9.606 seconds)
🏷 Tags: Denoising, Ultrasound
[`Download Jupyter notebook: demo_ultrasound_despeckling.ipynb`](https://deepinv.org/auto_examples/self-supervised-learning/_downloads/3ac45b8a7d0e656c52641779c7ac437c/demo_ultrasound_despeckling.ipynb)
[`Download Python source code: demo_ultrasound_despeckling.py`](https://deepinv.org/auto_examples/self-supervised-learning/_downloads/19dab0543560f4009cf3f0e8fff47d3a/demo_ultrasound_despeckling.py)
[`Download zipped: demo_ultrasound_despeckling.zip`](https://deepinv.org/auto_examples/self-supervised-learning/_downloads/b7d78bada90be43bfa1d48e93003a72d/demo_ultrasound_despeckling.zip)
[Gallery generated by Sphinx-Gallery](https://sphinx-gallery.github.io)