This repository contains pretrained weights and model descriptions for all of the SonoNet variations described in our recent submission:

Baumgartner et al., "Real-Time Detection and Localisation of Fetal Standard Scan Planes in 2D Freehand Ultrasound", arXiv preprint:1612.05601 (2016),

and prior work published here:

Baumgartner et al., "Real-time standard scan plane detection and localisation in fetal ultrasound using fully convolutional neural networks", Proc. MICCAI (2016).

Please acknowledge the first of the two papers above if you end up using the these weights for your work.

The networks were trained using theano and the lasagne deep learning framework.

The weights are saved in the respective .npz files, the model definitions are given in models.py. A minimal example for classifying the images in the folder example_images is given in example.py.

Running example.py requires the latest theano and lasagne versions. Follow the instructions here, under the section "Bleeding Edge".

Furthermore, numpy, scipy are required.

Once everything is set up you can simply run:

python example.py 

Demo videos are available on Youtube:

To run example.py, you need the latest Theano, the latest Lasagne, old version SciPy(version 0.17 works ok) with Python2.

to_pytorch_model.py helps convert Lasagne model to PyTorch model, PyTorch model is defined in SonoNet.py, you need PyTorch >= 0.4.1 with Python3. Converted PyTorch models is provided in pytorch.model.tar.gz for convenient.

crop_image.py helps crop example image according to example.py, and this requires imageio module.

Note: example_images/test/other/other.jpg is not provided by paper author, and is not cropped yet. Please also note that though I try my best to convert Lasagne model to PyTorch model, I could not make sure that the model works as expected in Lasagne. And after I write a simple script to run the example, I didn't get the same performance.

Saddly, I have to give up. The model is trained with grayscale image, and data is normalize to [0, 255] not [0, 1]. I don't think that I could use these pretrained model.