This software implements a Convolutional Neural Network (CNN) for automatic simultaneous localisation of multiple landmarks in 3D medical images (eg. fetal head ultrasound).
Tensorflow implementation of the MICCAI 2018 paper Fast Multiple Landmark Localisation Using a Patch-based Iterative Network.
- Python 2.7
- Tensorflow 1.8.0
- SciPy
- (Optional) NiBabel (for reading NIfTI input)
- (Optional) Matplotlib (for visualization)
- (Optional) MATLAB (for building shape model)
To train a CNN model:
$ python train.py
To test with an existing CNN model (with full evaluation):
$ python infer.py
To test with an existing CNN model (without evaluation)
$ python infer_simple.py
To build the statistical shape model, run ./shape_model/CreateShapeModel.m in MATLAB.
Pretrained CNN model available at ./cnn_model.
Pretrained shape model available at ./shape_model/shape_model.
We are not able to share the dataset we used for the paper due to sensitive patient information. However, we have provided a dummy data which is a random noise matrix to mimic the actual data we used.
Predicted landmarks (green) and ground truth landmarks (red) overlaid on the two standard scan planes of fetal brain.
Path taken by the landmarks for 10 iterations during inference. Each colour represent a different landmark. We initialise 6 instances for each landmark during inference and take the mean as the final prediction.
Yuanwei Li / [email protected]