Implicit Neural Distance Representation for Unsupervised and Supervised Classification of Complex Anatomies

Repository linked to publication: Implicit Neural Distance Representation for Unsupervised and Supervised Classification of Complex Anatomies (Accepted for MICCAI2021).

Authors: Kristine Aavild Juhl, Xabier Morales, Oscar Camara, Ole de Backer and Rasmus Reinhold Paulsen

Prepare training data

The preprocessing is a fully pythonic pipeline taking .vtk files as input and outputting arrays of point-coordinates and distances to be used for training. Meshes that are to be aligned using a set of pre-computed anatomical landmarks can be processed running:

python data_generator/make_train_data_LM.py

Meshes that should be aligned using iterative closest point without prealignment can be processed running:

python data_generator/make_train_data_LM.py

The following changes should be made in the respective files to preprocess your own dataset:

• l. 17: True/False indicates if the data should be prealigned
• l. 21: A list of fileids should be saved in a .txt file and the filepath should be specified.
• l. 22: A fileid should be given as the template for rigid alignemnt
• l. 43-47: Specify filepaths for the processed data
• l. 135+151 / 110+126: Specify the scale_factor so that the largest mesh fits the unit-sphere (number can be found by running l. 133 after aligning all shapes)
• l. 178+179 / 153+154: Set standard deviation on noise pertubation of sampling points

To visualize the mesh and the sampled points run 'check_trainingdata'.

The data is split into a training, validation and test set defined in the split files (see example of such files in examples/splits/XXX.json).

Train autodecoder model

This part of the code is largely based on deepSDF (https://github.com/facebookresearch/DeepSDF). Please see and cite this code for additional details on code organization etc. Parameters of training are stored in a "specification file" in the experiment directory. (See example in 'Experiments/Examples/specs.json'). The training is initiated as follows:

python train.py -e <experiment_directory> -c <optional: continue_from_epoch_number>

Find latent vectors from trained model and reconstruct unsigned distance field:

When a model is trained it can be used to generate latent codes for unknown shapes and to reconstruct the input mesh from code.

python reconstruct.py -e <experiment_directory> -c <checkpoint_number> -d <path_to_data> -s <split_file>

It is also possible to reconstruct a mesh with a known latent code (ie. interpolations in the latent space). A .pth file should be saved in '<experiment_directory>/reconstruction_from_latentcode/Codes/<latent_name>.pth'.

python reconstruct_from_latent_code.py -e <experiment_directory> -c <checkpoint_number> -d <path_to_data> -s <split_file>

Investigate latent space:

Visualize latent space:

The following script loads and visualize the latent vectors of the three datasets using PCA and tSNE as dimensionality reduction

python visualize_latentspace.py

Cluster latent vectors (unsupervised):

The following scripts loads the latent vectors and performs K-means clustering of the latent vectors.

python unsupervised_clustering.py

Train supervised classification model:

The following script trains and test a simple neural network learning the gender based on the self-optimized latent vectors

python supervised_classification.py

Reconstruct mesh from UDF

A mesh is extracted from the UDF by running:

python mesh_UDF.py

The method is dependent on the MRFSurface software which can be downloaded at http://www.imm.dtu.dk/~rapa/software.html