Learning a reconnecting model for curvilinear structures

We proposed to train a reconnecting model for curvilinear structures in 2D and in 3D that can be used as:

• a regularization term in an unsupervised plug-and-play segmentation approach
• a post-processing step for any type of vascular segmentation result

This repo contains the code to apply methods presented in two different papers :

[1] Sophie Carneiro-Esteves, Antoine Vacavant and Odyssée Merveille. "A plug-and-play framework for curvilinear structure segmentation based on a learned reconnecting regularization" Neurocomputing 2024

[2] Sophie Carneiro-Esteves, Antoine Vacavant and Odyssée Merveille. "Restoring Connectivity in Vascular Segmentations using a Learned Post-Processing Model" TGI3, MICCAI 2024 Workshop

This repository does not provide a single standalone script. Instead, it offers a set of reusable functions designed to be integrated into your existing pipeline. You can incorporate these functions at the appropriate stages of your workflow to customize and enhance your processing pipeline according to your specific needs.

Code structure

The code is stored in sources/.

• source_2D contains the python functions to process 2D images.
• source_3D contains the python functions to process 3D images.
• image_utils.py contains useful functions to open, normalize and save images (2D and 3D).

The files in source_2D and source_3D are:

• disconnect.py : contains the functions to disconnect a binary tree.
• train.py: contains the functions to train a reconnecting model with a generated dataset.
• post_treatement.py: contains the function to apply the reconnecting model as a post processing on a curvilinear segmentation.
• pretreatement.py: contains a function that subtracts the median filter from the image to remove the intensity variations.
• plug_and_play.py : contains the functions to apply an unsupervised plug-and-play segmentation approach using the reconnecting model.
• grad_div_interpolation.py: contains the functions to compute a discrete gradient.
• example.py :is a script demonstrating how to use the various functions that perform the different steps of the frameworks proposed in [1] and [2].
• compared_methods.py: includes various methods used to compare our approach with those presented in [1] and [2].

How to use it ?

To understand the order of operations, check out the example.py file. It demonstrates how to:

• Create a dataset.
• Train the neural network.
• Apply the trained model as a post-processing step.
• Use the trained model as a regularisation term for a plug-and-play segmentation.

If you don't have annotations or binary vascular trees to train a model, models used in [1] and [2] are available in the directory modeles.

Models

4 models are provided :

• 2D_model_CCO: a model trained on 2D synthetics binary trees generated with OpenCCO
• 2D_model_stare: a model trained on 2D manual annotations from the STARE Dataset (retinophotographies)
• 3D_model_vascu: a model trained on 3D synthetics binary trees generated with Vascusynth
• 3D_model_IXI: a model trained on 3D manual annotations made on the IXI dataset (Brain MRA)

Environment Conda

An python environment is given to run the code and can be installed with the following command :

conda env create -f environment.yml