SCONet

This repository contains the code for SCONet (Segmentation Convolutional Occupancy Network) proposed in the paper:

@inproceedings{Jouvencel2025,
    author = {Jouvencel, Maylis and Kéchichian, Razmig and Digne, Julie and Valette, Sébastien},
    title = {SCONet: Convolutional Occupancy Networks for Multi-Organ Segmentation},
    booktitle = {IEEE ISBI},
    year = {2025}
}

Installation

The necessary Python packages to generate point-clouds, train and evaluate the models can be installed by running:

bash env_setup.sh

The code is tested with PyTorch 2.5.0 and CUDA 12.4.

To generate the point-clouds, you also have to compile the code from the 3D implementation of SURF.

Usage

Demo

We provide some demo point clouds if you want to test SCONet. This data includes the pre-sampled point cloud corresponding to the patients in data/test.lst . You can download them here.

The pre-trained weights are stored in the folder out/AbdomenCT-1K_demo

To generate the segmentation maps with SCONet, you can run:

python generate.py configs/pointcloud/AbdomenCT-1K_demo.yaml

Train from scratch

Data preparation

You can download the AbdomenCT-1K from the official dataset repository.

The data should be organized like this:

AbdomenCT-1K/
  volumes/
    id.lst
    Case_00001_0000.nii.gz
    Case_00002_0000.nii.gz
    ...
  segmentations/
    id_seg.lst
    Case_00001.nii.gz
    Case_00002.nii.gz
    ...
  train.lst
  val.lst
  test.lst

The preprocessing includes:

• the extraction of the information of the dataset in an info_dataset.csv
• the contour extraction with Canny algorithm
• the generation of the pointclouds

You can run it with the command:

bash preprocess.sh path/vtkOpenSURF3D/surf3d configs/pointcloud/AbdomenCT-1K.yaml

Training

To train a network from scratch, you can run the command:

python train.py configs/pointcloud/AbdomenCT-1K.yaml

Inference

To run an inference, you can run the command:

python generate.py configs/pointcloud/AbdomenCT-1K.yaml

Evaluate

To run the evaluation of the network, you can run the command:

python eval_seg.py configs/pointcloud/AbdomenCT-1K.yaml