Joint Dense-Point Representation for Contour-Aware Graph Segmentation [MICCAI2023]

Kit Mills Bransby¹, Greg Slabaugh¹, Christos Bourantas^1,2, Qianni Zhang¹

¹ Queen Mary University of London, ² Department of Cardiology, Barts Health NHS Trust, London, United Kingdom

Accepted at MICCAI 2023 (top 14% of submissions). Paper PDF (Arxiv pre-print)

Installation

We use the same installation process as HybridGNet, and these dependencies can be installed in a fresh conda environment using:

conda env create -f environment.yml

To train and evaluate the Rasterize model, a differentiable rasterization pipeline is required, which can be installed by following install instructions for BoundaryFormer. We advise that this is created in a separate environment.

Datasets

Instructions for download and preprocessing datasets can be found in Datasets/README.md

Training

To train our joint dense-point network from scratch with a HCD loss on the JSRT & Padchest dataset, run the following command:

cd Train
python trainerLH_Joint_HCD.py

Trainers for all models and baselines are available in Train/, where LH (Lungs & Heart) = JSRT & Padchest dataset, and L (Lungs) = Montgomery & Shenzen dataset. Training weights will be saved to Results/ dir.

Paper Reproducibility

To reproduce the results in the paper, first download the model weights here, and place them in the weights/ directory. Run the evaluation scripts in Evaluate/, making sure that the directories described in Evaluate/README.md have been created.

Acknowledgements

Our codebase is adapted from HybridGNet. We thank Nicolas Gaggion for making this code open-source and publicly available. This research is part of AI-based Cardiac Image Computing (AICIC) funded by the faculty of Science and Engineering at Queen Mary University of London.