SwinUNETRv2 for 3D Medical Image Segmentation - The ASOCA Challenge

In this project, we present a novel 3D medical image segmentation model, SwinUNETRv2, which is based on the Swin Transformer and UNETR. We evaluate the model on the ASOCA Challenge dataset, which is a 3D medical image segmentation dataset for the segmentation of the aorta and pulmonary artery. With very little pretraining and fine-tuning, our model achieves competitive performance on the ASOCA Challenge dataset, with a Dice Score of 0.83 as seen on the official ASOCA leaderboard.

Requirements

• Pretraining and training the model requires a GPU with at least 22GB of VRAM (e.g. NVIDIA RTX4090).
• At least 32GB of RAM is recommended.

Installation

1. Clone the repository
2. Install the required packages from the requirements.txt file: pip install -r requirements.txt
3. Execute the notebooks in the numerical order they are in.

Data

The ASOCA Challenge dataset is available at this link. The dataset consists of 60 3D CT scans of the aorta and pulmonary artery, with corresponding segmentation masks. Out of the 60 scans, 36 are used for training, 4 for validation, and 20 for testing with the official ASOCA Challenge evaluation server and hidden labels.