Delin An¹, Pan Du¹, Pengfei Gu², Jian-Xun Wang^{1, 3}, and Chaoli Wang¹

University of Notre Dame¹, The University of Texas Rio Grande Valley², Cornell University³

This repository contains the implementation of Hierarchical LoG Bayesian Neural Network (LoGB-Net) for enhanced 3D aorta segmentation. The framework integrates Bayesian principles with a hierarchical Laplacian of Gaussian (LoG) module to achieve high geometric fidelity and multiscale blood vessel recognition, particularly for small-radius vessels.

Key Features:

• A Bayesian LoG module for uncertainty quantification and robust feature extraction.
• A UNet-inspired 3D architecture with multiscale encoder-decoder pathways.
• ASPP refinement for capturing multiscale contextual information.

This method suits medical imaging tasks requiring accurate segmentation with geometric preservation, such as aortic dissection analysis and computational fluid dynamics (CFD) simulation preparation.

We evaluate our method on the following dataset, with comparisons against baseline approaches.

Dataset

• VMR Aorta data
• AVT data

Qualitative Evaluation:

• Achieved superior Dice scores compared to state-of-the-art methods.
• Demonstrated robustness in detecting small vessels.

The code is developed by Python. After cloning the repository, follow the steps below for installation:

1. Create and activate the conda environment

conda create --name logb python=3.11
conda activate logb

2. Install dependencies

pip install -r requirements.txt

3. Train

python train.py

4. Test

python test.py

• Python (3.11), other versions should also work
• PyTorch (2.3.0), other versions should also work

Should you have any questions, please send emails to [email protected].