BAUN3D: Boundary-Attentive 3D-UNet for Automatic Segmentation of Tumor-Prone Organs in Volumetric CT

Overview

BAUN3D is a unique anatomy-aware deep learning radiomics for the localization and auto-segmentation of organs and tumors in volumetric CT images. Built specifically for contouring the challenging tumor-prone organs, the architecture comprises of: deformable cross attention mechanism, gated boundary refinement (GBR) module, and a composite loss objective function for handling curriculum learning, extreme class imbalance, small tumor targets, and contour structural continuity.

System requirements

• Python ≥ 3.8
• CUDA ≥ 11.8 (for GPU acceleration)
• 10GB+ GPU memory per GPU

Dataset

• This version of BAUN3D model was trained and validated with the medical segmentation decathlon (MSD) LiTS and Pancreas benchmark datasets.

• Download link to the model weights will be updated later.

Data Directory Structure

data/
├── lits/
│   ├── imagesTr/          # Training images (*.nii.gz)
│   ├── labelsTr/          # Training labels (*.nii.gz)
│   └── imagesTs/          # Test images
├── pancreas/
│   ├── imagesTr/
│   ├── labelsTr/
│   └── imagesTs/
└── ...

Train | Test | Inference

The training and inference source-codes, and running commands will be availed soon.

Outcomes

Quantitative results

Dataset	Organ Dice	Tumor Dice	Avg Dice	HD95 (mm)
LiTS	0.95	0.71	0.83	10.78
Pancreas	0.91	0.78	0.84	7.55

Qualitative results

Boundary segmentation sample (Liver/Tumor)

Boundary segmentation sample (Pancreas/Tumor)

Acknowledgements

Development of this software was sponsored by CAIM: Linkou, Chang Gung Memorial Hospital Research Project, under grant no. CLRPG3H0017