A Semi-Supervised Fracture-Attention Model for Segmenting Tubular Objects with Improved Topological Connectivity

This is the official code of A Semi-Supervised Fracture-Attention Model for Segmenting Tubular Objects with Improved Topological Connectivity. (Bioinformatics 2025.01)

Our Contributions

- Semi-Supervised Fracture-Attention Model (SSFA)

Overview of SSFA.

- More Intuitive Topological Evaluation Metric: Fracture Rate (FR)

Visualization of fractures in segmentation results.

$$ FR=\frac{N_F}{N_Y}\times 100% $$

Quantitative Comparison

Qualitative Comparison

Qualitative results on ER, SNEMI3D, CREMI and STARE-DRIVE. (a) Raw images. (b) Ground truth. (c) UNet. (d) TopoLoss. (e) clDice. (f) MT. (g) CPS. (h) SSFA. Red arrows highlight differences among the results.

Requirements

albumentations==0.5.2  
matplotlib==3.2.1  
networkx==2.3  
numba==0.44.1  
numpy==1.18.5  
opencv_python==4.2.0.32  
Pillow==9.3.0  
ripser==0.6.4  
scikit_image==0.19.1  
scikit_learn==1.2.0  
scipy==1.4.1  
skimage==0.0  
torch==1.8.0  
torchvision==0.9.0  
tqdm==4.32.1  
visdom==0.1.8.9

Usage

Data preparation Your datasets directory tree should be look like this:

to see tools/attention_map/initial_attention.py for structure1, structure2, semantic, attention.

dataset
├── train_sup
    ├── image
        ├── 1.tif
        └── ...
    ├── structure1
        ├── 1.tif
        └── ...
    ├── structure2
        ├── 1.tif
        └── ...
    ├── semantic
        ├── 1.tif
        └── ...
    ├── attention
        ├── 1.tif
        └── ...
    └── mask
        ├── 1.tif
        └── ...
├── train_unsup
    ├── image
    ├── structure1
    ├── structure2
    ├── semantic
    └── attention
├── val
    ├── image
    ├── structure1
    ├── structure2
    ├── semantic
    ├── attention
    └── mask

Training

python -m torch.distributed.launch --nproc_per_node=4 train_semi_TA.py

Testing

python -m torch.distributed.launch --nproc_per_node=4 test_TA.py

Citation

If our work is useful for your research, please cite our paper: