Deep Learning: Retinal Vessel Segmentation

Dependencies

We use PyTorch, TorchVision, OpenCV, PIL, and the albumentations library for extensive data augmentation.

conda install -c pytorch torchvision captum
conda install -c conda-forge imgaug
conda install -c albumentations albumentations

We use PyTorch's TensorBoard integration to see metrics and some segmentation results during training

tensorboard --logdir runs/

Training

General configuration parameters (training, validation slices, image input size) are located in config.py.

To launch the training script:

python train.py --model attunet --loss combined --lr 0.001 -E 80

Inference

python inference.py --model attunet --weights path/to/weights

Data

DRIVE data is located in

data/drive
    - training
    - test

Its mean is [0.5078, 0.2682, 0.1613], stdev is [0.3378, 0.1753, 0.0978]

We also use the STARE dataset with manual vessel annotations: http://cecas.clemson.edu/~ahoover/stare/probing/index.html.

wget http://cecas.clemson.edu/~ahoover/stare/probing/stare-images.tar
wget http://cecas.clemson.edu/~ahoover/stare/probing/labels-ah.tar

The layout of the data is

data/stare/
    - images  # images
    - labels/ # labels
        - labels_ah
        - labels_vk
        - results_hoover

CHASE dataset:

data/chase

The 1st manual annotation set is *_1stHO.png, the second manual annotations are *_2ndHO.png.

ARIA data is located in

data/aria
    - images
    - annotation 1
    - annotation 2
    - (markupdiscfovea) 

The database contains healthy subjects (61 images, "c" suffix), diabetic patients (59 images, "d" suffix), and patients with age-related macular degeneration (23 images, "a" suffix). We created a CSV file recording the

The data augmentation can be visualized in a notebook.

Visualization

Attention maps

The Attention U-Net model has attention maps that can be directly interpreted using the following script:

python visualize_attention.py --img-path path/to/img --weights path/to/model

check python visualize_attention.py -h for options.

Activation maps

See visualize_activations.py