Simulation-Based Segmentation of Vessels in Cerebral 3D OCTA Images

TL;DR: First, we selectively extract patches from vessel graphs that match the FOV and morphological characteristics of vasculature contained in cerebral OCTA images and transform them into voxelized volumes; second, we transform the voxelized volumes into synthetic cerebral 3D OCTA images by simulating the most dominant image acquisition artifacts; and third, we use our synthetic cerebral 3D OCTA images paired with their matching ground truth labels to train a blood vessel segmentation network, erasing the need for manual annotations.

Our work is described in detail in our manuscript (Simulation-Based Segmentation of Blood Vessels in Cerebral 3D OCTA Images).

Updates:

June 24: Added patch_sampling.py to provide additional information on patch sampling.

🟢 Installation

Create a virtual environment using, for example, anaconda

conda create -n syn-cerebral-octa-seg python=3.8

and install relevant packages:

pip install torch==1.13.1 --extra-index-url https://download.pytorch.org/whl/cu116

pip install monai==1.0.1

pip install scipy h5py SimpleITK scikit-image pyyaml tensorboard tqdm scikit-learn six

pip install -e .

🟢 Data

To download the synthetic cerebral 3D OCTA images & our manually annotated real 3D OCTA images, please follow the instructions below:

git lfs install                     # check whether git lfs is installed
git clone [email protected]:datasets/bwittmann/syn-cerebral-octa-seg

Folders containing synthetic cerebral 3D OCTA images and manually annotated real 3D OCTA images should be placed in ./dataset:

dataset/
└── manual_annotations/
    └── mx_0.nii                    # real cerebral 3D OCTA image
    └── mx_0_label.nii              # ground truth (manual annotations) 
        ...
└── synthetic_cerebral_octa/
    └── axxxx_0/
        └── sim/
            └── sim_data_xx.npy     # synthetic cerebral 3D OCTA image
            └── sim_seg_xx.npy      # ground truth
        └── ang.npy                 # metadata angle
        └── occ.npy                 # metadata occupancy below
        └── rad.npy                 # metadata radius
        └── seg.npy                 # voxelized volume
    └── axxxx_1/
        ...   

The data is described in more detail on Hugging Face 🤗.

🟢 Create your own customized synthetic cerebral 3D OCTA images

One can either make use of our provided synthetic cerebral 3D OCTA images (b421) or finetune the simulation parameters to match the characteristics of OCTA images at hand. To adjust the simulation, please run:

python syn_cerebral_octa_seg/preprocess_sim.py --name <name> --angle_correction --len_tail 1.0 --gran_noise_std 1.4

Here, we solely adjust two parameters, len_tail (tail length factor) and gran_noise_std (granular noise standard deviation), following our example in the manuscript (see Fig. 7).

After the customized dataset has been created, adjust sim_data_version in ./syn_cerebral_octa_seg/config.yaml to <name>.

🟢 Erase the need for manual annotations by training on synthetic images

To leverage the created synthetic cerebral 3D OCTA images for the sake of training a segmentation algorithm, simply run:

python syn_cerebral_octa_seg/train.py

Parameters can be adjusted by modifying ./syn_cerebral_octa_seg/config.yaml. The parameters in the provided config file were tuned on the validation set and used in our experiments.

🟢 Evaluate the performance on the test set

To finally evaluate the trained segmentation algorithm on our provided manual annotations, please run:

python syn_cerebral_octa_seg/test.py --run <name of experiment in ./runs folder> --last

Results can be accessed in <name of experiment in ./runs folder>.

🟢 Citation

If you find our work useful for your research, please consider citing our manuscript:

@misc{wittmann2024simulationbased,
      title={Simulation-Based Segmentation of Blood Vessels in Cerebral 3D OCTA Images}, 
      author={Bastian Wittmann and Lukas Glandorf and Johannes C. Paetzold and Tamaz Amiranashvili and Thomas Wälchli and Daniel Razansky and Bjoern Menze},
      year={2024},
      eprint={2403.07116},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}