VICCE
Variational Intensity Cross Channel Encoder for Unsupervised Vessel Segmentation on OCT Angiography (VICCE)
Variational intensity cross channel encoder is an vessel segmentation algorithm for 2D OCT angiography images. The associated publication is:
- Y. Liu, L. Zuo, A. Carass, Y. He, A. Filippatou, S.D. Solomon, S. Saidha, P.A. Calabresi, and J.L. Prince, "Variational intensity cross channel encoder for unsupervised vessel segmentation on OCT angiography", Proceedings of SPIE Medical Imaging (SPIE-MI 2020), Houston, TX, February 15 - 20, 2020. (PDF) (doi)
VICCE | |
Source code | 10kB |
Pretrained model | 80.2MB |
License (GPL v3.0) | 35kB |
Instructions
Prepare data
The data loader requires pairs of Spectralis and Cirrus scans as input. Name the Spectralis scan as "H.png" stand for Heidelberg and the Cirrus scan as "Z.png" for Zeiss, also provide a mask image "M.png" to outline the common field of view of the Spectralis scan and Cirrus scan. Those three images are stored in subfolder inside "data/train/" and "data/val/". The directory structure of the whole preject is as follows:
├──datasets │ ├──__init__.py │ └── vicce_dataset.py ├──train.py ├──evaluate.py ├──params.json ├──pytorch_env.yml ├──model │ ├──__init__.py │ ├──sampling.py │ ├──unet.py │ └──vicce_model.py | ├──utils │ ├──__init__.py │ └──utils.py | └──data ├──train | └──subject_* | ├── H.png | ├── Z.png | └── M.png └──val └──subject_* ├── H.png ├── Z.png └── M.png
Install packages
To install the required packages, run conda env create --name pytorch --file=pytorch_env.yml
Training/Evaluation
- If you have prepared registered pairs of Spectralis and Cirrus scans in "data/train/", you can use
python train.py --output vicce
to train VICCE for your own data. It will load hyper parameters from "params.json" and store checkpoints into folder "checkpoints/vicce/".
- To test images using existing models, run
python evaluate.py --output vicce
This will load parameters from "checkpoints/vicce/". This step also assume both "H.png" and "Z.png", as well as "M.png" exists in "data/val/*/", you might need to create dummy images for the data loader to work properly.
Use pretrained model
Create a folders "checkpoints/pretrained/" in the root directory. Download the pretrained models and put them in the created folder. Prepare your testing image and run python evaluate.py --output pretrained
to specify to use the pretrained model.
If you have questions regarding the method or software, please contact Yihao Liu