This repo is the code archive for An Accurate and Efficient Neural Network for OCTA Vessel Segmentation and a New Dataset (ICASSP 2024).
@misc{ning2023accurate,
title={An Accurate and Efficient Neural Network for OCTA Vessel Segmentation and a New Dataset},
author={Haojian Ning and Chengliang Wang and Xinrun Chen and Shiying Li},
year={2023},
eprint={2309.09483},
archivePrefix={arXiv},
primaryClass={eess.IV}
}
FRNet is a simple and efficient vessel segmentation network for OCTA images. It can achieve performance close to the SOTA method with only about 100k parameters. The model takes up less than 1MB of storage on the disk.
git clone https://github.com/nhjydywd/OCTA-FRNet
cd OCTA-FRNet
pip install -r requirements.txt
python run_benchmark.py
The current version of the code contains 2 models: FRNet-base and FRNet, and 3 datasets: ROSSA, OCTA_500 (3M) and OCTA_500 (6M).
By running run_benchmark.py, the 2 models on 3 datasets will be trained and evaluated at once (that is, a total of 2x3=6 results).
The results will be saved in json format to the result folder.
ROSSA is an retinal OCTA vessel segmentation dataset semi-automatically annotations created by us using Segmentation Anything Model(SAM). It contains 918 images, which are stored in the dataset/ROSSA folder of this repo:
train_manual contains 100 images (NO.1-NO.100) that we manually annotated, using as training set.
train_sam contains 618 images (NO.301-NO.918) that are semi-automatically annotated using SAM, also using as training set.
val contains 100 images (NO.101-NO.200) that we manually annotated, , using as validation set.
test contains 100 images (NO.201-NO.300) that we manually annotated, , using as test set.
If you want to run your own dataset, you can configure it in datasets.py, in function prepareDatasets:
def prepareDatasets():
all_datasets = {}
// Add your datasets here
// ......
return all_datasets
Note that your dataset should follow the given structure:
--dataset
|
|--Your Dataset
|
|--train
|--val
|--test
where each folder in train, val, test should follow the given format:( take train as an example)
--train
|
|--image
| |
| |--......(images)
| |--......
| |.......
|--label
|
|--......(labels)
|--......
|......
If you want to run your own model, please modify the models variable in settings_benchmark.py:
models = {
# More models can be added here......
}
Each item in models must be of type ObjectCreator, in which your model can be created.
The OCTA-500 Dataset: IPN-V2 and OCTA-500: Methodology and Dataset for Retinal Image Segmentation
Segmentation Anything Model (SAM): https://github.com/facebookresearch/segment-anything