Retinal Vessel Segmenation Model Benchmark

Code associated with paper: Benchmarking Retinal Blood Vessel Segmentation Models for Cross-Dataset and Cross-Disease Generalization.

This is a benchmark with the aim to investigate the performance of various deep learning model for retinal vessel segmentation It implements 5 backbone models on 3 datasets and compare the performances for different loss functions, image qualities and pathological conditions as well as the cross dataset generalization capabilities of the models.

We are actively expanding the benchmark to include more models, UQ methods, and datasets.

Backbones

Backbone Models	Paper	Official Repo
UNet	link
FR-UNet	link	link
MA-Net	link
SA-UNet	link	link
W-Net	link	link

Data

• FIVES
• CHASEDB1
• DRIVE

Environment Setup

Our code is developed with Python 3.10, The required packages are listed in requirements.txt and the environment can be created by running:

pip install -r requirements.txt

in your created environment.

Specify Arguments using .yaml Files

Specify the arguments in any of the files in the config folder with the respective model and loss function

Training

To run the code for training the baseline models:

# 
python train_baseline.py --config config/fives.yaml

This example is for fives dataset with the models specified inside.

To train the 3vs1 subgroup setting. It requires an additional argument with the train_disease.py file models on a particular dis:

# 
python train_disease.py --config config/fives.yaml --disease D

This example is for fives dataset with the models specified inside and to train on other group of disease except D Diabetic-Retinppathy