⭕ Contribution:

▶️The first attempt to joint the Representative Instance Pre-training (RIP) with the contextual analysis of local patch for retinal A/V segmentation.

▶️The RIP task is proposed to learn latent arteriovenous features from diverse spatial regions.

▶️Distance Aware and Patch Context Fusion modules are desgined to explore the relationship of patch and its context base on the sense of vascular structures.

▶️Validated against three datasets, outperforming state-of-the-art methods.

Conda environment settings:

conda create -n rip_av python=3.10 -y
conda activate rip_av
pip install poetry 
poetry install
pip install tensorflow==2.9.1
pip install torch==1.13.1 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117 --force

Download the public datasets: AV-DRIVE, LES-AV, HRF (v1, v2) and put in the ./data

├─dataset

​	├─test 
​	│  ├─av
​	│  └─images
​	└─training
​    	├─av
​    	└─images

There are two stage: RIP-stage and segmentation stage.

Run Preprocessing/generate_patch_selection_for_RIP.py to generate representative patches for RIP stage, or directly use the pretrained weights.

python generate_patch_selection_for_RIP.py 
	--dataset_path <dataset path>
	-- train_or_test <training|test>
    --out <output path>
cd ./RIP/
python train.py

Dataset	Pretrain weight
AV-DRIVE	RIP_pretrain_drive
LES-AV	RIP_pretrain_les
HRF	RIP_pretrain_hrf

AV-DRIVE

LES-AV

HRF

# train and evaluate  
cd AV/
python main config/config_train_general

# change config file and run test with visulization
python test_with_vis.py

where the hyper-parameters ${{\lambda }{a}}$, ${{\lambda }{s}}$, ${{\lambda }_{d}}$ are meticulously calibrated to balance these three losses and set as 0.01, 5 and 1, respectively, see ./AV/config/config_train_general.py.

Dataset	checkpoint
AV-DRIVE	RIP_checkpoint_drive
LES-AV	RIP_checkpoint_les
HRF	RIP_checkpoint_hrf

Full prediction results can be available at AV-DRIVE, LES-AV, HRF.

• training and test code
• visualization demo
• test on more retinal dataset