This is an official implementation of 'A Multi-task Network with Weight Decay Skip Connection Training for Anomaly Detection in Retinal Fundus Images'. (Accepted by MICCAI 2022)

• numpy>=1.17.0
• scipy>=1.5.2
• Pillow>=8.2.0
• pytorch>=1.7.1
• torchvision>=0.8.2
• tqdm>=4.59.0
• scikit-learn>= 0.24.2
• scikit-image>=0.17.2

The proposed method is evaluated on two publicly-available datasets, i.e.

• IDRiD
• ADAM

The proposed WDMT-Net method is trained through two steps:

• Data Preparation

  Generate the list of HOG image and Patches :

  python3 data_find.py \
  --dataset ['IDRiD'/'ADAM'] \
  --path {data dir}
  

  For example, python3 data_find.py --dataset 'IDRiD' --path './dataset/'

  And then you can get lists containing images and corresponding labels in './label/'.

• Training and testing model

  python3 main.py \
  --dataset ['IDRiD'/'ADAM'] \
  --datadir './labels/' \
  --hog \
  --lr 1e-4 \
  --batch_size 32 \
  --epochs 200 \
  --deta 0.05 \