Deep learning model designed for grading diabetic retinopathy

python = 3.8.18
torchvision = 0.16.0
torch = 2.1.0
timm = 0.9.12
tensorboard = 2.14.0
tqdm = 4.66.1

eye_pre_process： Preprocessing of retinal fundus images.
Encoder： Encoder training module.
modules: Contains model structure, loss function and learning rate reduction strategy.
utils：Contains some common functions and evaluation indicators.
BFFN：Binocular Features Fusion Network training module.
CAM： Category attention module.

1. Construct dataset

├── EyePACS dataset
    ├── train
        ├── class1
            ├── image1
            ├── ...
        ├── class2
            ├── image2
            ├── ...
        ├── ...
        ├── class5
    ├── valid
    ├── test

2. Configure the configs training file in the Encoder folder.
3. Add Category Attention Module before backbone's avgpool layer.
4. Run the main.py function in the Encoder folder to start training.

1. Construct a paired fundus image dataset (pair the left and right fundus images to generate a pkl file in the following format)

pkl_dict={
'train':[(path/1_left.png,path/1_right.png,left_label),(path/1_right.png,path/1_left.png,right_label),...],    
'valid':[...],
'test': [...]}

2. Fill in the generated pkl file address into the parameter configuration part of main.py

parser.add_argument('--data-index', type=str, default=r"path/pkl_dict.pkl", help='paired image data index path')

3. Run the main.py function in the BFFN folder to start training.