Image

| Paper | File |
| Lesion-aware Contrastive Learning for Diabetic Retinopathy Diagnosis | Contrstive Learning&Downstream Task |
| | |

Methodology

Requirements

• Python 3
• CUDA 11
• yaml
• PIL
• tqdm
• PyTorch=1.7.1
• torchvision

Dataset

All the used dataset are publicly-accessible:

EyePACS

IDRiD

Messidor

Train

Stage1: Construct of positive patch set and negative patch set

Stage2: Train the teacher model

$ python train.py

Stage3: Train the student model

$ python student_train.py

Evaluate

Fine-tuning the model in downstream tasks and conducting testing.

The dataset should be stored in the following file format:

eyepacs
|-- train
|   |-- 0
|   |   |-- 1.png
|   |   |-- 2.png
|   |   `-- 3.png
|   |-- 1
|   |   |-- 4.png
|   |   |-- 5.png
|   |   `-- 6.png
|   |-- 2
|   |   |-- 7.png
|   |   |-- 8.png
|   |   `-- 9.png
|   |-- 3
|   |   |-- 10.png
|   |   |-- 11.png
|   |   `-- 12.png
|   `-- 4
|       |-- 13.png
|       |-- 14.png
|       `-- 15.png
|-- val
|   |-- ...
|-- test
|   |-- ...

Execute the following commands:

$ python student_train.py -config='eyepacs.yaml'

Acknowledgment

Thanks for the Lesion_CL for the lesion detection network and the implementation of models, MoCo for the contrastive loss.