This repository contains codes for the autoencoder described in the paper Retinal Vein detection using Residual Block Incorporated U-Net Architecture and Fuzzy Inference System . I used the train set of DRIVE dataset for training and it's test dataset along with data from STARE for testing.
Here are the performance metrics measured on both datasets using my model.
| Metric | Drive | Stare |
|---|---|---|
| Accuracy | First manual: 0.9675 | A. Hoover: 0.9537 |
| Second manual: 0.9712 | V. Kouznetsova: 0.9314 | |
| Precision | First manual: 0.8453 | A. Hoover: 0.7689 |
| Second manual: 0.8486 | V. Kouznetsova: 0.8622 | |
| Sensitivity | First manual: 0.7690 | A. Hoover: 0.5582 |
| Second manual: 0.8013 | V. Kouznetsova: 0.4380 | |
| Specificity | First manual: 0.9865 | A. Hoover: 0.9862 |
| Second manual: 0.9869 | V. Kouznetsova: 0.9915 | |
| NPV | First manual: 0.9781 | A. Hoover: 0.9645 |
| Second manual: 0.9781 | V. Kouznetsova: 0.9645 | |
| AUC | First manual: 0.9818 | A. Hoover: 0.9223 |
| Second manual: 0.9848 | V. Kouznetsova: 0.8987 |
- Numpy
- OpenCV
- Keras
- Tensorflow
- Scikit-learn
All images were manually converted to *.tiff since OpenCV can not read images from DRIVE and STARE with other formats.
The images were extracted and kept under the directory ./data/.
Run train.py. Change label name with your labels in line 28 and 42
Run eval.py. Change label name with your lables in line 21