Vanilla U-Net based approach for retinal image segmentation

Even though this was not intended in paper but BatchNorm layers added before every ReLu layer . In new version because of that converging speeded up.

It is well known that U-Net is network which mainly used in Segmentation tasks. In this application i implemented network from scratch using U-Net paper U-Net paper

RITE dataset relatively small that is why data augmentation applied (5 degree rotation for every single image)(256x256x3)

Here is simplified network architecture

!

This repository doesnt includes any training dataset if you want to get some training material check out RITE Dataset

• Requirements :
  • OpenCV
  • PyTorch
  • NumPy
  • TorchVision

Here is a result from testset:

Learning Curve: