The available fundus images are not enouph to train Deep Learning DL model from scrach so will will used transfer learning technique. We used pretrained model witch trained on ImageNet dataset. To adapt the ResNet152 to lesion structures in fundus images I extracte lesions from anottated dataset. I agumenteed extracted ROIs of lesion with different sizes and differetn orientations. I re-initialized CONV layers of pre-trained model and fine-tuned it using extracted ROIs(3 classes). I load it and use for feature extraction Kaggle and Messidor datasets. Becuase I trained and fine-tuned the model on other dataset, I can used the all fundus iamges from both datasets as a testing set.I evaluate my method using Gradient boasting classifier and 10 fold CV.

Built With

Share this project: