Computer Aided Dignosis on Fundus Images

Bachelor Thesis project in SEU

Interface of the model

To start the interface, run:

1. Download models to the saved_models foler
2. run python main.py

Classification

1. A multi-label classification task using Resnet-50.
2. The trained dataset is: KaggleDR+ with 8 major classes and 53 minor classes.

Detection

1. Adopt weakly-supervised method for heatmap visulization which can serve the purpose of lesion detection in a way;
2. The methods include CAM, GradCAM and GradCAM++.

Segmentation

1. Use LadderNet for vessel and disc segmentation;
2. The model is trained on different datasets: DRIVE for vessel segmentation, IDRiD for optic disc segmentation.