In this project, I have tested the robustness of 3D UNet model for medical image segmenation. UNet model has been very succesful for medical image segmentation. Here, we look at its robustness for brain tumour segmenation. The model was trained on the BraTS dataset. BraTS dataset contains MRI images of brain of size (240, 240, 155). For more information visit here.
The colors used in GIFs correspond to following:
- Red is edema
- Green is a non-enhancing tumour
- Blue is an enhancing tumour.
Adversarial attacks have been done on patches of size (160, 160, 16) to reduce computational time. In all the methods below, the parameters were as follows:
- Iterations: 10
- Epsilon: 0.2
- Alpha: 0.02
- Dice Coefficient of prediction before attack: 0.7913155
- Dice Coefficient of prediction after attack: 0.43321887
Ground Truth:
Prediction before attack:
Prediction after attack:
- Dice Coefficient of prediction before attack: 0.8463957
- Dice Coefficient of prediction after attack: 0.53347206
Ground Truth:
Prediction before attack:
Prediction after attack:
- Dice Coefficient of prediction before attack: 0.840131
- Dice Coefficient of prediction after attack: 0.45113495
Ground Truth:
Prediction before attack:
Prediction after attack:
First half of the code dealing with building model in Notebook.ipynb and util.py has been borrowed from here.
For more information regarding the algorithms and model used in this project, one can read the papers below: