This is a Pytorch implementation of U-Net architecture following the paper:
"U-Net: Convolutional Networks for Biomedical Image Segmentation"
https://arxiv.org/pdf/1505.04597.pdf
The data is from 2012 ISBI Biomedical segmentaion challenge.
The model architecture is as follows:
Implementaion of this model is in model.py
High quality dataset is quite rare in the case of biomedical segmented data.
there is only 30 images in the training set.
since neural networks tends to work better when trained on large enough data, I use various image augmentation methods suggested by the author such as HorizontalFlip, VerticalFlip, ElasticTransform, RandomRotation, RandomResizedCrop etc.
All of this augmentation methods are implemented in augmentation.py
I used BCEWithLogitsLoss from torch.nn as my loss function.
I used Adam optimizer with a constant learing rate of 2e-4.
training code is in train.py
to train the model, you just need to run:
python train.py
Intersection Over Union (IoU): 0.8345
