In this project I've built a UNET model for liver tumor segmentation via PyTorch.
The dataset: Liver Tumor Segmentation.
For now it's just a jupyter notebook with:

• exploration of the initial dataset
• creation of new dataset of 2D images and their target masks
• defining Datasets and Dataloaders for our model
• defining UNET model
• experiments with fresh new UNET model
• experiments with small sanity-check dataset on fresh new UNET model

After understanding that this problem will take too much time to train a model from scratch, I've decided my next steps will be:

• refactor the code
• use pretrained ResNet layers for UNET's encoder
• train it on the whole dataset of scans
• maybe more experiments with loss function and over/undersampling duo to big class imbalance and importance of class 2