a b/README.md 

  1
  
# Meta-Learning: Finetuning of Pretrained Network with Attention





  
  

  2
  





  
  

  3
  
This repo contains the code submitted to Medival challenge on GI tract's poly segmentaion. The pre-trained





  
  

  4
  
weights of Unet model trained on brain MRI images were fine tuned ; which is a open source available at  https://pytorch.org/hub/mateuszbuda_brain-segmentation-pytorch_unet/,thanks to "mateuszbuda" .Attention mechanism was incorporated while finetuning weights transfered from Unet trained on MRI dataset.





  
  

  5
  





  
  

  6
  





  
  

  7
  





  
  

  8
  
## Docker





  
  

  9
  





  
  

  10
  
```





  
  

  11
  
docker build -f Dockerfile -t polypseg .





  
  

  12
  
```





  
  

  13
  





  
  

  14
  





  
  

  15
  





  
  

  16
  
## Data





  
  

  17
  
Dataset were made publicly available on  [/multimediaeval.github.io/editions/2020/tasks/medico/](https://multimediaeval.github.io/editions/2020/tasks/medico/). It consists of  images of  polyps in GI tract and its masks. Data folders should be  arranged as in the tree structure below to load data in different mode.





  
  

  18
  
![Data Folder Structure](./readme_fig/folder_tree.png)





  
  

  19
  





  
  

  20
  





  
  

  21
  
## Model





  
  

  22
  





  
  

  23
  
The pretrained model has been finetuned  with the polyp  dataset.Unet model is guided with attention mechanism.The dataset consists of 1000 images which was then splitted into 80:20 ratio of train and validation set. The test data consisted of 160 images .





  
  

  24
  





  
  

  25
  





  
  

  26
  
## Training steps





  
  

  27
  





  
  

  28
  
1. Download the dataset .





  
  

  29
  
2. Run docker container.





  
  

  30
  
3. Run `main.py` script.Root path is set to `./medico2020`. For  help run: `python3 main.py --help`.





  
  

  31
  





  
  

  32
  





  
  

  33
  





  
  

  34
  
## Test prediction





  
  

  35
  





  
  

  36
  
1. Download the test data





  
  

  37
  
2. Run docker container.





  
  

  38
  
3. Run `test.py`  which will load the provided  trained weights from `./weights/unet.pt` file  





  
  

  39
  
. Test predicted images will be saved in "./predictions". For  help run: `python3 test.py --help`.





  
  

  40
  





  
  

  41
  





  
  

  42
  





  
  

  43
  





  
  

  44
  
# Results





  
  

  45
  





  
  

  46
  
For viewing the graph,images and events; "./log" can be loaded and viewed in tensorboard.





  
  

  47
  





  
  

  48
  





  
  

  49
  
Mean DSC on validation=96.35 %  





  
  

  50
  





  
  

  51
  
Mean IOU on validation=87.91%





  
  

  52
  





  
  

  53
  
## Plot





  
  

  54
  





  
  

  55
  
![Validation DSC VS Iterations ](./readme_fig/val_dsc.png)





  
  

  56
  





  
  

  57
  





  
  

  58
  
## Sample predictions(Red) over groundtruth(Green)





  
  

  59
  





  
  

  60
  
![Green- GT, Red- Pred](./readme_fig/img.gif)