Meta-Learning: Finetuning of Pretrained Network with Attention

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

Docker

docker build -f Dockerfile -t polypseg .

Data

Dataset were made publicly available on /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.

Model

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 .

Training steps

Download the dataset .
Run docker container.
Run main.py script.Root path is set to ./medico2020. For help run: python3 main.py --help.

Test prediction

Download the test data
Run docker container.
Run test.py which will load the provided trained weights from ./weights/unet.pt file
. Test predicted images will be saved in "./predictions". For help run: python3 test.py --help.

Results

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

Mean DSC on validation=96.35 %

Mean IOU on validation=87.91%