https://github.com/byrkbrk/unet-implementation

PyTorch implementation of U-Net for segmentation of neural structures in electron microscopic stacks
https://github.com/byrkbrk/unet-implementation

deep-learning google-colab image-segmentation medical-imaging pytorch unet

Last synced: 22 days ago
JSON representation

PyTorch implementation of U-Net for segmentation of neural structures in electron microscopic stacks

• Host: GitHub
• URL: https://github.com/byrkbrk/unet-implementation
• Owner: byrkbrk
• Created: 2023-02-19T13:42:56.000Z (almost 2 years ago)
• Default Branch: main
• Last Pushed: 2023-10-08T14:53:09.000Z (over 1 year ago)
• Last Synced: 2024-11-13T16:51:53.936Z (3 months ago)
• Topics: deep-learning, google-colab, image-segmentation, medical-imaging, pytorch, unet
• Language: Jupyter Notebook
• Homepage:
• Size: 17.4 MB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# U-Net implementation using PyTorch on Google Colab

We implement the well-known image segmentatation architecture, [U-Net](https://arxiv.org/abs/1505.04597) for segmentation of neural structures in electron microscopic stacks.

Because [the segmentation challenge website](brainiac2.mit.edu/isbi_challenge/) indicated in 

[the website](https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/) 

of the authors (of the U-Net article) is (apparently) no longer accessible, we use the datasets (i.e., `volumes-train.tif`, `labels-train.tif`, `volumes-test.tif`) shared in [the repository](https://github.com/zhixuhao/unet/tree/master/data/membrane).

Our U-Net architecture is inspired by the Coursera course [Apply GANs](https://www.coursera.org/learn/apply-generative-adversarial-networks-gans/home/week/2).

---

## Requirements

This repository is designed to train the model and make inferences entirely on Google Colab. So, for successful training and inference, it suffices to

- open the notebook [`unet_cell_data.ipynb`](https://github.com/byrkbrk/unet-implementation/blob/22248e94a769afd2894ad695b7c64d89cfaaeadc/unet_cell_data.ipynb) 

on Colab (by using either this [link](https://colab.research.google.com/github/byrkbrk/unet-implementation/blob/main/unet_cell_data.ipynb) or the link *Open in Colab* at the top left of the notebook)

- sign in your Google account (if you haven't yet)

- run the cells (having short explanatory comments) one by one

## Training dataset

Our training dataset consists of 30 (512-by-512) input-label image pairs (i.e., `volumes-train.tif`, `labels-train.tif`) and one example of them is as follows:

![input-label-pair](./images-for-readme/input-label-pair.png)

## Test dataset and predictions

In like manner, our test dataset (viz., `volumes-test.tif`) comprises a total of 30 (512-by-512) images. Presented below are certain exemplars drawn from the test dataset, together with the corresponding predictions (of our trained U-Net):

![test-images](./images-for-readme/test-images.png)

![predictions](./images-for-readme/predictions.png)