Evaluation of Robustness of U-Net based Models to common Image Artefacts in Segmentation of Microscope Images - Machine Learning 2020

Authors: Andrea Oliveri, Célina Chkroun, Bernardo Conde

Introduction

This repository hosts our canonical implementation and training of a Neural Network called U-Net. U-Net is a neural network with the special purpose of segmentating various neuronal structures in electron microscopic stacks. We trained the net using three types of data generation from our original dataset and performed various types of analysis on the resulting models to their robustness against multiple types of image artifacts.

Folder Structure

• Dataset: directory storing the dataset.

• Models: directory containing three subdirectories in which the trained models are saved in the ProtoBuffer format.

• Scripts: contains all the Python3 code files. In this folder there are multiple important files:

  • UNET Train.ipynb is a Jupyter notebook the training of all the U-Net models is performed.

  • UNET Distortions Evaluate.ipynb is a Jupyter notebook where multiple evaluations of the trained networks are done, mostly recording how well the network performs with artificially altered images.

  • distortions.py contains all the methods used to alter the testing images.

  • image_processing_metods.py contains class definitions performing segmentation via classical image processing techniques, namely applying a filter to remove the distortions and performing otsu thresholding.

  • unet.py is where our canonical implementation of the U-Net resides.

  • utils.py contains simple helper functions used by the rest of the project.

  • plots.py contains helper functions used to show images and plot results measured in the Jupyter notebooks.

  • train_data_augmentation.py contains helper functions applying different types of distorition to training images and returning resulting images as a generator to be called during models' training.

Dependencies

Our project depends on python3, numpy, matplotlib.pyplot, notebook, opencv-python, tenserflow >= 2.1.0

Tensting environment

A .yml file was included in this repository describing the conda environment used for the training and testing. All tests were run on a Windows 10 machine with an Intel Core i7-7700HQ CPU and Nvidia GeForce GTX 1050 GPU.