Brain MRI Image Segmentation with U-Net

The network architecture in this project is based on the original U-Net publication U-Net: ConvolutionalNetworks for Biomedical Image Segmentation. The codes in this repo are mainly developed in Python and GoogleColab.

Table of contents

• Introduction
• Datasets
• Requirements
• Experiments
• Results
• Contacts

Introduction

The U-Net architecture of this project consists of an encoder and a decoder part that together givethe network an U-shaped form. The encoder part follows a traditional architecture of aconvolutional network. Each layer consists of a repeated 3x3 padded double convolutions, followed by batch normalization and a rectifier linear units (ReLU) activation. The output feature map is stored as a skip connection which will later be concatenated to the decoder part. Moving downwardsthrough the network, a down sampling is performed by a 2x2 max pooling operation with a stride of2, which doubles the number of feature channels.

Dataset

Brain MRI images

Requirements

• Python 3.9.1
• PyTorch 1.8.1
• CUDA toolkit 10.2
• Albumentations

Experiments

1. Optimiser Comparison - SGD vs Adam
2. Loss Comparison - Binary Cross-Entropy(BCE) Loss vs Dice Loss vs BCE + Dice

Results

In this project, it was found that Adam optimiser with Binary Cross-Entropy Loss produce the best test result.

SGD + BCE (Test accuracy: 68.13%)

Adam + BCE (Test accuracy: 76.48%)

Adam + Dice (Test accuracy: 74.13%)

Adam + BCE + BCE (Test accuracy: 73.81%)

Contact

• Chieh-Ju Wu (Jeremy) - jeremy.cjwukth@gmail.com
• Fredrik Mazur - fredrik@mazur.se
• Niclas Määttä - niclas.maatta@hotmail.com
• Daniel Grönås - daniel.gronas@hotmail.com