Brain Tumor Segmentation Using Convolution Neural Networks

Abstract/Project Description

Brain tumors treatment requires to know how extent the tumor was expanded. Without ionizing radiation magnetic resonance imaging(MRI) technique is one of the primary diagnostic tools for the brain tumor. Segmenting brain tumor extent from MRI consumes a huge amount of time if it is done manually. Using neural networks to segment the tumor extent will be an efficient way to examine brain tumor. But in traditional neural networks models for segmentation problem, Cross Entropy and Dice coefficient were commonly used as loss functions. Some type of brain tumors(like glioblastomas) are infiltrated and their borders are fuzzy and loss functions like cross entropy and Dice coefficient will not help in this situation. In this project, we propose a neural network model based on Active Contour models(ACM) and these ACMs will consider the geometrical information of the contours and takes them into consideration to avoid fuzziness during segmentation.

Introduction

One of the most dangerous types of cancers is brain tumor, because a person who is suffering with brain tumor will lose their cognitive functions and their quality of life will be poor. Most common brain tumor is gliomas, and there are two type of gliomas, one is low grade gliomas( a person diagnosed with this will have several years life expectancy) and another one is high grade gliomas( a person diagnosed with this will have 2 years life expectancy). A surgery is required for treating the brain tumor even though radiation and chemotherapy were used to slow down the growth of the tumor. Magnetic resonance imaging(MRI) helps by providing detailed pictures of brain, and most widely used method for diagnosing brain tumors. It will be hard to detect the tumors because they were poorly contrasted, tentacle like structures, and often diffused. One more issue with segmenting brain tumors is they can be formed anywhere in the brain in different size and shape. Brain typically consists of three tissues: the grey matter, white matter and cerebrospinal fluid. To segment a brain tumor, we need to identify abnormal areas rather than normal tissues. Glioblastomas which is one of the most difficult brain tumors to segment because their border are fuzzy, and it will be difficult to distinguish them with normal tissue. To overcome this issue, we propose a neural network model with active contour model, and that contour model will take geometric information of the contour into consideration while performing loss function. In that way the edges were also taken as loss during gradient descent.

 

Background/Related Work

In this section, we will discuss about different related works, CNN based segmentation methods and use of different loss functions which are related to this work.

CNN- based Segmentation Methods

In many of the computer vision tasks, CNNs showed a remarkable performance. The end to end working nature of CNN is one of its strength because this approach can extract hierarchical and multi-resolution features during learning process. In one of the previous works[4], authors used small 3×3 kernels so that it will help to design a deep neural architecture and intensity normalization in pre processing step, which together proved to be very effective in brain tumor segmentation. Also, authors used high level extracted features from CNN using hough transform technique and the tumors which were detected were segmented with a set of fully connected layers and the segmented mask is classified through FCs[5].In another works related to this topic[6], authors used a different implantation of CNN by making it exploit both local and global features simultaneously, their network use a final layer which is actually convolution implementation of fully connected layer which will allow 40 fold speed up and they used two phased training to tackle imbalances in labels of brain tumors. To avoid computational costs of CNN, authors[7] implemented a CNN based model which will efficiently combines advantages long range 2D context and short-range 3D context, by using voxelwise voting strategy, they merged the outputs of several cascaded 2D-3D models so that this will overcome the limitations of specific choices of neural networks and they implemented a network architecture in which separate subnetworks are used to process the different MR sequences so that model is more robust to the problem of missing MR squences. CNN can be categorized into pixel based and imaged based approaches when it comes into segmentation tasks. Each pixel in the image will be classified into different objects in pixel-based approach. The image-based approaches like U-net is simple and have good performance when compared to pixel-based approaches. But, lack of consideration on outside the target so that small segmented object occurs around the boundaries. To Overcome this issue, researchers prove that developing different loss function will improve the performance of the U-net. This will be done by introducing shape-aware term in the segmentation loss function. The performance results of cervical X-ray images were increased by 12% by using this approach. Inspired from this progress, we borrowed a novel loss function to improve the segmentation performance.

Get Help With Your Essay
If you need assistance with writing your essay, our professional essay writing service is here to help!
Essay Writing Service

 

LOSS Functions

Loss functions play a key role while training a CNN model. Cross-Entropy and Dice co-efficient are commonly use loss functions. But they have limitations: one is they are pixel-based loss functions and they measure the similarity between ground truth and prediction result on pixel based, but the geometric information will not be utilized. The proposed loss function will consider the region and length of the boundary so that it will help in preserving the shape of the contour.

Goal/Purpose of the Project

 

The goal of the project is to build Neural network based on active contour models to segment brain tumor, this method will consider geometrical information(length and area of the region) while calculating the loss function. This approach will help preserving shape of the contour which plays an important role while examining the tumor.

 

Features and Methodology

In this project, we use U-net architecture to segment the brain tumor and we will use Cross Entropy loss, Dice Co-efficient loss and Active contour loss while training models and compare the results.

Method 1: U-net architecture

•         For biological microscopy images O. Ronneberger et al. (2015) the FCN of J. Long et al(2015).

•         This Networks is composed of two parts

The contracting part which will compute features

The expanding part to spatially localize patterns.

•         The contracting part will extract features with 3×3 convolutions just like FCN.

•         The expanding part will reduce feature maps while increasing their width and height.

•         To avoid losing pattern information cropped features from contracting part were copied.

•         In the end, to generate a segmentation map a 1×1 convolution will process all the feature maps.

 

 

 

Image Source: O. Ronneberger et al. (2015)

 

 

Unique Contributions:

 

We will build a U-net CNN with active contour loss(which considers geometric information) for segmenting brain tumor. Up to my knowledge this approach is never used for this problem and this will help in segmenting the tumor without loss edges.

 

Technology Stack:

•         U-Net Convolution Network

•         Active Contour models

•         Python

•         Jupyter Notebook

Data Set:

 

The data which we will be using is provided by 2019 MICCAI BraTs Challenge [3]. Each data set consists of four different MRI pulse sequences, each of which consists of 155 brain slices. Professional clinicians participated in providing ground truth labels for each case

Timelines:

Time Line (2019)

No of Weeks

Task

Sep 23rd to Oct 4th , 2019

2 weeks

BraTs data set analysis

Oct 5th to November 2nd ,2019

5weeks

Image pre processing and making data ready for the model

November 4th,2019 to January 6th ,2020

10 weeks

Building model and using different loss fucntions to compare results

January 7th to February 7th,2020

4 weeks

Report Writing

February 10th to March 10th ,2020

3 weeks

Submit Report And approvals

References:

[1] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. Unet: Convolutional networks for biomedical image segmentation.

[2] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. International Journal of Computer Vision.

[3] SM Masudur Rahman Al Arif, Karen Knapp, and Greg Slabaugh. Shape-aware deep convolutional neural network for vertebrae segmentation.

[4] Sérgio Pereira, Adriano Pinto, Victor Alves and Carlos A. Silva. Brain Tumor Segmentation Using Convolutional Neural Networks in MRI Images.

[5] Mina Rezaei, Haojin Yang, and Christoph Meinel. Instance Tumor Segmentation using Multitask Convolutional Neural Network.

[6] Mohammad Havaei, David Warde-Farley, Aaron Courville, Chris Pal, Hugo Larochelle, Axel Davy, Antoine Biard, Y. Bengio, Pierre-Marc Jodoin. Brain Tumor Segmentation with Deep Neural Networks.

[7] Pawel Mlynarski, Hervé Delingette, Antonio Criminisi, Nicholas Ayache. 3D Convolutional Neural Networks for Tumor Segmentation using Long-range 2D Context.