a comparative study of deep learning based...

A comparativestudy of deeplearning based

methods for MRIimage processing

RobertDadashi-Tazehozi

rd2669

Introduction

Articles

Motivation

Medicalbackground

Neurological Diseases

MRI

Image processingpipeline

Datasets

Preprocessing

Random Forest forClassification

Results

A comparative study of deep learning basedmethods for MRI image processing

Robert Dadashi-Tazehozird2669

Department of Computer ScienceColumbia University

Deep Learning for Computer Vision and NaturalLanguage Processing

EECS 6894




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Outline

IntroductionArticlesMotivation

Medical backgroundNeurological DiseasesMRI

Image processing pipelineDatasetsPreprocessingRandom Forest for ClassificationResults




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Articles

Deep Learning for Cerebellar Ataxia Classification andFunctional Score RegressionZhen Yang, Shenghua Zhong, Aaron Carass, Sarah H. Ying,and Jerry L. PrinceJohns Hopkins University

Deep Learning of Image Features from Unlabeled Datafor Multiple Sclerosis Lesion SegmentationYoungjin Yoo, Tom Brosch, Anthony Traboulsee, David K.B.Li, and Roger TamUniversity of British Columbia




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Outline







rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Motivation

I MRI data

I Deep Learning + Machine Learning

I Different goals and barriers while same data type, hencea comparative study- Data description- Preprocessing- Methods and Algorithms used- Results

I Project: Diabetic Retinopathy Detection

I Personal reasons




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Outline







rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Ataxia

Overview A neuro-degenerative disease- Affects the cerebellum- Symptoms: lack of muscular coordination

https://www.youtube.com/watch?v=5eBwn22Bnio

Goals:Classify different types of Ataxia: HC, SCA2, SCA6, ATQuantify functional loss based on structural change




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Multiple sclerosis

Overview A inflammatory disease- Brain cells damaged: Demyelination- Symptoms: Mental, Physical, Psychatric troubles- Cause: Genetic? Environmental factors?

https://www.youtube.com/watch?v=qgySDmRRzxY

Goals:Automatic segmentation of lesionned areas




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Outline







rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

MRI imaging

I Human body composed of small magnets

I Magnets aligned then excited by pulses

I Magnets go back to their lowest energy state,electromagnetic waves emitted

I Processing of these waves enable to reconstruct 3Dstructure, differentiate muscles tissues from fat, whitematter from grey matter in the brain




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

MRI imaging

Figure: Initial state

Figure: Excitation

Figure: Energy emission




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Outline







rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

DatasetsAtaxia

I 168 scansI 3D Images projected on 9 plans (Resulting in 32 * 32

images)

Not enough labelled data ⇒ Interest of generatingsynthetic data

Multiple SclerosisI 1450 scansI Resolution 256 * 256 * 50I Only 100 scans segmented

Lot of unlabelled data ⇒ Interest of unsupervisedmethods




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Outline







rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Preprocessing

Ataxia

I 3D Images projected on 9 plans (Resulting in 32 * 32images)

I Generate translated and rotated images

I Reduce dimensions using a Stacked Auto-Encoder foreach plan

Multiple Sclerosis

I Images of resolution 256 * 256 * 50

I Patches 9 * 9 * 3 (low-scale features)⇒ Train Restricted Boltzmann Machines for featureextraction

I Patches 15 * 15 * 5 (high-scale features)⇒ Train Deep Belief Network for feature extraction




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Stacked Auto-Encoder

I Target vector is the output

I Along the way, compression of the data

I Trained layer after layer (greedy)




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Restricted Boltzman Machines

I Visible and Hidden Units

I Energy Based Method

E (v , h) = −vTWh − aT v − bTh

P(v , h) =1

Ze−E(v ,h)




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results





rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results


How to train RBM ? [Hinton 2002]

I Generate hidden units from visible units

I Generate visible units from these very hidden units

I Update weight:

∆Wi ,j = α(< vj , hi >input − < vj , hi >generateddata)




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Deep Belief Networks [Hinton 2009]

I Stacked RBMs

I Greedy training




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Outline







rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

ClassificationAtaxia

Multiple Sclerosis




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Random Forests

I Bagging Data

I Selecting Features

I Generate Decision Tree

Averaging all the decision trees ⇒ Classifier




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Outline







rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Results

Ataxia

Multiple SclerosisDice Similarity Measure:Q : Lesionned voxels (manually segmented)P : Lesionned voxels (automatically segmented)

d =2|Q ∩ P||Q|+ |P|

Weiss (state of the art) : mean 29, std 13Method: mean 38, std 19




rd2669

Introduction

Articles

Motivation

Medicalbackground


MRI


Datasets

Preprocessing


Results

Thank you for your attention !

a comparative study of deep learning based...

Documents