LV Contour Segmentation in

TMR Images Using Semantic

Description of Tissue and Prior

Knowledge Correction

Student: Albert Andaluz González

Advisors: Débora Gil Resina & Jaume Garcia i Barnés

Master in Computer Vision and Artificial Intelligence

Introduction Our Contribution Results Conclusions

Index

1. Introduction

2. Our Contribution

3. Results

4. Conclusions

Introduction Our Contribution Results Conclusions

Clinical problem

Requirements: accurate contour estimation

Treatment: diagnose of the heart function

Method: Extraction of clinical scores from regional

wall motion

The problem: 30% of global deaths are caused by

heart diseases

Introduction Our Contribution Results Conclusions

Contour Segmentation

Manual:

slow

Inter/intra observer variability -> requires high

expertise.

Automatic

Objective for all subjects

Requires accurate estimation of the LV

boundaries

Introduction Our Contribution Results Conclusions

Tagged Magnetic ResonanceB

ase

Mid

Apex

Long Axis (LA)

LV(*) TMR in SA view

Short

Axis (SA)

cuts

(*)LV = Left Ventricle

Introduction Our Contribution Results Conclusions

Problems in Tagged Magnetic Resonance

…and few algorithms for TMR automatic segmentation currently exist

Blood pool and tissue appearance is similar at time=0 …

Tag pattern misleads some common image descriptors…

LV Boundaries

Contours

Introduction Our Contribution Results Conclusions

1. Introduction

2. Our Contribution

3. Results

4. Conclusions

Introduction Our Contribution Results Conclusions

Our Contribution

Goal: Algorithm for automatic segmentation

for the extraction of clinical regional scores

of the LV

Visual Computational

Tags Gabor Filter Amplitude

Movement Motion analysis

Semantic definition of the LV

Two steps:

1. Shape approximation

2. Shape correction

Introduction Our Contribution Results Conclusions

Classic snakes formulation

Internal energy External energy Elasticity Rigidity

External energy: attracts the snakes towards the LV contours

Internal energy: avoids deviation from anatomical LV shapes

Snakes minimize the energy potential given by the following equation:

Introduction Our Contribution Results Conclusions

External energy computation

1. Define contour for image potential

2. LV Contour extraction

3. Distance map to the LV contours

Introduction Our Contribution Results Conclusions

Semantic descriptors

Amplitude1 Amplitude 2 Motion moduleOriginal

t=0

t=1

t=end

…

Introduction Our Contribution Results Conclusions

Energy definition

mean mean mean

mean

*

*

*

*

*

*

*

*

*

*

*

*

Amplitude 2 Motion moduleAmplitude1

Introduction Our Contribution Results Conclusions

Contour extraction

2 Detection 3 Selection1 Clustering from E1|E2

K-Means Contours+ Binary Morphology Region area Filtering

Epicardium (outer boundary of the LV)

Endocardium (inner boundary of the LV)

Introduction Our Contribution Results Conclusions

Distance map

Epicardium Endocardium

Distance

map

Vector

Field

Distance to

LV contours

Introduction Our Contribution Results Conclusions

Shape Correction Converged snake shape is corrected using PCA and GOPA:

Incoming shape Mean shape Variation modes

If for PCA eigenvalues then

Converged snake

Correction

Correction applied

No correction

needed

External energy

Introduction Our Contribution Results Conclusions

1. Introduction

2. Our Contribution

3. Results

4. Conclusions

Introduction Our Contribution Results Conclusions

Test set

29 real cases from 15 healthy patients

2 cuts (basal and mid)

2 energies for automatic

segmentation

TMR sequences from the Clinica La Creu

Blanca

Introduction Our Contribution Results Conclusions

Segmentation error

distance between contours

Accuracy of Clinical Scores for clinicial

aplicability

Compare global & regional rotation auto vs

manual

Validation protocol

Introduction Our Contribution Results Conclusions

Segmentation error

Manual

Automatic

External energy

Base

Mid

Accuracy (pixels)

Accuracy (pixels)

Introduction Our Contribution Results Conclusions

Rotation (low error)

ManualAutomatic

LV Contours

Global score

Regional

I

IL

IS

A

AL

AS

LV

Introduction Our Contribution Results Conclusions

Rotation (mild error)

I

IL

IS

ManualAutomatic

LV Contours

A

AL

AS

LVGlobal score

Regional

Introduction Our Contribution Results Conclusions

Absolute error (in degrees)

Introduction Our Contribution Results Conclusions

1. Introduction

2. Our Contribution

3. Results

4. Conclusions

Introduction Our Contribution Results Conclusions

Conclusions

Promising support tool for clinical analysis

However, our method still shows high accuracy in

the scores:

< 0.2º mean global rotation

< 1º mean regional rotation

There is some mild error between segmentations…

We have presented a method that incorporates motion in

the description of the myocardium....

Introduction Our Contribution Results Conclusions

Future work

Validate in apical cuts

Use Mean shift for improved clustering

Validate our method in pathologic

patients

Dr. Herman K.Hellerstein - Cardiologist

1916- 1993 (Ohio, USA)

“Coronary heart disease is a

silent disease and the first

manifestation frequently is

sudden death.”