addison elliott , akhila karlapalem amy givan maria ... · addison elliott1, akhila karlapalem1,...

Post on 08-Aug-2020

20 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

RESEARCH POSTER PRESENTATION DESIGN © 2015

www.PosterPresentations.com

y = 1.308x - 158.33R² = 0.7585

0

500

1000

1500

2000

2500

3000

0 1000 2000 3000

Co

mp

ute

r A

rea

(mm

2)

Observer Area (mm2)

y = 1.3156x - 154.96R² = 0.7094

0

500

1000

1500

2000

2500

0 500 1000 1500 2000 2500

Co

mp

ute

r A

rea

(mm

2)

Observer Area (mm2)

Figure 4 – (3-4) Edge converted polar image with optimal path

shown in blue. (5) Cartesian image with optimal path in blue

• Dice Similarity Coefficient

• End-Systolic: 0.750±0.074

• End-Diastolic: 0.802±0.079

• Hausdorff Distance

• End-Systolic: 9.03±2.71mm

• End-Diastolic: 9.09±2.37mm

• 10 female subjects, pre- and post-intervention, 18-40 years old

• Select end-systolic (ES) and diastolic (ED) frames from scans

• 2 apex, 7 papillary muscles, 6 mitral valve (randomly chosen)

• B-mode images created from custom software from IQ files

• Wavelet’s transform used for speckle reduction

Algorithm

1. Smooth image using anisotropic diffusion

2. Convert to polar domain (origin = endocardial center)

3. Robert’s edge detection filter applied to polar image

4. DP Dijkstra’s Algorithm

1. Cost function = inverted edge image

2. Moves from left->right angular dimension (loops back

around), stops at start point

5. Convert path back to Cartesian domain

METHODS

RESULTSDISCUSSION AND CONCLUSIONS

• Simple validation of algorithm to segment endocardium

• Limitations

• Algorithm purely based on image content, messes up for

missing or weak boundaries

• Only one observer, not providing any context of inter-

observer variability

• Future Work

• Factor in temporal dimension (easier to see with motion)

• Use shape model or ASMs to segment boundaries

• Integrate algorithm in one software package to eventually

fit a shape model to

• Coronary heart disease is the #1 leading cause of death

globally

• Echocardiography is the first step in assessing cardiac health

• Ultrasound is inexpensive, portable and non-invasive

compared to MRI

• Measures such as LVEF, LV mass/volume difficult to measure

manually

• Automatic segmentation allow easy calculation of measures

• Objective: Create automatic segmentation algorithm to

segment the endocardium in PSAX views. Validate to 1

expert tracer

Addison Elliott1, Akhila Karlapalem1, Amy Givan2, Maria Fernandez-del-Valle2, Jon Klingensmith1

1 Department of Electrical and Computer Engineering, Southern Illinois University Edwardsville, Edwardsville, Illinois, 2 Department of Applied Health, Southern Illinois University Edwardsville, Edwardsville, Illinois

Endocardium Segmentation in Parasternal Short-axis Echocardiograms using Dynamic Programming

Figure 1 – Diagram of parasternal short-axis views

Figure 5 – Left: Subject 8 ED mitral valve PSAX. Right: Plot of

computer & observer CSA with black line Hausdorff Dist.

INTRODUCTION

Figure 2 – Apical PSAX view

-800

-600

-400

-200

0

200

400

600

800

1000

0 500 1000 1500 2000

Dif

fere

nce

in A

rea

(mm

2)

Mean Area (mm2)

-800

-600

-400

-200

0

200

400

600

800

1000

0 500 1000 1500 2000 2500

Dif

fere

nce

in A

rea

(mm

2)

Mean Area (mm2)

Figure 6 – Bland-Altman charts (bottom) and linear regression

(top) shown for ES (left) and ED (right)

ACKNOWLEDGEMENTS

Funding provided by AHA Grant #17AIREA33670361

(Volumetric assessment of epicardial adipose tissue using

echocardiography)

IRB ID: #17-0817-1C

RESULTSMETHODS

Figure 3 – (1) Smoothed PSAX apical view. (2) Polar image from

smoothed image in (1). Endocardial center shown in (1) as red

point.

top related