6. Context
Research proposal – Quick overview
Quick overview 1
1 2 3
1 Fundus photography 2 Vessel extraction 3 Artery/Vein labelling
Automated analysis of blood vessels
with deep learning
Focus area
7. 7
Analysis of the retina offers a non-intrusive way of diagnosing ocular
aberrations, but also systemic diseases
7
Medical motivationContext 2
Ocular diseases
1
Systemic diseases
2
Image of the human retina
Fovea
Optic disc
• Diabetic retinopathy
• Glaucoma
• Ocular hypertension
• Cardiovascular disease (CVD)
• Alzheimer
• Atherosclerosis
Rods and cones - zoom
8. 8
Analysis of the retina offers a non-intrusive way of diagnosing ocular
aberrations, but also systemic diseases
8
Medical motivationContext 3
Ocular diseases
1
Systemic diseases
2
Image of the human retina
Fovea
Optic disc
• Diabetic retinopathy
• Glaucoma
• Ocular hypertension
• Cardiovascular disease (CVD)
• Alzheimer
• Atherosclerosis
Hard exudate
Hemorrhage
9. 9
Analysis of the retina offers a non-intrusive way of diagnosing ocular
aberrations, but also systemic diseases
9
Medical motivationContext 4
Ocular diseases
1
Systemic diseases
2
Image of the human retina
Fovea
Cupping of
optic disc
• Diabetic retinopathy
• Glaucoma
• Ocular hypertension
• Cardiovascular disease (CVD)
• Alzheimer
• Atherosclerosis
Narrowing arterioles
10. 10
Analysis of the retina offers a non-intrusive way of diagnosing ocular
aberrations, but also systemic diseases
10
Medical motivationContext 5
Ocular diseases
1
Systemic diseases
2
Image of the human retina
Fovea
Optic disc
• Diabetic retinopathy
• Glaucoma
• Ocular hypertension
• Cardiovascular disease (CVD)
• Alzheimer
• Atherosclerosis
Narrowing arterioles
Narrowing venules
11. 11
Analysis of the retina offers a non-intrusive way of diagnosing ocular
aberrations, but also systemic diseases
11
Medical motivationContext 6
Ocular diseases
1
Systemic diseases
2
Image of the human retina
Fovea
Optic disc
• Diabetic retinopathy
• Glaucoma
• Ocular hypertension
• Cardiovascular disease (CVD)
• Alzheimer
• Atherosclerosis
Narrowing arterioles
Widening venules (Hypertension, atherosclerosis)
Narrowing venules (Alzheimer)
SIMILAR SYMPTOMS
OBSERVABLE IN
VASCULATURE
12. 12
Analysis of the retina offers a non-intrusive way of diagnosing ocular
aberrations, but also systemic diseases
12
Medical motivationContext 7
Ocular diseases
1
Systemic diseases
2
Fovea
Optic disc
LINKED
CRAE = Average artery width
CRVE = Average vein widthA/V Ratio
How to compute A/V ratio in automated setting?
= Image segmentation problem
13. 13
Image segmentation methods based on deep learning are currently
outperforming other methods…
13
Technological motivationContext 8
PIXEL-WISE CLASSIFICATION
COMPETITION TO LABEL 20 TYPES OF OBJECTS,
DOMINATED BY DEEP LEARNING ARCHITECTURES
14. 14
But has not been applied on the A/V classification of retinal blood
vessels up to now
14
Technological motivationContext 9
1 2 3
PIXEL-WISE CLASSIFICATION
ALREADY APPLIED NOT APPLIED THUSFAR…
15. 15
Deep learning is a set of machine learning architectures inspired by
the structure and function of the human brain
15
Deep learning explainedDeep learning applied 13
X W + b Y
SIMPLE NEURAL NETWORK
Size
Price
Y = w * X +b
IS THIS THE BEST LINE?
One perceptron
16. 16
Deep learning is a set of machine learning architectures inspired by
the structure and function of the human brain
16
Deep learning explainedDeep learning applied 13
X W + b Y
SIMPLE NEURAL NETWORK
Size
Price
Y = w * X +b
OPTIMIZE LOSS FUNCTION
Here: Mean Square Error (MSE)
One perceptron
17. 17
Deep learning is a set of machine learning architectures inspired by
the structure and function of the human brain
17
Deep learning explainedDeep learning applied 13
X W + b Y
IMAGE
SIMPLE NEURAL NETWORK
Size
Price
Y = w * X +b
OPTIMIZE LOSS FUNCTION
Here: Mean Square Error (MSE)
HOW DOES THIS APPLY ON IMAGES?
Dim: 512x512x3
One perceptron
18. 18
Deep learning is a set of machine learning architectures inspired by
the structure and function of the human brain
18
Deep learning explainedDeep learning applied 13
X W + b Y
IMAGE
SIMPLE NEURAL NETWORK
Size
Price
Y = w * X +b
OPTIMIZE LOSS FUNCTION
Here: Mean Square Error (MSE)
HOW DOES THIS APPLY ON IMAGES?
Dim: 512x512x3
X
X
X
X
…
EVERY PIXEL IS AN INPUT…
MEANS LOTS OF DATA TO PROCESS IN
THE NETWORK
Solution: CNN
One perceptron
19. 19
Deep learning is a set of machine learning architectures inspired by
the structure and function of the human brain
19
Deep learning explainedDeep learning applied 14
CNN APPLIED TO IMAGE SEGMENTATION
Many hidden layers: deep learning
20. 20
In order to obtain CRAE and CRVE in an automated way, deep learning
can be applied in two ways (here: overview of the first)
20
SegmentationDeep learning applied 16
CNN APPLIED TO IMAGE SEGMENTATION
Backpropagation to alter network weights
Manual A/V labeling – GROUND TRUTH
Fundus image
C
O
M
P
A
R
E
Model prediction
21. 30th of March, 2017
Thank you for your attention!
Ruben Hemelings
KU Leuven
rhemelings@me.com
Contact details Bart Elen
VITO
bart.elen@vito.be