Contribu8ons to Image Processing For Medical Image Analysis From Computer-‐Aided-‐Diagnosis to Embedded Systems For In Situ Diagnosis TODAY YESTERDAY
Aymeric Histace HDR Defense
TOMORROW
28-11-2014
Roadmap • General Presenta+on • Who am I?
• Research Ac+vi+es
• What are my contribu3ons?
• Perspec+ves • What is my research project? 28/11/14
Aymeric Histace -‐ HDR Defense
2
Roadmap • General Presenta+on • Who am I?
• Research Ac+vi+es
• What are my contribu3ons?
• Perspec+ves • What is my research project? 28/11/14
Aymeric Histace -‐ HDR Defense
3
Who am I?
CV
Certificate of Birth • 4th Sept. 1977, Lyon • Parents: B. Histace and D. FerreyreHistace
1977
2001
2001
2004 PhD
MSc Master of Engineering • EIGSI La Rochelle • Industrial Systems Engineering
28/11/14
• Signal and Image in Biology and Medecine
• “Detection and tracking of structure in image sequences: application to tagged cardiac MR images”
• University of Angers
• Mention: “Très Honorable avec les félicitations du jury”
• Wavelet Compression of Thoravision images (X)
• University of Angers
Aymeric Histace -‐ HDR Defense
4
Who am I?
Professional Experience 2001
2004
PhD, “Moniteur”
2013 28/11/14
2006 Lecturer
Associate Professor (UCP, ETIS) PES
TODAY
Research • Member of ASTRE team • In charge of the “Embedded Systems for Health” axis • Elected Member of the Laboratory Council
28/11/14
Teaching Administrative • Co-Head of MSc MADOCS (Methods for Complexe Data Analysis)
Aymeric Histace -‐ HDR Defense
• Institute of Technology • Dpt of Electrical Engineering and Industrial Informatic (GEII) • Member of the “Commission de choix”
5
Who am I?
Teaching and Responsabili8es
Sept 2004 2006
IUT GEII
28/112014
Electrical Engineering, Energy, Control (L1, L2+L2 app, L3P)
MSc
Sept 2007
Image Processing (Basic, Variationnal approach, Statistical Learning) Lecturer, University of Angers
Sept 2010 L2 app (08-14): Industrial relationships and pedagogical coordination
L3P Instru. (07-08) 2007
28/11/14
2008
ENSEA
Signal Processing for ECG (ENSEA) Sept 2013 MSc SIC (11-14): Electronics for Biosensors CMI Pedagogical Coordination « Multimedia Indexing …L3P ESH MSc and Data Mining » MADOCS 2011 2012 >2014
Aymeric Histace -‐ HDR Defense
6
Who am I?
Research Overview
Next Step
Step 1
1 Noise rms
28/11/14
Aymeric Histace -‐ HDR Defense
18
In Details: Stochas8c Resonance
Useful Noise Effect
Application to Image Restoration Gaussian noise
Objectives : • Noise reduction • Contour integrity No
Method
# I(x, y, 0) = I 0 % $ ∂I %& = div gη ( ∇I ) ∇I ∂t
(
28/11/14
with
)
gη (u ) = g (u + η ( x, y )) Stochastic Variant of Perona-Malik process
Aymeric Histace -‐ HDR Defense
Yes
• η a noise assumed independent and identically distributed (rms amplitude ση) • g a decreasing monotonic function 19
In Details: Stochas8c Resonance
Useful Noise Effect
Experiment 1
gη (u) = g(u + η (x)) with
We consider a case where k parameter is badly tuned regarding PM approach : (left) Original noisy contour
⎧1 if s ≥ k g ( s) = ⎨ ⎩0 if s k
(right) Restored one using PM approach with k=0.6
Hard-threshold
Result S-PM
Purposely injection of η noise in g function can randomly retune the function PM PM
28/11/14
S-PM
Aymeric Histace -‐ HDR Defense
20
In Details: Stochas8c Resonance
Useful Noise Effect
Experiment 2
gη (u) = g(u + η (x)) with
We consider a case where k parameter is badly tuned regarding PM approach : (left) Original noisy contour
⎧1 if s ≥ k g ( s) = ⎨ ⎩0 if s k
(right) Restored one using PM approach with k=0.6
Hard-threshold
Result
(Quantification)
Variation of the ratio of well restored function of the noise rms amplitude (1000 attempts). The non-diffusion ratio is maximum for a non zero amount of noise
28/11/14
Aymeric Histace -‐ HDR Defense
21
In Details: Stochas8c Resonance
Useful Noise Effect
Extension to image restoration Gaussian Noise (PM, S-PM)
# I(x, y, 0) = I 0 % $ ∂I %& = div gη ( ∇I ) ∇I ∂t
(
with and
)
gη (u ) = g (u + η ( x, y )) g (u ) = e
−
u
2
k2
Multiplicative Noise (PM, S-PM)
28/11/14
Impulsive Noise (PM, S-PM)
Aymeric Histace -‐ HDR Defense
22
In Details: Stochas8c Resonance
Useful Noise Effect
Result (quantification) Normalized Cross Covariance wrt the amount of injected noise (1000 attempts)
Question 1: Answer Gaussian Multiplicative Impulsive
In each case the similarity measure is maximum for a non-zero amount of injected noise Elec. Letters 2006, PSIP 2007, IEEE SocPar 2010, Int Journal of Comp .Infor. Systems and Industrial Management 2012
28/11/14
Aymeric Histace -‐ HDR Defense
23
And Then…
CAD: Main Contribu8ons
Stochastic Resonance Non-Linear PDE ∂I = div gη ( ∇I ) ∇I ∂t
(
)
Active Contour With Shape Prior E = E prior + Eimage
ChanVese
Alpha-Divergence Based Active Contour
E = Ehistogram + Eregularisation
Shape learning
Shape descriptor (Legendre)
gη (u ) = g (u + η ( x, y ))
Gaussian Noise
Coll. CREATIS David Rousseau
28/11/14
Up : Foulonneau, Down : Our Approach Coll. UCLan B. Matuszewski
Aymeric Histace -‐ HDR Defense
Joint Optimization: -Segmentation Process -Metric of the divergence Coll. I3S PhD Leila Meziou
24
In Details: Segmenta8on
Ac8ve Contours
Principle
Ω = Ωin ∪ Γ ∪Ωout
Gradient-based approach:
E(∂Ω) =
!
min ( Eimage + Eregularisation +...) 28/11/14
∫ k (x, y)ds b
∂Ω
Region-based approach:
E(Ωi ) =
Aymeric Histace -‐ HDR Defense
∫ k (x, y, Ω )ds b
i
Ω
25
In Details: Ac8ve Contour
Shape Prior
Cardiac MRI
Main Idea • In Medical Image, the shape of the structure to segment is often “known” • Question: Knowing the “mean” shape of an object, can we integrate it in an AC segmentation process ?
SA
LA
Microconfocal images
X-Ray Radiography (hip bone)
Proposal
( )
( )
( )
E λr = E prior λr + Eimage λr 28/11/14
Aymeric Histace -‐ HDR Defense
with λr a reduced shape descriptor (statistical learning) 26
In Details: Ac8ve Contour
Shape Prior
Shape Space Representation Shape descriptor
Statistical Learning (PCA)
Shape Ω
Moments
λi
Test image N0
such as N0 = 5
∑λ
i
i=1 N 0 = 20 N 0 = 40 The chicken image set used to build the statistical shape model
Zernike moments
28/11/14
Legendre moments
1 λ= NS
Aymeric Histace -‐ HDR Defense
NS
∑ i =1
1 λi Q = NS
NS
∑ (λ i =1
T
i
− λ )( λ i − λ )
λ r ,i = P T ( λ i − λ )
27
In Details: Ac8ve Contour
Shape Prior
Shape Space of Moments Eigen Shapes
How it works
( )
( )
( )
E λr = E prior λr + Eimage λr Final Result
NS
NS
p ( λ r ) = ∑ N ( λ r ; λ r ,i , σ 2 )
p ( λ r ) = ∑ N ( λ r ; λ r ,i , σ 2 )
i =1
i =1
λ = λr ,1 ⋅ p1 + λr ,2 ⋅ p2 + λ
⎛ N S 2 ⎞ E prior ( λ r ) = − ln ⎜ ∑ N ( λ r ; λ r ,i , σ ) ⎟ ⎝ i =1 ⎠
28/11/14
Iterations shown in the feature space spanned by the first two principal axes
Aymeric Histace -‐ HDR Defense
Different from Template Matching
28
In Details: Ac8ve Contour
Shape Prior
Results 1 (Legendre Moments)
Chan-Vese method
Multi-reference shape prior method (Foulonneau 09)
The proposed method
Original binary image
28/11/14
Image with Gaussian noise
Image with structural noise
Image with Gaussian and structural noise
Aymeric Histace -‐ HDR Defense
29
In Details: Ac8ve Contour
Shape Prior
Results 2
Question 2: Answer 1
CAIP 2011, Journal of Mathematical Imaging and Vision 2013
28/11/14
Space Shape of Legendre Moments makes possible Shape Prior integration different from template matching
Aymeric Histace -‐ HDR Defense
30
And Then…
CAD: Main Contribu8ons
Stochastic Resonance Non-Linear PDE ∂I = div gη ( ∇I ) ∇I ∂t
(
)
Active Contour With Shape Prior E = E prior + Eimage
ChanVese
Alpha-Divergence Based Active Contour
E = Ehistogram + Eregularisation
Shape learning
Shape descriptor (Legendre)
gη (u ) = g (u + η ( x, y ))
Gaussian Noise
Coll. CREATIS David Rousseau
28/11/14
Up : Foulonneau, Down : Our Approach Coll. UCLan B. Matuszewski
Aymeric Histace -‐ HDR Defense
Joint Optimization: -Segmentation Process -Metric of the divergence Coll. I3S PhD Leila Meziou
31
In Details: Ac8ve Contour
Alpha-‐divergence
Histogram-Based Active Contour
E = Ehistogram + Eregularisation
Probability density function
Divergence
D ( p1 || p2 , Ω) =
∫ ϕ ( p , p , λ )d λ 1
2
with
ℜm
" ϕ a similarity function $ pi the pdf of Ωi # $ $% λ the quantization level
Questions: • How to model the pdf pi (derivation constraints)? • Which φ function? 28/11/14
Aymeric Histace -‐ HDR Defense
32
In Details: Ac8ve Contour
Alpha-‐divergence
Histogram-Based Active Contour pdf modeling
Divergence
(Parzen Window)
Usually
pˆ i ( λ, Ωi ) =
1 Ωi
∫ g ( I(x) − λ ) dx σ
Ωi
with gσ a Gaussian kernel of variance σ
28/11/14
• Kullback-Leibler • Hellinger • Ki2
α={0,1} α=0.5 α=2
Our proposal • Alpha-divergence
Aymeric Histace -‐ HDR Defense
33
In Details: Ac8ve Contour
Alpha-‐divergence
Joint Optimization Maximization of the divergence
Optimization of alpha-parameter
argmax Γ ( Dα ( pin || pout , Ω))
argmaxα ( Dα ( pin || pout , Ω))
$ ∂α & = −∂Dα ( pin || pout , α ) & ∂t % & ∂Γ = −∂ pin , pout Dα ( pin || pout , α ) &' ∂t 28/11/14
Aymeric Histace -‐ HDR Defense
Algorithm 1. αt+1 (αinit = 1) 2. Γt+1 34
In Details: Ac8ve Contour
Alpha-‐divergence
Result 1 (Synthetic images)
28/11/14
Aymeric Histace -‐ HDR Defense
35
In Details: Ac8ve Contour
Alpha-‐divergence
Result 2 (Synthetic images)
28/11/14
Aymeric Histace -‐ HDR Defense
36
In Details: Ac8ve Contour
Alpha-‐divergence
Result 3 (Natural Images) Microconfocal images of cells
Question 2: Answer 2 Alpha-divergence are a flexible tool to cop with different noise scenarios in medical image analysis (but not only)
Videocapsule (coloscopy)
X-Ray images
ICIP 2011, ICASSP 2012, MIUA 2012 (Best Student Paper Award), Annals of BMVA 2013, ICIP 2014
28/11/14
Aymeric Histace -‐ HDR Defense
37
CAD to in situ Diagnosis Energy
Interaction
Computation
Size
28/11/14
Aymeric Histace -‐ HDR Defense
38
2011-‐>2014: CAD to In Situ Diagnosis
Cyclope Project
Main idea (induced by Question 3) To develop a smart autonomous videocapsule with embedded image processing capabilities
In situ detection of intestinal pathologies
Polyp
Ulcer
Angioma 28/11/14
Chrone disease Aymeric Histace -‐ HDR Defense
Colorectal Cancer 39
2011-‐>2014: CAD to In Situ Diagnosis
Cyclope Project
Main idea (induced by Question 3) To develop a smart autonomous videocapsule with embedded image processing capabilities
Context
polyp
Colorectal Cancer 28/11/14
Aymeric Histace -‐ HDR Defense
Advantages
Drawbacks
•
Total control
•
Anaesthesia
•
Possibility of biopsy's
•
Hospitalization
•
Real-time analysis
Advantages
Drawbacks
•
Painless
•
Battery life
•
No sedation
•
Low resolution
•
No hospitalization
•
No control
•
Just swallow it!
•
~150k images
40
2011-‐>2014: The Cyclope Project
The Cyclope Pill
A Multispectral WCE
Infrared (Active Stereo Vision)
3D feature-based detection
Visible
28/11/14
2D feature based detection
Aymeric Histace -‐ HDR Defense
41
2011-‐>2014: The Cyclope Project
2D Detec8on (Compa8ble With Embedding Constraints)
Gray-‐level Image
Edge detec8on
Hough Transform
ROI extraction
Learning by boosting Training examples
Classifiers
Co-occurrence matrix (26 features)
Feature extrac8on (Texture)
Candidate Polyps Database : 300 Positives examples, 1200 Negatives ones
28/11/14
Aymeric Histace -‐ HDR Defense
42
2011-‐>2014: The Cyclope Project
2D Detec8on Results and Performance
28/11/14
Aymeric Histace -‐ HDR Defense
43
2011-‐>2014: The Cyclope Project
2D Detec8on Results and Performance (Real Time Tracking) Question 3: Answer It is possible to design low complexity detection/recognition algorithms in accordance with: (i) embedding constraints, and (ii) expected performance Bernal 88% 91%
Our Approach 99% 95%
1% 5% DCIS 2012, IEEE EMBC 2013, GRETSI 13, International Journal of Computer Assisted Radiology and Surgery, 2014
28/11/14
Sensibility
Aymeric Histace -‐ HDR Defense
Specificity
False Positive Rate
44
Synthesis of contribu8ons 2005
28/11/2014
From CAD…
Question 1 • Stochastic Resonance • Non-linear PDEbased image restoration process • Double Well potential 28/11/14
…to in situ Diagnosis
Question 2 Active contour with: • Shape Space moments • Alphadivergence • Fractional entropy
Aymeric Histace -‐ HDR Defense
Question 3 • Early Detection of Colorectal Cancer • Real-Time tracking of colonic polyps • In Situ Diagnosis
45
What are my contribu8ons?
Related Publica8ons
(2006-‐2014)
9 Journals, 1 Patent: PRL*, JMIV*, EL*, IJCARS*+, IJBI+, Annals of BMVA,…
12 10 8
28 Int. Conferences: ICIP, ICASSP, CAIP, EMBC, MIUA, CARS,…
6 4 2 0
2010 2011 2012 2013 2014
Indexed by JCR + Indexed by PubMed *
28/11/14
Others:
Journals, Patents
3 Book Chapters, 1 Book Editing, 1 Proceeding Editing, 6 National Conferences (GRETSI)
Inter. Conf.
Aymeric Histace -‐ HDR Defense
Others
46
Roadmap • General Presenta+on • Who am I?
• Research Ac+vi+es
• What are my contribu3ons?
• Perspec+ves • What is my research project? 28/11/14
Aymeric Histace -‐ HDR Defense
47
Research Ac8vi8es-‐Evolu8on
ECSON, TERAFS, SIMBAD
Cyclope In Situ Diagnosis
CAD
Signal and Image Processing
Embedded Systems
Smart-Aided Diagnosis Medecine - Biology
Electronic, Data-Base
Understanding of Living mechanisms
e-diagnosis SmartEEG PAPILLON
FibroSES iFib
28/11/14
Aymeric Histace -‐ HDR Defense
48
…and The Story Does Not End
Research Project (Short and Mid Terms)
Image Processing / Pattern recognition Electronics, Embedded Systems
Alpha-Divergence in Pattern Detection
Electrical Characterization of Fibrosis Induced by Implants
• Collaboration with D. Rousseau (CREATIS Lyon)
• IMS, INL, LIP6, ERRMECe
Smart-Videoendoscopy • PhD of Quentin Angermann • SATT • Evolucare Medical Imaging Cell shape characterization
Smart EEG • LIP6, INL, SME Cyclope • H2020 Project (Greece, Switzerland, UK, Spain, France, Belgium)
• Collaboration with UCLan and UB
28/11/14
Aymeric Histace -‐ HDR Defense
49
Long term
Research Project Cellular level TERAFS, iFib, FibroSES
ECSON, SIMBAD
Cyclope
Image processing
Fluorescence
RF Physiological data
Localization
NAUTILES (H2020)
SmartEEG NeuroCOM (ANR) SPOT (FUI)
CyberPill
Drug Treatment
Nanoparticles
Control
28/11/14
NAUTILES (H2020)
Aymeric Histace -‐ HDR Defense
50
THANKS TO THEM Q. Angermann (2014-…), C. Azib (2013), E. Bonnefoye (2012), M. Breuilly (2009), N. Cazin (2014), M. Degaudez (2007), M.-C. Desseroit (2013), H. Diouane (2014), C. Fouquet (2011-2014), M. Garnier (2009, 2011), C. Georgel (2013), M. Ibouchichene (2014), A. Izard (2013), L.-A. Latchimy (2013), T. Longret (2013), L. Meziou (2010-2013), S. Mouzai (2014), C. Nzuzi (2013), M. Rémignon (2014), A. Riaz (2013), H. Saidi (2014), J. Silva-Quintero (2012), A. Sittadannam (2014), Y. Zhang (2009-2010).
28/11/14
Aymeric Histace -‐ HDR Defense
51
THANK YOU FOR YOUR ATTENTION
28/11/14
Aymeric Histace -‐ HDR Defense
52
28/11/14
Aymeric Histace -‐ HDR Defense
53
Research Project From Nano to Macro
Mathematics
Biology
Imaging Technic
Signal and Image Processing
Electronic
To understand manifestations of a same pathologic phenomenon from cell to organ scales (even nano) and draw some possible connections 28/11/14
Aymeric Histace -‐ HDR Defense
54
Synthesis
•
In charge of the SES axis of ASTRE team
•
Elected member of the Research Council of ETIS
•
8 research projects
•
Co-head of MSc MADOCS
•
Active inter. and Nat. Collaborations
•
Different types of experiences (IUT, UCP)
•
8 Journal papers, 28 Int. Conf.
•
2 defended PhD, 2 on-going
2006-2014
•
L1, L2, L3, MSc, ENSEA
•
Interaction Teaching/ Research
•
In charge of the industrial relationship (2008-2014)
28/11/14
And what about the future now ? Aymeric Histace -‐ HDR Defense
55
Snapshot
Today-‐Teaching • Neuville In+tute of Technology, UCP, ENSEA Students
Classes
Lecture and Tuto (h) Labs (h)
1st year (L1)
Electrical Engineering and Energy
32
2nd year (L2)
Electrical Engineering and Energy
48
2nd year (L2)
Control Theory
28
2nd year App. (L2)
Control Theory
54
Prof. Licence (L3)
Electrical Engineering and Energy
30
MSc SIC
Image Processing
15
MSc MADOCS
Sta8s8cal Learning
6
ENSEA 3rd year (EIB)
Signal Processing for ECG
6
Total
28/11/14
171
Aymeric Histace -‐ HDR Defense
32
80
56
Snapshot
Today-‐Supervising
Synthesis
3 PhD students: 120%+50% • 2 Defended (11/2013, 06/2014) • 1 just started (50%)
17%
25% PhD MSc
7 MSc students: 330% • MSc SIC, ESA
Others 58%
Others: • 1 Post-doc (50%-1 year) • 1 Research Engineer (50%) •
28/11/14
Several Initiation-to-Research projects
Aymeric Histace -‐ HDR Defense
57
In Details: Stochas8c Resonance
Useful Noise Effect
Theory (particular case)
gη (u) = g(u + η (x))
•
gη : a static or memory less non-linear function
with
•
Shaping by noise of the input-output characteristic:
⎧1 if s ≥ k g ( s) = ⎨ ⎩0 if s k
+∞
geff (s) = E !"g ( s + η (x, y))#$ =
η
−∞
Hard-threshold
Experiment 1
∫ g(u) f (u − s) du
Theory
fη(u): pdf of purposely injected noise (uniform)
Real
Comparison between real stochastic diffusion process and theoretical one.
28/11/14
Aymeric Histace -‐ HDR Defense
58
In Details: Stochas8c Resonance
Useful Noise Effect
Theory (particular case)
gη (u) = g(u + η (x))
•
gη : a static or memory less non-linear function
with
•
Shaping by noise of the input-output characteristic:
⎧1 if s ≥ k g ( s) = ⎨ ⎩0 if s k
+∞
geff (s) = E !"g ( s + η (x, y))#$ =
∫ g(u) f (u − s) du η
−∞
Hard-threshold
Experiment 2
Theory
Real
Comparison of the evolution of normalized cross-covariance (1000 attempts)
28/11/14
Aymeric Histace -‐ HDR Defense
59
In Details: Stochas8c Resonance
Useful Noise Effect
Result (quantification) Texture ROI
Normalized Cross Covariance wrt the iteration number of the diffusive process
Gaussian
28/11/14
Multiplicative
Impulsive
Aymeric Histace -‐ HDR Defense
PM S-PM
60
In Details: Stochas8c Resonance
Useful Noise Effect
Extension to Image Restoration
# I(x, y, 0) = I 0 % $ ∂I %& = div gη ( ∇I ) ∇I ∂t
(
with and
)
gη (u ) = g (u + η ( x, y )) g (u ) = e
−
u
2
k2
(a) Additive zero-mean Gaussian noise with ψ0 = ψori + ξ (b) Multiplicative Gaussian noise of mean unity with ψ0 = ψori + ξ . ψori (c) Impulsive noise (d) Original image
28/11/14
Aymeric Histace -‐ HDR Defense
61
In Details: Ac8ve Contour
Shape Prior
Shape Space of Legendre Moments Legendre Moments
Reconstruction p + q ≤ N0
Shape Ω
λqp
1 = Ω
∑ p ,q
λ pq Lpq ( x, y )
∫ L pq (x, y, Ω)dxdy
Ω
with
N0 = 5
N 0 = 20
N 0 = 40
⎛ x − x ⎞ ⎛ y − y ⎞ ⎟ L ⎜ ⎟ L pq (x, y, Ω ) = L p ⎜⎜ q ⎜ ⎟ 1 2 1 2 ⎜ Ω ⎟ ⎜ Ω ⎟⎟ ⎝ ⎠ ⎝ ⎠
Ln (x ) =
n ⎤ 2n + 1 1 d n ⎡ 2 x − 1 ⎥⎦ 2 2 n n! dx n ⎢⎣
28/11/14
(
)
Images reconstructed from the Legendre moments for different moments’ order N0
Aymeric Histace -‐ HDR Defense
62
In Details: Ac8ve Contour
Shape Prior
Shape Space of Legendre Moments Greedy algorithm
How it works Starting shape
•Ω( k ) → Ωʹ′( k ) 2 2 ∂ϕ = ( I − µ Ω ) + ( I − µ Ω c ) ∇ϕ ∂t ⎛ ∇ϕ ⎞ + γ∇ ⎜⎜ ⎟⎟ ∇ϕ ⎝ ∇ϕ ⎠
)
(
Final Result
k k •Ωʹ′( ) → λ (r )
λ (rk ) = PT ( λ r − λ )
Eimage (Chan Vese)
•λ (rk ) → λ ʹ′r( k ) λ ʹ′r( k ) = λ (rk ) − β
NS
p ( λ r ) = ∑ N ( λ r ; λ r ,i , σ 2 ) i =1
∂E prior ∂λ r
k λ r = λ (r )
•λ ʹ′r( k ) → Ω( k +1)
Iterations shown in the feature space spanned by the first two principal axes
Ω( k +1) = ⎛ p + q ≤ N0 ( k ) ⎞ ⎪⎧ ⎪⎫ ʹ′ L pq x, y, Ωʹ′( k ) ⎟ > 0.5⎬ ⎨( x, y ) : ⎜ ∑ λ pq ⎪⎩ ⎪⎭ ⎝ p ,q ⎠
(
28/11/14
Different from Template Matching
)
Aymeric Histace -‐ HDR Defense
63
In Details: Ac8ve Contour
Shape Prior
Results 2
28/11/14
Aymeric Histace -‐ HDR Defense
64
In Details: Ac8ve Contour
Alpha-‐divergence
Result 1 (Synthetic images)
28/11/14
Aymeric Histace -‐ HDR Defense
65
2011-‐>2014: The Cyclope Project
2D Detec8on Results and Performance Bernal 88% 91%
Our Approach 99%
95%
100% 80% 60% 40% 1%
20%
5%
0% Sensibility
28/11/14
Specificity
False Positive Rate
Aymeric Histace -‐ HDR Defense
66
Important Dates
From 2006 to 2014 09/2006
11/2008
10/2010
09/2011
11/2013
10/2014
PES
Associate Professor
SIMBAD Activity
First PhD Student
• ETIS Lab- ICI Team • Collaboration with F. • Preciosio (MIDI • Cergy-Pontoise team) • Institute of Technology • Beginning of the collaboration with the • GEII Department University of Central Lancashire, UK
Cyclope Project
L. Meziou’s PhD Defense
Member of ASTRE Team
• Embedded Image Processing for • 1 Journal Paper • 3rd PhD Student ``Histogram-Based Smart • 8 Int. Conf. (1 • ``Smart Active Contour Videocapsule Best Student videoendoscopy Using AlphaPaper Award) for CRC early Divergence: detection’’ Application to • 1 Nat. Conf. Medical Image (GRETSI) • Co-supervised Segmentation’’ with Olivier • ASTRE • 1 GDR Romain • Co-supervised with Collaboration presentation F. Precioso
28/11/14
Leila Meziou
Aymeric Histace -‐ HDR Defense
67
Who am I?
Research/Teaching Computer-Aided Diagnosis 2014