Soft Plaque Detection and Automatic Vessel Segmentation Georgia - PowerPoint PPT Presentation

Soft Plaque Detection and Automatic Vessel Segmentation Georgia Tech Shawn Lankton - Arthur Stillman - Emory University Paolo Raggi - Emory University Allen Tannenbaum - Georgia Tech and Emory PMMIA, MICCAI Workshop September 20, 2009

Vessel Analysis • Heart disease • Diagnosis and risk • Plaque detection /44 2

CTA Imagery • In-vivo, 3-D scan • X-ray attenuation • Contrast agent /44 3

Soft Plaque Detection • Coronary plaques • Dangerous • Hard to see • No good tools /44 4

Soft Plaque Detection • Coronary plaques • Dangerous • Hard to see • No good tools /44 4

Prior Work • Saur et. al. “... detection of calcified coronary plaques...” MICCAI. 2008 • Brunner et. al. “… classification of calcified arterial lesions.” MICCAI 2008 • Renard and Yang: “… detection of soft plaques …” {SSIAI,ISBI,ICIP} 2008 /44 5

Objective • From a simple initialization… • Segment the vessel… • Detect all soft plaques. /44 6

Some Definitions /44 7

Active Surfaces • Level set implementation • Iterative optimization • Simple, flexible, principled /44 8

Definitions • An Image I : R N → R on the domain Ω • A Surface Γ embedded in φ : R N → R • Such that Γ = { x ∈ Ω | φ ( x ) = 0 } Sethian. Level Set Methods and Fast Marching Methods. 1999 /44 9

Definitions  1 φ < − ǫ inside  0 φ > ǫ H φ = outside smooth otherwise   1 φ = 0 the surface  | φ | < ǫ δφ = 0 the rest smooth otherwise  /44 10

Localized Active Contour Model /44 11

Localizing Lankton and Tannenbaum “Localized Region-Based Active Contours,” TIP, 2008 /44 12

Localizing x Lankton and Tannenbaum “Localized Region-Based Active Contours,” TIP, 2008 /44 12

Localizing x B ( x, y ) Lankton and Tannenbaum “Localized Region-Based Active Contours,” TIP, 2008 /44 12

Localizing x B ( x, y ) B ( x, y ) · H φ ( y ) Lankton and Tannenbaum “Localized Region-Based Active Contours,” TIP, 2008 /44 12

Localizing x B ( x, y ) B ( x, y ) · H φ ( y ) B ( x, y ) · (1 − H φ ( y )) Lankton and Tannenbaum “Localized Region-Based Active Contours,” TIP, 2008 /44 12

Localizing x B ( x, y ) B ( x, y ) · H φ ( y ) B ( x, y ) · (1 − H φ ( y )) Lankton and Tannenbaum “Localized Region-Based Active Contours,” TIP, 2008 /44 12

Localizing x B ( x, y ) ≤ r . B ( x, y ) · H φ ( y ) B ( x, y ) · (1 − H φ ( y )) Lankton and Tannenbaum “Localized Region-Based Active Contours,” TIP, 2008 /44 12

Local Regions in 3-D (a) Example Surface (b) Local Region (c) Local Interior (d) Local Exterior /44 13

Localized Contours E ( φ ) = /44 14

Localized Contours � E ( φ ) = δφ ( x ) dy dx Ω x every point on the contour /44 14

Localized Contours � � E ( φ ) = δφ ( x ) B ( x, y ) dy dy dx Ω x Ω y all information in a local ball around x /44 14

Localized Contours � � E ( φ ) = B ( x, y ) · F ( I, φ , x, y ) dy δφ ( x ) dy dx Ω x Ω y compute an internal energy F /44 14

Localized Contours � � E ( φ ) = B ( x, y ) · F ( I, φ , x, y ) dy δφ ( x ) dy dx Ω x Ω y � + λ δφ ( x ) �∇ φ ( x ) � dx Ω x /44 14

Localized Contours � � E ( φ ) = B ( x, y ) · F ( I, φ , x, y ) dy δφ ( x ) dy dx Ω x Ω y � + λ δφ ( x ) �∇ φ ( x ) � dx Ω x � ∇ φ ( x ) � ∂φ � ( x ) = δφ ( x ) B ( x, y ) · ∇ φ ( y ) F ( I, φ , x, y ) dy + λδφ ( x ) div �∇ φ ( x ) � | ∇ φ ( x ) | ∂ t Ω y /44 14

Internal Energies • Local Uniform Modeling • Local Mean Separation /44 15

Localized Means � � � · y H φ ( y ) · I ( y ) dy Ω y � µ in ( ) = � � � y H φ ( y ) dy Ω y � � � · (1 − H φ ( y ) ) · I ( y ) dy − H · Ω y � µ out ( ) = � · (1 − H φ ( y ) ) dy Ω y � /44 16

Localized Means � � � · y B ( x, y ) · (1 y H φ ( y ) · I ( y ) dy Ω y � µ in ( in ( x ) = ) = � � � � y H φ ( y ) dy y B ( x, y ) · (1 Ω y � � � � · (1 − H φ ( y ) ) · I ( y ) dy y B ( x, y ) · (1 − H · Ω y � µ out ( out ( x ) = ) = � � y B ( x, y ) · (1 · (1 − H φ ( y ) ) dy Ω y � � /44 16

Local Vessel Segmentation /44 17

Vessel Segmentation • Simple initialization • No leaks • Branch handling • No shape information /44 18

Local Means /44 19

Local Uniform Modeling Energy Assumption: The foreground and background are approximately constant locally . F um = H φ ( y ) ( I ( y ) − µ in( x ) ) 2 + (1 − H φ ( y ) )( I ( y ) − µ out( x ) ) 2 /44 20

Why Uniform Modeling? • Enforce similarity • Surface expands quickly • Move to capture the “vessel wall” /44 21

Energy Minimization d φ ( x ) � �� � 2 � � 2 − � = δφ ( x ) B ( x, y ) · δφ ( y ) · I ( y ) − µ in ( x ) I ( y ) − µ out ( x ) dy dt ˜ Ω y � ∇ φ ( x ) � + λ div | ∇ φ ( x ) | | ∇ φ ( x ) | /44 22

Energy Minimization d φ ( x ) � �� � 2 � � 2 − � = δφ ( x ) B ( x, y ) · δφ ( y ) · I ( y ) − µ in ( x ) I ( y ) − µ out ( x ) dy dt ˜ Ω y � ∇ φ ( x ) � + λ div | ∇ φ ( x ) | | ∇ φ ( x ) | Domain restriction: ˜ Ω = Ω ∩ ( I < − 600 HU) /44 22

Vessel Segmentation LAD RCA /44 23

Vessel Segmentation LAD RCA /44 23

Parameters • Radius r = 5 λ = 0 . 1 max ( | d φ • Smoothness dt | ) /44 24

Soft Plaque Detection /44 25

Soft Plaque Detection • Two-front approach • Inside moves out • Outside moves in /44 26

Local Mean Separation Energy Assumption: The foreground and background are different locally . F ms = − ( µ in( x ) − µ out( x ) ) 2 Yezzi et al. “A Fully Global Approach to Image Segmentation... ,” JVCIR 2002 /44 27

Why Means Separation? • Enforces differences • Interior won’t grow out • Exterior won’t move in /44 28

Detection Energy � 2 � 2 �� � � I ( y ) − µ out ( x ) I ( y ) − µ in ( x ) d φ ( x ) � = B ( x, y ) · δφ ( y ) dy − A out ( x ) A in ( x ) dt ˜ Ω y � ∇ φ ( x ) � + λ div | ∇ φ ( x ) | | ∇ φ ( x ) | Clever initializations are required /44 29

Initial Surfaces � � � E shrink ( φ ) = ( B ( x, y ) · H φ ( y ) ) y ) ) dy dx + λ δφ ( x ) �∇ φ ( x ) � dx δφ ( x ) Ω x Ω y Ω x � dx + λ δφ ( x ) �∇ φ ( x ) � dx E grow ( φ ) = − H φ ( x ) Ω x /44 30

Initial Surfaces � � � E shrink ( φ ) = ( B ( x, y ) · H φ ( y ) ) y ) ) dy dx + λ δφ ( x ) �∇ φ ( x ) � dx δφ ( x ) Ω x Ω y Ω x � dx + λ δφ ( x ) �∇ φ ( x ) � dx E grow ( φ ) = − H φ ( x ) Ω x /44 30

Steps for Detection • Segment the Vessel • Create Initialization • Run Local Mean Separation • Check for Differences /44 31

3-D Example /44 32

3-D Example /44 32

3-D Example /44 32

3-D Example /44 32

3-D Example /44 32

Detection Results /44 33

Coronary Anatomy • Coronaries • RCA • LAD • LCX /44 34

Coronary Anatomy • Coronaries • RCA • LAD • LCX /44 34

Coronary Anatomy • Coronaries • RCA • LAD • LCX /44 34

Coronary Anatomy • Coronaries • RCA • LAD • LCX /44 34

2-D Results (a) Initial Surfaces (b) Result of Evolution (c) Expert Marking (d) Detected Plaque /44 35

2-D Results (a) Initial Surfaces (b) Result of Evolution (c) Expert Marking (d) Detected Plaque /44 36

3-D Results (LCX) /44 37

3-D Results (LAD) /44 38

3-D Results (RCA) /44 39

3-D Results (RCA) /44 40

Results Summary Table 5.1: Results of soft plaque detection in Figures 5.8 and 5.9. Plaque ID Remodeling Vessel Segment Confirmed Detected #1 LAD negative × #2 LAD positive × × #3 LAD positive × × #4 LCX negative × × #5 LCX positive × × #6 RCA negative × × #7 RCA negative × × #8 RCA positive × × /44 41

Summary & Conclusions • Localized analysis makes sense • Vessel segmentations are satisfactory • Detection identifies 87.5% of plaques /44 42

Continuing Research • Coupling contour evolution • Acquiring additional data • Performing more tests /44 43

Thank You. /44 44