COMBINATION OF SHAPE-CONSTRAINED AND INFLATION DEFORMABLE MODELS, WITH APPLICATION TO THE SEGMENTATION OF THE LEFT ATRIAL APPENDAGE 1
Pol Grasland-Mongrain, Jochen Peters, Olivier Ecabert
2
[1]
[email protected] Philips Research Europe–Aachen Weisshausstr. 2, 52066 Aachen, Germany [2]
[email protected]
Goal
Conclusion
Develop an automatic method to segment highly deformable structures like the Left Atrial Appendage (LAA)
General algorithm combining two energies: - one external, to make the LAA to grow, - one internal, to preserve the shape of the LAA, with a protection against loops. It could be applied to other complex structures.
(1) What is the left atrial appendage ?
(4) Loops : an annoying problem
Fig 4.1. Apparition of loop under the mesh
Loops may appear during adaptation. => intersecting triangles are smoothed with the internal energy
Characteristics of the Left Atrial Appendage (LAA): - linked to the Left Atrium (LA); - sizes from 1 to 19 cm3; - highly variable shape, often tubular and hooked
(5) Example of mesh adaptation 2D CT slice (Z axis)
3D view of the mesh
Fig 5.1. Initial state
(2) Inclusion in segmentation algorithm New Image 1. Heart Detection
Segmentation Chain 2. Parametric Adaptation (Similarity)
3. Parametric Adaptation (Piecewise Affine)
Fig 5.2. Five iterations further, α = 0.2
Segmented Image
4. Deformable Adaptation
5. LAA Inflation
Fig 5.3. Five iterations further, α=1
Included in a global heart-segmentation C++ algorithm => base location of the LAA is known; => surroundings substructures are already segmented. Fig 5.4. Five iterations further, α=2
(3) External and internal energies combination Fig 5.5. Five iterations further, α=5
3.1. Combination of two antagonist energies:
Etotal = Einternal + α Eexternal (with α ponderation factor) Fig 5.6. Five iterations further, α = 10
3.2. Internal Energy: - At each step, penalizes all moves of the triangle vertices. - The further, the bigger is the penalization.
(6) Qualitative and quantitative results 100% 80
3.3. External Energy: - At each step, moves all triangles center to target points. - Target points are taken between candidate points with: - correct gray value: under (resp. above) threshold if triangle center is under (resp. above) threshold; - no previous segmentation (= don't belong to other structures. - Among the candidate points, the farthest of the triangle center, in the direction of the interface is taken as the target point.
60 40 20 0 1
2
3
4
5
6
7 8 9 10 11 12 13 14 15 16 17 Patient n° Fig 6.1 Sensitivity and Positive Prediction Value (PPV) of the inflated mesh Patient Number compared to manually drawn groundtruth for 17 patients
Sensitivity
true positive = true positive + false negative true positive
PPV = true positive + false positive
Triangle center Target point
- Difficulties to reach the tip of the LAA (low sensitivity); - Very few segmentation errors, with a good adaptation to the shape of the LAA (high PPV); - Failures (patient 2,3,4, and 14) are due to small inaccuracies during the first segmentation phases occurring near the LAA base.
Acknowledgments: The research leading to these results has received funding from the European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreement n. 224495 (euHeart project). We would like to thank H. Lehmann and R. Kneser from Philips Research Europe–Aachen for their support. In addition, we thank P. Cignoni and his team for creating the MeshLab software. Finally, we thank H. Delingette for creating the connection between us. This presentation has been supported by the Institut Langevin and Philips Research Europe