Segmentation of the left atrial appendage from 3D images Pol Grasland-Mongrain 20/04/2009 – 28/08/2009

Views of the Left Atrial Appendage

Views of the Left Atrial Appendage • • • •

Variable shapes 1 to 19 cm3 Function ? Has to be ablated sometimes

Motivation

Current implementation: • LAA is represented as a short trunk in the model • Current framework not flexible enough to grow into highly variable shape

Motivation

Motivation of my master thesis: • Addition of an Automatic Segmentation Algorithm of the Left Atrial Appendage in the Philips Framework

Plan I. Current Method at Philips II. Actual Work III. Results and Future Work

Philips Aachen method

Segmentation Chain

New Image

1. Heart

Detection

2. Parametric

Adaptation (Similarity)

3. Parametric Adaptation (Piecewise Affine)

Segmented Image

4. Deformable Adaptation

Philips Aachen method

Segmentation Chain

New Image

1. Heart

Detection

2. Parametric

Adaptation (Similarity)

3. Parametric Adaptation (Piecewise Affine)

vadap = T[v] E = Eext[T]

Segmented Image

4. Deformable Adaptation

Parametric Adaptation, Deformable Models • Use External Energy :

Philips Aachen method

Segmentation Chain

New Image

1. Heart

Detection

2. Parametric

Adaptation (Similarity)

3. Parametric Adaptation (Piecewise Affine)

vadap = T[v] E = Eext[T]

Segmented Image

4. Deformable Adaptation

Free motion for vadap E = Eext[v]+ α Eint[v]

Deformable Models • Internal Energy

Plan I. Current Method at Philips II. Actual Work 1. Segment manually 17 patients LAA 2. Modify Philips models 1. 2.

Interface Left Atrium - Left Atrium Appendage Mesh which inflate

3. Code an automatic mesh-inflation algorithm 1. 2. 3.

External Energy Threshold between LAA – Background Internal Energy

III. Results and Future Work

Plan I. Current Method at Philips II. Actual Work 1. Segment manually 17 patients LAA 2. Modify Philips models 1. 2.

Interface Left Atrium - Left Atrium Appendage Mesh which inflate

3. Code an automatic mesh-inflation algorithm 1. 2. 3.

External Energy Threshold between LAA – Background Internal Energy

III. Results and Future Work

Plan I. Current Method at Philips II. Actual Work 1. Segment manually 17 patients LAA 2. Modify Philips model 1. 2.

Interface Left Atrium - Left Atrium Appendage Mesh which inflate

3. Code an automatic mesh-inflation algorithm 1. 2. 3.

External Energy Threshold between LAA – Background Internal Energy

III. Results and Future Work

Model Modification

Plan I. Current Method at Philips II. Actual Work 1. Segment manually 17 patients LAA 2. Modify Philips model 1. 2.

Interface Left Atrium - Left Atrium Appendage Mesh which inflate

3. Code an automatic mesh-inflation algorithm 1. 2. 3.

External Energy Threshold between LAA – Background Internal Energy

III. Results and Future Work

External and Internal Energies Internal Energy

External Energy Edgebased

Regionbased

Mesh Reference

Triangle Regularization

Curvature

N-Gon Regularization

External Energy : Edge-Based • No specific features

External Energy : Region-Based •Gray Value Above or Under ?

External Energy : Region-Growing •Gray Value Still Above (Under) ? •Already Annotated ?

External Energy : Region-Growing •GrayValue ValueStill StillAbove Above(Under) (Under)?? Gray •AlreadyAnnotated Annotated?? Already

External Energy : Region-Growing •GrayValue ValueStill StillAbove Above(Under) (Under)?? Gray •AlreadyAnnotated Annotated?? Already

Threshold LAA-Myocardium

Threshold LAA-Myocardium

Threshold LAA-Myocardium

o Minimization of classification error o Stop when Area1 = Area2

Internal Energy : Mesh Reference • Updated Mesh

Internal Energy : Triangle Regularization • Approximate each triangle by a rotated and scaled equilateral triangle

Internal Energy : Triangle Regularization • Approximate each triangle by a rotated and scaled equilateral triangle

Internal Energy : Curvature • Remove the peaks

Internal Energy : Curvature • Remove the peaks

Internal Energy : N-Gon Regularization • Approximate each “N-Gon” by a rotated and scaled regular N-Gon

External and Internal Energies Internal Energy

External Energy Edgebased

Regionbased

Mesh Reference

Triangle Regularization

Curvature

N-Gon Regularization

Plan I. Current Method at Philips II. Actual Work III. Results and Future Work

Results

Results • Main problem : loops -> repair

Results

Specificity = True Pos. / (True Pos. + False Neg.) Quality = True Pos. / (True Pos. + False Pos.)

Results Left Atrial Appendage Inflation Results Specificity = True Pos. / (True Pos. + False Neg.)

Quality = True Pos. / (True Pos. + False Pos.)

100

80

60

40

20

0

Sum up: • almost all segmented voxels really belong to LAA • but the mesh doesn’t inflate enough

Results

(1)

(5)

(10)

(7)

Majors Fails

(11)

(14)

Possible future works • Improve the loop repair : – Freeze vertices – Better correction

• Find a new internal energy ?

Thank you for your attention ! Any Questions ?