Presentation is loading. Please wait.

Presentation is loading. Please wait.

Segmentation of 3D Tubular Structures Paul Hernandez-Herrera Computational Biomedicine Lab Advisor: Ioannis A. Kakadiaris and Manos Papadakis 1.

Similar presentations


Presentation on theme: "Segmentation of 3D Tubular Structures Paul Hernandez-Herrera Computational Biomedicine Lab Advisor: Ioannis A. Kakadiaris and Manos Papadakis 1."— Presentation transcript:

1 Segmentation of 3D Tubular Structures Paul Hernandez-Herrera Computational Biomedicine Lab Advisor: Ioannis A. Kakadiaris and Manos Papadakis 1

2 Motivation Tubular structures appear in biomedical images – Neuron – Vessels – Coronary arteries – Airways 2

3 Challenges 3 Size

4 Intensity 4 Challenges

5 Noise 5 Challenges

6 Contrast 6 Challenges

7 1.Develop a binary segmentation algorithm able to – handle different sizes – work with any acquisition modality – deal with noise in the image – handle anisotropic images – do a fast segmentation – have minimum or null user interaction 7 Thesis Objectives

8 8 2.Develop a centerline algorithm able to – Correctly extract the morphology Handle overlapping structures connect gaps – Fast extraction

9 9 Pipeline Input: 3D image stack Radius Step 1: Background voxels detection Step 2: Feature extraction Step 3: Background enhancement Step 4: Segmentation

10 10 Segmentation as one-class classification Input: 3D image Radius Detect voxels in background Voxels with unknown label Train a model (Cost function) Feature vectors Get cost value Accepted as Background Rejected as Background These are foreground voxels

11 11 Step1: Background voxel detection Compute the Laplacian of the 3D image The output has the following properties 1.Negative values in the foreground 2.Value close to zero in the boundary 3.It is positive near but outside the TS 4.Ringing (positive and negative) in the background

12 12 Step 2: Feature extraction Feature vector  Eigenvalues of Hessian matrix

13 13 Step 3: Cost function Approximate feature vectors distribution for background voxels Normalize the distribution Smooth the normalized distribution

14 14 Step 3: Background enhancement Input imageEnhanced image

15 15 Step 4: Segmentation Enhanced imageSegmentation

16 16 Results: Multiphoton Input Segmentation

17 17 Results: Confocal Input Segmentation

18 18 Results: Brain vessels Input Segmentation

19 19 Ongoing work Automatic radius estimation Allow the proposed method to handle any number of features Centerline extraction

20 20 Thanks Thanks for your attention QUESTIONS?


Download ppt "Segmentation of 3D Tubular Structures Paul Hernandez-Herrera Computational Biomedicine Lab Advisor: Ioannis A. Kakadiaris and Manos Papadakis 1."

Similar presentations


Ads by Google