Tools for Shape Analysis of Vascular Response using Two Photon Laser Scanning Microscopy

1. Tools for Shape Analysis of Vascular Response using Two Photon Laser Scanning Microscopy By Han van Triest Committee: Prof. Dr. Ir. B.M. ter Haar RomenyDr. M. A. M. J. van ZandvoortDr. Ir. H. C. van Assen A. Vilanova i Bartrolí, PhD R.T.A. Megens, MSc

2. Overview • Biological Introduction • Technical Introduction • Vessel Radius Estimation • Cell counting • Conclusion and Recommendations

3. Biological Introduction Vascular diseases are a big problem in the western world. It is estimated that arteriosclerosis is the underlying cause of 50 % of all deaths in the western world To unravel the underlying mechanisms more research is required

4. Biological Introduction – Vessel Anatomy

5. Biological Introduction – Remodelling

6. Technical Introduction – Fluorescence • Excitation • Energy loss processes • Emission Energy of Photon: 2 1 3

7. Technical Introduction – Confocal Laser Scanning Microscopy

8. Technical Introduction – Confocal Laser Scanning Microscopy Advantages: • Optical sectioning Disadvantages: • Excitation of out-of-focus regions • High energy of excitation photons • Low penetration depth

9. Technical Introduction – Two Photon Laser Scanning Microscopy

10. Technical Introduction – Two Photon Laser Scanning Microscopy

11. Technical Introduction – Two Photon Laser Scanning Microscopy Advantages: • No pinhole to block out-of-focus light required • Increased penetration depth • Excitation photons of lower energy • Imaging of viable tissue • Multiple dyes usable for targeting of different structures Disadvantage: • Higher wavelength limits maximal achievable resolution

12. Technical Introduction – Imaging

13. Processing – Description of Vessels Features: • Radius of the vessel • Ratio vessel wall thickness – vessel radius • Cell volume fraction Needed: • Vessel Radius • Number of Cells

14. Radius Estimation

15. Radius Estimation – Methods • Statistical methods: Least squares estimators • Robust statistics: Reduction of the influence of outliers • Hough Transform

16. Radius Estimation – Hough Transform Line through a point in image space Set of parameters that describe the point

17. Radius Estimation – Hough Transform

18. Radius Estimation – Circular Hough Transform Circle can be described by:

19. Radius Estimation – Hough Transform Advantages: • Robust against noise • Able to find partly occluded objects Disadvantages: • Expensive, both computational and memory cost

20. Radius Estimation – Proposed method A circle is defined by three non co-linear points. • Store only center coordinates • Weight vote by average distance between p1, p2 and p3 Find r by voting for most likely value of the radius

21. Radius Estimation – Finding Edge Points A global threshold is infeasible due to differences in optical paths for emitted photons

22. Radius Estimation – Finding Edge Points Modified Full Width at Half Maximum: Outside Inside

23. Radius Estimation – Experiments • 20 images, 10 single slices, 10 taken from three dimensional stacks • Test images have both sides of the wall vissible • Groundtruth given by the average estimate of 12 volunteers • Results compared with common least squares estimator • Tests are performed for values of α between 0.2 and 0.8 in steps of 0.05, and using 20 to 250 points in steps of 10 • In total 24960 estimates are made

24. Radius Estimation – Influence of α xz-scan: z-stack slice: Blue line: LSE Red line: MHT

25. Radius Estimation – Influence of number of points xz-scan: z-stack slice: Blue line: LSE Red line: MHT

26. Radius Estimation – Conclusion • Proposed method outperforms least squares fitting method for xz-scans • Proposed method performs equally compared to least squares fitting method for z-stack slices • The best value for α used in the proposed method is α = 0.4 • At least 100 points is required for a stable result using the proposed method

27. Cell counting

28. Cell Counting – Algorithm Noise Reduction Potential Center Extraction Potential Edgepoint Extraction Edgepoint Selection Ellipsoid Fitting Oversegmentation Reduction

29. Cell Counting – Noise Reduction Edge-preserving filtering: Median Filtering Each pixel is replaced by the median of its surrounding Purple line: Original object Blue line: Degraded object Red line: Median filter, kernel width 5 pixels Black line: Median filter, kernel width 25 pixels

30. Cell Counting – Potential Center Detection Assumption: Blob-like structures Center is maximum of the blob Local maxima within a region are potential centers.

31. Cell Counting – Potential Edgepoint Extraction • Sample rays from each potential center • Rays intersect points along a generalized spiral

32. Cell Counting – Potential Edgepoint Extraction Constraint: Points on a downward flank These points can be found at points in which the second order derivative switches from negative to positive. Blue line: Image intensity along ray Purple line: First order derivative Sienna line: Second order derivative

33. 2 5 9 7 4 Cell Counting – Dynamic Programming 3 a c 2 6 A B 2 5 2 b d 4 6 Shortest Route: AbcB

34. Cell Counting – Edgepoint Selection Find set of most likely edge points Cost function:

35. Axes proportions Orientation Position Size Cell Counting – Ellipsoid Fitting Ellipsoid can be described by a quadric, a general polynomial in three dimensions of order two: Fitted on the data using a least squares fitting procedure

36. Cell Counting – Oversegmentation Reduction • Find overlapping nuclei • Check wether nuclei are parallel • Merge the sets of edgepoints of parallel overlapping nuclei • Perform Ellipsoid fitting on the combined data sets

37. Cell Counting – Results Before Merging

38. Cell Counting – Results After Merging

39. Cell Counting – Discussion Three types of frequent mistakes: A Incorrect merging of two blunt nuclei B Center of cell not found C No distinct directions

40. Cell Counting – Discussion A another problem is due to leakage of light from other colors

41. Cell Counting – Conclusion Although the method only has been tested on a single dataset, the results show to be promising. Most of the cells are found while there is a relatively small amount of false negatives and false positives

42. Recommendations • Test the algorithm on more datasets • Investigate the influence of parameters • For the calculation of the cost during the dynamic programming step, take into account more points on the surface • Remove outliers in the selected set, as outliers have great effect on the least squares algorithm • Optimize the imaging parameters to get as litle non cellular structures as possible • Classify the cells into subclasses

43. Questions?