1. Vision and Image Processing Group University of Waterloo Justin Eichel, Akshaya Mishra, Paul Fieguth, David Clausi, Kostadinka Bizheva

2.  UHROCT  ultra high resolution optical coherence tomography  47,000 A-scans/s  3um x 10um (axial x lateral) resolution  Dataset  Corneal hypoxia study  2 healthy subjects  Contact inducted

3.  Issues  Low contrast  Noise

4.  Issues  Low contrast  Noise Stroma Bowman’s membrane Epithelium Endothelium Descemet’s membrane

5.  Artifacts  Eye lashes

6.  Artifacts  Eye lashes ▪ Different Eyelashes

7.  Artifacts  Eye lashes ▪ Different Eyelashes ▪ Timing

8.  Artifacts  Eye lashes ▪ Different Eyelashes ▪ Timing  Lower Contrast

9.  Active Contours  Designed to engulf an object  Gradient information  Parametric Active Contours  Geometric Active Contours  Edge-free Active Contours

10.  Active Contours  Designed to engulf an object  Gradient information  Parametric Active Contours  Geometric Active Contours  Edge-free Active Contours

11.  Failed  Noisy image  Noisy image gradient

12.  Intelligent Scissors (Mortenson et al, 1995)  User guided boundary identification  Noisy gradient  Discontinuities due to imaging artifacts

13.  Intelligent Scissors (Mortenson et al, 1995)  User guided boundary identification  Noisy gradient  Discontinuities due to imaging artifacts  Unfair example?

14. Proposed MethodIntelligent Scissors Few discontinuities

15. Proposed MethodIntelligent Scissors With artifact

16. Proposed MethodIntelligent Scissors Low contrast image

17. Proposed MethodIntelligent Scissors Well conditioned image

18.  Enhanced Intelligent Scissors (Mishra et al, 2008)  Better than Intelligent Scissors  User guided boundary identification  Noisy gradient  Discontinuities due to imaging artifacts  Better for upper and lower curves  Still not great for inner curves

19. SourceEnhanced Intelligent Scissors Few discontinuities

20. With artifact Enhanced Intelligent ScissorsSource

21. Low contrast image Enhanced Intelligent ScissorsSource

22. Well conditioned image Enhanced Intelligent ScissorsSource

23.  Semi-automated boundary identification  Identify high contrast outer boundaries  Develop model of cornea  Parameter estimation  Local optimization

24. Well conditioned image Close up of source

25.  Preprocessing  Create a smooth gradient  Morphological operators ▪ Set of structuring elements to enhance the arch ▪ Creates higher contrast upper and lower curves  Blur to reduce noise

26. SourcePreprocessed image Many discontinuities

27. SourcePreprocessed image Few discontinuities

28. With artifact SourcePreprocessed image

29. Low contrast image SourcePreprocessed image

30. Well conditioned image SourcePreprocessed image

31.  User input  Enhanced Intelligent Scissors  2 points on upper curve  2 points on lower curve User input

32.  User input  Enhanced Intelligent Scissors  2 points on upper curve  2 points on lower curve  Fit data to polynomial  >250 data points  4 th order polynomial  filters “sloppy input” Polynomial fitting

33.  User input  Enhanced Intelligent Scissors  2 points on upper curve  2 points on lower curve  Fit data to polynomial  >250 data points  4 th order polynomial  filters “sloppy input” Polynomial fitting

34.  Corneal Model  Shortest distance between curves ▪ medial axis transform  Define alpha, s, theta, and Omega  Let’s have a closer look

35. Source

36.  Parameter Estimation  Find inner curves  Modify alpha and theta to generate search path Omega  Look at points in the neighborhood of the path

38.  Parameter Estimation  False peaks  Use a prior knowledge  Gaussian mixture model  Use statistics from datasets alpha01

39.  Parameter Estimation  Select path with largest difference in intensity  Keep corresponding values of alpha and theta  Future work  Currently only focusing on alpha

42.  Fully Automated Method  Local optimization  Use model to provide initial values for local optimization  3D reconstruction