MultiModal Retinal Imaging

1
Multi-Modal Retinal Imaging
Analysis and diagnosis of the Optic Nerve Head
Phil LeggSchool of Computer ScienceCardiff
University
2
Current Hypothesis

• To successfully register retinal images of
  different modalities to provide further analysis
  of the Optic Nerve Head that would aid an
  Ophthalmic clinician in the diagnosis of a
  patient.

3
Multi-Modal Registration
4
What's the problem?

• Images of different modalities may have little or
  no distinct intensity relationship.
• Established method, Mutual Information, fails for
  these modalities.
• Intensity relationship is too weak
• Many possible transformation to be considered.

5
Successful Registrations
6
Unsuccessful Registrations
7
Solution?

• Appearance may differ but structure remains the
  same.
• Use gauge co-ordinate feature derivatives to
  extract features of these images.
• Incorporate these images with Mutual Information
  to provide suitable registration.

8
Gauge Co-ordinates
Lw
Original
Scale 2
Lwww
LwL2w
Scale 4
Scale 2
9
Incorporate Features in MI

• Generate intensity matrix P m x n
• m number of feature images
• n number of overlapping pixels
• We find the co-variance matrix to show
  relationship between intensities and reduce
  complexity.
• MI is determined by the entropy of the
  co-variance matrix.

10
Results
11
Progress

• Our registration method performs well we wish
  to test on more data sets though to clarify
  this.
• Registration takes approx. 2 minutes can this
  be improved upon?
• Automated analysis of images we wish to segment
  image to extract Optic Disc boundaries.

12
Multi-Modal Segmentation

• Uni-modal versus Joint Segmentation for Region
  Based Image Fusion J. J. Lewis et al.
• A comparison of segmentation techniques for
  multi-modal data sets.
• Should we segment individually and then merge?
• Should we create a joint image and segment this
  instead?

13
Multi-Modal Segmentation
Uni-modal IR Image
Uni-modal Visible Image
Uni-modal segmentations
Joint segmentation
14
How does it relate?

• Segmentation of multi-modal images to clearly
  define features within a scene.
• We wish to segment our images to analyse features
  of the Optic Disc.
• Important information represented in each
  modality.

15
How does it differ?

• Segmentation of different modalities
• (Visible and IR images examples shown are
  landscape images)
• Retinal images are much harder to clearly define
  features such as Optic Disc.
• Previously registered data sets.

16
How I will use this in my thesis?

• Segmentation is a wide area of Computer Vision
  that we are now investigating.
• This reference highlights the benefits of
  multi-modal data for segmentation.
• Potentially aim to extend on their work to
  perform segmentation of the Optic Disk.
• A difficult task in comparison to examples

17
Conclusion

• We have proposed a suitable scheme for performing
  registration of retinal image
• Need to test on other registration problems
• Do we need to train feature images again when
  using other modalities?
• Segmentation of Optic Nerve Head to provide
  structural analysis
• Multi-Modal approach exists in the literature
• Can this be adapted to perform Optic Disk
  extraction?