Retinal Image Fusion and Registration

1
Retinal Image Fusion and Registration

• Libor Kubecka, Jiri Jan
• Department of Biomedical Engineering, FEEC,
  Brno University of Technology, Czech Republic

• 2. Image Fusion
• via discrete wavelet decomposition (DWT)
• fusion decision map orthogonalisation by
  computation of maximal direction of linear
  operator L(I) (e.g. gradient of the vector-valued
  image I) from the norm of the first differential
  of L(I)
• Final orthogonalised value of the linear
  operator (given wavelet), l highest eigenvalue
  of matrix G, q its relevant eigenvector

Introduction HRT (Heildelberg Retina Thomograph)
is widely used for imaging and following
examination of the shape of the optic nerve head
(ONH) and therefore for diagnosis glaucoma. For
this purpose, correct segmentation of ONH is
necessary but when manually performed also highly
subjective and time consuming task.
Unfortunately, current automatic segmentation
algorithms are not sufficiently robust especially
because the information of the ONH contour is
sometimes missing in HRT images. Therefore, we
hope that fusion of HRT image and colour fundus
photograph will provide additional useable
information. Here, the algorithms of image
registration and fusion are described and results
of ONH segmentation in fused images are presented.

• 1. Image Registration
• Type 2-dimensional, bi-modal, multi-resolutional,
  performed by optimization of global similarity
  metrics.
• Spatial transformation model affine,
  perspective.
• Criterion mutual information.
• Optimizers controlled random search (CRS),
  Powell.
• Interpolation nearest neighbour, partial volume.
• Registration results

Fusion schema
Registration schema
3. Segmentation For the purpose of segmentation,
we modified Chrasteks method for the case of
fused image. This method consists of
morphological operations for image preprocessing,
Hough transform for circle fitting and anchored
active contour model for the final fine
segmentation of the optic nerve head.
Total number of images 334
Group I (precisly registered) 316
Group II (sligthly mis-registered) 13
Group III (mis-registered) 5
Sufficiently registered ( III) 329
Rate of succes 98.5
Average mark 0.13

• The qualitative evaluation of
  this step should be performed in the future.

4. Results We have successfully designed a
registration method making use of robust
optimization (controlled random search) of
modified mutual information similarity criterion.
Quality of the registration step has been
evaluated by human observer and reaches 98 of
successfully registered images. Further we
applied a method for fusion of registered images
based on wavelet transform and computation of
maximal length of gradient of vector image.
Finally, successful results of segmentation of
optic disc contour has been presented.
Mosaic (HRT, CFP)
HRT with overlaid edges from CFP
Acknowledgements Authors sincerely acknowledge
the contribution of Prof. G. Michelson,
Augen-Klinik Erlangen (Germany) who provided Kowa
fundus camera images, data from Heidelberg Retina
Tomograph and valuable consultations. Also Radim
Chrasteks contribution in optic disc
segmentation phase is highly acknowledged. This
project has been completed with support by the
grant FRVŠ 3118/2005 (Ministry of Education,
Czech republic) and also by the support of the
research centre DAR (Ministry of Education, Czech
Republic), proj. no. 1M6798555601.
EMBEC 2005, Prague, CZECH REPUBLIC

20.-25.11. 2005