ULTRA-FAST LOCALIZATION OF THE OPTIC DISC IN RETINAL

FUNDUS IMAGES

Ahmed E. Mahfouz

Dr. Ahmed S. Fahmy

Objectives

Project Goal:

Automatic Quantitative Analysis of Retinal Images

To achieve this goal, localization of retinal anatomical structures is required as an initial step

In this work, we propose an ultrafast localization of the Optic Disc (OD)

Optic Disc (OD) in Retina Images

Colored Retina Fundus Image, STARE Database

Optic Disc (OD)Fovea

Vessels Tree

Importance of OD localization

Localization of the Optic Disc (OD) is useful for:1. OD segmentation2. Locating other structures (e.g. fovea)3. Retinal image registration4. Identification of left & right eyes5. Tracking retinal vessels

OD Localization: Simple Threshold

Retina Fundus Image, STARE Database

Retina Fundus Image, STARE Database

OD Localization: Simple Threshold

OD Localization: sophisticated methods

Several OD localization methods are available:1. Vasculature convergence

A. Hoover et al. (2003), A. Youssif et al. (2008)

2. Geometrical model fitting M. Foracchia et al. (2004)

3. Template matching M. Lalonde et al. (2003)

4. Hough Transform F. ter Haar (2005)

Vasculature Convergence

1Aliaa Youssif, Atef Ghalwash, and Amr Ghoneim "Optic Disc Detection from Normalized digital Fundus Images by means of a Vessels' Direction Matched Filter" IEEE Trans. Medical Imaging, vol.27, no.1, pp. 11–18, Jan. 2008.

13.5 min

Original Image Contrast Enhancement

Histogram Equalization

Vessels ExtractionThinningFinal Result

Geometric Model Fitting

22 minOriginal Image Contrast

EnhancementHistogram Equalization

Vessels ExtractionModel FittingFinal Result2M. Foracchia, E. Grisan, A. Ruggeri, "Detection of Optic Disc in Retinal Images by Means of a Geometrical Model of Vessel Structure," IEEE Trans. Medical Imaging, vol. 23, no. 10, pp. 1189 – 1195, Oct. 2004.

Motivation

Although available techniques have high success rate, they are impractical due to their long computation times. OD localization is a preprocessing step that provides a seed

for further processing steps (segmentation, registration, …etc)

New fast localization techniques are required

Ultra-Fast Localization

Search Space Dimensionality Reduction

×

Ultra-Fast Localization

Horizontal Localization

Projection of the image features on the horizontal-axis

Projected Features:1) Difference

between vertical and horizontal edges (MAX)

2) Pixels’ intensity (MIN)

• Window height = image height• Window width = 2 X main vessel thickness

Ultra-Fast Localization

Ultra-Fast Localization

Vertical Localization

Projection of the image features on the vertical-axis

Projected Features:1) Summation of vertical and horizontal edges (MAX) 2) Pixels’ intensity (MAX)

Ultra-Fast Localization

Ultra-Fast Localization

Directionality of the vessels is described by their corresponding edges No matched filters are used for vessels extraction

The horizontal and vertical edge maps are acquired by applying a 3 X 1 gradient mask [1 0 -1] and its transpose to the retinal image

Evaluation

The new technique is evaluated using four publicly available databases (mostly used in literature):STARE, DRIVE, DIARETDB0, DIARETDB1

The new technique correctly located the OD in 321 images out of 340 images (94.4% accuracy)

The new technique achieved a computation time of 0.2 – 0.7 seconds depending on the image resolution

Results (success)

Results (failure)

This work has been accepted for presentation in 2009 IEEE Int. Conf. on Image Processing (ICIP’09)

Accuracy ImprovementIn the previous technique, only the following image

features have been used:1. Edge information of the retinal vessels.2. Pixels’ Intensity

In order to improve the accuracy of the technique, we propose using additional image features such as “Geometric Shape of the Optic Disc”

Accuracy Improvement

1. Construct a set of candidate locations 2 candidates

2. Define a scoring index for each candidate Index = Value of cand_H at this location

3. Weight each index with a factor Geometrical

features

cand_H

Accuracy Improvement

Evaluation

Technique Success Rate Computation Time

Without Improvement 94.4% 0.2 - 0.7 sec

With Improvement 97% 0.3 - 1 sec

Comparing the performance the proposed technique with/without the improvement step on the 4 databases (340 images)

This work has been accepted for presentation in the Int. Conf. of Medical Image Computing and Computer Assisted Intervention (MICCAI’09)

Conclusion A novel technique for fast and accurate OD localization

is developed.

The technique is based on reducing the search space dimensionality by projecting image features onto orthogonal axes.

The new technique is much faster (fraction of a second) than currently available techniques (few minutes)

Next Step Apply the new technique to Fluorescien

Angiograms (FA) of the retina (high contrast images)

Use the new idea of projecting the image features in other applications, such as:

1. Registration of retinal images 2. Determining the orientation of the retinal images

(left or right eye)

Thanks