Novel algorithm to differentiate astigmatism

from keratoconus

By: Sarah Ali Hasan801251019

Under the Guidance of :Dr. M.D .Singh

Outline: Anatomy of Human eye & Facts about Cornea

Modern Corneal Topography

Computerized video keratoscopy.

Measurement of the Corneal Surface.

Corneal maps.

Characterizing Corneas

Objective of the Algorithm

An Algorithm to distinguish Astigmatism from Keratoconus

Methods, results and discussion

Input Data

Testing the behavior of the images and choosing the best refining technique

Shape matching technique.

Anatomy of human eye:

External structures: eyelids, eyelashes, tears and fat glands, extra ocular muscles, conjunctiva.

Internal structures: cornea, sclera, iris, ciliary body, choroid, retina, lens, anterior and compartment, optic nerve.

A typical human eye can see wavelengths from about 380 to 750 nm and about 790 to 400 terahertz.

Spheroid structure with an average diameter of 24mm.

A human eyeball is like a simple camera: Sclera: outer walls, hard,

like a light-tight box. Cornea and crystalline

lens (eyelens): the two lens system.

Retina: at the back of eyeball, like the film.

Iris: like diaphragms in a camera.

Pupil: camera aperture. Eyelid: lens cover.

Focusing: The cornea and eyelens form a compound lens system, producing a real inverted image on the retina.

Refraction: From air to cornea:

(n=1.376): large bending, the main focusing.

From cornea to eyelens: (n=1.406), less focusing power. (Eyelens can develop white cloudiness when getting old: Cataracts.)

Facts about Cornea:

Cornea scatters almost 10% of the incident light.

Cornea’s average of curvature is 7.8 mm.

Cornea is accounting for about 43-44 diopters at corneal Apex (that is: 2/3 of total dioptric power of human eye).

Corneal Geography: Apex: central zone (4mm), overlies the pupil & responsible For HD vision.

Paracentral Zone: Where the cornea begins to flatten.

Peripheral zone. Limbal zone.

Facts about Cornea: (Contd.)

Computerized Videokeratoscopy(Modern Corneal Topographers):

A non-invasive medical imaging technique for mapping the surface curvature of the cornea.

The corneal topography and color coded maps derived from quantitative analysis of numerous points on the mires obtained from video photokeratoscopy.

Physicians interpret the color coded images to diagnose & treat patients with eye refracting issues.

Instruments to Measure the Corneal Surface: Cornea has a smooth surface

acting as convex transparent mirror, reflects part of the incident light.

Different instruments developed to measure and access this corneal reflex.

These non-invasive instruments use a light target (lamp, LED, Plasido disc, etc.) and a microscope or other optic system to measure corneal reflex of these light targets.

A handheld keratoscope, sometimes known as Placida's disk can provide a simple non invasive visualization of the surface of the cornea by projecting a series of concentric rings of light onto the cornea.

(a) (b)

Determination of corneal image forming properties: Height: The fundamental way of describing any surface

mathematically is to define the distance of each of its points from a reference surface.

Height maps with a flat reference plane show the overall shape of the cornea.

Finer detail can be provided by the use of a spherical reference plane, with the height being expressed as the ‘difference from a sphere’.

Radius of curvature (ROC): Corneas with a steep surface slope have a small radius of curvature and those which are flatter, have a relatively large radius of curvature. The radius of curvature is usually expressed in mm.

Power :Power (in dioptres(D)) is a measure of the refractive effect of the anterior corneal surface

Radius of curvature is converted to power using the standard keratometric index (SKI =1.3375).

This is an approximate figure derived from assumptions about the thickness and refractive index of the cornea, and the shape of its posterior surface.

Common Corneal Maps:

Axial map: It shows variations in corneal curvature as projections and uses colors to represent dioptric values(Global view).

Tangential map: It also displays

the cornea as a topographical illustration,(Local view).

(A)

(B)

Refractive map: It’s based on axial map, utilizes the measured dioptric power and applies Snell's law to describe the cornea's actual refractive power.

Elevation map: Represents the height of the corneal surface at different positions relative to either a spherical or elliptical surface.

3-D reconstruction map: used to visualize the overall shape of the cornea in more realistic way.

Topographic scales:

Two basic scales are commonly used :

Relative, Normalized or Adaptive color scale:

Different scale for each map, The computer determines the maximum & minimum curvatures for the map and automatically distributes the range of the colors.

Absolute, Standardized or International standard scale:

Same scale for every map produced, Good for comparison between different maps.

Characterizing Corneas: Round cornea with

a nice example of the theoretical corneal cap.

The color changes look as if they are concentric to fixation.

No clearly defined pattern within the pupillary margin.

Refractive astigmatism would not be expected on this patient.

On an oval cornea the color rings are egg-shaped but still do not have any color changes within the pupillary margin.

Corneal refractive astigmatism would be unlikely in this patient as well.

Astigmatic corneas are treated with refractive surgeries (LASEK, LASIK or PRK); while Keratoconic corneas are treated with totally different procedures (Mayo-ring, Riboflavin UVA- induced cross linking Or Collagen Cross linking).

Oculists make the final decision on diagnosing Keratoconus manually by interpreting data of axial and elevation maps.

Our aim is to suggest a computerized technique to help making this decision.

Objective of the Algorithm

Tools needed :

Shape recognition mechanism .

Alignment recognition mechanism.

Pixel value access.

An Algorithm to distinguish Astigmatism from Keratokonus:

Algorithm:

Input the topographic image. Apply shape recognition algorithm to identify the bowties

in the image. If the bowtie shape found fits to shape (1), Or (2) (single

closed end or extended shape); apply alignment recognition algorithm to identify the axis of both of the bowties.

Access the pixel and assign them to diopter vales to calculate the maximum & minimum diopter in the map.

According to shape, alignment & pixel values found combinations return one of the next diagnostic result:

1-”Typical astigmatism”:(with the rule), in case of extended bowties shape add ( limbus-to-limbus).

2-”Typical against the rule astigmatism”. in case of extended bowties shape add ( limbus-to-limbus).

3- Asymmetric astigmatism, here the higher valued diopter should be inside the closed loop regardless of the alignment.

4-In case of not orthogonal axis, return the result : ”Irregular Astigmatism”.

5-In case of closed loop shape with an alignment of -30degrees to 30 degrees vertical alignment& higher diopter values inside the lower loop return the result “Typical keratoconus”.

6- Same in last case but vertically flopped shape, return the result “Atypical keratoconus” .

Methods, results & discussion :

Data base is a selected group of (151) PETACAM topographic images.

All the samples are for patients of Dr.G.S Randhawa Eye Hospital.

The axial front map is used for the purpose of classification.

1.Input Data:

Sample images

Input Data:

2.Testing the behavior of the images:

All the image processing techniques can be done in grayscale level; no lose of data occurred while converting from RGB to Grey & even to B&W.

A. Conversion & thresholding

Thresholding:

Every image has its own unique response to the known thresholding techniques .

The rule is “no rule”.

Randomly logical and morphological operations has been applied to the images to get prior impression on their behavior.

After certain tests; Morphological operations has been selected as the classification operators for the input images.

B. Logical & Morphological Operations:

Objective: To examine the effect of every “basic” morphological operation separately on all the images.

No of Morphological. Operations: Seven; erode, dilate, open, close, thin, thick & skeleton.

Conclusion:

1) Individual comment on “Image response vs. Diagnoses” for every image has been recorded for the next experiments.

2) The foreground (black)& background (white) of the images should be swapped in the binary image; to get better results.

Test no.1:

Objective: Same as in test(1) but :to include two thresholding values instead of one & adjusting the binary representation according to conclusions of test (1).

No of Morphological. Operations: Seven; erode, dilate, open, close, thin, thick & skeleton.

Conclusion:

1) The number of thresholding values & combinations of Boolean relation among them has no effect on the response of the images.

2) Position of the selected threshold value affects the response; for all the images: position of thresholding values are:1;anywhere on meridians & 2; anywhere on the boundary region of the (circular/ bowtie) shape .

Test no.2:

Objective: Same as in test(2) but :to find out the best neighborhood index for each morphological operation & eliminating the test morphological operation for the next experiment.

No of Morphological. Operations: Seven; erode, dilate, open, close, thin, thick & skeleton.

Conclusion:

1) Dilating at 5-7 pixels and closing at 10-20 gave the finest results for all images.

2) Dilating, closing and skeletonising are elected for the next experiment .

Test no.3:

Objective: To examine the effect of closing and dialating and choosing the best neighborhood for each one.

No of Morphological. Operations: two; dilate, & close.

Conclusion:

1) The best morphological operation in refining the topographic pattern is closing.

2) The best neighborhood value is 25 pixel, till no change occurs is the enhancement technique chosen to refine the images .

Test no.4:

Objective: To examine the effect of closing at 25, dialating at 5 & 7 and skeletonising at 5and choosing the best neighborhood for each.

No of Morphological. Operations: three; dilate, close & skeleton.

Conclusion:

Closing at 25 is the best, it refines the detected patterns better than any combinations of morphological operations.

Test no.5:

3.Shape matching: Cross correlation is used to determine the

similarity between the ideal patterns of each disease and the

input image. Alignment is

automatically

included in by

using orthogonal

shapes.

  The corresponding confusion matrix is :

Publications:

[1] Sarah Ali Hasan, Mandeep Singh “An Algorithm to differentiate Astigmatism from Keratoconus in Axial Topographic images”, BEATS-2014 International Conference, (Published on 14-15 Feb).

[2] Sarah Ali Hasan, Dr. Mandeep Singh, “Automatic Diagnosis of Astigmatism for Pentacam Sagittal Maps”, Third International Workshop on Recent Advances in Medical Informatics (RAMI-2014) (Accepted to be published on (27-29Sep). IEEE conference.

THANK YOU !