WO2010134889A1

WO2010134889A1 - Methods and systems for pathological myopia detection

Info

Publication number: WO2010134889A1
Application number: PCT/SG2009/000179
Authority: WO
Inventors: Jiang Liu; Damon Wong Wing Kee; Joo Hwee Lim; Huiqi Ll; Ngan Meng Tan; Zhuo Zhang; Shijian Lu; Tien Yin Wong; Seang Mei Saw; Louis Tong
Original assignee: Singapore Health Services Pte Ltd; Agency For Science, Technology And Research; National University Of Singapore
Priority date: 2009-05-19
Filing date: 2009-05-19
Publication date: 2010-11-25

Abstract

A system for the detection of pathological myopia is proposed. The system makes use of a computerized algorithm for detecting peripapillary atrophy (PPA) around the optic disc region, which has been found to be correlated with the occurance of pathological myopia. Various techniques for detecting PPA may be used in different embodiments of the invention, such as a disc difference analysis, a grey level analysis, and/or a texture analysis, in which a numerical measure is obtained of the texture of a portion of the image identified as subject to PPA, and a determination is made from this measure of whether PPA is present.

Description

Methods and Systems for Pathological Myopia Detection

Field of the invention

The present invention relates to methods and systems for automatic detection of myopia, such as methods and systems which can be used to provide a low- cost screening operation, identifying individuals for whom further screening should be performed.

Background of the invention

Myopia is a major public health problem, with enormous and often underestimated economic and medical costs, and the early detection and management of degenerative eye diseases is of utmost importance in the care of myopic adults. High myopia is associated with ocular disorders such as retinal breaks, chorioretinal atrophy, Fuch's spot, lacquer cracks, pigmentary degeneration, lattice degeneration, posterior staphyloma, white without pressure, and larger, tilted rotated discs. Lesions such as myopic maculopath, myopic retinoschisis, retinal detachments and choroidal neovascularization are potentially visually disturbing. Myopia-related visual impairment may affect the productivity, mobility, quality of life and activities of daily living of individuals. Potentially blinding myopia-related pathologies are often irreversible in nature, especially if diagnosed late. The risks of visual loss in myopia are sufficiently high to warrant measures to prevent pathologic myopia.

A few studies have developed comprehensive and consistent methods for grading pathologic myopia. In the Blue Mountains Eye Study, myopic retinopathy included the presence of staphyloma, lacquer cracks, Fuch's spot, myopic chorioretinal thinning or atrophy, beta-peripapillary atrophy, cytotorsion or tilting of the optic disc, and the T-sign found in central retinal vessels. Avila et al reported a grading system for myopic chorioretinopathy consisting of 6 grades of increasing severity[1]: • MO: normal posterior pole with no tessellation pattern in the macular area

• M1 : tessellation and choroidal pallor pattern in the macular area • M2: choroidal pallor and tessellation, and the border of an ecstasia posteriorly was visualized

• M3: pallor and tessellation with several yellowish lacquer cracks in Bruch's membrane and posterior staphyloma

• M4: choroidal pallor and tessellation, with lacquer cracks with posterior staphyloma and focal areas of deep choroidal atrophy

• M5: choroidal pallorand tessellation with lacquer cracks, posterior staphyloma, geographies areas of atrophy of retinal pigment epithelium and choroids, and choroidal neovascularization were visualized

M3 and higher grades were defined as "with maculopathy".

Summary of the Invention

All the current methods for assessment of a patient for pathologic myopia described above are still largely reliant on manual efforts. This limits the applicability of such methods for screening efforts. An automatic method for the assessment of retinal images for pathologic myopia has yet to be reported and yet is very much needed, due to the large social impact of myopia in many societies today.

In general terms, the present invention proposes a system and/or method for the detection of pathological myopia by making use of a computerized algorithm for detecting peripapillary atrophy (PPA) around the optic disc region.

The term PPA refers to thinning of the retina and retinal pigment epithelium in the region immediately surrounding the optic nerve head, and results from misalignment of the edges of the neurosensory retina, retinal pigment epithelium, choroids and sclera. In a recent study, the presence of PPA was reported to be highly correlated to pathological myopia. PPA appears in a retinal fundus image as temporal choroidal or scleral crescents or rings around the optic disc.

Various techniques for detecting PPA may be used in different embodiments of the invention, targeted towards detecting different characteristics of PPA. These techniques may include one or more of: a disc difference analysis, in which inner and outer starting contours are defined, and allowed to grow using a level set algorithm, and then a test is performed on the region between the contours (i.e. the intersection region of the discs which they respectively enclose), or one or more portions of the intersection region, such as the left and right halves of the intersection region (the test may be of the size of the halves of the intersection region, e.g. in relation to an estimate of the size of the optic disk), a grey level analysis, in which statistics are collected of grey scale levels in multiple regions identified as subject to PPA, and a determination is made from these statistics of whether PPA is present, by comparing the statistics relating to different regions, and/or a texture analysis, in which a numerical measure is obtained of the texture of a portion of the image identified as subject to PPA, and a determination is made from this measure of whether PPA is present.

Preferably, synthesis of the detection results is performed to produce a meaningful report. Optionally, this may employ an adaptive learning-based method, such as neural networks or support vector machines.

Embodiments of the invention make it possible to provide a clinically significant, fast, objective tool for the automatic assessment of retinal images for pathological myopia which can be used for population screening. These tools preferably have the following characteristics: • Automatic analysis for pathological myopia

• Objective and consistent performance

• Designed though integration of clinical expertise • Employs knowledge-based heuristics for increased accuracy and relevance

• Multi-factorial analysis and decision improves detection accuracy

• Readily applicable to existing hardware

Embodiments of the invention have the potential to be used in healthcare institutions for the detection of pathological myopia, an important condition resulting in visual loss.

The technology can also be integrated by retinal fundus manufacturers into their hardware as a built-in function, extending the capability of such devices.

Employing advanced image processing with clinical expertise in the design of the framework and detection parameters, and via knowledge-based heuristics, embodiments of the invention are able to assess retinal images for pathological myopia automatically, providing an objective and consistent evaluation for screening and early detection.

Brief description of the drawings

An embodiment of the invention will now be described for the sake of example only with reference to the following drawings in which:

Figure 1 is a flow diagram of the steps of the embodiment; Figure 2 illustrates a Gray level analysis method used in the embodiment, and is composed of Figures 2(a) to (d), of which Figure 2(a) shows identified contours, Figure 2(b) the extracted local ROI, Figure 2(c) the left quadrant, and Figure 2(d) the right quadrant;

Figure 3 shows sub-steps of step 7 of Fig. 1 ; Figure 4 illustrates three images obtained in step 3 of Fig. 1 , and comprises Figure 4(a) which shows an extracted disc ROI, Figure 4(b) which shows a texture map, and Figure 4(c) which shows a classified texture map;

Figure 5 illustrates the detected disc (in green) and the ROI considered around the disc centre (white), (a) Superior, (b) nasal, (c) inferior and (d) temporal. Note zonal descriptions are for a right eye and the temporal/nasal zones are to be reversed for a left eye; and

Figure 6 illustrates knowledge-based constraints on peripapillary zonal analysis, and comprises Figure 6(a) which shows the zonal limitations and Figure 6(b) which includes the texture map.

Detailed description of the embodiments

The steps of an embodiment of the invention will now be explained with reference to Figure 1. The embodiment is given the title PAMELA (PAthological Myopia dEtection through peripapilLary Atrophy). The first step (numbered step 1 in Figure 1) is to obtain a retinal fundus image, for example using standard retinal imaging equipment, or by reading data describing a pre-existing image from a database. Typically, only a single image is employed per eye.

In step 2, a region of interest including the optic disc is delineated, as a first step in obtaining the optic cup and disk. The optic disc generally occupies less than 5% of the pixels in a typical retinal fundus image, so while the disc and cup extraction explained below can alternatively be performed on the entire image, localizing the ROI helps to reduce the computational cost as well as improve segmentation accuracy.

The optic disc region is usually of a brighter pallor or higher color intensity than the external part of the fundus image. This characteristic is exploited by automatically identifying the 0.5% of the pixels in the image with the highest intensity. Next, the retinal image is subdivided into 64 regions, and an approximate ROI centre is selected based on the region containing the highest number of identified pixels. Following this, the ROI is defined as a rectangle around the ROI centre with dimensions of twice the typical optic disc diameter

Since, as mentioned above, PPA usually exists in the temporal side of the retina, as temporal choroidal or scleral crescents or rings around the optic disc, the embodiment compares the texture of the two sides of the disc boundary region to estimate whether atrophy exists. If no atrophy exists, there would be little difference in the texture of the two sides of the disc boundary region. If there is atrophy, the texture would be very different.

The embodiment uses three methods to detect PPA in the fundus images: a disc difference analysis (step 3); a grey level analysis (step 6); and a texture analysis (step 7). These three methods are explained in detail below. The grey level method and texture method both employ data which has been subject to optic disc extraction (step 4) and boundary smoothing (step 5).

The outputs of steps 3, 6 and 7, and optionally also further algorithms are performed in parallel, and the results combined in step 8 by a decision engine, to output in step 9 an assessment of whether PPA is present.

We first describe the set of steps 4 to 6. In the step of optic disc extraction (step 4), the optic disc is first detected within the ROI by using at least two methods. One is based on the variational level set method, which is a segmentation method based on a global analysis of the image [2]. The second method is based on thresholding of the image. Thresholding is performed by an analysis of the color channels of the ROI. The color-based histograms of the ROI are analyzed, and a value is used to select the high intensity pixels in the ROI, which represent the optic disc. The level-set method is generally used due to improved accuracy. However when the level-set determined optic disc is too- large or too-small, as determined from heuristics, the thresholding method is used. After detection of the optic disc, the detected boundary contour can potentially be "rough" due to the presence of blood vessels around the disc boundary, which can cause an erroneous disc boundary. To smoothen the boundary, ellipse fitting is used (step 5). Ellipse fitting involves determining an ellipse which would fit the data points while minimizing error between the ellipse and the unfitted points.

The grey level analysis (step 6), starts by defining a region of interest RO!2 (different from the ROI found in step 2) by drawing a circle 31 which has a radius of 1.2 times of the semi-major axis and a center at the center of the optic disc 33 (Figure 2(a)). The optic disc region is then excluded from ROI2, as shown by Figure 2(b). That is, ROI2 is the region between the lines 31 and 33.

The next sub-step is to further fine tune the ROI by drawing two lines that pass through the optic disc center found by ellipse-fitting, and are at an angle of +45 degrees with the horizontal direction. The top and bottom quadrants are filtered, and only the right and left quadrants are left, as shown by Figures 2(c) and 2(d) respectively. The final sub-step of step 6 is to analyze the grey scale image of the two quadrants. The mean intensity and standard deviation of each quadrant are measured. Thus, we can obtain the difference in the mean intensity and standard deviation of the two sides. After doing experiments on several images, we found that the images with PPA have much larger difference in mean intensity and standard deviation compared with images without PPA. Therefore, we make some assumptions as the conditions to detect PPA. If the difference in mean intensity of the two sides is greater than a threshold such as 30 (this Is this assuming 255 grey scale levels in total). The patient is confirmed to have PPA. If the difference in mean intensity is greater than 5 and difference in standard deviation is greater than 0.6, the patient has a high probability to have PPA (or confirmed to have PPA). Otherwise, the patient does not have PPA.

We now turn to an explanation of the disc difference analysis (step 3 of Fig. 1). The basic idea for this method is to define inner and outer starting contours, and grow each of them separately using a level set method. The level set method referenced in [2] is used. Two starting contours were defined. The outer starting contour is typically a circle which is larger than a typical the optic disc, and is allowed to grow inwards. The inner starting contour is a circle smaller than a typical optic disc, and is allowed to grow outwards. Two different contours for the disc will be obtained. Region which is the intersection of the disks within the respective two contours is considered as possible suspect atrophy regions. A vertical line is drawn to partition the intersection regions into right and left parts, and they are analyzed separately. If the area of either side of the residue is greater than 10% of the inner contour the patient is considered to have PPA.

We now turn to a discussion of the texture analysis of step 7. This employs the ellipse-fitted optic disc obtained in steps 4 and 5. The sub-steps of step 7 are shown in Fig. 3. The disc segmentation and ROI extraction step 71 of Fig. 4 the same as the ROI extraction of step 2 of Fig. 1 and follows the method described in Fig 1. The subsequent sub-steps 72-76 utilize regional pixel textures, in order to output an indication (indicated as "PPA") of whether PPA is present.

In a first sub-step 72, the image is first prepared for analysis by performing contrast enhancement. This is done via adjusting the pixel intensities such that there is minimal saturation at the intensity extremities. This is a simple method using a standard function from MATLAB.

Following this, in sub-step 73, the image is processed for entropy analysis. Entropy is a measure of statistical randomness that can be used to characterize the degree of randomness or 'roughness' in an image. To determine the local texture for each pixel in the image, a 9x9 pixel square neighbourhood is analyzed for entropy for each pixel. Entropy ε defined as ε = Σ(p*log2(p)), where p refers to the normalized intensity histogram counts from the 9x9 area, and the sum is over the pixels of the 9x9 square. Subsequently, a texture map is generated which indicates the entropy values at each pixel in the image. To further classify the results, the entropy values, which range from 0 to 5, with a higher value indicating a higher degree of roughness, were classified as in the following table.

Table 1 Classification of values

An example is shown in Figure 4. Figure 4(a) shows the region of interest obtained in sub-step 71. The ROI is the entire block. Figure 4(b) is the image of Figure 4(a) but with corresponding pixels given a darkness indicating the corresponding values of ε , with darker pixels having higher respective values. Figure 4(c) is the image of Figure 4(a), but after the entropy values ε have been classified into levels 1 to 5 according to table 1; again, darker pixels have higher classified entropy values.

Steps 74-76 employ knowledge-based constraints. Firstly, it is known that PPA generally occurs with a high frequency in the temporal aspect of the peripapillary region. Therefore, in step 74 the peripapillary region is sub-divided into 4 sectors, centered at the disc centre, and with an angular width of 90°. The four sectors are positioned according to general inferior, superior, nasal and temporal zones, as shown in Figure 5.

Secondly, it is known that PPA occurs in the peripapillary region outside the optic disc. To take this into account, a buffer region is defined from the detected disc boundary, as shown in Fig. 6(a). The disc boundary typically has high contrast, which can potentially affect the texture-based analysis for PPA. The buffer region ensures that the regions analyzed are free from edge effects due to the disc boundaries, and that only the peripapillary region is analyzed.

In step 75, entropy results in each zone are each calculated to give the following metrics

Z = {inferior, superior, nasal, temporal}

d ^U4Z = -* Z = {inferior, superior, nasal, temporal}

^W4 _Z

5Z Z = {inferior, superior, nasal, temporal}

Where the curly brackets indicate that the values are to be calculated successively for each zone Z, n is the total number of pixels in that zone, s is the total entropy score, μ_z\s the average entropy score in that zone; n₄ is the number of pixels with an entropy score of 4, d₄ is the pixel density of the n₄ pixels, n₅ is the number of pixels with an entropy score of 5 and cfe is the pixel density of n₅ pixels. Concurrently, the ratios of each zone relative to the

temporal zone for each metric, e.g. ^μtemp°^ral I ^μ"^asal are also calculated to obtain a comparative analysis across the zones, due to the higher incidence rate of PPA within the temporal zone.

More than 20 features are thus calculated using the above methods. Employing context-based clinician knowledge on PPA occurrence, such as a greater average roughness score in the temporal zone compared to the nasal zone and the highest absolute score in the nasal zone, each retinal image was analyzed and the risk of PPA was determined in step 76, to give a PPA determination.

The decision engine receives the PPA determinations output in steps 3, 6 and 7 as shown in Table 2, and fuses them for Pathological Myopia Risk Assessment. It can be observed that each method produces slightly different assessments, due to the various techniques employed for each. In order to make the best use of the available results, the results were fused using the following rules, based on a modified voting system:

1. Decide on a positive assessment for pathological myopia if all three methods report positively

2. If any two methods report a positive assessment, determine the risk of pathological myopia by depending on the results from texture analysis.

3. If no two methods report a positive assessment, the overall determination is negative.

Experimental tests

To evaluate the performance of the embodiment, an experiment was conducted using images obtained from the Singapore Cohort study Of the Risk factors for Myopia (SCORM), a study conducted by the Singapore Eye Research Institute. This cohort study has enrolled 1979 children aged 7 to 9 years from 3 schools since 1999. Yearly eye examinations are being conducted and the children will be followed for the next 10 years. A sample of 40 images was obtained from the SCORM study. The image set was analysed by a senior ophthalmologist from the Singapore eye Research Institute and was accessed for the presence of PPA. This assessment from the ophthalmologist was then used as the ground truth.

Using the embodiment, the same test set of 40 images was also processed using the framework and methods described in Sections 2.1 and 2.2. As shown in Table 2, an overall 95% accuracy of correct assessment for pathological myopia was attained, with a sensitivity of 0.91 and a specificity of 1.0.

Table 2. Presence of PPA for individual methods and overall decision.

Discussion

The embodiment is based on the detection of peripapillary atrophy. Other embodiments include also analysis of one or more further features, such as any one or more of:

1) Staphyloma: scleral ectasia with thinning

2) Lacquer cracks: healed fissures in the Bruch's membrane

3) Fuch's spot: circular black spot surrounded by macular haemorrhage 4) Choroiretinal atrophy: thinning of the choroids and choriocapillaris loss

5) Disc tilt: the ratio of the longest and shortest diameter of the optic disc

6) Torsion of the optic disc: longest axis of the optic disc rotated more than 15 degress outside vertical meridian 7) T-sign: distal bifurcation of the central retinal vessel at least 0.5mm beyond the lamina cribosa

8) Pallor and tessellation: large choroidal vessles visualized through the retinal pigment epithelium 9) Cup and disc vertical and horizontal diameters 10) Choroidal neovascularization

The presence of one of more of these features can be taken into account by the decision engine in step 8, e.g. by (i) first deciding whether there is PPA using the outputs of steps 3, 6 and 7, and then (ii) then combining this result with determination of features (1) to (10), or (ii) by using the results of steps 3, 6 and 7, and the determination of features (1) to (10) simultaneously.

Definitions

The term "computerized" is used in this document to describe a process which is performed using a computer system, such as a general computer including software designed to cause a processor of the computer to operate the method. The computer system may for example include a CPU and one or more data storage device for storage of the software and/or the images, as well as a display device for displaying visually the results of the algorithm and/or a data interface for outing the results in electronic form. The term "computer program product" refers to a software product, such as one stored on a tangible data storage medium. The term "automated" is used in this document to describe a computerized method which is performed without human interaction, except possibly for human initiation of the method. By contrast, the term "semi- automated" is used to describe a computerized method that includes automatic process steps, but may include reverting to a human user for input during the performance of the semi-automated method.

References [1]Avila, M.P., Weiter, JJ. , JaIk, A.E., Trempe, C.L., Pruett, R.C., Schepens, CL. , Natural history of choroidal neovascularization in degenerative myopia, Ophthamology, 91 , 1573-1581.

[2]Li, C, Xu, C, Gui, C₁ Fox, M. D. Level set evolution without re-initialization: a new variational formulation. In Proc. of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2005.

Claims

1. A method for the detection of pathological myopia in an eye, the method comprising: (i) obtaining a fundus image of an eye;

(ii) automatically identifying one or regions of the fundus image potentially subject to peripapillary atrophy (PPA);

(iii) automatically obtaining numerical characteristics of the identified region or regions; and (iv) automatically determining whether PPA is present from the numerical characteristics.

2. A method according to claim 1 which includes one or more of:

(a) a disc difference analysis, in which in step (ii) inner and outer starting contours are defined, and grown using a level set algorithm, and the identified region is an intersection region which is the intersection of the disks which the contours respectively enclose, and in step (iii) one of the numerical characteristics relates to the area of at least one or more portions of the intersection region; (b) a grey level analysis, in which in step (iii) statistics are collected of grey scale levels in a plurality of said regions, and in step (iv) a comparison is made of the statistics from different ones of the regions to determine whether PPA is present, and

(c) a texture analysis, in which in step (iii) a numerical measure is obtained of the texture of a portion of the image identified as subject to PPA, and in (step iv) a determination is made from this measure of whether PPA is present.

3. A method according to claim 2 in which the grey level analysis is of two regions horizontally spaced to either side of an optic disc centre, the corresponding determination of whether PPA is present depending on differences between respective numerical characteristics of the two regions.

4. A method according to claim 2 in which the texture analysis includes obtaining an entropy statistic for each of multiple ones of the regions.

5. A method according to any preceding claim in which a plurality of determinations are made of whether PPA is present, and in step (iv) the results of the plurality of determinations are combined to form an overall assessment of whether PPA is present.

6. A method according to claim 5 when dependent upon claim 2, the plurality of determinations including three determinations made respectively according to analyses (a) to (c).

7. A system for the detection of pathological myopia in an eye, the system comprising a processor arranged to:

(i) receive a fundus image of an eye;

(ii) automatically identify one or regions of the fundus image potentially subject to peripapillary atrophy (PPA);

(iii) automatically obtain numerical characteristics of the identified region or regions; and

(iv) automatically determine whether PPA is present from the numerical characteristics.

8. A system according to claim 7 in which the processor is arranged to perform one or more of:

(a) a disc difference analysis, in which in step (ii) inner and outer starting contours are defined, and grown using a level set algorithm, and the identified region is the intersection of the disks which the contours respectively enclose, and in step (iii) one of the numerical characteristics relates to the area of at least one or more portions of the intersection region;

(b) a grey level analysis, in which in step (iii) statistics are collected of grey scale levels in a plurality of said regions, and in step (iv) a comparison is made of the statistics from different ones of the regions to determine whether PPA is present; and

9. A system according to claim 8 in which the grey level analysis is of two regions horizontally spaced to either side of an optic disc centre, the corresponding determination of whether PPA is present depending on differences between respective numerical characteristics of the two regions.

10. A system according to claim 8 in which the texture analysis includes obtaining an entropy statistic for each of multiple ones of the regions.

11. A system according to any of claims 7 to 10 in which a plurality of determinations are made of whether PPA is present, and in step (iv) the results of the plurality of determinations are combined to form an overall assessment of whether PPA is present.

12. A system according to claim 11 when dependent upon claim 8, the plurality of determinations including three determinations made respectively according to analyses (a) to (c).