confocal corneal endothelium dystrophy’s analysis …

Journal of Engineering Science and Technology Vol. 15, No. 2 (2020) 1338 - 1356 © School of Engineering, Taylor’s University

1338

CONFOCAL CORNEAL ENDOTHELIUM DYSTROPHY’S ANALYSIS USING PARTICLE FILTER

KAMIREDDY VIJAY CHANDRA1, 2, BHASKAR MOHAN MURARI3,*

1School of Electronics Engineering, Vellore Institute of Technology, Vellore, India 2Department of Electronics and Instrumentation Engineering,Vallurupalli Nageswara Rao

Vignana Jyothi Institute of Engineering and Technology, Hyderabad, India 3Department of Sensor and Biomedical Technology

School of Electronics Engineering, Vellore Institute of Technology, Vellore, India *Corresponding Author: [email protected]

Abstract The cornea of the human eye is the main functional element for vision. The visual data information is transferred from all the layers of the cornea to the focal point of the retina. optical nerve acts as media between the cornea and brain. The image data sets are acquired from the confocal microscope, which improves the penetration depth in order to extract from all five layers. Two datasets are acquired from confocal endothelium image and pre-processed and analysed. The data set ‘1’consists of ‘2’ normal and ‘3’ abnormal images. And dataset ‘2’ consists of five abnormal images. The image is segmented for identifying the cell contours and boundary conditions of the cell. The contrast levels of the image are enhanced with histogram equalization. The high-frequency noise signal is curtailed with a gaussian filter. The image is further processed for identifying the dark objects using auto metric threshold method. The cell density, cell structure, Pleomorphism, Polymegathism, Elongation Factor (EF), Compactness Factor (CF), and Heywood Circularity Factor (HCF) has been meticulously estimated with Particle Filter (PF) approach for each endothelium cells. The estimated results are compared with the existing morphological operation 1(MO1), morphological operation 2 (MO2) and S-PSO (Snake Particle Swarm Optimization) Approaches. From the experiment, the proposed approach obtains satisfactory results in part of (average time inspection) is 18.49 m/s and Standard Deviation (SD) is 0.51 m/s per image is meticulously calculated.

Keywords: Corneal endothelium, Morphological, Particle filter, Pleomorphism, Polymegethism.

1339 K. V. Chandra and B. M. Murari

Journal of Engineering Science and Technology April 2020, Vol. 15(2)

1. Introduction The cornea is the most transparent part of the eye. visual data is processed through the cornea to the retinal basement. In cornea typical layers are embedded to actively function. Five layers are very sensitive in nature any damage to the layers results in severe visual impairment. Sharif et al. [1] proposed that in the cornea, various types of proteins, fibrils are in a systematic order to get function smoothly. The cornea protects the iris, lens and retinal structure from UV light. Cornea physical structure in the horizontal direction around 12.6 mm and vertically 11.7 mm, with a centre thickness of 520 μm and 650 μm at its periphery [2]. The outermost corneal layer is epithelium layer, Bowman’s layer and stroma layer is the middle layers. Descemet's layer and endothelium layer are the innermost layers. Among five layers in the cornea, endothelium layer is a very sensitive membrane, and it helps aqueous humour liquid to flow inside and outside through corneal layers as shown in Fig. 1.

Fig. 1. Corneal layers in eye (source: www.allaboutvision.com).

At each scanning with a duration of the 20 s nearly 350 images are acquired from a confocal microscope, all are 2-Dimensional images. these images are helpful for diagnosing the condition of the cornea. Due to the spherical structure of the cornea, the light illuminated on it is non-uniform, which leads darken the image, it is difficult to process the image. These images required processing techniques to eliminate significant noises. Confocal Microscope gives a detailed vision of corneal layers to identify dystrophies, Bacterial, viruses and fungal infections [3].

Endothelium layer is a very sensitive layer and last layer in cornea it has hexagonal structural cells, which are a hyper-reflective pattern. It is transparent in nature, no muscular (or) nerves present in endothelium cells, sometimes nuclei may be seen in it [4-6]. Dystrophies, degenerations, injuries, dry eye, infections cause the blurriness. Therefore, it is important to monitor the vision of the eye. Fuch’s Dystrophy (FD), Advanced Fuch’s Dystrophy (AFD), edema and other ocular diseases may directly affect the endothelium layer. Therefore, cell density reduced, due to this vision may get fuzzy/hazy [7]. Thus, it is important to monitor cell density in EL (endothelium layer). to identify various diseases.

From CM-4 350, images/scan are acquired to analyse all images by ophthalmologists/clinicians it consumes long time. Developing an intelligent and

Confocal Corneal Endothelium Dystrophy’s Analysis using Particle . . . . 1340


robust algorithm, which provides clinical parameters by less processing time, fetches more advantages to the ophthalmologists to speed up the treatment with many numbers of patients. In this paper, a novel algorithm is proposed and inspecting endothelium layer to get rid of the above problem. The proposed algorithm, which is automatic can provide clinical parameters like cell density, cell size, cell shape, endothelium cell HCF, endothelium cell CF, endothelium cell EF. Confoscan 4 microscope and HRT Rostock module having the capability to analyze the endothelium cell structure. Confoscan 4 is a fully automated microscope able to analyze minimum cell area, maximum cell area, average cell area, and cell density. Whereas HRT Rostock module is semi-automated microscope average area cells and other parameters could not possible to collect because of the internal functions not framed according to the standard ophthalmologist’s view.

Existing methodologies Endothelium cell density is analysed with the inverse transpose of a matrix. In this method, Cell density is analyzed with a perfect tessellation of hexagonal shaped cells. Different cell shape like polygon (or) another shape can’t be traced out with this algorithm. Different patients have different shapes of cells depending upon the dystrophies stage. usually, Normal endothelial cells are hexagonal in nature. FD, AFD, PPCD makes the cells into a different shape. When dis-oriented shape cell is analyzed with inverse transpose matrix, matrix values are varied accordingly, which is unable to detect exact endothelial cells. In this method, they acquired corneas in berlin cornea bank. Grisan et al. [8] mentioned that the proposed work, direct images are considered from CM-4 [8].

Initially, Endothelium cell density and cell shape are analyzed with morphometric analysis, a technique used here cellular neural network (CNN) algorithm. Image segmentation is done to get the customized cell boundary part and rebuilding it to get the original image. According to Tervo et al. [9] Endothelium cell area, cell elongation could not identify [9].

Later 2D-Discrete Fourier Transform (DFT) is applied to the images and Acquired Automatic endothelium cell density. In an image, all spatial frequency components are acquired. the repetitive spatial component structures are retrieved from DFT. In this method, all images are acquired from cornea bank Berlin(www.karger.com). Cell contour could not be traced in this method. Based on studies by Ruggeri et al. [10], Endothelium cell images are acquired after culturing of the cornea, with 2% fetal calf serum (FCS) containing 6% dextran 5000, to get a low number of folds on Descemet's membrane, and to focus largely on the endothelium layer [11].

Another approach was the neural network segmentation module (NNSM). The skeletonized cell contours were used to process the image. Splitting of the image and merging of the image is performed to get Estimate ECD. This method requires external experts to get correct the data from NNSM [12].

Further (S-PSO) approach was also applied to the endothelial cell contours. Endothelial cell density is also estimated. And, it provided minimum cell area, maximum cell area, Average cell area and effective analysis is possible with this statistical index. In order to effectively analyse the dystrophies along with cell area, cell size, cell elongation factor, cell compression factor, cell circularity factor also



estimated. These cell parameters are helpful for accurate diagnosis of dystrophy by ophthalmologists. It is accomplished by the PF.

The main limitation of the existing methodologies is Endothelium cell completer contour cannot identify. Elongation of the cell, compression of the cell, Circularity structure of cell not able to identify with existing methods. With proposed algorithm limitations are swamped and accuracy is accomplished.

2. Methodology Considering Fig. A-1 (Appendix A), convolving the image with kernel filter with size (3*3) produces (4*4*2) image. Convolving the image with kernel filter enhanced the image resolution by eradicating the unnecessary edge layers. The vertical edge and horizontal edge of the endothelium image is extracted for further processing.

The convolution of the image was carried out by following Eq. (1).

𝑦𝑦𝑦𝑦 × 𝑦𝑦𝑦𝑦 × 𝑦𝑦𝑦𝑦 ) × (𝑧𝑧𝑦𝑦 × 𝑧𝑧𝑦𝑦 × 𝑧𝑧𝑦𝑦) = 𝑦𝑦 − 𝑧𝑧 + 1 (1)

where yc, zc, are the same numbers. y-z+1 *y-z+1* y'c yc, zc are colour channels

The features of the image are extracted as: height of the image (yx) width of the image (yy) height of the filtered image (zx) width of the filtered image (zy)

Considering Fig. A-2 (Appendix A), metric threshold differentiates abnormal colour variation endothelial cells to normal colour endothelium cell.

∑∑−

+−=

−+=1

1|)(|)(|)|

L

jiziik σσ

j

0ii-( ik( i)y)F( x, (2)

where ‘j’ is the pixel value for F (x, y), No. of rows (x), No. of columns (y) ‘ iσ ’ is the average of all pixel quantities, which lies between the range of 0 and j.

‘zσ ’ is the average of all pixel quantities lies in the range between ‘j+1’ and

255 pixels.

J}y)F(x, if 0 Jy)F(x, If 1

≤≥= {y)F( x, (3)

‘J’ = Thresholding value is ‘117’.

Below 117 pixels are clustered to identify the dark objects present in the endothelium image.



The proposed algorithm (PF) is the main heart of the algorithm. The endothelium cells are considered as particles in this algorithm. The following steps are followed in order to track the endothelial cell and its contour.

�𝑃𝑃𝑡𝑡−1 < 𝑦𝑦𝑡𝑡−11 ,𝑃𝑃𝑊𝑊𝑇𝑇−1𝑘𝑘 > 𝑉𝑉𝑡𝑡 ,𝑋𝑋𝑡𝑡� (4)

11−ty is the previous state/position of the particle.

kWT

P1−

is the weight (or) pixel of the particle.

Vt is the new action movement to forward Xt is the new pixel measurement.

• 𝑝𝑝𝑡𝑡 = φ, σ = 0,'p𝑡𝑡 ′ is the set of particles (or) cells, and ϕ is the empty cells, initially. σ is the normalization constant, Initially, it is ‘zero’

• For k = 1 …m resampling new pixels (generating kth new pixel samples) where ‘m’ is the no. of. particles we need.

• l(k) is a sampling of the particle from the previous distribution according to the weights pwt-1.

• 𝑝𝑝𝑡𝑡𝑘𝑘 is the new state, 𝑝𝑝𝑡𝑡𝑘𝑘sampling using old particle, i.e., pt-1, 𝑎𝑎nd have the action 𝑣𝑣𝑡𝑡. This leads to the next level pt using the previous sample yt-1 (k) and pixel control vt.

• 𝑦𝑦𝑡𝑡𝑘𝑘 = particle(𝑦𝑦𝑡𝑡/𝑦𝑦𝑡𝑡|) ‘𝑦𝑦𝑡𝑡𝑘𝑘’ is the new pixel weight.

• 𝜎𝜎 = 𝜎𝜎 + 𝑦𝑦𝑡𝑡| updating normalization factor (or) updating new pixel quantities.

• Pt is the new state, 𝑝𝑝𝑡𝑡 ∪ �< 𝑦𝑦𝑡𝑡|,𝑝𝑝𝑤𝑤

| >� inserting new pixel quantities, and repeat step 2 to step 6 till complete cell track.

• Dividing weights by normalization factor 'p𝑤𝑤| /𝜎𝜎′to get normalization weights.

A (PF) is the effective filter for estimation of new pixel quantity in each image, based on the old pixel values. The PF Calculating mean pixel densities estimates new pixel quantity in an image. pixel strength is stable for normal endothelium cell structure. Due to FD’s, AFD’s, PPCD, ICD the cell density reduced from 2400 cells/mm2, consequently pixel density is also getting reduced. Estimating the new pixel quantities based on old pixel quantity particles can be identified. PF is an effective algorithm to accurately diagnosis the normal and abnormal images. By analysing the particles (or) cells, which distributed in an image the condition of the cornea is estimated. Statistical parameters like cell density, endothelium CF, endothelium EF, HCF, the cell area is accurately measured with less processing time as shown in the result section.

3. Results and Discussion Endothelium images acquired with CM-4. All the images are acquired with stabilization of Z-ring standard to avoid misalignment due to manual artefacts. The confocal microscope is magnified with (40X) zoom. Original image size is 710*533-pixel quantities. All images are interpolated with quadratic interpolation to get the standard dimensional structure. Total ‘10’ images are acquired from the CM-4, Organized into ‘2’ datasets ((Ref -http://rod-rep.com -Rotterdam



Ophthalmic Data Repository). The first dataset consists of ‘2’ normal endothelium images and ’3’ abnormal endothelium images. The second dataset consists of ‘5’ abnormal images. All ‘10’ images are processed through a particle filter (PF) algorithm to estimate the cell density and cell quantitative parameters. Table 1 represents a pre-processing stage of images; the original image is segmented to get a histogram of the image.

Normal endothelium images cell density range between 2400 cells/mm2 to 3200 cells/mm2. Dystrophies like FD, AFD, PPCD, and ICD, causes cell density to fall below the normal range. In PF algorithm original image is segmented for clear identification of cell boundaries for its edge detection. To improve the brightness level of the segmented image histogram equalization is performed. Low and high-frequency noises are eliminated with Gaussian filter, corresponding PSNR and MSE values of the filtered image are tabulated.

Metric thresholding is performed to identify dark objects present in the image. Below ‘117’ threshold values are clustered and above are diminished. The particle filter is applied at minimum parameter value ‘1’ and maximum parameter value ‘2.4’. All the endothelial cell particles can be identified in between ‘1’ and ‘2.4’, above this range cells are segregated.

Cell density is estimated with a number of particles in ROI of an image as shown in Eq. (7).

Cell density = No.of particle cells in distinctive areaTotal area of an endothelium image(𝐴𝐴)

(5)

Area = Height of the image* Length of the image (6)

Error deviation is calculated with E%

E% = �Manual cell density(Standard measurement)-PF cell densityManual cell density

� (7)

Table 1. Pre-processing and processing stages of images (data set 1).

Serialno.

Segmented image

Histogram of the image

Gaussian filtered image

Auto metric

threshold

Particle filtered image

CD cells/mm2

I1

1184



I2

2621

I3

2455

I4

324

I5

601



Considering Fig. 2, the mean value of images I2 (0.90 mm2), and I3 (0.88 mm2) are in close approximation to the standard value, i.e., 1 mm2 for normal images, SD: I2 (3.01 mm2), I3 (3.33 mm2) are close to the standard value 3.5 mm2 for normal images. The Mean value for images I1 (0.64 mm2), I4 (0.35 mm2), I5 (0.47 mm2) are close to the standard value, i.e., 0.8 mm2 for abnormal images. SD for I1 (2.90 mm2), I4 (2.15 mm2), I5 (2.48 mm2) is also close to the standard value < ‘2.80’ mm2 for abnormal images. The area of the acquired image is 0.2611*105, remains unchanged during the processing stage. as the images are quadratically interpolated from I1 image to I5.

Figure 3 represents the result of an auto-metric threshold, the lower limit is ‘10’ and the upper limit is ‘245’ the images I1 (119), I2 (120), I3 (121), I4 (117), I5 (117) are the resultant metric threshold outputs, which ranges from ‘117-121’ below this range concludes dark object identification. Above specified range is bright object identification, which has never achieved in I1-I5.

Fig. 2. Variation of mean, SD, and area I1-I5.

Fig. 3. Auto metric threshold of I1-I5.

0.640.9 0.88

0.35 0.47

2.5

3.013.33

2.152.48

0.2611 0.2611 0.2611 0.2611 0.2611

0

0.5

1

1.5

2

2.5

3

3.5

I1 I2 I3 I4 I5

Mean value Standard Deviation Area mm2

10 10 10 10 10

245 245 245 245 245

119 120 121 117 117

0

50

100

150

200

250

300

I1 I2 I3 I4 I5

Lower limit Upper limit Result



Confocal microscope images consist of spurious noises and all the noise components present in images I1(46.8 dB), I2(51.2 dB), I3(50.2 dB), I4(55.1 dB), I5(52.3 dB) were exterminated with Gaussian filter. valid information is restored in endothelium images. MSE and PSNR values as shown in Table 2 is accomplished with following Eqs. (8) and (9).

𝑀𝑀𝑀𝑀𝑀𝑀 = ∑ ∑ [𝐼𝐼(𝑖𝑖, 𝑗𝑗) − 𝐼𝐼|(𝑖𝑖, 𝑖𝑖)𝑧𝑧−1𝑗𝑗=0 ]𝑦𝑦=1

𝑖𝑖=0 (8)

𝑃𝑃𝑀𝑀𝑃𝑃𝑃𝑃 𝑑𝑑𝑑𝑑 = 10 𝑙𝑙𝑙𝑙𝑙𝑙 10 2552

𝑀𝑀𝑀𝑀𝑀𝑀 (9)

Number of rows (y)

Number of columns (z).

The input of the image I (i, j).

The output of the image I' (i, j).

Table 3. shows complete processing stages of data set 2 (I6-I10) processed images like that of data set 1(I1-I5).

Table 2. PSNR vs. MSE values of (I1-I5) images. Serial no. PSNR value dB MSE value

I1 46.8 0.28 I2 51.2 0.38 I3 50.2 0.32 I4 55.1 0.48 I5 52.3 0.40

Table 3. Pre-processing and processing stages of (data set 2).

Serial no.

Segmented image

Histogram of the image

Gaussian filtered image

Auto metric threshold

Particle filtered image

Cd cells/mm2

I6

1941



I7

2236

I8

518

I9

604

I10

529



Considering Fig. 4, the mean value of two datasets out of ‘5’ images, I6 (0.77 mm2), I7 (0.85 mm2), I8 (0.42 mm2), I9 (0.44 mm2), I10 (0.50 mm2) are close to the standard value, i.e., 0.8 mm2 for abnormal images, SD: I6 (2.71 mm2), I7 (2.84 mm2), I8 (2.03 mm2), I9 (2.11 mm2), I10 (2.22 mm2) are also close to the standard value < ‘2.80’ mm2 for abnormal images. The area size of the acquired image is 0.2611*105, remains unchanged during processing the images. As the images are quadratically interpolated from I1 image to I5 image.

Figure 5 represents the result of the auto-metric threshold, the lower limit is ‘10’ and the upper limit is ‘245’ the images I1 (119), I2 (117), I3 (117), I4 (118), I5 (118) are the resultant metric threshold outputs, which ranges from ‘117-119’ below this range concludes dark object identification. Above specified range is bright object identification, which has never achieved in I1-I5.

Fig. 4. Variation of mean, SD, area for I6-I10.

Fig 5. Auto metric threshold of I6-I10.

0.77 0.85

0.42 0.44 0.5

2.712.84

2.03 2.11 2.22

0.2611 0.2611 0.2611 0.2611 0.2611

0

0.5

1

1.5

2

2.5

3

I6 I7 I8 I9 I10

Mean value Standard Deviation Area mm2

10 10 10 10 10

245 245 245 245 245

119 117 117 118 118

0

50

100

150

200

250

300

I6 I7 I8 I9 I10

lower limit upper limit lower value



Figure 6 shows comparative cell density with existing methods to the proposed algorithm. NAVIS is the software used to estimate cell density. Manual cell density is the standard measurement carried by ophthalmologists. ‘MCD’ value considered as the reference value. The particle filter is the proposed automatic algorithm to estimate cell density. In this study ‘10’ images are processed simultaneously through NAVIS software and Particle filter algorithm, both resultants are compared with standard measurement method, i.e., MCD.

Table 4 shows images(I1-I10) cell density comparison. For I1, I2 NAVIS cell density is very close to the manual cell density when compared with proposed particle filter cell density. For the remaining images (I3-I10) particle filter cell density is very close to the manual cell density. When compared with existing software particle filter algorithm shows superior results than NAVIS cell density approach.

PF error deviation value is compared with MCD as shown in Table 4. In I1 to I10 images I3, I5, I8, I9 harvested below 10% error. I1, I2 harvested below 15% error, I4, I6, I7, I10 harvested above 20% error due to misalignment of Z-ring stabilizer in CM.

Fig. 6. Cell density deviation between MCD, NCD, PFCD.

Table 4. Various statistical parameters of (I1-I10).

Images NAVIS density cells/mm2

Manual density (MCD) cells/mm2

Particle filter density cells/mm2

Error Elongation factor

(α)

Compactness factor

(δ)

Heywood circularity

factor (ζ)

I1 1403 1335 1184 0.11 3.04 0.67 1.21 I2 3073 3033 2621 0.13 1.06 0.85 1.11 I3 2633 2404 2455 0.10 1.23 0.88 1.13 I4 2276 546 324 0.40 2.65 0.61 1.29 I5 1836 633 601 0.05 2.50 0.65 1.23 I6 2008 554 1941 0.69 2.34 0.66 1.15 I7 2399 456 2236 3.90 2.43 0.65 1.18 I8 2000 526 518 0.01 2.44 0.64 1.24 I9 2108 653 604 0.07 2.77 0.65 1.26 I10 1978 863 529 0.38 2.42 0.67 1.18

0

500

1000

1500

2000

2500

3000

3500

I1 I2 I3 I4 I5 I6 I7 I8 I9 I10

Manual cell density Navis cell density Particle filter cell density



The main advantage of the particle filter algorithm is each individual cell contour cell parameters are analysed along with cell density. PF identified each individual endothelial cell of EF, CF, HCF. The healthy endothelial cell has a perfect hexagonal structure, FD’s, AFD, PPCD, ICD causes irregular cell shape, which results in cell elongation (or) cell compactness (or) irregular circularity shape. Figure 7 shows (I1-I10) images average cell EF, average cell CF, average cell HCF values. EF of I2 (1.06), I3 (1.23) and CF is I2 (0.85), I3 (0.88), HCF is I2 (1.11), I3 (1.13), which are normal range. EF of I1 (3.04), I4 (2.65), I5 (2.50), I6 (2.34), I7 (2.43), I8 (2.44), I9 (2.77), I10 (2.42), CF and HCF for (I1, I4-I10) are tabulated in Table 5. Due to any abnormalities (FD and AFD) the EF value increases CF and HCF values decrease significantly.

Heywood circularity factor determines the dimensions of the individual cell, as per its classification, If (HCF) = 1, Perfect circle. Else HCF >1 (or) HCF<1, Polygon cell structure.

For endothelial cells, HCF should be >1. If it gets equal to 1 the cell possesses exact circular in shape, which is ideally not possible in endothelial cell structure.

Endothelium cells are in hexagonal shape, sometimes cell structure may differ when cells are infected with FD, AFD, ICD. The deformation of cell structure enhances and widely spreads due to dystrophy attack. Elongation may be more in this type of cell structure. Elongation factor reveals the healthy endothelium cell structure.

As per its classification, If the elongation factor lies within the lower limit (1) and upper limit (3) values,

i.e., 1 >EF ≤ 2 represents healthy endothelial cells. EF < 1 or EF >2 represents abnormal Endothelial cells

In this paper, the average EF of the cell is calculated from each image in two data sets. Total of 10 images average EF is calculated and found abnormal due to FD, AFD, PPCD.

Fig. 7. EF, CF, HCF of (I1-I10).

0

0.5

1

1.5

2

2.5

3

3.5

I1 I2 I3 I4 I5 I6 I7 I8 I9 I10

Elongation factor Compactness factor Heywood circularity factor



PF is an effective filter to analyse the endothelium cell particles in an image. In this work, all endothelium cell particles in an image parameter range from 1 to 2.4. ‘1’ is min cell diameter ‘2.4’ is the maximum cell diameter. All the particle cell values range in between 1 to 2.4 as shown in Table 5. I2 minimum value is 0.75 and maximum value is 1.6, I3 minimum value is 0.94 and the maximum value is 1.63 for normal images. Remaining abnormal images particle parameter ranges from 1.73 to 2.87.

Three images are investigated with existing methods MO1, MO2, and S-PSO. Resultant error is compared with the PF algorithm, Comparative plot is shown in Fig. 8. The error is drastically negligible in the PF algorithm and it produced favourable results and supportive hand for clinical examinations to ophthalmologists and clinicians. Along with significant accuracy, PF algorithm processing time is reduced, i.e., the average processing time is 18.49 m/s.

Table 5. Particle filter.

Serial no. Parameter range Current parameter

Mean value Min value Max value Min value Max value

I1 1 2.4 1.05 1.92 1.21 I2 1 2.4 0.75 1.6 1.21 I3 1 2.4 0.94 1.63 1.18 I4 1 2.4 1.02 1,64 1.22 I5 1 2.4 1.05 2.23 1.23 I6 1 2.4 1.04 2.21 1.15 I7 1 2.4 1.02 2.87 1.19 I8 1 2.4 1.06 1.73 1.24 I9 1 2.4 1.08 2.08 1.26

I10 1 2.4 1.04 1.70 1.18

Fig. 8. Comparison of maximum errors with MO1 vs. MO2, S-PSO and PA [1].

13.1912.28

10.96

22.2420.93

18.23

12.2

9.557.45

10.02

7.5 6.7

0

5

10

15

20

25

I1 I2 I3

MO1 MO2 S-PSO PF



5. Conclusions This scientific study we report results of ‘10’ abnormal images ‘3’ normal images are processed with a PF algorithm to estimate cell contour structure and Endothelium Cell Density (ECD). The scope of this study is very much helpful to ophthalmologists/clinicians for accurate diagnosis of disease in other corneal layer and it accomplishes healthy cornea for the human eye. The ‘10’ abnormal images include Fuch’s Dystrophic images, AFD images, and Posterior Polymorphous Corneal Dystrophy (PPCD) images and Iridocorneal Dystrophy (ICD) images. All images are processed through PF, and able to extract clinical parameters such as ECD, minimum cell area, maximum cell area, average cell area, EF, CF, HCF.

Proposed algorithm main advantage is its ability to get each individual cell profiling in an image instead of collective data. The PF algorithm is compared with existing method S-PSO in all statistical level. PF algorithm gives depth vision and each cell contour and pleomorphism and polymegethism of the cell. All these statistical indexes could help clinicians/ophthalmologists to better analysis of dystrophies. The developed system can process a large number of images with instant statistical indexes like optical coherent tomographers, corneal topographers. The processing time may reduce drastically with better efficiency. Usually, to process the image, it takes 10 to 12 minutes time. With advanced algorithms like Snake-Particle Swarm Optimization (S-PSO) approaches, it takes 2-3 minutes of time. With the PF algorithm the average time is less than 20 milliseconds and it is also providing a clear graphical representation, which is helpful to the patients for better understanding and better explanation by the clinicians.

Along with the above quantitative parameters, the algorithm processing time is reduced drastically. The ‘10’ images are segmented and applied Gaussian filter to eliminate high-level frequency noise components. With an Auto-metric threshold and particle filter cell density is meticulously estimated. The estimated parameters are compared with existing methodologies MO1, MO2, S-PSO, NAVIS cell density, and the PF algorithm achieved better results than existing methods. PF also estimates EF, CF, HCF parameters, which are more favourable for ophthalmologists/clinicians to accurately diagnosis the condition of the cornea. The average processing time ‘18.49’ m/s is achieved with PF algorithm. The errors deviation 12.14%, 20.46%, 9.73%, of MO1, MO2, and S-PSO, is dominated 7.75% with PF and considerable accuracy is achieved. PF algorithm is a fully automatic algorithm and a considerable amount of time is reduced for ophthalmologists/clinicians in busy clinical scheduling time in hospitals. PF algorithm is very much helpful to diagnosis FD’s, AFD’s, ICD, PPCD and condition of the cornea in an effective manner. Further, the research is carried out for analysis of diseases encountered in other corneal layers like stroma layer, Descemet's layer, Bowman’s layer, and Descemet's layer. The scope of work is very much helpful to ophthalmologists/clinicians for accurate diagnosis of disease in other corneal layer and it accomplishes healthy cornea for the human eye.

Nomenclatures J Threshold value 𝑃𝑃𝑤𝑤𝑡𝑡−1𝑘𝑘 Weight (or) pixel of the particle

Vt New action movement to forward



Xt New pixel measurement 𝑦𝑦𝑡𝑡−1

| Previous state yx Height of the image yy Width of the image zx Height of the filtered image zy Width of the filtered image Greek Symbols γ Image contrast/brightness λ Pixel quantity of smoothness σi Average of all pixel quantities Abbreviations

AFD Advanced Fuch’s Dystrophy CD Cell Density CF Compactness Factor ECD Endothelium Cell Density EF Elongation Factor FD Fuch’s Dystrophy FECD Fuch’s Endothelium Cell Density HA Hybrid Algorithm HCF Heywood Circulatory Factor HSL Hue, Saturation Luminance I Image ICD Irido Corneal Dystrophy IDO Identified Dark Object MO1 & MO2

Morphological Operations 1 & 2

MSE Mean Square Error PF Particle Filter PPCD Posterior Polymorphous Corneal Dystrophy PSNR Peak Signal to Noise Ratio QI Qubic Plane Interpolation ROI Region of Interest SM Specular Microscope S-PSO Snake Particle Swarm Optimization

References 1. Sharif, M.S.; Qahwaji, R.; Shahamatnia, E.; Alzubaidi, R.; Ipson, S.; and

Brahma, A. (2015). An efficient intelligent analysis system for confocal corneal endothelium images. Journal of Computer Methods Programs in Biomedicine, 122(3), 421-436.

2. Morgan, S.R. (2014). Understanding the structural basis of corneal refractive function and its modification via novel therapeutic approaches. Ph.D. Thesis. College of Biomedical and Life Sciences, Cardiff University, Cardiff, Wales, United Kingdom.

3. Klotz, S.A.; Penn, C.C.; Negvesky, G.J.; and Butrus, S.I. (2000). Fungal and parasitic infections of the eye. Clinical Microbiology Review, 13(4), 662-685.



4. Johnson, Z.K.; Narayanan, M.; Siah, W.F.; Anwar, H.; and Figueiredo, F.C. (2015). An analysis of the incidence, management, clinical outcomes and risk factors of Acanthamoeba keratitis infections in a tertiary hospital in the UK over the last 5 years. Investigative Ophthalmology & Visual Science, 56(7), 1889.

5. Guthoff, R.F.; Zhivov, A.; and Stachs, O. (2009). In vivo confocal microscopy, an inner vision of the cornea - a major review. Clinical and Experimental Ophthalmology, 37(1) 100-117.

6. Niederer, R.L.; and McGhee, C.N.J. (2010) Clinical in vivo confocal microscopy of the human cornea in health and disease. Progress in Retinal and Eye Research, 29(1), 30-58.

7. Chiou, A.G.-Y.; Kaufman, S.C.; Kaufman, H.E.; Beuerman, R.W. (2006). Clinical corneal confocal microscopy. Survey of Ophthalmology, 51(5), 482-500.

8. Grisan, E.; Paviotti, A.; Laurenti, N.; and Ruggeri, A. (2005). A lattice estimation approach for the automatic evaluation of corneal endothelium density. in Engineering in Medicine and Biology Society. Proceedings of the IEEE 27th Annual Conference in Engineering in Medicine and Biology. Shanghai, China, 1700-1703.

9. Tervo, T.; and Moilanen, J. (2003). In vivo confocal microscopy for evaluation of wound healing following corneal refractive surgery. Progress in Retinal and Eye Research, 22(3) 339-358.

10. Ruggeri, A.; Grisan, E.; and Jaroszewski, J. (2005). A new system for the automatic estimation of endothelial cell density in donor corneas. The British Journal Ophthalmology. 89(3), 306-311.

11. Salerno, M.; Sargeni, F.; Bonaiuto, V.; Amerini, P.; Cerulli, L.; and Ricci, F. (1998). A new CNN based tool for an automated morphometry analysis of the corneal endothelium. Proceedings of the Fifth IEEE International Workshop on Cellular Neural Networks and Their Applications. London, United Kingdom, 169-174.

12. Foracchia, M.; and Ruggeri, A. (2000). Cell contour detection in corneal endothelium in-vivo microscopy. Proceedings of the 22nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society. Chicago, Illinois, United States of America, 1033-1035.

Appendix A

Computer Programme

A.1. Programme flow The proposed algorithm in this paper uses a PF as shown in Fig. A-1. In this paper, the same statistical indexes are projected with better efficiency along with Endothelium EF, CF, HCF.

Dark edges and noise intervened from the endothelium images acquired from the CM-4. The Gaussian filter has been applied to the images in order to eliminate the noise components. Mean square error (MSE) and Peak signal to noise ratio (PSNR) estimated for all the data sets of the images. The variation in geometric sizes in the image is smoothened with Quadratic interpolation method. Then the



size was normalized to (885*664). Smoothing the image improve the analysis further. Bright structural view of the endothelium cells has been harvested with histogram equalization. Mean and standard deviation estimated for image brightness analysis. The image is further processed to transform the low contrast level pixels into high contrast level pixels using Logarithmic mapping. Hue saturation Luminance (HSL) applied on the image for adjusting the brightness and intensity levels.

Fig. A-1. Main Flow Chart of the Particle Filter Algorithm used in this study.

Pre-processing

Image acquisition

Pre-sampling image (Quadratic interpolation)

Gaussian filter (kernel size ‘3’)

Histogram equalization

Processing

Logarithmic mapping

Hue saturation luminance Plane (colour plane extraction)

Convolution (Kernal size ‘15’

Auto metric thresholding (Dark object identification)

Particle filter (Minimum value = 0, Maximum value = 1.9

Endothelium particle analysis



Fig. A-2. A Flow chart for Automatic Metric Threshold.

Curtail pixel quantities from undistinguished intensity of ROI

Curtail medium intensity of ROI

from pixel quantities

Corrected endothelium image

Metric clustering

Binary image inversion

Looking for dark objects

Confocal Endothelium image

confocal corneal endothelium dystrophy’s analysis …

Documents