2.1.

Optical Design for Fundus Camera

In our designed fundus camera, the oblique illumination was created by the transmission of light through the sclera using two LEDs with an 850-nm nominal wavelength and spectral bandwidth of 30 nm full-width at half-maximum. The light that reaches the retina is not direct; only a part of the light is transmitted through the sclera (40%) and is in accordance with ISO eye safety standards for light safety limits. The two near-infrared (NIR)-LEDs could be illuminated at the same time or separately to get an image with one or both illumination points. The LED lights were placed at desired angle positions to avoid illumination via the pupil and to create an oblique illumination through the sclera. Figure 1(a) shows the basic design of the fabricated fundus camera, and Fig. 1(b) demonstrates the schematic illustration of illumination used in the fabricated fundus camera to get non-mydriatic wide-field fundus photography. The beam angle for each LED was around 20 deg (0.35 rad). When both LEDs were used, the light passing through the sclera was diffused, and it illuminated the intraocular area homogenously. The illumination light path in the eyes is shown in Fig. 1(c). The white bright part enclosed in the red circle in Fig. 1(c) shows the illumination area. Two small dots on the pupil are the reflection from the ophthalmic lens. The pupil size increased as the eyeshade covered the eye. The pupil size in Fig. 1(c) is 5.4 mm, and as the camera came closer to the eye, the pupil size increased to $\sim 7\mathrm{mm}$ in diameter [Fig. 1(d)], showing that it is nonmydriatic; the system requires no drug for pupil dilation and will be less traumatizing to the patients. For demonstration purposes, the author used her eye to show the illumination path through the sclera.

Fig. 1

(a) Basic optical design of the fabricated wide angled fundus camera, (b) schematic illustration of trans-scleral illumination using IR-LEDs, (c) photograph of the eye showing the illumination light path through the sclera and pupil size, (d) photograph of eye showing a bigger pupil size of 7 mm in front of the camera without using any pupil dilation drug, (e) photograph of the prototype fundus camera, and (f) photograph showing the self-imaging capability of the fundus camera.

Figure 1(e) shows a photograph of the fabricated prototype fundus camera. For the ergonomic design, portable fundus camera components were placed in a 3D-printed ABS plastic enclosure designed in SolidWorks (Dassault Systemes) and printed using a 3D printer (Flash Forge Inventor II). A black-colored eyeshade with a rubber eyepiece was attached to the printed enclosure to allow for a natural pupil opening during NIR imaging. The camera is palm sized and lightweight. For wide-field photography, a 78D ophthalmic lens (Righton, Japan) was used to collect light coming out of the pupil. An achromatic lens (lens 2) was used to relay the fundus image on the sensor of the camera. A small web-camera (Logitech HD Pro webcam C920) with full HD 1080p video recording capability was used to get fundus images. The web-camera was modified for NIR–imaging by replacing the color filter on the camera sensor to an NIR filter. Our fabricated portable fundus camera prototype was able to be used in a handheld and portable manner by connecting the camera to the laptop computer. The still-image or live-video can be recorded by positioning the camera in front of and close to the eye with the person looking directly into the camera. The imaging position can easily be checked by looking at the computer screen by oneself. Figure 1(f) shows the self-imaging capability of the fundus camera.

The field-of-view with respect to the internal eye angle was measured using a fabricated model eye. According to the reported literature, the equatorial diameter of the retina in the human eye is about 24 mm, and at ora-serrata (end of retinal tissue), the average internal diameter of retina is about 20 mm.³⁶ Thus, a 180 deg visual angle field-of-view corresponds to the whole retinal area of $1148{\mathrm{mm}}^2$. Figure 2(a) is the schematic diagram of the eye showing the peripheral retina, ora-serrata, and total internal eye angle for the retina (241 deg). The size of the model eye used for measurement was 24 mm in diameter, and we used an achromatic lens ($f=24\mathrm{mm}$) as the cornea. Figure 2(b) is the photograph of the model eye. To show the field-of-view of the fabricated fundus camera, we pasted a measurement scale in millimeters on the posterior surface of the model eye. The image taken from the fundus camera is shown in Fig. 2(c). The marked dotted line shows the boundary of the equator of the model eye. To measure the maximum angular field-of-view, the clearest and most readable marker on the vertical (50 mm) and horizontal (45 mm) scales was determined. The average readable scale on either side was taken as 47.5 mm, and the eye angle was calculated. The equator of the globe and line representing the ora serrata (4 mm above the equatorial diameter) is clearly visible. According to the reported literature, the visual angle and eye angle can be converted to each other.^37,38 Using the conversion factor³⁸ (1.37) for the eye angle to the visual angle, the measured angular field-of-view for the fundus camera was 165.6 deg.

Fig. 2

(a) Schematic illustration of the eye showing the locations of peripheral retina, ora-serrata, and total internal angle for retina, (b) photographs of the model eye, and (c) image of the model eye captured by the fabricated fundus camera showing a 165.6 deg ultra-wide angular field-of-view. The dotted line shows the equator of the globe.

Enucleated swine eyes were used to check the imaging ability of the fabricated fundus camera. The enucleated swine eyes for the experiment were bought from Sendai Central Meat Whole Sale Market Co. Ltd. For experiments using animal parts, we followed the regulations specified by Tohoku University. For initial imaging, we used an enucleated swine eye. The depth measurement results obtained from our fabricated device were compared with those obtained from a commercial OCT device by NIDEK company (Model -RS-3000 advance). The axial resolution of the employed OCT device was $4\mu \mathrm{m}$, and the operating wavelength was 880 nm. The stage of the OCT was modified for animal eye imaging. The eye was fixed on a sample holder that mimicked the human face, and the camera was placed in front of and close to the eye. The eye was illuminated through the sclera by two NIR LEDs attached on the outer cover of the ophthalmic lens at pre-specified angled positions.

3.1.

Near-Infrared Fundus Imaging with Portable Fundus Camera

The color fundus image is considered the gold standard for ophthalmic documentation, but NIR/IR imaging has some advantages over colored fundus imaging.³⁹ Oxygenated hemoglobin, hemoglobin, and water are the main fundus molecules that absorb IR light, and melanin is a strong reflector. NIR imaging is possible with less light as NIR can penetrate into deeper layers. IR reflectance imaging provides information on alterations in the retinal structure, sub-retinal lesions, and accumulation of fluid within and beneath the retina and the retinal epithelium. Another potential benefit of IR illumination is that it minimizes the baffling effects of intra-individual differences in the pigmentation.⁴⁰ Moreover, NIR/IR imaging can help detect chronic changes even with the presence of a hemorrhage in the eye that may be undetected on other imaging systems.

The LED light operation can be controlled separately for oblique illumination to get quasi 3D images of the retina and optic disc. Figure 3(a) shows the photograph of an enucleated swine eye (sample 1) used for imaging. Figure 3(b) shows the large field-of-view coverage by the constructed fundus camera. The optic disc and choroidal vessels are clearly visible. Imaging of choroidal vessels is important to check the eye’s health condition. The choroidal thickness varies among the disease states and changes with increasing age and axial length. Choroidal vessels are seen instantly with NIR imaging compared with other wavelengths.⁴¹ Figure 3(c) gives another image (sample 2) taken by the same fundus camera. Blood vessels in and around the optic disc are visible.

Fig. 3

(a) Photograph of the enucleated swine eye showing the pupil size, (b) image of the retina of this swine eye (sample 1) showing the optic disc and choroidal vessels, (c) retina of another enucleated swine eye (sample 2) showing the blood vessels and optic disc, and (d) photograph of sample 2 after removing the cornea and lens to verify the operation of the fundus camera.

To confirm the reliability and ultra-wide field imaging capability of our fundus camera, we removed the cornea and crystalline lens of the enucleated swine eye and took the photo of the cut eye by a smart phone camera. Figure 3(d) is the photograph of the enucleated swine eye taken after removing the cornea and crystalline lens. Figures 3(c) and 3(d) look similar in features, confirming ultra-wide field coverage by our fundus camera. The contrast of the blood vessels is not very high in the NIR image. The reason for variation is that there are several factors that influence the appearance and size of retinal vessels as a function of wavelength. The most important factor is the optical absorption by blood columns and vessels, as they have low contrast against the fundus background for wavelengths longer than 600 nm. Another reason is that, with NIR illumination, there is scattering multiple times over a wide region of the fundus, so only heavily pigmented structures are visible.

The scattered light that acted as noise in the image was removed using an aperture at the pupil conjugate plane. Figure 4 shows the various images taken with and without the aperture. Figure 4(a) shows an image taken without the aperture. Figures 4(b) and 4(c) are the images taken with varying sizes of the aperture. Images became clearer and contrast in the image was improved by reducing the noise.

Fig. 4

(a) Retina image without using any aperture. (b) and (c) Images of retina taken with varying sizes of aperture.

3.2.

Oblique Illumination and Imaging

Oblique illumination is a well-known technique that has been described by many scientists to improve the contrast of transparent specimens by introducing a pseudo-relief effect.^42,43 In oblique or anaxial illumination, the specimen is unevenly illuminated, and it can be interpreted in the context of Abbe’s theory of diffraction.⁴² According to Abbe’s theory, if a narrow axial illuminating beam is used instead of broad illuminating beams, then directionally shadowed differential phase contrast images that are very similar to differential contrast images are produced. In our designed fundus camera, using an LED with a narrow beam angle of 20 deg, anaxial illumination was created at the retinal plane.

When both LEDs are illuminated at the same time, a wide field-of-view coverage was obtained. By changing another lens with a slightly narrower field-of-view, the area near the optic disc was observed more clearly. With the use of only one-sided illumination, quasi 3D images were obtained. Figure 5(a) shows the wide-angled image of the swine eye’s retina. Figures 5(b) and 5(c) are the quasi 3D images obtained using illumination from only one LED. Figure 5(d) shows more detailed features around the optic disc, and Fig. 5(e) is the corresponding quasi 3D image of the pig’s retina with a slightly narrow field-of-view. This quasi 3D effect itself can be brought about in a gradual manner without having to replace the LED. For the detection of glaucoma, images of the optic disc are very important. The shape of the optic disc in glaucomatous eye changes with progression of the disease.⁴⁴ The quasi 3D effect is more prominent in Figs. 5(c) and 5(e). The depth of field is greater in both quasi 3D images compared with the image in Fig. 5(a), where the depth of the optic disc of the swine eye can be estimated by the shadow length measurement.

Fig. 5

(a) Wide-angled image of the swine eye. (b) and (c) Quasi 3D images obtained using oblique illumination. (d) Narrow-field image showing the detailed area near the optic disc and (e) corresponding quasi 3D image of the narrow-field image.

3.3.

Measurement of Optic Disc’s Depth

We measured the depth of the optic disc from the shadow created by oblique angle illumination. Generally, the method that we used in the present study is applied for the measurement of the depth of a crater in planetary bodies. This method is complementary to other methods as it requires only one image to measure the depth. The shape of the crater is considered to be a paraboloid. When the light (for example sunlight) falls on the crater, a shadow is formed inside the crater. The size of the shadow or the shadow-length (L) depends on the incident angle of sun rays, i.e., the illumination angle ($\theta$), and it changes from larger to smaller with increasing $\theta$. The shadow disappears when the light falls directly over the crater, i.e., $\theta =90\mathrm{deg}$. The depth of the crater “$d$” is measured by measuring the length of the shadow using the following equation:⁴⁵

Fig. 6

(a) Schematic of depth measurement setup using a model eye and oblique illumination, (b) schematic of the pattern resembling the optic disc inside the model eye, $D$ is the diameter of the pattern, $d$ is the depth, and $\mathrm{\Theta }$ is the illumination angle (for this measurement, $O=45\mathrm{deg}$). The diameter of the pattern resembling the optic disc was 3 mm. (c) An overhead view of the shadow created by the illumination angle, where $L$ is the length of the shadow. (d) Image of the model eye with the pattern resembling the optic disc taken by the fundus camera, where the diameter of the structure resembling a crater is 3 mm. (d)(ii) Enhanced image of the area marked on panel d(i), showing the shadow created by oblique illumination and mathematical calculation of the shadow length.

The NIR light was chosen in view of its lower absorption compared with visible light. On the other hand, $O=45\mathrm{deg}$ allows for direct imaging of the optic disc and its shadow through the eye lens. The calculation of the depth becomes simple when the incident angle is 45 deg as the trigonometric tangent function ($\tan \mathrm{\Theta }$) value varies sharply with slight changes in the angle. But with tan 45 deg, a variation of $\pm 5\mathrm{deg}$ does not change the calculation values significantly. So minor a change in angle will not change the depth measurement values. [Fig. 6(c)]. Based on the geometry of our model eye’s optic disc and the illumination angle ($\mathrm{\Theta }$), the formation of shadow and its length can be predicted from Eq. (1). The expected length ($L$) of the shadow should be half of the diameter of the optic disc [as shown in Fig. 6(c)]. Figure 6(d) shows the mathematical calculation of the shadow length on the model eye. The dotted area marked in Fig. 6(d)(i) was enhanced using imageJ to clearly see the boundaries of the shadow. The surface of the model eye is rough as it was printed by a 3D printer. Figure 6(d)(ii) shows the enhanced image with marked dimensions for calculating the shadow length. The distance between $A$ ($x=-R$) and $B$ ($x=X_c+R$) is the shadow length ($L$), where $R$ is radius of the cup like structure. The calculated depth using Eq. (1) matched with the designed depth.

In digital photography, it is often difficult to recognize shadows in an image. Shadow regions are assumed to be formed when there is obstruction along the illumination path, resulting in changes of image intensities. A shadow consists of two regions, umbra and penumbra.⁴⁶ On non-flat/curved surfaces, shadows with wide boundaries, known as penumbrae, are observed. So it is important to detect shadows accurately in digital images while maintaining original local and textural information. Several methods for the detection of shadow have been reported.^47,48 Clustering-based algorithms can partition image pixels into a certain number of similar/dissimilar groups and can be applied for the detection of shadows in an image. K-means clustering is one of the most popular algorithms,⁴⁹ and it has already been implemented in the Open CV2 module of the Python 3 programing language. We used it for the detection of shadow in the present study. Before applying K-means clustering, the colored image [as shown in Fig. 7(a)] was converted to an 8-bit image, and pixel intensity was normalized in the range of 0 to 255. Contrast limited adaptive histogram equalization was applied to the image to make it clearer [Fig. 7(b)].

Fig. 7

(a) Original RGB image of the optic disc like pattern of the model eye, (b) intensity normalized image of the original image of the model eye, (c) image obtained with the K-means clustering algorithm, and (d) differential K-means image to calculate the shadow length.

K-means clustering was performed by considering the different numbers of clusters. Iterating the image for 1000 times with the number clusters in the range of $\sim 10$ to 20 is good enough for detecting the shadow of the optic disc in the image. For example, the image shown in Fig. 7(c) is obtained by considering a total of 10 clusters. To precisely measure the shadow length, the image obtained with K-means clustering was differentiated [Fig. 7(d)]. The area marked by the dotted circle in Fig. 7(d) shows the actual shadow region consisting of the umbra and penumbra regions. The length of the shadow measured in the center is $\sim 1.49\pm 0.02\mathrm{mm}$. The depth calculated using Eq. (1) is 1.50 mm, which matches well with the designed depth of the model eye. By varying the incident angle to $45\pm 0.5\mathrm{deg}$, the measured results showed good accordance with the actual value. These results clearly confirmed that it is possible to know the depth of an optic disc from the shadow.

For qualitative analysis and to compare the depth measurement results with optical coherent topography (OCT), we imaged enucleated swine eyes using the OCT device. Figure 8 shows OCT images of three pig eye samples with measured depths at different cross-sections. The arrows in Figs. 8(a), 8(d), and 8(g) denote the location and direction of the cross-sectional OCT scan. Figures 8(b), 8(e), and 8(h) are cross-sectional images of the optic nerve head. The measured depth of these samples is shown in Figs. 8(c), 8(f), and 8(i). The depth at different cross-sectional points varied from 183 to $490\mu \mathrm{m}$ for various enucleated pig eyes.

Fig. 8

(a)–(c) Optic disc images of three different enucleated swine eyes. (b), (e), and (h) The cross-sectional images of the optic nerve head taken in the directions and positions of the arrow marked in panels (a), (b), and (c). The depths at the marked arrows are shown in panels (c), (f), and (i).

Figure 9(a) shows the cropped image of the optic disc of an enucleated swine eye taken using angled illumination. A shadow in the optic disc that was created by oblique illumination can be easily observed in the image. The shape of the optic disc of the swine eye is not circular or paraboloid; it is elliptical in shape. The shadow obtained by imaging the pig’s optic disc fits well with an ellipse. For elliptical shapes, the depth from the shadow is calculated using following equation:⁴⁵

Fig. 9

(a) Cropped original RGB image of the enucleated swine eye obtained using the fundus camera. (b) Intensity normalized image of the original RGB image. (c) Image obtained with the K-means clustering algorithm and (d) differential K-means image to calculate the shadow length.

Figure 9(b) is the intensity normalized image with 256 levels of intensity. Similar to the model-eye analysis, to detect the shadow area more clearly, the K-means clustering algorithm was applied to the normalized swine eye image. Figure 9(c) is the image obtained after performing 1000 iterations of the clustering algorithm. As a result, the shadow clusters merged among other color clusters with identical normalized intensity. The clusters with lower intensity values belong to the shadow. To accurately measure the length of the shadow, the image obtained after K-means clustering was differentiated. Figure 9(d) is the differential K-means image. The area marked with dotted circles (1) and (2) are the shadow regions. Dotted circle (1) is the shadow area inside the optic disc that was used to measure the shadow length. To compare the results with the OCT measurement, the shadow length was measured at different cross-sections similar to the OCT scan. The diameter of the optic disc of the swine eye was measured manually by dissecting the eye and using an optical microscope. The diameter was about 2.1 mm. The angle of oblique illumination for this image was $45\mathrm{deg}\pm 5\mathrm{deg}$. Using Eq. (2), the depth of the optic disc at various cross-sections was calculated. Table 1 shows a comparison of the actual values obtained with the OCT device and our proposed method. We used five enucleated pig eyes for measuring the depth of the optic disc with the OCT device. The same eyes were used to measure the depth of the optic disc with our proposed method. The depth values varied from 0.179 to 0.350 mm, which are in good accordance with depth values obtained from OCT measurements. The precision of a set of measurements can be determined by calculating the standard deviation for a set of data. In our proposed method, we calculated the standard deviation of the shadow length measurements. The standard deviation value was 0.028 mm for each set of shadow length measurements. In each set, 11 values were taken.

Table 1

Comparison of depth values of the optic disc obtained by the OCT device and proposed method.

The depth of the elliptical/parabolic optic disc can be successfully estimated using this simple method. These depth values are useful for calculating the cup volume and area. According to the reported literature, the shape and dimensions of structures can be obtained from the cast shadow.⁵⁰ The shapes obtained from the shadows can be used for 3D surface reconstruction.⁵¹ Estimating the shape and 3D reconstruction of the optic disc could be helpful in detecting the progression of various eye diseases, including glaucoma.

3.4.

Light Safety

During screening and regular medical checkups of the eye, radiation may have to be directed toward the eye, which can impose potential photochemical and thermal hazards to the retina. For our prototype, most of the light passes through the sclera and illuminates the retina. For an 850-nm wavelength, transmission through the sclera is $\sim 40\%$ and absorption is $\sim 5\%$.⁵² To ensure the safety of the prototype, photochemical and thermal limits should be according to ISO standards, ISO 15004-2 and ISO 10940 (ISO 15004-2,2007: ISO 10940, 2009). For ocular tissues, we measured light exposure values at the sclera, lens, cornea, and retina and compared them with their respective exposure limits as mentioned in ISO 15004-2, 2007 standards.⁵³ For skin exposure limits, guidelines given in 2013 ICNIRP for limits of exposure to incoherent visible and IR radiation for skin were followed.⁵⁴

4.1.

Limitations

Despite having several advantages of directly obtaining quasi 3D images by oblique illumination and the ease of calculating the depth values from the shadow length using a single image, our proposed method has some limitations. For example, certain structures in the image may easily be misinterpreted when using off-axis illumination alone. Due to the greater depth of field and the presence of the visual cues to depth structure in the images, this method is useful only for the initial screening of the diseases. For practical applications to the clinics, imaging with and without oblique illumination should be analyzed. For further improvement in glaucoma diagnosis, the implementation of artificial intelligence using images obtained by our fundus camera and depth measurement is being considered.