A computer-implemented method for automatically evaluating the open profile of the gonioscopy angle through processing of digital gonioscopy images representing sectors of the anatomical layer of the iris-corneal interface.

Abstract

Automatically evaluates the opening profile of corners. [Solution] A computer implementation method for substantially continuous and automatic evaluation of the angle open profile by processing multiple digital gonioscopy images, typically RGB images, each representing a sector of the anatomical layer of the iris-corneal interface extending 360° as a whole.

Description

Technical Field

[0001] The present invention relates to a computer-implemented method for substantially continuously and automatically evaluating the opening profile of the eye angle by processing anatomical layers of the iris-corneal interface, i.e., digital gonioscopy images (typically RGB images) representing the eye angle.

[0002] In particular, the present invention relates to the use of a deep learning neural network (also referred to as artificial intelligence), and more particularly to a computer-implemented method including a learning stage of the deep learning neural network and a stage of continuously evaluating the opening profile of the eye angle using the trained network. Further, the present invention also relates to a computer processing system configured to continuously and automatically evaluate the opening profile of the eye angle by processing digital gonioscopy images representing the iris-corneal interface layer, and a computer program product configured to perform such evaluation by the system.

Background Art

[0003] Glaucoma is widely recognized as a major cause of irreversible blindness, and currently more than 70 million people worldwide are affected. The onset of this disease may be due to the dysfunction of a specific structure of the eye, i.e., the trabecular meshwork (hereinafter referred to as "TM") that surrounds the iris in the anterior chamber of the eye and is located along the interface with the cornea. The TM regulates intraocular pressure by draining aqueous humor.

[0004] More specifically, the trabecular meshwork separates the iris and the sclera and is located within the eye angle. Exactly, considering an image of the eye angle (also called a gonioscopy image) (assuming the angle is open), the following anatomical components (hereinafter referred to as "layers") are arranged in the following order along the surface of the angle: iris, ciliary body, scleral spur, pigmented trabecular meshwork, non-pigmented trabecular meshwork, Schwalbe line, cornea.

[0005] A cross-sectional view of the iris-corneal interface is shown in Figure 1, and an example of an RGB gonioscopy digital image representing the circumferential sector of the iris-corneal interface is shown in Figure 2. The angular surface is continuous and regular, but there may be a fairly extensive anatomical structure protruding from this surface, namely synechiae (adhesion) or iris adhesion. This can cause angle closure in some or more locations at the iris-corneal interface along the aforementioned sector.

[0006] When the effectiveness of TM decreases, intraocular pressure may increase, and consequently, risk factors may also increase. Gonioscopy, a standard examination method, involves manually manipulating a multifaceted mirror placed in the patient's eye to allow the physician to observe structures present at the iris-corneal interface. The visibility and composition of these structures are crucial for determining risk and etiology. Because this interface is outside the directly observable area, a mirror is required, as shown in Figure 3. However, this system has several limitations, is time-consuming, and requires special training with a long and difficult learning curve.

[0007] In particular, diagnoses based on this type of test are subjective, and different clinicians often present conflicting views.

[0008] Another disadvantage is that acquiring images for follow-up purposes, such as detecting changes over time or performing clinical examinations, presents many problems.

[0009] Furthermore, it is now known that some ophthalmic devices, such as the NIDEK GS-1 ophthalmic device, have been developed to perform semi-automatic gonioscopy, store a large amount of RGB digital images of the iris-corneal interface layer, and provide a complete 360° view as shown in Figure 4. [Overview of the project] [Problems that the invention aims to solve]

[0010] As described above, each RGB digital gonioscopy image represents the iris-corneal interface layer, i.e., the circumferential sector of the corneal angle. In particular, it should be noted that each RGB digital gonioscopy image represents the three-dimensional sector of the iris-corneal interface or corneal angle layer as an actual two-dimensional color image.

[0011] However, algorithms used to process this type of image can only classify the open profile state of the angle with a single value for each quadrant into which the anatomical region is divided during image acquisition. Typically, the anatomical region is divided into four quadrants, each representing a 90° angle at the iris-corneal interface, and therefore known processing algorithms output only four classification values ​​related to the open profile state of the angle in these four quadrants.

[0012] A disadvantage is that, compared to the aforementioned manual method, while semi-automated gonioscopy systems can provide more objective results regarding the open profile of the angle, they cannot identify localized angle closure in situations where the iris is locally adhered to the trabecular meshwork, a condition known as intraocular synechia.

[0013] Therefore, conventional technology requires a semi-automated system that can support diagnostic evaluation by clinicians and enable more accurate and detailed assessment of the angle open profile across the entire 360° of the iris-corneal interface. [Means for solving the problem]

[0014] The present invention aims to overcome the above-mentioned drawbacks and limitations.

[0015] For this purpose, the first object of the present invention is to provide a computer implementation method for automatically analyzing digital gonioscopy images, typically RGB images, which will provide an output that more accurately and in detail evaluates the open profile of the angle over 360° of the entire iris-corneal interface.

[0016] In particular, an object of the present invention is to provide a computer implementation method that can produce results that allow even non-specialist clinicians to improve the current standard examination for evaluating the risk of angle closure, namely gonioscopy. Such closure can cause elevated intraocular pressure and increase the risk of developing glaucoma.

[0017] Another objective of the present invention is to provide a computer-implemented method that enables clinicians, especially non-specialist clinicians, to clearly understand which areas of the 360° iris-corneal interface should be intervened in and how.

[0018] The above objective is achieved by performing the computer implementation method described in claim 1.

[0019] Other features of the computer implementation method covered by the present invention are described in the dependent claims.

[0020] Furthermore, the above objectives can also be achieved through the computer processing system described in claim 12 and the computer program product described in claim 13.

[0021] In addition to the purposes described above, the advantages described below are illustrated only in this specification and are not intended to limit the invention, and will be clarified with reference to the accompanying drawings. [Brief explanation of the drawing]

[0022] [Figure 1] Figure 1 is a schematic diagram showing a partial cross-section of the 360° interface between the iris and the cornea. [Figure 2] Figure 2 shows some examples of RGB digital gonioscopy images representing the layers of the iris-corneal interface. [Figure 3] Figure 3 is a schematic diagram of a prism mirror that allows observation of the iris-corneal interface. [Figure 4]FIG. 4 shows an example of a 360° view of the iris-cornea interface layer of an eye obtained using an ophthalmic device. [Figure 5] FIG. 5 is a diagram schematically showing a learning procedure of a deep neural network. [Figure 6] FIG. 6 is a diagram showing an annotation stage of a learning digital gonioscopy image. [Figure 7] FIG. 7 is an example of a graphic showing the result of substantially continuously automatically evaluating the open profile of a gonion obtained by the method of the present invention. MODE FOR CARRYING OUT THE INVENTION

[0023] As described above, the iris-cornea interface layer, that is, the digital gonioscopy image representing the gonion, is obtained using a semi-automatic ophthalmic device such as the NIDEK GS-1 ophthalmic device, and as shown in FIG. 4, it is possible to define a complete view over the entire 360° of the iris-cornea interface layer of the eye.

[0024] The computer-implemented method of the present invention is configured to substantially continuously automatically evaluate the state of the open profile of the gonion by processing the digital gonioscopy image. Each image represents a circumferential sector over 360° of the iris-cornea interface layer. Preferably, these digital gonioscopy images are in RGB format.

[0025] More specifically, in the RGB digital gonioscopy image, each R, G, and B channel is represented by 1 byte for each color pixel.

[0026] However, the original color digital image representing the iris-cornea interface layer may be defined in a different color model, the digital capacity assigned to each channel of each pixel may be other than 1 byte, and further, it is not excluded that the digital image is obtained as a grayscale image.

[0027] Furthermore, according to a preferred embodiment of the present invention, each digital gonioscopy inspection image is acquired by focus stacking technology. That is, in this technology, according to its definition, images are taken sequentially multiple times for different focal planes for the same composition, and a final image with a deeper depth of field than the conventional technology can be obtained. The aim is to obtain a high-resolution image of the trabecular meshwork region.

[0028] However, this does not rule out the possibility that the aforementioned digital gonioscopy images may be acquired in a single imaging session.

[0029] Regarding the computer implementation method of the present invention, firstly, this method utilizes a deep neural network (DNN), and will be described below based on preferred embodiments.

[0030] Furthermore, the method of the present invention includes a training stage for the deep neural network and an evaluation stage for evaluating the aforementioned profile using the trained deep neural network.

[0031] In particular, the learning stage includes the following series of processes.

[0032] First, the learning stage in the method of the present invention includes the step of acquiring multiple training image sets that represent circumferential sectors of the anatomical layer of the iris-corneal interface, covering a total of 360°.

[0033] In this regard, to optimize the conditional distribution of the open profile state of the angioscopy angle and to optimally train the deep neural network, a set of training digital gonioscopy images is selected in which 33% of the angioscopy angles are open, 35% are occludable, and 32% are occluded.

[0034] Furthermore, during the learning stage, a first preprocessing step is performed on each digital gonioscopy image.

[0035] According to a preferred embodiment of the present invention, the preprocessing step includes a process of rotating the data relating to each digital gonioscopy image so that the training digital gonioscopy images are aligned in the same direction, i.e., the layers of the iris-corneal interface are aligned along the same directional axis (preferably the horizontal direction). Preferably, although not required, each rotated digital gonioscopy image displays the iris layer at the bottom of the image and the corneal layer at the top.

[0036] In this regard, it must also be considered that each digital gonioscopy image is accompanied by data regarding the precise direction of acquisition.

[0037] Alternatively, the direction of acquisition could be roughly determined by referring to the image itself.

[0038] According to a preferred embodiment of the present invention, each digital gonioscopy image represents a circumferential sector corresponding to 22° of the iris-corneal interface, an example of which is shown in Figure 2. Therefore, according to this preferred embodiment, the representation of the angle over the entire 360° of the iris-corneal interface is composed of 16 digital gonioscopy images corresponding to 16 consecutive circumferential sectors of the same iris-corneal interface.

[0039] However, according to other embodiments of the present invention, the case is not excluded in which each digital gonioscopy image represents a circumferential sector of the iris-corneal interface corresponding to an angle smaller or larger than 22°. In this case, naturally, the number of images constituting the entire 360° of the iris-corneal interface will be more or less than 16.

[0040] According to a preferred embodiment of the present invention, the preprocessing step also includes, after rotation, the extraction of a region of interest (ROI) centered on the trabecular meshwork region represented within the same image. This process facilitates the annotation of the image by the clinician, as will be detailed below.

[0041] In particular, in a preferred embodiment of the present invention, the annotation process is carried out in two stages to define a first vector of length X2.

[0042] In the first stage, pre-processed training digital gonioscopy images are presented to the clinician, and as shown in the upper image of Figure 6, the clinician is tasked with manually and graphically marking the transition points between the open and occludable states of the angle, between the occludable and occluded states, between the occluded and occludable states, or between the occludable and open states.

[0043] After the transition point is marked on the digital image, in the second stage of annotation, the first vector is defined by automatically processing the digital gonioscopy image marked by the clinician, as shown in the lower image of Figure 6.

[0044] In particular, the first vector is defined to include a numerical value that identifies the state of the angle's open profile for each of a plurality of consecutive predetermined sampling points. The numerical value corresponds to each sampling point along the directional axis in the digital gonioscopy image and indicates either an open state, a closable state, or a closed state.

[0045] According to a preferred embodiment of the present invention, the sampling points for each digital gonioscopy image are 60 points relative to the first vector. Therefore, according to the preferred embodiment described herein, the total number of sampling points for the entire 360° of the iris-corneal interface is 960 points.

[0046] In this way, it becomes possible to acquire high-density information regarding the open profile of the angle across the entire 360° of the iris-corneal interface, which has the advantage of providing a substantially continuous approximation of the profile.

[0047] However, this does not rule out the possibility that a larger or smaller number of sampling points than those indicated above may be selected.

[0048] More preferably, though not required, the numerical value representing the open state of the corner is set to 0, the numerical value representing the closable state is set to 1, and the numerical value representing the closed state is set to 2. However, even in this case, the possibility that the numerical values ​​assigned to each of the three states above may differ from the values ​​mentioned above cannot be ruled out.

[0049] In any case, the first annotation vector represents the "ground truth (correct data)" for the target digital gonioscopy image.

[0050] According to the method of the present invention, an annotation file containing a corresponding first vector is created for each rotated training digital gonioscopy image. Then, each rotated training digital gonioscopy image and the corresponding annotation file are supplied as input data to a deep neural network, and the network is trained.

[0051] However, according to a preferred embodiment of the present invention, a second preprocessing step is performed on each of the rotated training digital gonioscopy images before inputting them into the deep neural network. This second preprocessing step includes the following: Noise filtering; Adaptive flattening of histograms of digital gonioscopy images; Reducing the size of digital gonioscopy images; Divignetting (correction of peripheral light falloff) in digital gonioscopy images; Grayscale conversion of digital gonioscopy images.

[0052] However, according to other embodiments of the present invention, it is not excluded that the second pretreatment step is not performed, or that a process other than those described above is performed.

[0053] Regarding the configuration of the deep neural network, it is defined as a fully convolutional neural network, preferably comprising at least one convolutional block (CB) and at least one multiple convolutional block (MCB), and is configured to output a second vector for each digital gonioscopy image having the same length X2 as a first vector associated with the same image.

[0054] More specifically, the second vector represents a prediction of the angle open profile of a rotated training digital gonioscopy image, and at each of a plurality of consecutive sampling points, along the directional axis, includes a numerical value that identifies the predicted state of the angle open profile at the same sampling point of the digital gonioscopy image. This numerical value represents either an open, occludable, or occluded state of the angle.

[0055] For the second vector as well, the number of sampling points in each digital gonioscopy image is 60, with a value of 0 for an open state, 1 for a state where occlusion is possible, and 2 for an occluded state.

[0056] More specifically regarding the configuration of a fully convolutional neural network, a convolutional block (CB) is preferably composed of a 2D convolutional layer, a 2D batch normalization layer, and a Leaky ReLU activation function in that order, while a multiple convolutional block (MCB) consists of two CBs followed by a 2D dropout layer and optionally a 2D max pooling layer.

[0057] However, according to other embodiments of the present invention, a fully convolutional neural network may consist only of CBs or only of MCBs, or the CBs and MCBs may have configurations different from those described above, as long as a second vector is output. Furthermore, according to a modified version of a preferred embodiment of the present invention, as long as the second vector is the output of one or more blocks, both the convolutional blocks (CBs) and the multiple convolutional blocks (MCBs) may have structures different from those described above, as long as they are not excluded.

[0058] After the second vector is obtained by processing training digital gonioscopy images, in the training phase, the first vector and the same second vector for the same digital gonioscopy images are input into a regression loss type cost function.

[0059] Thus, the method of the present invention offers the advantageous effect that the identification and evaluation of the angle open profile state is treated as a regression problem rather than a classification problem as assumed by the known art. In other words, the method of the present invention makes it possible to define the angle open profile substantially continuously over the entire 360° of the iris-corneal interface.

[0060] According to a preferred embodiment of the present invention, the regression loss type cost function is not essential, but is preferably an L1 loss regression function. In fact, L1 loss regression functions are known to have the advantageous effects of being less sensitive to outliers and noise, suppressing overfitting, and ensuring better generalization performance.

[0061] However, as long as it is a regression loss type, the possibility of using a cost function different from the aforementioned L1 loss regression function cannot be ruled out.

[0062] As described above, in the method of the present invention, after the deep neural network has been trained, an evaluation step is also performed to evaluate the angle opening profile for multiple sets of digital gonioscopy images representing the entire configuration in which the iris-corneal interface layer extends over 360°. Each digital gonioscopy image represents one sector of the anatomical layer of the iris-corneal interface.

[0063] According to the preferred embodiment described above, the set of digital gonioscopy images consists of 16 digital gonioscopy images, each representing a 22° circumferential sector of the iris-corneal interface, and by arranging these images sequentially, the entire 360° layer of the iris-corneal interface can be defined.

[0064] In particular, this evaluation stage includes the following processes: A process that receives each image from a digital gonioscopy image set, performs a first preprocessing step on each image, and also performs a second preprocessing step if specified during the training phase. This process involves supplying a set of digital gonioscopy images as input to a trained deep neural network and receiving a corresponding second vector from the trained deep neural network as output for each image. This process involves graphically overlaying a second set of vectors obtained from the same set of digital gonioscopy images onto a diagram that graphically represents the structure of the iris-corneal interface layer extending over 360°. This provides a final 360° digital gonioscopy image in which the open profile of the angle over the entire 360° of the iris-corneal interface is graphically displayed in a substantially continuous manner.

[0065] In particular, the final digital gonioscopy image is a 360° digital gonioscopy image, in which three consecutive circular profiles are drawn at the level of the circular profile of the angle, each indicated by three different colors or three different geometric shapes. These represent the open angle, the potentially closable angle, and the closable angle, respectively, and correspond precisely to the points on the profile where the respective states were confirmed by the method. Therefore, the graphical representation visually represents the state of the angle, which precisely corresponds to the evaluated profile represented by a set of second vectors obtained from processing the image set by a trained deep neural network. Preferably, the open angle is assigned green, the potentially closable angle yellow, and the closable angle red. An example of the graphical representation is shown in Figure 7, where, if necessary, the numerical values ​​assigned to the three states are also shown corresponding to each color.

[0066] In addition to providing a superior and spatially precise assessment of the angle open profile, this method enables clinicians to make easier, more intuitive, and immediate decisions when evaluating the angle open profile with subsequent pathological diagnosis in mind.

[0067] The present invention also relates to a computer processing system configured to substantially continuously and automatically evaluate the angle open profile for each image representing a circumferential sector over the entire 360° of the anatomical layer of the iris-corneal interface by processing digital gonioscopy images, typically RGB images.

[0068] In particular, the computer processing system includes a readable storage medium in which computer programming code is stored, the computer programming code, and at least one processor (preferably a microprocessor). The processor is configured to operate in conjunction with the readable storage medium to execute the computer programming code and carry out the learning and evaluation stages according to the method of the present invention.

[0069] Furthermore, the present invention also relates to a computer program product that includes a portion of software code configured to perform a learning phase and an evaluation phase according to the method of the present invention when executed in the memory of a computer.

[0070] Therefore, based on the above explanation, the computer implementation method, computer processing system, and computer program product that are the subject of the present invention all achieve their intended objectives.

[0071] In particular, the present invention achieves the objective of providing a computer implementation method for automatically analyzing digital gonioscopy images (typically RGB images). This method can provide a more accurate and precise evaluation of the angle open profile over 360° of the entire iris-corneal interface as output. Furthermore, the present invention achieves the objective of providing a computer implementation method that allows even non-specialist clinicians to improve the current standard examination, gonioscopy, and output results that enable the assessment of the risk of angle occlusion. Such occlusion can cause an increase in intraocular pressure and increase the risk of developing glaucoma.

[0072] The present invention also aims to provide a computer-implemented method that enables clinicians, especially non-specialist clinicians, to clearly understand which areas to intervene in and how, across the entire 360° of the iris-corneal interface.

Claims

1. A computer implementation method for substantially continuous and automatic evaluation of the angle open profile by processing multiple digital gonioscopy images, typically RGB images, each representing a sector of the anatomical layer of the iris-corneal interface extending 360° as a whole, comprising the use of a deep neural network (DNN) and comprising the following steps: a) A training stage in which the deep neural network is trained, comprising the following steps: The method involves receiving a training image set having a plurality of digital gonioscopy images and performing a first preprocessing step on the digital gonioscopy images, the first preprocessing step including at least one rotation process on the data relating to each digital gonioscopy image so that the digital gonioscopy images are aligned in the same direction, i.e., the layers of the iris-corneal interface are aligned along the same directional axis, preferably horizontally, preferably with the iris layer at the bottom and the corneal layer at the top of each digital gonioscopy image; The system receives an annotation file containing a first vector of length X2 corresponding to each training digital gonioscopy image, wherein the first vector represents the ground truth of the digital gonioscopy image and, for each of a plurality of consecutive sampling points, includes a numerical value that identifies the state of the angle's open profile at the corresponding sampling point of the digital gonioscopy image along the directional axis in terms of openness, wherein the numerical value represents either an open state, a closable state, or a closed state of the angle; Each rotated training digital gonioscopy image and associated annotation file are supplied as input data to a deep neural network, wherein the deep neural network is a fully convolutional neural network comprising at least one convolutional block (CB) and / or at least one multiple convolutional block (MCB), the CB and / or MCB being configured to output a second vector having the same length X2 as a first vector, the second vector representing a prediction of the angle open profile in the rotated training digital gonioscopy image and including a numerical value that identifies, in terms of openness, the state of the angle open profile at the same sampling point of the digital gonioscopy image along the directional axis at each predetermined point among a plurality of consecutive sampling points, the numerical value representing either an open state, a closable state or a closed state of the angle; The first vector and the second vector are input into a regression loss type cost function to train the deep neural network; b) An evaluation step of evaluating the angle open profile of a set of digital gonioscopy images that represent the entire anatomical layer of the iris-corneal interface over a 360° circumference, wherein each digital gonioscopy image represents a sector of the anatomical layer of the iris-corneal interface, and the step includes the following processing: Receiving each digital gonioscopy image and performing the first preprocessing step; The digital gonioscopy images are supplied as input to the trained deep neural network, and the second vector is received as output from the trained deep neural network for each digital gonioscopy image; The present invention provides a 360° digital gonioscopy image in which the open profile state of the angle of the iris-corneal interface is substantially continuously graphically displayed, by superimposing the set of second vectors obtained for the set of digital gonioscopy images onto the graphical representation defined by the set of digital gonioscopy images.

2. In the computer implementation method described in claim 1, A computer implementation method characterized in that the regression loss type cost function is an L1 loss regression function.

3. In the computer implementation method according to claim 1 or 2, A computer implementation method characterized in that a numerical value representing the open state of the corner is set to 0, a numerical value representing the closable state is set to 1, and a numerical value representing the closed state is set to 2.

4. In the computer implementation method according to any one of claims 1 to 3, The aforementioned fully convolutional neural network includes at least convolutional blocks (CBs) and multiple convolutional blocks (MCBs), The aforementioned convolutional block (CB) includes, in order, a 2D convolutional layer, 2D batch normalization, and a Leaky ReLU activation function. The computer implementation method is characterized in that the multiple convolutional block (MCB) includes two convolutional blocks, a 2D dropout following the two convolutional blocks, and an arbitrary 2D maximum pooling layer following the two convolutional blocks.

5. A computer-aided mounting method according to any one of claims 1 to 4, characterized in that the first preprocessing step includes, following the rotation process, a process for extracting a region of interest (ROI) centered on the trabecular meshwork region represented in the digital gonioscopy image.

6. A computer implementation method according to any one of claims 1 to 5, characterized in that the digital gonioscopy inspection image is an RGB type digital gonioscopy inspection image.

7. The computer implementation method according to claim 6, wherein a second preprocessing step for the digital gonioscopy image is provided between the first preprocessing step and the processing of the digital gonioscopy image by the deep neural network, and the second preprocessing step includes the following processing: Noise filtering, adaptive flattening of the histogram of the digital gonioscopy image, size reduction of the digital gonioscopy image, devignetting of the digital gonioscopy image, and grayscale conversion of the digital gonioscopy image.

8. A computer implementation method according to any one of claims 1 to 7, characterized in that each digital gonioscopy inspection image is acquired by focus stacking technology.

9. A computer implementation method according to any one of claims 1 to 8, characterized in that the sampling points for each digital gonioscopy image are 60 points for the first vector and the second vector.

10. In the computer implementation method according to any one of claims 1 to 9, A computer implementation method characterized in that the first vector is defined by an automated process that starts from each training digital gonioscopy image preprocessed through the first preprocessing step, wherein the transition points between an open state and a closable state of the angle, between a closable state and a closed state, between a closed state and a closable state, or between a closable state and an open state of the angle are manually and graphically marked.

11. In the computer implementation method according to any one of claims 1 to 10, Each digital gonioscopy image represents the 22° sector of the iris-corneal interface. A computer implementation method characterized in that, in order to evaluate the open profile of the angle, the set of digital gonioscopy images taken as input to the trained deep neural network includes 16 digital gonioscopy images relating to 16 consecutive sectors of the iris-corneal interface.

12. A computer processing system configured to substantially continuously and automatically evaluate the open profile of the gonioscopy angle by processing multiple digital gonioscopy images, typically RGB images, each image representing a sector of the anatomical layer of the iris-corneal interface extending 360° as a whole. A readable storage medium containing computer programming code, The aforementioned computer programming code, A computer processing system comprising: at least one processor, preferably a microprocessor, configured to operate in conjunction with the readable storage medium and execute the computer programming code to carry out the learning stage and the evaluation stage described in any one of claims 1 to 11.

13. A computer program product comprising a portion of software code configured to be executed in the memory of a computer to perform the learning stage and the evaluation stage described in any one of claims 1 to 11.

Images