SYSTEMS AND METHODS USING WEIGHTED-ENSEMBLE SUPERVISED-LEARNING FOR AUTOMATIC DETECTION OF OPHTHALMIC DISEASE FROM IMAGES

Abstract

Disclosed herein are systems, methods, and devices for classifying ophthalmic images according to disease type, state, and stage. The disclosed invention details systems, methods, and devices to perform the aforementioned classification based on weighted-linkage of an ensemble of machine learning models. In some parts, each model is trained on a training data set and tested on a test dataset. In other parts, the models are ranked based on classification performance, and model weights are assigned based on model rank. To classify an ophthalmic image, that image is presented to each model of the ensemble for classification, yielding a probabilistic classification score—of each model. Using the model weights, a weighted-average of the individual model-generated probabilistic scores is computed and used for the classification.

Claims

1. A method for weighted-ensemble training of machine-learning models to classify ophthalmic images according to features such as disease type and state; where the method comprises of: a) an ensemble of machine-learning models each of which consists of: i. a feature extraction mechanism ii. a classification mechanism b) a step to split the input data into training and test sets c) a step to initialize the weights d) for each model, a step in which the feature extraction mechanism yields a feature vector or other object encoding the ophthalmic image features e) for each model, a step in which the feature vector is passed into the classifier to yield a class prediction f) for each model, a mechanism to iteratively update the weights to reduce class prediction error g) for each model, a stopping mechanism for the iteration h) a step to compare and rank the models based on their performance on a test dataset i) a step to assign weights to the various models in the ensemble j) given a subject ophthalmic image, a step to compute the weighted-average of the class predictions of the plurality of models, and to choose the ophthalmic image class based on this weighted-averaging step.

2. The method of claim 1 wherein some model of the ensemble is a convolutional neural network

3. The method of claim 1 wherein some model of the ensemble is a recurrent neural network

4. The method of claim 1 wherein a rectified linear unit (ReLU) or leaky ReLU is used as the activation function of hidden layers

5. The method of claim 1 wherein a softmax function is used as the activation function of the output layer

6. The method of claim 1 wherein batch normalization is performed

7. The method of claim 1 wherein drop out regularization is performed in the input layers

8. The method of claim 1 wherein the weight initialization step utilizes a pre-trained model

9. The method of claim 1 wherein the weight initialization step is based on random assignment

10. The method of claim 1 wherein the iterative weight update mechanism is back-propagation

11. The method of claim 1 wherein the stopping mechanism is to proceed iteratively till a preset number of iterations or till a preset prediction performance threshold is reached

12. The method of claim 1 wherein the method for assigning weights to models is based on model performance rank

13. The method of claim 1 wherein a pooling step is performed between feature extraction or classification layers

14. A combined imaging and computing system, consisting: a) a system to capture or retrieve an ophthalmic image b) a computer or computing envirnomnent consisting of processing and storage components c) a trained weighted-ensemble of machine learning models stored on the storage component d) executable commands stored on the storage component such that, upon command, i. a ophthalmic image is obtained ii. the ophthalmic image is stored in the storage components iii. the ophthalmic image is retrieved and a classified by passage through the trained weighted-ensemble iv. the image class such as disease state and stage is provided as output v. the image class can be transmitted over a network to a third party for storage, further interpretation, and/or appropriate action.

15. The method of claim 14 wherein the ophthalmic image is obtained by an integrated local device which captures the image of an eye or some of its parts in real time

16. The method of claim 14 wherein the ophthalmic image is obtained by retrieval from a remote imaging system or database

17. The method of claim 14 wherein some of the models in the ensemble are convolutional neural networks

18. The method of claim 14 wherein some of the models in the ensemble are recurrent neural networks

19. The method of claim 14 wherein the trained weighted-ensemble is trained as follows: a) a database of labeled ophthalmic images is split into training and test sets b) each model in the ensemble is trained and tested c) the models are ranked based on their performance on the test dataset d) a model weight is assigned to each model based on its performance rank

20. The method of claim 19 wherein classification of an ophthalmic image is done as follows: a) the image is passed through each model, generating probabilistic class scores for each b) using the model weights, a weighted-average of the probabilistic class scores is computed across models c) the weighted-average of class scores is used to classify the image

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In the following detailed description of the invention, we reference the herein listed drawings and their associated descriptions, in which:

[0026] FIG. 1 is a schematic of tomogram pre-training processing;

[0027] FIG. 2 Feature extraction and classification scheme;

[0028] FIG. 3 Convolutional Layers;

[0029] FIG. 4 Convolution Operation;

[0030] FIG. 5 Fully Connected Layer;

[0031] FIG. 6 View of inter-layer connection;

[0032] FIG. 7 is an example of two connected sublayers of a fully connected layer;

[0033] FIG. 8 is an example of a convolutional neural network;

[0034] FIG. 9 is an example of the iterative training scheme;

[0035] FIG. 10 Weighted Averaging of Ensemble;

[0036] FIG. 11 Weighted Ensemble Class Score Computation; and

[0037] FIG. 12 Computing environment.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0038] The illustration in FIG. 1 is a preferred embodiment of the pre-training processing steps carried out on the data. The schematic includes an unlabeled set of images 100. In step 110, the unlabeled data in 100 is labeled by an expert or some other entity with sufficient knowledge to do so competently. This labeling yields a labeled data set depicted in 120. In the step 130 the labeled data set 120 is partitioned into a training set, 150, and test data set, 140. The choice of partitioning fraction is itself a learnable hyper-parameter—in the sense that various fractions can be tried empirically to determine the fraction with best most generalizable results. Various forms of pre-processing such as data augmentation and random shuffling can be done to the data set of labeled images 120 to yield a data set of processed tomograms. The processed and labeled images are then partitioned into a training set, 150, and a test set, 140. In turn, the training and test sets are entered as input into each of the machine learning models that comprise the ensemble.

[0039] The depiction in FIG. 2 is an exemplary schema of a model of the ensemble. In this one embodiment, the ophthalmic image, 200, is accessed by a feature extraction mechanism, for example by convolutional operations characteristic of convolutional neural networks. In FIG. 2, the feature extraction mechanism is represented by 210. The output of the feature extraction mechanism is a feature vector or other mathematical object which encodes the features of the ophthalmic image. In what preceded and in what follows, the term feature vector is understood to mean either a mathematical vector or any other mathematical object that encodes the features of the ophthalmic image. In FIG. 2 the feature vector is represented by 220. The feature vector is then acted on by a classification mechanism, for example, the fully connected layers characteristic of convolutional neural networks or other multilayer perceptron based schemes. In FIG. 2 the classification mechanism is represented by 230. The output of the classification process is a choice of image class, as depicted in 240. Of note, this choice of image class can take on various forms including probabilistic as in cases where the softmax activation function is used in the output.

[0040] In some embodiment of the invention, some of the members of the ensemble can be convolutional neural networks (CNNs). An exemplary illustration of a feature extraction scheme of a CNN is depicted in FIG. 3. The scheme takes an ophthalmic image as input, as depicted in 300. A set of feature extraction operations are then carried out on the ophthalmic image. This typically would involve some generalization of the dot product, but can use other process instead. In the case of convolutional neural networks, for example, this dot product step uses the convolution operation. In FIG. 3, the first of these set of feature extraction operations is depicted by 310 and yields 320. Of note, 320 is a set of feature maps whose number is equal to the number of convolutional filters. Each of the feature maps shown in 320 is generated by doing convolution operations on 300 using a distinct filter. The convolution operation is further exemplified in FIG. 4 below. Each of the feature maps in 320 in turn serves as an input image for a similar type of dot-product operation 330 and in turn yields a next layer of feature maps as depicted in 340. The dot product operation yielding the subsequent feature map can be done any arbitrary—up to a point—number of times as depicted e.g. as in 350 yielding 360, and so on. The actual number of times the dot product can be done may be limited by the relative size of the input image in comparison to the filter. The number of times the operation is applied and the forms of the operation can be chosen to optimize the classification performance of the architecture. The end product of the feature extraction steps is a feature vector or other object encoding the image features. Here, that object is represented by 370.

[0041] Depicted in FIG. 4 is a convolution operation. In this example, the operand is the ophthalmic image depicted in 400. At the top left corner of this ophthalmic image is an illustration of the positions of a given convolutional filter. Each position is offset from the next by a prescribed stride. A prescribed dot product operation is then conducted, such as pixel-wise multiplication followed by summation of all the products, as in the following equation:

[00001]$\begin{array}{cc}{c}_k=\underset{i}{.Math.}.Math..Math.u_i.Math.v_{i,k},& \left(1\right)\end{array}$

[0042] where u.sub.i is the ith pixel value in the filter, v.sub.i,k is the ith pixel value of the portion of the ophthalmic image that overlaps the filter when the filter is in the kth position. And c.sub.k is the value of the kth pixel of the generated feature map. The multiple overlapping positions of the filter can be thought of as the filter scanning over the ophthalmic image and performing the aforementioned computations as it does so. In FIG. 4 this scanning is represented by 410, and the generated feature map is represented by 420.

[0043] In some embodiment of the invention, the ensemble contains some machine learning models whose classification mechanisms are multilayer perceptrons—also known as fully connected layers. An exemplification of such a fully connected layer is depicted in FIG. 5. The input is a feature vector represented by 500. This is connected to the first hidden layer 510. Each neuron of this layer is connected to every neuron of the next hidden layer, 520. This pattern continues sequentially into the output layers 540. The weighted interconnections between nodes are depicted by lines as exemplified in 550.

[0044] The depiction in FIG. 6 illustrates an exemplary configuration of a preferred embodiment of a single sublayer of the fully connected layer. The representation in FIG. 6 contains examples of network nodes or “neurons” depicted by 600, 630, and 640. The representation in FIG. 6 also contains examples of weights or multiplicative coefficients associated with each connection between any two network nodes. In particular, one sees weight 620 between nodes 600 and 630, and weight 610 between nodes 600 and 640.

[0045] The depiction in FIG. 7 illustrates an exemplary computation during a forward pass. In particular, the labeled equations demonstrate the linear combination of weighted inputs between sublayers in the fully connected layer. The general mathematical expression for this step is given by:

[00002]$\begin{array}{cc}\underset{i=1}{\overset{n}{.Math.}}.Math..Math.w_{\mathrm{ij}}.Math.x_i,& \left(2\right)\end{array}$

[0046] where x.sub.α denotes the output from neuron X.sub.α, w.sub.ij is the weight connecting neuron X.sub.i to neuron X.sub.j, and n is the number of neurons providing input into neuron X.sub.j, such as is depicted in 710 of FIG. 7. Similarly, the expression for the input into neuron X.sub.k is shown in 700.

[0047] Equation (2) and its type are then subsequently fed as input into an activation function σ(x) such as ReLU for example but not limitation, yielding the following form:

[00003]$\begin{array}{cc}\sigma \left(\underset{i=1}{\overset{n}{.Math.}}.Math..Math.w_{\mathrm{ij}}.Math.x_i\right).& \left(3\right)\end{array}$

[0048] An exemplary method by which an individual model of the ensemble performs feature extraction and subsequent classification is depicted in FIG. 8. This particular example is a convolutional neural network. Other architectures like recurrent neural networks, convolutional recurrent neural networks, and various hybrids and ensembles of diverse architectures can be used. In this particular example shown in FIG. 8, the feature extraction part is depicted in 810 while the classification part is depicted in 830. The direction of operations is depicted by 840. The initial forward pass consists of the following steps: [0049] 1. The filter weights and the fully connected layer weights are initialized either randomly or using some prior knowledge such as a pre-trained model. [0050] 2. Using the initialized filter weights, a dot product of the ophthalmic image, 800, and the filter is taken. [0051] 3. This yields the feature maps shown, upon whom sequential applications of a dot product yields the feature object depicted in 820. [0052] 4. The feature object is acted upon by the classification scheme to yield an estimate of the image class, as depicted by 850. [0053] 5. The image class estimated by the algorithm is compared to target values stored in the label. The net extent of the estimation error across classes is quantified by a loss function, for example hinged loss or other variant. We then proceed to iteratively minimize the loss or net error, as described in FIG. 9 below.

[0054] The error computed above is the objective function which we seek to minimize. An example is as follows:

[00004]$\begin{array}{cc}\mathrm{Loss}\left({\left\{{\hat{y}}_p-\rho \left(\underset{t}{.Math.}.Math..Math.w_{\mathrm{tp}}(.Math.\mathrm{.Math.}.Math..Math.\underset{k}{.Math.}.Math..Math.w_{\mathrm{kl}}\left(\gamma \left(\underset{j}{.Math.}.Math..Math.w_{\mathrm{jk}}\left(\sigma \left(\underset{i}{.Math.}.Math..Math.w_{\mathrm{ij}}.Math.x_j\right)\right)\right)\right).Math..Math.\mathrm{.Math.}.Math.)\right)\right\}}_p\right),& \left(4\right)\end{array}$

[0055] where x.sub.i are the input features; w are weights; σ, γ, ρ are activation functions; and ŷ.sub.p is the target value of the pth class. Of note L is a composite function consisting of the weighted linear combinations of inputs into each successive layer. The effect of any given weight on the net loss can therefore be computed using the chain rule. For instance, we can re-write the loss function in the notationally concise functional form

L(w)=b(c(d( . . . i(j(w))))),  (5)

[0056] where w is a weight and b, c, d, . . . , i, j are functions describing the network. Then the effect of weight w on loss L, denoted

[00005]$\frac{\partial L}{\partial w},$

is given by

[00006]$\begin{array}{cc}\frac{\partial L}{\partial w}=\frac{\partial L}{\partial b}.Math.\frac{\partial b}{\partial c}.Math.\frac{\partial c}{\partial d}.Math..Math.\mathrm{.Math.}.Math..Math.\frac{\partial i}{\partial j}.Math.\frac{\partial j}{\partial w}.& \left(6\right)\end{array}$

[0057] This is done in a computationally efficient manner using the well-known back-propagation algorithm. In some preferred embodiment of the invention disclosed herein, an ophthalmic image input is obtained and the training procedure is carried out in an iterative manner as shown in FIG. 9. An ophthalmic image is shown in 900—in this case illustrated, it is an OCT image of the macula. The forward pass is done as shown in 910. A loss is determined and used as input into the back-propagation phase depicted in 920. The back-propagation determines the influence of each weight on the loss. This information is then used in phase 930 to update the weights in the indicated direction, i.e. to decrease the loss. Once the weights are updated, the forward pass is repeated to determine the new loss, which in turn is passed again as input into the back-propagation phase and so on. The procedure proceeds iteratively as noted till a prescribed stopping point, i.e. till the loss is below a prescribed amount or till the cycle has repeated a preset number of times.

[0058] FIG. 10 illustrates by way of example, a weighted averaging procedure of the invention disclosed herein. 1000 depicts a subject ophthalmic image to be classified by the disclosed method. There are N number of models—1010, 1020, 1030, 1040—in the ensemble, where N can be any number. Training of the models are as exemplified above. Of note, the models in the ensemble can be chosen or designed based on any number of criteria including but not limited to level of performance on a test dataset, heuristic criteria such as depth and complexity of model architecture, known good performance on other types of datasets and problem domains—where “good” can be defined as desired. 1050 is the probability predicted by model 1, 1010, that ophthalmic image u 1000 is of class t.sub.j. We represent this with the notation,

P(uεt.sub.j|m.sub.1).  (7)

[0059] Similarly, 1060 is the probability predicted by model 2, 1020, that ophthalmic image u 1000 is of class t.sub.j, 1070 is the probability predicted by model 3, 1030, that ophthalmic image u 1000 is of class t.sub.j, and 1080 is the probability predicted by model N, 1040, that ophthalmic image u 1000 is of class t.sub.j. Model weights are determined based on performance of the individual models on test data. Any number of order preserving weight assignment schemes can be applied, such that the better the relative performance of a model, the higher its assigned weight. The weight assignment scheme can include a performance threshold below which a weight of zero is assigned. i.e. models with low enough performance can be excluded from the voting. In FIG. 10 the weights 1090, 1092, 1094, and 1096 are associated with models 1010, 1020, 1030, and 1040 respectively. The weighted average as shown in 1098 is

[00007]$\begin{array}{cc}\frac{1}{N}.Math.\underset{i=1}{\overset{N}{.Math.}}.Math..Math.w_i.Math.P\left(u\in t_j|m_i\right)& \left(8\right)\end{array}$

[0060] In FIG. 11, 1100 represents the process for computing the weighted average of the probabilities that the subject image u belongs to class t.sub.a. Details of the computation 1100 are as exemplified in FIG. 10. Similarly, 1110 represents the process for computing the weighted average of the probabilities that the subject image u belongs to class t.sub.b, and 1120 represents the process for computing the weighted average of the probabilities that the subject image u belongs to class t.sub.z. The respective weighted averages are depicted in 1130, 1140, and 1150. The weighted averages are passed along in steps 1160, 1170, and 1180 for normalization and computation of the class scores predicted by the weighted-ensemble. By way of example, 1190 depicts the probability that image u belongs to class t.sub.k, i.e.,

[00008]$\begin{array}{cc}P\left(u\in t_k|m_1,m_2,m_3,\mathrm{.Math.}.Math.,m_N\right)=\frac{\underset{i=1}{\overset{N}{.Math.}}.Math..Math.w_i.Math.P\left(u\in t_k|m_i\right)}{\underset{j}{.Math.}.Math..Math.\underset{i=1}{\overset{N}{.Math.}}.Math..Math.w_i.Math.P\left(u\in t_j|m_i\right)}& \left(9\right)\end{array}$

[0061] The denominator in the above equation is the normalization factor that makes weighted-ensemble class scores a distribution, i.e. sum to unity. In contrast to the loss function—whose evaluation can be negative, and hence can require for exponentiation (or similar mechanism) to ensure positivity and to allow for the formation of a distribution. Here, each of the individual model predictions are typically already probabilities, i.e. non-negative and in [0, 1].

[0062] Ones skilled in the art will recognize that the invention disclosed herein can be implemented over an arbitrary range of computing configurations. We will refer to any instantiation of these computing configurations as the computing environment. An exemplary illustration of a computing environment is depicted in FIG. 12. Examples of computing environments include but are not limited to desktop computers, laptop computers, tablet personal computers, mainframes, mobile smart phones, smart television, programmable hand-held devices and consumer products, distributed computing infrastructures over a network, cloud computing environments, or any assembly of computing components such as memory and processing—for example.

[0063] As illustrated in FIG. 12 the invention disclosed herein can be implemented over a system that contains a device or unit for processing the instructions of the invention. This processing unit 16000 can be a single core central processing unit (CPU), multiple core CPU, graphics processing unit (GPU), multiplexed or multiply-connected GPU system, or any other homogeneous or heterogeneous distributed network of processors.

[0064] In some embodiment of the invention disclosed herein, the computing environment can contain a memory mechanism to store computer-readable media. By way of example and not limitation, this can include removable or non-removable media, volatile or non-volatile media. By way of example and not limitation, removable media can be in the form of flash memory card, USB drives, compact discs (CD), blu-ray discs, digital versatile disc (DVD) or other removable optical storage forms, floppy discs, magnetic tapes, magnetic cassettes, and external hard disc drives. By way of example but not limitation, non-removable media can be in the form of magnetic drives, random access memory (RAM), read-only memory (ROM) and any other memory media fixed to the computer.

[0065] As depicted in FIG. 12, the computing environment can include a system memory 16030 which can be volatile memory such as random access memory (RAM) and may also include non-volatile memory such as read-only memory (ROM). Additionally, there typically is some mass storage device 16040 associated with the computing environment, which can take the form of hard disc drive (HDD), solid state drive, or CD, CD-ROM, blu-ray disc or other optical media storage device. In some other embodiment of the invention the system can be connected to remote data 16240.

[0066] The computer readable content stored on the various memory devices can include an operating system, computer codes, and other applications 16050. By way of example not limitation, the operating system can be any number of proprietary software such as Microsoft windows, Android, Macintosh operating system, iphone operating system (iOS), or Linux commercial distributions. It can also be open source software such as Linux versions e.g. Ubuntu. In other embodiments of the invention, imaging software and connection instructions to an imaging device 16060 can also be stored on the memory mechanism. The procedural algorithm set forth in the disclosure herein can be stored on—but not limited to—any of the aforementioned memory mechanisms. In particular, computer readable instructions for training and subsequent image classification tasks can be stored on the memory mechanism.

[0067] The computing environment typically includes a system bus 16010 through which the various computing components are connected and communicate with each other. The system bus 16010 can consist of a memory bus, an address bus, and a control bus. Furthermore, it can be implemented via a number of architectures including but not limited to Industry Standard Architecture (ISA) bus, Extended ISA (EISA) bus, Universal Serial Bus (USB), microchannel bus, peripheral component interconnect (PCI) bus, PCI-Express bus, Video Electronics Standard Association (VESA) local bus, Small Computer System Interface (SCSI) bus, and Accelerated Graphics Port (AGP) bus. The bus system can take the form of wired or wireless channels, and all components of the computer can be located remote from each other and connected via the bus system. By way of example and not of limitation, the processing unit 16000, memory 16020, input devices 16120, output devices 16150 can all be connected via the bus system. In the representation depicted in FIG. 12, by way of example not limitation, the processing unit 16000 can be connected to the main system bus 16010 via a bus route connection 16100; the memory 16020 can be connected via a bus route 16110; the output adapter 16170 can be connected via a bus route 16180; the input adapter 16140 can be connected via a bus route 16190; the network adapter 16260 can be connected via a bus route 16200; the remote data store 16240 can be connected vis a bus route 16230; and the cloud infrastructure can be connected to the main system bus vis a bus route 16220.

[0068] In some embodiment of the invention disclosed herein, FIG. 12 illustrates that instructions and commands can be input by the user using any number of input devices 16120. The input device 16120 can be connected to an input adapter 16140 via an interface 16130 and/or via coupling to a tributary of the bus system 16010. Examples of input devices 16120 include but are by no means limited to keyboards, mouse devices, stylus pens, touchscreen mechanisms and other tactile systems, microphones, joysticks, infrared (IR) remote control systems, optical perception systems, body suits and other motion detectors. In addition to the bus system 16010, examples of interfaces through which the input device 16120 can be connected include but are by no means limited to USB ports, IR interface, IEEE 802.15.1 short wavelength UHF radio wave system (bluetooth), parallel ports, game ports, and IEEE 1394 serial ports such as FireWire, LLINK, and Lynx.

[0069] In some embodiment of the invention disclosed herein, FIG. 12 illustrates that output data, instructions, and other media can be output via any number of output devices 16150. The output device 16150 can be connected to an output adapter 16170 via an interface 16160 and/or via coupling to a tributary of the bus system 16010. Examples of output devices 16150 include but are by no means limited to computer monitors, printers, speakers, vibration systems, and direct write of computer-readable instructions to memory devices and mechanisms. Such memory devices and mechanisms can include by way of example and not limitation, removable or non-removable media, volatile or non-volatile media. By way of example and not limitation, removable media can be in the form of flash memory card, USB drives, compact discs (CD), blu-ray discs, digital versatile disc (DVD) or other removable optical storage forms, floppy discs, magnetic tapes, magnetic cassettes, and external hard disc drives. By way of example but not limitation, non-removable media can be in the form of magnetic drives, random access memory (RAM), read-only memory (ROM) and any other memory media fixed to the computer. In addition to the bus system 16010, examples of interfaces through which the output device 16150 can be connected include but are by no means limited to USB ports, IR interface, IEEE 802.15.1 short wavelength UHF radio wave system (bluetooth), parallel ports, game ports, and IEEE 1394 serial ports such as FireWire, i.LINK, and Lynx.

[0070] In some embodiment of the invention disclosed herein some of the computing components can be located remotely and connected to via a wired or wireless network. By way of example and not limitation, FIG. 12 shows a cloud 16210 and a remote data source 16240 connected to the main system bus 16010 via bus routes 16220 and 16230 respectively. The cloud computing infrastructure 16210 can itself contain any number of computing components or a complete computing environment in the form of a virtual machine (VM). The remote data source 16240 can be connected via a network to any number of external sources such as imaging devices, imaging systems, or imaging software.

[0071] In some embodiment of the invention disclosed herein, an imaging system which captures and pre-processes images, e.g. 16060, is attached directly to the system. Stored in the memory mechanism—16020, 16240, or 16210—is a model trained according to the machine learning procedure set-forth herein. Computer-readable instructions are also stored in the memory mechanism, so that upon command, images can be captured from a patient in real time, or can be received over a network from a remote or local previously collated database. In response to command such images can be classified by the pre-trained machine learning procedure disclosed herein. The classification output can then be transmitted to the care provider and/or patient for information, interpretation, storage, and appropriate action. This transmission can be done over a wired or wireless network as previously detailed, as the recipient of the classification output can be at a remote location.

[0072] Illustrating the invention disclosed herein, an anonymized database of 3000 ocular coherence tomograms (OCTs) of the macula was compiled. Binary labels were assigned by an American board-certified ophthalmologist and Retina specialist. The labels were ‘actively exudating age-related macula degeneration’ or ‘not actively exudating age-related macula degeneration’. The database was split into one dataset for training and a separate dataset for validation. 400 OCT images were used for validation—200 ‘actively exudating’ and 200 ‘not actively exudating’. The algorithm achieved 99.2% accuracy in distinguishing between ‘actively exudating’ and ‘not actively exudating’.

[0073] The objects set forth in the preceding are presented in an illustrative manner for reason of efficiency. It is hereby noted that the above disclosed methods and systems can be implemented in manners such that modifications are made to the particular illustration presented above, while yet the spirit and scope of the invention is retained. The interpretation of the above disclosure is to contain such modifications, and is not to be limited to the particular illustrative examples and associated drawings set-forth herein.

[0074] Furthermore, by intention, the following claims encompass all of the general and specific attributes of the invention described herein; and encompass all possible expressions of the scope of the invention, which can be interpreted—as pertaining to language—as falling between the aforementioned general and specific ends.