Evaluation of the Clinical Significance of Vitreous Floater by Machine Learning

Abstract

Certain embodiments disclosed herein provide a method of training a machine learning model to estimate the clinical significance of floaters in a patient's eye. One or more images, such as SLO images or frontal retinal OCT images, are evaluated to identify shadow regions corresponding to floaters. The shadow regions are measured and the measurements are processed using a machine learning model to obtain an estimated significance. The machine learning model is then updated according to a comparison of the estimated significance and a clinically assigned significance by a human. The machine learning model may be updated additionally or alternatively by evaluating an estimated category with respect to visibility threshold data, such as one or more visibility threshold surfaces defined with respect to two or more variables.

Description

Technical Field

[0001] Cross - Reference to Related Applications This application claims priority to U.S. Provisional Patent Application No. 63 / 388,901, filed Jul. 13, 2022, the entire content of which is incorporated herein by reference.

Background Art

[0002] Light received by the human eye passes through the transparent cornea that covers the iris and pupil of the eye. The light is transmitted through the pupil and focused by the lens located within a structure called the lens capsule on the posterior side of the pupil. The light is focused by the lens and the cornea onto the retina containing rods and cones that can generate nerve impulses in response to light. The space between the lens and the retina is filled with a transparent gel known as the vitreous humor.

[0003] For various reasons, there may be floating substances in the vitreous humor. Floating substances are typically formed from clumps of cells, collagen fibers, or other tissues and are more opaque than the surrounding vitreous humor. Floating substances cast shadows on the retina, causing visual impairment in patients, and in some patients, there is a risk of exacerbation.

[0004] Facilitating the treatment of floating substances would be an advancement in the art.

Summary of the Invention

Means for Solving the Problems

[0005] The present disclosure generally relates to systems and methods for evaluating the clinical significance of vitreous floaters. In one aspect, a computing device receives one or more images of a patient's eye and identifies one or more shadow regions within the one or more images. The shadow regions are processed to obtain one or more measurements of the shadow regions. The computing device processes the one or more measurements using a machine - learning model to obtain an estimated clinical significance of the floaters in the patient's eye.

[0006] The following description and the related drawings show specific exemplary features of one or more embodiments in detail.

[0007] The accompanying drawings show specific aspects of one or more embodiments and should not be considered as limiting the scope of the present disclosure.

Brief Description of the Drawings

[0008]

Figure 1A

Figure 1B

Figure 2A

Figure 2B

Figure 3

Figure 4

Modes for Carrying Out the Invention

[0009] For clarity, the same reference numerals are used, where possible, to indicate common identical elements in multiple drawings. The elements and features of one embodiment can be advantageously incorporated into other embodiments without special mention.

[0010] Referring to FIG. 1A, a human eye 100 includes a cornea 102, which is a spherical transparent layer through which light entering the eye 100 passes. The light then passes through each of the anterior chamber 138, pupil 104, and lens 106 of the eye 100. The remaining portion of the globe 108 of the eye 100, known as the posterior chamber or vitreous chamber 140, is filled with a transparent gel known as the vitreous 110. The light is focused by the cornea 102 and lens 106 through the vitreous 110 onto the retina 112 at the back of the eye 100.

[0011] Vitreous floaters 114 are clumps of cells, collagen fibers, or other impurities within the vitreous 110. If vitreous floaters 114 are present, they cast a shadow 116 on the retina 112. The shadow 116 may occupy an angular range 118 of the visual field of the eye 100. If the floaters 114 are large enough, opaque, and / or numerous, they can significantly reduce the patient's vision. If the shadow 116 of the floaters 114 crosses the fovea, the patient's vision may be reduced.

[0012] Referring to FIG. 1B, an image 120 of the retina 112 can be acquired, for example, by using a scanning laser ophthalmoscopy (SLO), a visible light camera, an optical coherence tomography (OCT) microscope, or other imaging modalities. For example, the image 120 may be a frontal image acquired using an SLO or an OCT microscope. A portion of the light irradiated onto the retina 112 to acquire the image 120 is scattered by any of the floaters 114 present in the vitreous 110, thereby creating a shadow region 122 in the image 120.

[0013] The shadow region 122 can be identified as having a pixel intensity with a strong contrast relative to the region surrounding the shadow region 122. Since the floating object 114 has motility, the shadow region 122 in the image 120 may be identified as having an intensity that is relatively different from the intensity of the shadow region 122 in the preceding or subsequent images in a series of video image frames including the image 120. For example, each image in a series of images may be aligned with respect to reference features of the retina, such as the pattern of blood vessels (e.g., veins) in the retina, to track and correct for eye movement. In such an example, the change from one aligned image to another may thus correspond to the shadow 116 of the floating object 114. The shadow region 122 can be identified using any approach for detecting an object that is moving relative to a stationary background, and such an approach corrects for eye movement in the same way as it corrects for camera movement. The still or moving image 120 can be analyzed using a machine learning model trained to identify the shadow region 122 corresponding to the floating object 114.

[0014] Various metrics of the shadow region 122 of each floating object 114 may be extracted from the image 120. For example, the shadow region 122 has a size. The size can be measured as the area (e.g., number of pixels) within the boundary 124 of the shadow region. The size of the shadow region 122 may also be more easily obtained as the size of a two-dimensional bounding box that is orthogonal (parallel to the rows and columns of the pixels of the image 120) or oriented to fit the shadow region 122. The size can be represented by the two-dimensional dimensions of the floating object 114, such as the longest line that can be contained within the shadow region 122.

[0015] The shadow region 122 may be characterized by the contrast of the shadow region 122 relative to the region surrounding the image 120. For example, the contrast is obtained by dividing the average pixel intensity of the pixels of the image of the shadow region 122 by the average intensity of the pixels outside the shadow region 122 (e.g., a pixel band with a depth of one pixel or more around the boundary 124 of the shadow region 122). Any approach for characterizing the contrast between regions of an image may be used.

[0016] The movement of the shadow region 122 from one image to the next within a series of video frames 120 can also be measured or characterized. The movement can be detected using any motion tracking approach known in the art, such as a Kalman filter or a similar algorithm. The movement is characterized based on the speed and / or direction of the movement. For example, it can be characterized as a component 126 of the speed of the shadow region 122 directed towards or away from the fovea 128 (i.e., the representation of the fovea of the eye 100 in the image 120). In some implementations, the movement of the shadow region 122 is stationary (e.g., the movement is below a threshold), or is directed towards the fovea 128 or away from the fovea 128.

[0017] The shadow region 122 is characterized by the position of the shadow region 122 relative to the fovea 128. For example, the distance between the shadow region 122 and the fovea 128 is calculated as the shortest distance from the center of the fovea 128 to the point of the shadow region 122 that is closest to the center of the fovea 128. In another example, the shadow region 122 is characterized as overlapping with the fovea 128, the perifovea 130, the parafovea 132, the macula 134, or the peripheral region 136 of the retina 112. For example, the shadow region 122 may be characterized as the innermost region of the fovea 128, the perifovea 130, the parafovea 132, the macula 134, or the peripheral region 136 that overlaps at least a portion of the shadow region 122. The position of the fovea 128, the perifovea 130, the parafovea 132, the macula 134, or the peripheral region 136 of the eye within the image 120 can be estimated based on the position of the image 120, i.e., assuming that the center of the image 120 is the center of the fovea, and the accepted values can be used for the fovea 128, the perifovea 130, the parafovea 132, the macula 134, and the peripheral region 136. Those skilled in the art can easily recognize the eye socket as a dark spot having a characteristic shape of retinal blood vessels with a dimension of about 0.3 mm (field of view of about 1°).

[0018] The above metrics for the shadow regions 122 are merely exemplary, and other characteristics of each shadow region 122 may also be measured. For example, in the case of multiple floaters, the metrics may include one or more separation distances between shadow regions, the average separation distance between floaters, the spatial frequency of the shadow pattern formed by the shadow regions 122 of the floaters (e.g., two-dimensional Fourier transform of the shadow pattern), and the spatial and temporal frequencies of the shadow pattern (e.g., three-dimensional Fourier transform of the shadow pattern of a series of video frames 120).

[0019] FIG. 2A shows a system 200a for training a machine learning model 202 that characterizes the clinical significance of floaters 114 in an eye 100 based on one or more shadow regions 122 in one or more images 120 of the retina 112 of the eye 100. The machine learning model may be implemented as a neural network, a deep neural network, a convolutional neural network, a multiple linear regression model, a random sample consensus regression model, a multiple polynomial regression model, a support vector regression model, a Bayesian neural network, a genetic algorithm, or any other type of machine learning model.

[0020] The machine learning model 202 may be trained by one or more training algorithms 204 to output a category 206 that estimates the clinical significance of the floaters 114. The clinical significance may be one of a set of discrete values corresponding to invisibility (e.g., imperceptible), prominent, uncomfortable, severity of vision loss, etc., or clinical symptoms (e.g., treatment required to maintain vision).

[0021] The machine learning model 202 may employ metrics of the shadow regions 122, such as size 208, contrast 210, direction of movement 212, and position 214, as inputs. These metrics may be as described above with respect to FIG. 1B, or obtained by other approaches. Similarly, any other metrics described above with respect to FIG. 1B may be used in place of or in addition to those shown in FIG. 2A.

[0022] In some implementation examples, the size of the shaded region 122 is assigned to one of a set of bins representing a range of possible sizes referred to herein by the following exemplary identifiers: extra-small (XS), small (S), large (L), and extra-large (XL). The input to the machine learning model 202 may thus be an identifier of a bin mapped to a size range that includes the size of a given shaded region 122 within the image 120. Any number of bins may be used, and any range of sizes may be assigned to each bin such that the ranges of sizes of different bins do not overlap.

[0023] Similarly, a range of possible contrasts may be divided into non-overlapping sub-ranges, each assigned to a bin having a bin identifier, e.g., high, medium, low. The contrast 210 input to the machine learning model 202 for the shaded region may be an identifier of a bin having an assigned sub-range that includes the contrast of the shaded region 122.

[0024] The input movement direction 212 to the machine learning model 202 may be one of two values, one indicating movement towards the fovea 128 and the other indicating movement away from the fovea. Other characterizations of the movement of the shaded region 122 may be input to the machine learning model 202 alternatively or additionally.

[0025] The position 214 input to the machine learning model 202 may be an identifier of the closest region of the retina 112 that overlaps the shaded region 122. For example, the identifier may correspond to a part or all of the fovea, perifovea, parafovea, macula, and peripheral region of the retina 112. Alternatively, the distance between the center of the fovea and the point of the shaded region 122 closest to the fovea may be used. An identifier of a bin mapped to a range of distances including the said distance may also be used as the position 214 input to the machine learning model 202.

[0026] The training of the machine learning model 202 may be performed using a plurality of training data entries. Each training data entry may include a set of inputs including some or all of the metrics described herein, such as size 208, contrast 210, direction of movement 212, position 214 of the shadow region 122, or any other exemplary metrics described herein. Each training data entry further includes an assigned category as the desired output. The assigned category may be one of the output categories 206 assigned by a professional to the training data entry based on the observation of the image 120 including the shadow region 122, or based on the observation of the patient in whom the eyes are represented in the image 120. The assigned category may also be assigned by a patient who has evaluated the severity of the floating object 114 that generated the shadow region 122.

[0027] Each training data entry may be processed by inputting the input matter into the machine learning model 202, receiving the estimated category, and comparing the estimated category with the assigned category of the training data entry. The training algorithm 204 may then update the machine learning model 202 according to the difference, or "loss", between the estimated category and the assigned category of the training data entry. For example, an integer value i may be assigned to each category such that the loss function evaluated by the training algorithm is |i1 - i0| or a function thereof, where i1 is the subscript of the assigned category and i0 is the subscript of the estimated category.

[0028] In use, the image 120 is received, the shadow region 122 is identified, and its metrics are calculated. The metrics are processed using the machine learning model 202 to obtain an estimated category, and the estimated category is then output to a display device or the like for the user.

[0029] Figure 2B shows an alternative system 200b for training a machine learning model that characterizes the clinical significance of floaters using visibility threshold data. The visibility threshold data includes experimental data that characterizes an observer's ability to distinguish features. The experimental data may include data specific to the floaters and may additionally or alternatively include data not specific to the floaters, such as data used to characterize the effectiveness of camouflage. In particular, the experimental data can define a threshold surface with respect to three or two variables, and the threshold surface defines a boundary between combinations of values of two or more visible variables and combinations of two or more non-visible values.

[0030] Examples of non-limiting sources of visibility threshold data may include any of the following, all of which are hereby incorporated by reference: “Motion and Vision.II.Stabilized Spatio-Temporal Threshold Surface,” D.H. Kelly, J. Opt. Soc. Am., Vol. 69, No. 10 (October 1979); “Contrast Sensitivity of the Human Eye and its Effects on Image Quality,” Barten, P.G. (1999); “The contrast sensitivity gradient across the human visual field:” Pointer, J.S., Hess, R.F. Vision Research, 29(9), 1133-1151 (1989); “Visual Processing of Moving Stimuli,” D.H. Kelly, J. Opt. Soc. Am., Vol. 2, No. 2 (February 1985); “Retinal Inhomogeneity.I.Spatiotemporal Contrast Sensitivity,” J. Opt. Sec. Am., Vol. 1, No. 1 (January 1984); “Retinal Inhomogeneity.II.Spatial Summation,” J. Opt. Soc. Am., Vol. 1, No. 1 (January 1984); “Retinal Inhomogeneity.III.Circular-Retina Theory,” D.H. Kelly, J. Opt. Soc. Am., Vol. 2, No. 6 (June 1985); “Motion and Vision.II.Stabilized Spatio-Temporal Threshold Surface,” J. Opt. Soc. Am., Vol. 69, No. 10 (October 1979); “Contrast Sensitivity of the Human Eye and Its Effects on Image Quality,” P.G.J. Barten, SPIE Optical Engineering Press (1999).

[0031] Non-limiting examples of variables that can define a threshold surface may include spatial frequency (cycles per visual angle), temporal frequency (Hz), speed (per second), modulation (e.g., amplitude of spatial or temporal contrast), and eccentricity (distance from the fovea measured in degrees).

[0032] Examples of threshold surfaces include a first surface defined with respect to spatial frequency, speed, and / or modulation amplitude, a second surface defined with respect to spatial frequency, eccentricity, and modulation amplitude, and a third surface defined with respect to spatial frequency, temporal frequency, and / or modulation amplitude.

[0033] System 200b may include a visibility threshold data comparator 216 that can evaluate the shadow pattern of one or more shadow regions 122 with respect to experimental data. For example, for each threshold surface of one or more threshold surfaces defined by two or more variables, the visibility threshold data comparator 216 calculates the values of two or more variables of the shadow pattern and evaluates the values with respect to the two or more variables to determine whether the values of the two or more variables indicate that the shadow pattern is visible.

[0034] The visibility threshold data comparator 216 may perform one or more evaluations for each threshold surface. The one or more evaluations may include an evaluation of whether the values of two or more variables of the shadow pattern exceed the threshold surface, i.e., whether they are visible. The one or more evaluations may include the calculation of the distance (D T (j)) between the values of two or more variables that define the threshold surface j and the closest point on the threshold surface j, where j is the index assigned to each threshold surface.

[0035] The machine learning model 202 may receive the same inputs as in the case of system 200a and output the same estimated categories as in the case of system 200a. Similarly, the training data entries used for training the machine learning model 202 may be the same as those described above with respect to system 200a. However, the training algorithm 204 may be modified compared to that described with respect to system 200a. More specifically, the training algorithm 204 can evaluate the estimated categories obtained by processing the inputs of the learning data entries with respect to the output of the visibility threshold data comparator 216. For example, assume that each output category 206 has a subscript i. Each image 120 of the set of images may be assigned to the category i assigned by a trained medical expert or an evaluator such as the patient for whom each image 120 was taken. The values of the variables defining each threshold surface j may be calculated for each image 120. The distance D T (j) from each threshold surface j for each image can be calculated. Thus, for each output category i and threshold surface j, there is a distribution of the distances of the images 120 assigned to the output category i. Using this distribution, the probability distribution P ij (D T (j))(j)) can be defined such that for any given distance D T (j), P ij (D T (j)) is the probability that the image 120 assigned to the category i is at a distance D T (j) from the threshold surface j.

[0036] For a given set of one or more input images 120, assume the estimated category is i0. The loss function used in the training algorithm 204 calculates D T (j) of the shadow pattern of the input image 120 and obtains P i0j (D T j)) for each threshold surface J.

[0037] The loss function may be a function of P i0j (D T (j)) for all threshold surfaces j, and the loss function is P i0j (D T(j) increases as the value of () decreases. For example, the loss function is [Number] or [Number] or it may be that function, where N is the number of threshold surfaces. The above formula is merely illustrative, and other functions of P i0j (D T (j)) may also be used such that the loss function increases as the value of P i0j (D T (j) decreases. Note that the loss function may also be a function of the difference between the estimated category and the assigned category of the training data entry. For example, the loss function may also be a function of |i1 - i0|, where i1 is the subscript of the assigned category and i0 is the subscript of the estimated category.

[0038] The training algorithm 204 can train the machine learning model 202 to minimize the loss function within any constraints (e.g., available data, possible number of iterations).

[0039] Using the above approach, the training algorithm 204 trains the machine learning model 202 to assign each image 120 to the output category 206 such that the distance D T (j) of the image 120 is closer to the peak of the probability distribution of the said category 206 than to the peaks of the probability distributions of other possible categories 206. Since the probability distribution is more continuous than a finite number of categories (e.g., 5 shown in the figure), the training algorithm 204 can easily converge with parameters suitable for the machine learning model 202.

[0040] During training, the machine learning model 202 may be used in the same way as described above for the system 200a. In some embodiments, the visibility threshold data is not used during the use of the machine learning model 202.

[0041] FIG. 3 shows a method 300 for training a machine learning model 202 to determine the clinical significance of floaters using visibility threshold data. Method 300 may be performed with respect to an image 120. For example, for each eye of a plurality of patients, each training data entry can be generated using a series of two or more images 120. Each series of two or more images 120 has a corresponding category 206 (“assigned category”) assigned by a person such as a patient or an expert who has received training.

[0042] Method 300 may include, at step 302, identifying a shadow region 122 within two or more images 120 corresponding to a floater 114. As described above, the shadow region 122 may be identified based on the movement shown in two or more images 120 or using other approaches.

[0043] Method 300 may include, at step 304, measuring the shadow region 122. Measuring the shadow region 122 may include obtaining any of the metrics (size 208, contrast 210, movement direction 212, position 214) described above with respect to FIG. 2A. Measuring the shadow region may include calculating any of the values of the variables used to define any of the visibility threshold surfaces used in method 300.

[0044] Method 300 may include, at step 306, processing some or all of the metrics described above with respect to FIG. 2A using the machine learning model 202 to obtain an estimated category. Method 300 further includes, at step 308, measuring the shadow region 122 to obtain the values of two or more variables that define one or more visibility threshold surfaces.

[0045] Method 300 may include, at step 310, calculating a loss function for the estimated category and two or more values. For example, the loss function may be a function of the probability distribution of two or more values and the estimated category as described above. The machine learning model 202 is then updated by a training algorithm 204 at step 312.

[0046] FIG. 4 shows an exemplary computing system 400 that at least partially implements one or more of the functions described herein with respect to FIGS. 1A - 3. The computing system 400 may be integrated with an imaging device such as an SLO or may be a separate computing device that receives an image of a patient's eye from an imaging device.

[0047] As shown, the computing system 400 includes a central processing unit (CPU) 402, one or more I / O device interfaces 404 that can connect various I / O devices 414 (e.g., keyboard, display, mouse device, pen input, etc.) to the computing system 400, a network 490 (which may be a local network, intranet, Internet, or any other group of computing systems communicatively connected to each other as described in relation to FIG. 1), a memory 408, a storage 410, and a network interface 406 for connecting to an interconnect 412.

[0048] If the computing system 400 is an imaging system such as a digital microscope, OCT microscope, or SLO, the computing system 400 may further include one or more optical elements for performing ophthalmic imaging of a patient's eye and any other elements known to those skilled in the art. If the computing system 400 is a surgical microscope, the computing system 400 may further include many other elements known to those skilled in the art for performing the ophthalmic surgeries described herein, as is known to those skilled in the art.

[0049] The CPU 402 can retrieve and execute programming instructions stored in the memory 408. Similarly, the CPU 402 can retrieve and store application data within the memory 408. The interconnect 412 transfers programming instructions and application data between the CPU 402, the I / O device interface 404, the network interface 406, the memory 408, and the storage 410. The CPU 402 is included to represent a single CPU, multiple CPUs, a single CPU having multiple processing cores, and the like.

[0050] The memory 408 represents volatile memory such as random access memory and / or non-volatile memory such as non-volatile random access memory, phase change random access memory, etc. As shown, the memory 408 can store the machine learning model 202 during training and utilization. In the computing system 400 used for training the machine learning model 202, the memory 408 can further store the training algorithm 204.

[0051] The storage 410 may be non-volatile memory such as a disk drive, a solid state drive, or a group of storage devices distributed across multiple storage systems. Optionally, the storage 410 can store the training data entries 416 to train the machine learning model 202 using the approach described above. The storage 410 can store the visibility threshold data 418, e.g., data describing one or more visibility threshold surfaces and an algorithm for calculating values of variables that define one or more visibility threshold surfaces from the shadow patterns.

[0052] Additional Considerations The above description provides the various embodiments described herein to enable those skilled in the art to implement them. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein can also be applied to other embodiments. For example, changes to the functions and arrangements of the discussed elements may be made without departing from the scope of the present disclosure. In various examples, various procedures or elements may be omitted, substituted, or added as necessary. The features described in some examples may be combined in some other examples. For example, an apparatus may be implemented or a method may be executed using any number of aspects described herein. Also, the scope of the present disclosure is intended to cover, in addition to the various aspects of the disclosure described herein, or other than those, similar apparatuses or methods implemented using structures, functions, or a combination of structures and functions. It should be understood that any aspect of the disclosure disclosed herein may be implemented by one or more of the elements recited in the claims.

[0053] As used herein, the phrase referring to "at least one of" a list of items refers to any combination of those items that includes a single member. As an example, "at least one of a, b, or c" is intended to cover not only a, b, c, a-b, a-c, b-c, and a-b-c, but also any combination with multiple identical elements (e.g., a-a, a-a-a, a-a-b, a-a-c, a-b-b, a-c-c, b-b, b-b-b, b-b-c, c-c, and c-c-c or any other order of a, b, and c).

[0054] As used herein, the term "determine" encompasses a wide range of operations. For example, "determine" may include calculating, computing, processing, deriving, investigating, querying (e.g., querying a table, database, or other data structure), verifying, etc. Also, "determine" may include receiving (e.g., receiving information), accessing (e.g., accessing data in a memory), etc. Also, "determine" may include resolving, selecting, choosing, establishing, etc.

[0055] The methods disclosed herein include one or more steps or operations for implementing the methods. The method steps and / or operations may be interchangeable with each other without departing from the claims. In other words, the order and / or use of specific steps and / or operations may be changed without departing from the claims, unless a specific order of steps or operations is specified. Further, the various operations of the methods described above may be performed by any suitable means capable of performing the corresponding functions. These means may include various hardware and / or software elements and / or modules including, but not limited to, circuits, application specific integrated circuits (ASICs), or processors. Generally, where there are operations shown in the drawings, those operations may include corresponding means-plus-function elements numbered similarly.

[0056] The various exemplary logical blocks, modules, and circuits described in connection with the present disclosure may be implemented or executed by a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic element (PLD), discrete gate or transistor logic, discrete hardware elements, or any combination thereof designed to perform the functions described herein. The general purpose processor may be a microprocessor, but in the alternative, the processor may be any commercially available processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0057] The processing system may be implemented in a bus architecture. The bus may include any number of interconnecting buses and bridges depending on the particular application of the processing system and overall design constraints. The bus can interconnect various circuits, particularly including processors, machine-readable media, and input / output devices. A user interface (e.g., keypad, display, mouse, joystick, etc.) may also be connected to the bus. The bus can also connect various other circuits known in the art and not described further herein, such as a timing source, peripherals, voltage regulators, power management circuits, etc. The processor may be implemented with one or more general-purpose and / or dedicated processors. By way of example, this includes microprocessors, microcontrollers, DSP processors, and other circuits capable of executing software. Those skilled in the art will recognize how to optimally implement the described functions of the processing system depending on the particular application and overall design constraints imposed on the overall system.

[0058] When implemented in software, the above-described functions may be stored or transmitted as one or more instructions or codes on a computer-readable medium. Software shall be construed broadly as meaning instructions, data, or any combination thereof, regardless of how it is called, such as software, firmware, middleware, microcode, hardware description language, etc. The computer-readable medium includes both computer storage media and communication media, such as any medium that facilitates transfer of a computer program from one place to another. The processor may play a role in managing buses and general processing, including execution of software modules stored on a computer-readable storage medium. The computer-readable storage medium may be coupled to the processor such that the processor can read information from, and write information to, the storage medium. Alternatively, the storage medium may be integrated into the processor. As an example, the computer-readable medium may include a transmission line, a carrier wave modulated by data, and / or a computer-readable storage medium in which instructions are stored separately from a wireless node, all of which may be accessed from the processor via a bus interface. Alternatively or additionally, the computer-readable medium, or any portion thereof, may be integrated into the processor, as in the case of a cache and / or a general-purpose register file. Examples of machine-readable storage media include, for example, RAM (Random Access Memory), flash memory, ROM (Read Only Memory), PROM (Programmable Read Only Memory), EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read Only Memory), registers, magnetic disks, optical disks, hard drives, or any other suitable storage medium or any combination thereof. The machine-readable medium may be implemented in a computer program product.

[0059] A software module may contain a single instruction or multiple instructions, and may be distributed across several different code sections, across different programs, and across multiple storage media. A computer-readable medium may contain a number of software modules. A software module, when executed by a device such as a processor, contains instructions that cause a processing system to perform various functions. A software module may contain a transmitting module and a receiving module. Each software module may reside in a single storage device or may be distributed across multiple storage devices. For example, when a trigger event occurs, a software module may be loaded from a hard drive into RAM. During the execution of a software module, the processor may load a portion of the instructions into a cache to increase the access speed. Then one or more cache lines may be loaded into the general-purpose register file for execution by the processor. When referring to the functions of a software module, it should be understood that such functions are realized by the processor when executing the instructions from the software module.

[0060] The following claims are not limited to the embodiments shown in this specification and shall be construed to cover the full scope consistent with the claim language. In the claims, when an element is recited in the singular, it does not mean "only one" unless specifically stated otherwise, but rather "one or more." Unless specifically stated otherwise, the term "some" refers to one or more. No element of any claim shall be construed under 35 U.S.C. § 112, paragraph (f) unless the element is expressly recited using the phrase "means for" or, in the case of a method claim, the element is recited using the phrase "step of." All structural and functional equivalents of the elements of the various aspects described throughout this disclosure that are known or later become known to those of ordinary skill in the art are hereby expressly incorporated by reference and are intended to be encompassed by the claims. Further, the disclosure of this specification is not intended to be dedicated to the public regardless of whether such disclosure is expressly recited in the claims.

Claims

1. A method of training a machine learning model to characterize floating objects, comprising: For each of a plurality of training data entries, Processing, by a computing device, input data from each training data entry having a machine learning model to obtain an estimated clinical significance of the floating object, the input data describing one or more shadow regions in one or more images of a patient's eye retina; (a) Evaluating, by the computing device, an estimated clinical significance with respect to output data from each training data entry and visibility threshold data; Updating, by the computing device, the machine learning model according to the result of (a); A method comprising the above steps.

2. The one or more images include a series of video frames, The input data includes The size of the one or more shadow regions, The contrast of each of the one or more shadow regions, The moving direction of each of the one or more shadow regions, and The position of each of the one or more shadow regions relative to the representation of the fovea of the patient's eye in the one or more images, The method according to claim 1, including one or more of the above.

3. The method according to claim 1, wherein the visibility threshold data includes a visibility threshold surface defined by two or more variables.

4. The two or more variables include Spatial frequency, Temporal frequency, Modulation amplitude, and Eccentricity, The method according to claim 3, including two or more of the above.

5. The method according to claim 2, wherein the visibility threshold data includes a visibility threshold surface defined by two or more variables.

6. The two or more variables include spatial frequency, speed, and modulation amplitude, the method according to claim 5.

7. The two or more variables include spatial frequency, eccentricity, and modulation amplitude, the method according to claim 5.

8. The two or more variables include spatial frequency, temporal frequency, and modulation amplitude, the method according to claim 5.

9. The method according to claim 1, wherein the one or more images include images from a scanning laser ophthalmoscope.

10. The method according to claim 1, wherein the estimated clinical significance includes one selected from a set of possible output categories.

Images