Method, system, and computer program product for visual field mapping

Abstract

To provide a method, system, and computer program product for visual field mapping.SOLUTION: To measure the quality of vision in an eye's visual field, deviations between the gaze position and associated stimulus positions where a stimulus to be tracked is displayed in detecting the gaze position and the magnitude of the registered deviations are determined during a measurement period. For each field portion in the visual field map of the vision, the quality of vision is determined according to an estimate value of the quality of vision associated with the registered deviations located relative to the gaze position such that the associated stimulus position is in that field portion. For each field portion of the visual field map, the quality of vision is determined according to an estimate value of the quality of visual field associated with the registered deviations located relative to the gaze position such that the associated stimulus position is in that field portion. For each associated registered deviation, the quality of vision is estimated according to the magnitude of the associated registered deviation and the magnitude of at least the registered deviations before and after that.SELECTED DRAWING: None

Description

[Technical Field]

[0001] The present invention relates to a method, system, and computer program product for human visual field mapping. Visual field mapping allows for determining a map indicating the quality of vision in a human visual field around the optical axis of one eye or around the optical axis of both eyes. The map of vision quality in the acquired visual field indicates the presence of visual field defects in a portion of the visual field and, if present, the location within the visual field of the portion of the visual field affected by the visual field defect. Traditionally, visual field mapping involves presenting stimuli at several locations within the (potential) visual field of one eye and registering whether the stimuli are viewed. Locations where the stimuli are not reliably viewed receive a low visual score in the visual field map. [Background technology]

[0002] The problem with traditional visual field mapping methods is that they are time-consuming (Standard Automated Perimetry - SAP), insensitive to small defects (Frequency Doubling Technique - FDT), and cannot be performed reliably on people who are unable to concentrate, maintain a fixed gaze position throughout the test, and follow instructions on how to respond to viewed stimuli.

[0003] U.S. Patent Application Publication No. 2006 / 0114414 discloses measuring the field of view of the eye by presenting stimuli at different positions relative to the current direction of gaze, measuring the reaction time of saccadic eye movements in response to stimuli presented in a direction slightly off the current direction of gaze, and calculating the visual sensitivity in the direction in which the stimuli were presented as a function of the measured response time.

[0004] WO 2012 / 141576 discloses a method for determining visual field defects by presenting successive visual stimuli in the visual field of one eye, capturing eye movements in response to the stimuli, and determining saccadic reaction time to new stimuli. If the change in gaze corresponds to the location of the presented stimulus, the stimulus is registered as being viewed. Evaluating saccadic reaction time allows for accurate determination of visual field defects.

[0005] U.S. Patent Application Publication No. 2012 / 0022395 discloses the analysis of abnormal eye movements in a human or animal subject by presenting stimuli, capturing eye movements in response to the stimuli, determining the value of at least one parameter of the saccadic eye movement, and determining abnormality based on the determined value and a predetermined parameter value utilizing an artificial intelligence module.

[0006] US Patent No. 9,730,582 discloses the assessment of eye movement response behavior and the deterioration of visual processing, especially under the influence of diseases affecting visual sensitivity and resolution (such as retinal disease and glaucoma), by measuring characteristics of pursuit behavior including pursuit initiation, pursuit accuracy and speed, and the cloverleaf-shaped anisotropy of the directional acquisition of pursuit responses. Summary of the Invention [Means for solving the problem]

[0007] The object of the present invention is to provide a simple and accurate solution that allows mapping of vision quality in monocular and binocular fields.

[0008] According to the present invention, this object is achieved by providing a method as defined in claim 1. The present invention can also be embodied in a system as defined in claim 14 and in a computer program product as defined in claim 15.

[0009] It is sufficient for the person whose eyes are to be measured to follow the movement of the stimulus to be followed, since for each field portion there is a corresponding stimulus position relative to the gaze position so that the corresponding stimulus position is located in that field portion, the quality of the vision is determined depending on the estimate of the quality of vision of the relevant registered deviations, and for each relevant registered deviation the quality of vision is estimated depending on the magnitude of the relevant registered deviation and the magnitude of at least the registered deviations before and after it, which allows the measurement to be carried out quickly and does not require accurate gaze fixation on the stimulus before displaying another stimulus at the position in the visual field whose visual acuity is to be determined. This method is particularly insensitive to inaccuracies in the direction of gaze and to disturbances due to the concentration and reaction time of the person whose eyes are to be measured.

[0010] Because the eyes are more strongly induced to move their gaze toward the stimulus the farther the stimulus is from the gaze position, a large deviation within a given duration is a stronger indication of a visual defect than a small deviation within the same duration, and therefore the quality of vision at a given location in the visual field is determined by the magnitude of the deviation when the stimulus is at that location in the visual field, as well as the deviations before and / or after.

[0011] The duration of the group of deviations of which a deviation at a given location is a part indicates the slowness of fixation of the gaze position relative to the current location of the stimulus, and therefore the quality of vision at a given location in the visual field is further determined by the duration of the series of deviations that includes that deviation and that begin and end with a deviation of magnitude below a minimum threshold magnitude. Slow fixation at a given location indicates a visual field defect at that location in the visual field.

[0012] To obtain useful indicators of various deviations and types of visual dysfunction, it is preferable to move the stimulus to be tracked at various speeds. The speed variations preferably include gradual increases and decreases in speed (continuous acceleration and deceleration curves) to uniformly generate deviations of various magnitudes. The speed variations may also include speed jumps to generate relatively large deviations. Furthermore, continuously moving the stimulus position during the measurement period is particularly effective in avoiding artifacts caused by inaccurate gaze fixation on the stimulus to be tracked.

[0013] To avoid disturbances in measurement results due to blinking or missing measurements, gaze positions measured while the gaze position moves at a speed exceeding a threshold (e.g., exceeding 300 degrees / s), stops, and then moves again at a speed exceeding the threshold are replaced with interpolated gaze positions.To effectively remove such disturbances, positions within a predetermined range of time (e.g., 1 / 60 to 1 / 15 seconds) or number of samples (e.g., 2 to 10 samples) before and after the gaze position measured while the gaze position moves at a speed exceeding the threshold, stops, and then moves again at a speed exceeding the threshold are replaced with interpolated gaze positions.

[0014] In order to carry out the measurement according to the invention particularly accurately, it is preferable that there is only one moving stimulus to be followed, more particularly that there is only one moving stimulus or that exactly one stimulus is displayed.

[0015] According to a preferred embodiment, the method includes the use of a recurrent neural network, for example a recurrent neural network trained with gaze positions obtained by measuring the gaze positions of a healthy eye following a displayed stimulus to be followed.

[0016] According to this embodiment, it is possible to determine an index of the type of visual impairment in the entire visual field using a neural network from the registered gaze position and stimulus position. For example, it is possible to simulate a visual impairment by suppressing the display of the stimulus to be followed in a predetermined part of the field.

[0017] Indicators of types of visual dysfunction across the visual field, such as slowed eye movements, nystagmus, under- or over-measured saccadic eye movements, which provide potentially useful diagnostic checkpoints, can be determined in a particularly accurate way from the registered gaze and stimulus positions by using a recurrent neural network trained with gaze positions obtained by measuring the gaze positions of healthy eyes following displayed stimuli to be followed, and visual impairments can be simulated by suppressing the display of the stimuli to be followed in parts of the field.

[0018] In embodiments of the present invention, indicators of the type of visual dysfunction that provide potentially diagnostic checkpoints can be determined particularly reliably if the data input to the recurrent neural network during training and during use of the trained recurrent neural network includes at least one of the following categories of data: - The maximum correlation between gaze position velocity and stimulus position velocity. - The temporal offset (delay) between gaze position and stimulus position. -Temporal precision of gaze position relative to stimulus location. - Variance explained by a Gaussian model fitted to the correlation plot of gaze position velocity and stimulus position velocity versus time delay. - The number of occurrences of the gaze position most likely to deviate from the stimulus position. - The average spatial offset (bias) between gaze position and stimulus position. - Mean deviation of gaze position from stimulus position. - Variance of deviation of gaze position from stimulus position explained by a Gaussian model fitted to a graph of deviation occurrence in multiple ranges.

[0019] In a particularly precise embodiment, the determination of the quality of vision for each relevant one of the registered deviations is performed by integrating over the series of registered deviations, the series being Starting from the beginning of the series of registered deviations whose magnitude is less than or equal to the minimum threshold h0, Among the relevant ones of the registered deviations, a first successive registered deviation having an increasing magnitude up to the magnitude h of the relevant deviation; Among the deviations registered, including relevant ones, and a second successive registered deviation having a magnitude decreasing from a magnitude h of the relevant one of the registered deviations to the last deviation in the series having a magnitude less than or equal to the minimum threshold magnitude h0 of the registered deviations; and Except for the first one of the series of registered deviations and the last one of the series of registered deviations, an uninterrupted series of registered deviations whose magnitude is greater than the minimum threshold magnitude h0 are formed.

[0020] In order to appropriately weight the magnitude of the series of deviations for which the integral is calculated, before integration of the series of registered deviations, the magnitude of the series of registered deviations that is greater than h is preferably reduced to an upper limit value corresponding to the magnitude h, which is preferably equal to h.

[0021] The integration at each time point from the start of deviation from a threshold magnitude h0 at h, back to the threshold magnitude h0, and capping at magnitude h can be performed by cluster analysis, e.g., calculation using a threshold-free cluster enhancement algorithm.

[0022] In another embodiment, a particularly accurate determination of the quality of the field of view for each field portion is performed by: a learning input including a series of learning time points each including a stimulus position that is the position of a stimulus to be followed and a gaze position of a normal eye that follows the stimulus to be followed, and an indicator of whether, at each learning time point, the stimulus position relative to the gaze position is in a field portion where the display of the stimulus to be followed is suppressed; a learning output comprising a map of field portions where the presentation of the to-be-followed stimulus was suppressed during the measurement session to obtain the learning input; This is performed using a recurrent neural network that is trained to obtain a recurrent neural model having

[0023] The method according to this embodiment preferably comprises: inputting a series of time points each including a stimulus position, which is the position of the stimulus to be tracked, and a gaze position of the eye to be tested that follows the stimulus to be tracked; Preferably, the trained recurrent neural network classifies the quality of vision at a series of time points into scotoma time points at locations within the visual field where vision is classified as functional and non-scotoma time points at locations within the visual field where vision is classified as impaired; determining, for each of the field portions, a time point at which the deviation of the stimulus position relative to the gaze position is in that field portion; generating, for each of the field portions, a visual field map showing an aggregated visual field quality value according to the visual quality estimated at the time points having the stimulus location determined to be in that field portion; Includes during operation.

[0024] If the recurrent neural network includes a fully connected layer in which all output units are connected to all input units and vice versa, and at least one gated recurrent unit with long short-term memory to capture time dependencies and that processes sequential information in a recursive manner, the recurrent neural network can be particularly effective in taking into account the duration of each deviation that occurs in the registered stimulus position and the associated gaze position.

[0025] In order to estimate a particularly accurate visual field map, the method Acquiring and inputting stimulus luminance and tracking type data during the measurement period; At least two fully connected layers that process the luminance and tracking data into categorical data indicating the type of scotoma (e.g., binasal hemianopsia, bitemporal hemianopsia, blind spot, cortical spreading depression, scintillating scotoma); inputting the combination of time points and categorical data from the gated recurrent unit into a softmax classifier of a recurrent neural network; A softmax classifier predicts, for each time point, whether the stimulus position at that time point is located in the visual field at a position on the scotoma; Advantageously, it further comprises:

[0026] Further features, advantages and details of the present invention will become apparent from the detailed description and drawings. [Brief explanation of the drawings]

[0027] [Figure 1] 1 is a flow chart of steps for acquiring data regarding eye movements during a test. [Figure 2] 1 is a schematic diagram of a system according to the present invention, with an example in which a visual tracking stimulus is displayed; [Figure 3] 1 is a graph showing the horizontal position of the stimulus over time sh(t). [Figure 4] 1 is a flowchart of steps for extracting spatiotemporal features and categorizing visual field defects based thereon. [Figure 5A] 10 is a graph of observed gaze direction over time, before and after filtering to account for missing data and blink-related displacements. [Figure 5B] 10 is a graph of observed gaze direction over time, before and after filtering to account for missing data and blink-related displacements. [Figure 6A] 10 is a graph showing the rate of change of registered gaze direction over time. [Figure 6B] 1 is a graph of the cross-correlation between stimulus velocity and delayed eye velocity versus delay time (cross-correlation curve). [Figure 7A]Graph of registered stimuli and gaze position over time in the x direction. [Figure 7B] 10 is a graph of the probability density distribution of the positional shift between the gaze position and the stimulation position. [Figure 8] 1 is a flowchart of a stadia survey of a deep recurrent neural network architecture to determine category (visual defect type) classifiers and timepoint classifiers. [Figure 9] 1 is a flowchart of steps for determining a visual field map for one eye from data regarding eye movements during a test. [Figure 10A] 1 is an example of a visual field map without visual field defects. [Figure 10B] This is a visual field map obtained from measurement data obtained from testing of a normal eye (i.e., the map shown in Figure 10A is applied) without suppressing the stimulus presentation at a predetermined location by threshold-free cluster enhancement (TFCE) processing. [Figure 10C] This is a visual field defect map (i.e., the map shown in Figure 10A) obtained from measurement data obtained from a performance test of a normal eye without suppressing stimulus presentation at a predetermined location by recursive neural network (RNN) processing. [Figure 10D] Map of the location of pseudoscotoma (peripheral visual field defect pattern type) formed by the location of suppressed stimulus presentation when testing normal eye performance. [Figure 10E] This is a visual field map obtained by TCFE processing from measurement data obtained from a performance test of a normal eye in which stimulus presentation at a predetermined position was suppressed as shown in Figure 10D. [Figure 10F] This is a visual field defect map obtained by RNN processing from measurement data obtained from a performance test of a normal eye in which stimulus presentation at a predetermined location was suppressed, as shown in Figure 10D. [Figure 10G]Map of the location of pseudoscotoma (hemianopsia pattern type) formed by the location of suppressed stimulus presentation when testing the performance of the healthy eye. [Figure 10H] This is a visual field map obtained by TFCE processing from measurement data obtained from a performance test of a normal eye in which stimulus presentation at a predetermined location was suppressed as shown in Figure 10G. [Figure 10I] As shown in Figure 10G, this is a visual field defect map obtained by RNN processing from measurement data obtained from a performance test of a healthy eye in which stimulus presentation at a predetermined location was suppressed. DETAILED DESCRIPTION OF THE INVENTION

[0028] The invention will be further explained with reference to examples of methods according to the invention and examples of tests for validating the methods according to the invention.

[0029] The following hardware is available for the method according to the invention and for testing to validate the method according to the invention: A display screen 52 (see FIG. 2), such as an LCD monitor. An eye tracker 53 for tracking the gaze direction of the eyes on the display, such as the monitor-integrated eye tracker Eyelink 1000 (SR-Research, Ottawa, Canada). a data processing system 54 connected to the display screen 52 for controlling the display screen 52 to display the visual stimulus 1 at a stimulus position (in this example the centre of the stimulus blob) that moves across the display screen 52, and connected to the eye tracker 53 for receiving data representing gaze position from the eye tracker. The data processing system 54 is programmed to carry out the method according to the example described below, in which data from the eye tracker 53 is acquired at a sampling rate of 1000 Hz and downsampled to match in time the 240 Hz refresh rate of the display screen 52.

[0030] In this example, the visual stimulus to be tracked is a uniform gray background (∼140 cd / m 2 ) (see Figure 2). Its motion includes components in the vertical (Y) and horizontal (X) directions. The Gaussian blob 1 can be presented within the following range of contrast levels: maximum luminance of ~385 cd / m at maximum contrast (50%); 2 , whereas when presented at minimum contrast (5%), the maximum luminance is ~160 cd / m 2 The size (full width at half maximum) of Gaussian blob 1 is 0.83 degrees of visual angle, which corresponds to size III in the stimulus of a Goldmann perimeter, a commonly used visual field measurement device. The person whose visual field is being tested is instructed to follow stimulus 1 with the line of sight of the eye being tested. In addition to stimulus 1 to be followed, other visual stimuli can also be displayed, and they may be moving and / or stationary. In this example, no other stimuli are presented.

[0031] In step 3 (Fig. 1), a stimulus trajectory is created, consisting of a random path with the following constraints: The stimulus locus 4 must be within the boundaries of the screen. The stimulus trajectory cannot contain periodic autocorrelation. The stimulation trajectory 4 in this example is constructed by generating the following velocity vectors:

[0032]

number

[0033] At each time point, the velocity values ​​for the horizontal (vx) and vertical (vy) components are described by a Gaussian distribution with a mean of 0 and separate standard deviations for the horizontal (σh) and vertical (σv) directions. For example, for a standard screen resolution of 1920 x 1080 pixels, the horizontal stand deviation can be σh = 64.45 deg / s and the vertical stand deviation can be σv = 32.33 deg / s. Preferably, these σ values ​​may be adjusted based on the screen dimensions and / or the specific application, e.g., obtaining a representation of the visual field map under specific conditions (e.g., taking measurements from individuals with known illnesses or known symptoms).

[0034] The velocity vectors can be low-pass filtered by convolution with a Gaussian kernel to obtain a cutoff frequency, preferably around 10 Hz. Afterwards, the velocities are converted to the position of the stimulus 1 via time integration.

[0035]

number

[0036] To induce a visual acuity test subject to perform saccades, the trajectory can include rapid displacements in random directions, as long as the displayed stimulus to be followed does not leave the boundaries of the screen. This can be achieved by adding such displacements after a fixed number of seconds, preferably 2 seconds, or after random time intervals. Figure 3 shows an example of a graph 4' of the horizontal component sx of the stimulus trajectory's position over time. As shown in Figure 1, this created trajectory 4 is stored in storage 5. In this embodiment, all data is stored in the same storage 5, but multiple data storages may store sets of dates.

[0037] For data acquisition, the person whose vision is to be tested is positioned in front of the screen, e.g., at a viewing distance of 60 cm. Note that other display technologies, such as holography, can be used instead of a screen. In step 6 (Fig. 1), stimulus 1 is displayed and moved on screen 52, with x-direction movement following stimulus trajectory 4 retrieved from storage 5 and y-direction movement following another stimulus trajectory, preferably with rapid x-direction displacements and rapid y-direction displacements simultaneously. Head movements are preferably minimized, for example, by placing the head on a chin rest, or filtered out. Before each data acquisition session, eye tracker 53 is preferably calibrated using a standard procedure, such as 9-point calibration. Eye tracking step 13 preferably includes multiple trials. In this example, each trial lasts 20 seconds and is repeated six times. The acquired results are horizontal and vertical gaze coordinates (expressed in pixels) at each time point. The pixel values ​​are converted to visual angle, and time series data of gaze position 7 are obtained.

number

[0038] As a preprocessing step, the time series of gaze positions 7 are read from storage 5. In step 15 of the preprocessing step of this embodiment, the first time period (e.g., the first 250 milliseconds) of each time series is discarded to avoid artifacts while the subject is settling down.

[0039] Other artifacts that should be avoided are artifacts due to blinking and missing data. The blink period can be determined by, for example, the time it takes for the gaze position 7(

number

[0040] Thus, in this embodiment, a graph 7' (see FIG. 5B) of the filtered data 14 (see FIG. 1) is obtained, and this graph is also stored in the storage 5.

[0041] The time series of position error is

number

number

[0042] In step 17, the gaze position 7' ​​of the filtered data 14

number

number

number

number

number

number

[0043] Filtered stimulus position velocity

number

number

number

number

[0044] CCG ·amplitude:

number

number

[0045] PDD · Amplitude: Most likely to be misaligned. · μ: spatial offset (bias) between the eye and the stimulus. · σ: Spatial accuracy of tracking performance. ·R2: Explains the variance of the spatial Gaussian model.

[0046] The eight spatiotemporal features listed above representing the relationship between stimulus position and associated gaze position in one measurement are calculated in both the horizontal and vertical directions, resulting in a total of 16 spatiotemporal features being calculated and stored in storage 5. In step 45 (FIG. 4), these spatiotemporal features are input to and processed by category classifier 45 to determine an estimate of the overall visual field classification 35.

[0047] The spatiotemporal integral of the displacement between the stimulus and gaze positions can be used to determine a quality map of the visual field of the eye whose eye movements are being measured. A person with reduced oculomotor function, e.g., reduced vision in one eye's visual field, will have their gaze position deviate from the target stimulus more often and for longer periods of time than a person with normal oculomotor function. The magnitude of the displacement 46 (Figure 9) as a function of time, h(t), can be defined as:

[0048]

number

[0049] The spatiotemporal integration 47 is performed using a threshold-free cluster enhancement (TFCE) algorithm. Such an algorithm is described in Smith, S. M. & Nichols, T. E., 2009, Threshold-free cluster enhancement: addressing problems of smoothing, threshold dependence and localization in cluster inference, Neuroimage, 44, 83-98. In this embodiment of the algorithm, for each registered combination of stimulus position and gaze position, the TFCE score is given by the sum of the magnitudes of all the "supports" of the cluster formed by the deviation curve below it. First, the magnitude h is gradually increased from h0, and at a certain time point t, the magnitude h is calculated. tWhen the deviation curve is increased to , the time lapse reaches a threshold value of h, and the contiguous cluster containing t ends when the magnitude returns to h0. The surface area under the deviation curve defines the space-time integral score for that magnitude h. This score is the magnitude h (to the H power) multiplied by the time e (to the E power), and is expressed as follows:

[0050]

number

[0051] This integral is implemented as a discrete sum with a finite step size dh (e.g., dh = 1 / 2500 of the maximum value of h). h0 is typically the minimum value of h, and E and H can be set in advance (optimized by simulated field effects). The resulting spatiotemporal integrated property: D STI Thus, for each displacement, a time series of weighted displacements 48 is obtained. STI is determined by the magnitude of the deviation and the time it takes for the deviation curve, which is a part of the deviation, to return to the magnitude h0. Also, the value D assigned to the positional deviation 48 STI can also be defined as forming the integral of each time point from the beginning of the deviation to magnitude h until the deviation returns to magnitude h0 and caps at magnitude h.

[0052] To determine the quality of the visual field of the visual field map 51 of the eye where the measurement was performed, the input time series D STI Each occurrence of is associated with its horizontal and vertical components D'x = px(t) - sx(t) and D'y = py(t) - sy(t), respectively, which form the x and y components of the stimulus's position relative to the center of the visual field and contain an estimate of the quality of the visual field at that position relative to the center of the visual field for each time point.

[0053] In step 50, these components are mapped onto a Cartesian coordinate plane with its origin at the center of the line of sight of the eye. Thus, the Cartesian coordinate plane represents the visual field, and its center represents the fovea.

[0054] In step 50, spatial binning may be applied (e.g., where the bin size corresponds to one degree of field of view), so that the value of each bin is the D STI This is the average value of the values.

[0055] Finally, the calculated map 51 is stored in storage 5 and can be displayed, for example, on a 50x40 grid covering a ±25 degree field of view. The map indicates the severity of the visual field defect (e.g., color-coded or grayscale), making the map easy to interpret by an ophthalmologist.

[0056] Figures 10B, 10E, and 10H show example visual field maps resulting from an experimental implementation of the spatiotemporal integration described using the TFCE method.

[0057] Instead of the spatiotemporal integration method, a trained recurrent neural network can be used to estimate the visual quality map of the eye's visual field from the measurements. By training the recurrent neural network, we can obtain the artificial intelligence classifiers 34 and 45 (Figure 8). The training input x is the gaze position.

number

number

[0058] The deep recurrent neural network first determines the gaze position.

number

number

[0059] Every time point of the sequential data (from the GRU layer 41) is merged with the categorical data (obtained using the luminance level and tracking type information from the FC layer 42) and processed by two FC layers 43 with 32 and 2 nodes, respectively. The output of these FC layers 43 is the input of a softmax classifier 34, which predicts whether the stimulus position is above the scotoma, relative to its position in the visual field, i.e., the gaze position, at each time point.

[0060] The other stream merges the last time point of the sequential data (from the GRU layer 41) with the categorical data (from the FC layer 42), and these data are processed by two FC layers 44, one with 32 nodes and the other with 4 nodes. The output of these FC layers 44 is the input of another softmax classifier 45, which predicts the state of the visual field (categorical classifier 45).

[0061] Cross-entropy loss is used to define the cost function of the model.

[0062]

number

[0063] where M is the number of classes, y is the ground truth label (obtained from the simulated visual field map), and p is the predicted probability distribution, i.e., the output of each softmax classifier. The subscript s refers to the pointwise scotoma classifier 34, and the subscript d refers to the visual field defect classifier 45. To prioritize optimizing the visual field quality map within the visual field, α and β may be set to, for example, the following values: α = 0.75, β = 0.25. The model parameters θ can be trained using, for example, mini-batch gradient descent with RMSprop for 15,000 iterations with a batch size of B = 128.

[0064] A training batch can be formed by first selecting B distinct sequences from a set of trials, e.g., 20 seconds long, initially sampled at 240 Hz. Then, from each sequence, one subsequence of 4.17 seconds (1000 time steps) can be randomly sampled, and finally downsampled to 60 Hz (250 time steps). The brightness level and tracking type of the corresponding sequence are also added to the training batch.

[0065] To determine the visual field defect classification 35 (for the entire visual field), a decision tree (DT) algorithm can be trained to classify and reduce the dimensionality of the feature space. Each node of the DT divides the feature space into subspaces, ultimately resulting in one of a set of possible visual field defects. At each division, a decision is made based on the Gini's Diversity Index (GDI), a decision criterion given below:

[0066]

number

[0067] Here, f(i) is the fraction of samples in the training set of class i that reach a particular node. If a node contains only samples of a unique class, the GDI value is 0, the node is assigned to that class, and the decision is complete. The performance of the classifier is evaluated using a 10-fold cross-validation method, randomly dividing the entire dataset into 10 equally sized subsets. Nine subsets then constitute the training set, and the remaining set is used as the test set. This process is repeated until all subsets have been used once as the test set. The estimated overall accuracy is the average of the accuracies measured after each iteration. This analysis using the categorical classifier 45 leads to a categorical classification 35 of the observer's visual field defect (e.g., no defect, central defect, peripheral defect, hemi-defect), which is stored in storage 5 and can be used by ophthalmologists as a preliminary screening tool to determine what further steps may be most needed to make a diagnosis.

[0068] Models 34 and 45 can be considered as a mapping of y = f(x;θ),

number

[0069] For example, data acquisition for one eye can be performed using six 20-second trials for each luminance / tracking combination, and the predicted output probability distributions for multiple subsequences can be averaged. This allows averaging predictions for 6 x 2 x 2 = 24 downsampled sequences. The predicted visual field defect35 for the eye whose eye movements were measured is then:

[0070]

number

[0071] where M is the number of subsequences in trial set S for eye s whose eye movements were measured.

[0072] In particular, given the time series px(t), sx(t), py(t) and sy(t) along with the intensity and tracking, the classifier model 34 (FIG. 4) calculates the scotoma overlap p for each downsampled subsequence. s Then, successive subsequence predictions are concatenated to produce a classified misalignment time series,49,D labeled Form.

[0073] In the visual field mapping step 50, the deviation D labeled The set of is divided into two subsets according to the labeling: if the label has a value of 1, the particular data point is classified as occluded by a scotoma, while if the label has a value of 0, the particular data point is classified as not occluded by a scotoma.

[0074] To determine the visual field quality map 51 of one eye where the measurement was taken, the input time series D labeled Each occurrence of is associated with its horizontal and vertical components, respectively, D'x = px(t) - sx(t) and D'y = py(t) - sy(t), using a trained recurrent neural network, which form the x- and y-components of the stimulus's position relative to the center of the visual field, including an estimate of the visual field quality at that position relative to the center of the visual field, for each time point.

[0075] In step 50, these components are mapped onto a Cartesian coordinate plane with its origin at the center of the line of sight of the eye. Thus, the Cartesian coordinate plane represents the visual field, and its center represents the fovea.

[0076] In step 50, spatial binning may be applied (e.g., so that the bin size corresponds to one degree of field of view), so that the value of each bin is the D labeled This is the average value of the values.

[0077] Finally, the calculated map 51 may be stored in storage 5 and can be displayed, for example, on a 50x40 grid covering a ±25 degree field of view. The map indicates the severity of the visual field defect (e.g., color-coded or grayscale), making the map easy to interpret by an ophthalmologist.

[0078] 10C, 10F, and 10I show example visual field maps resulting from an experimental implementation of the recurrent neural network method described above.

[0079] 10A, 10D, and 10G are maps showing the location of the scotoma where the presentation of the visual stimulus to be tracked is suppressed (black indicates suppression, white indicates no suppression, and gray indicates increasing partial suppression as it gets darker). More specifically, FIG. 10A shows an example of an unmasked test data map for simulating an eye without a visual field defect, FIG. 10D shows an example of a test data mask for simulating an eye with a peripheral defect, and FIG. 10G shows an example of a test data mask for simulating an eye with a hemi-defect.

[0080] Figures 10B, 10E, and 10H show maps reconstructed using the values ​​provided by the TFCE algorithm described above from measurements obtained with a normal eye when the stimulus to be tracked is not masked (Figure 10B), when it is masked as in Figure 10D (Figure 10E), and when it is masked as in Figure 10G (Figure 10H), respectively. The maps reconstructed using the values ​​provided by the TFCE algorithm described above demonstrate the accuracy with which a visual field map of a priori known scotoma can be reconstructed based on measurements of an eye in which the scotoma is simulated by (virtual) masking using test data.

[0081] Figures 10C, 10F, and 10I show maps obtained using the predictions provided by the RNN model from measurements taken in a normal eye when the stimulus to be followed is not masked (Figure 10C), when it is masked as in Figure 10D (Figure 10F), and when it is masked as in Figure 10G (Figure 10I), respectively. As can be seen, for both eyes, the maps reconstructed using the predictions provided by the RNN model are nearly identical to the maps of the simulated scotoma locations.

[0082] Some features are described as part of the same or separate embodiments, however, it will be understood that the scope of the present invention also includes embodiments having all or any combination of these features other than the specific combination of features embodied in the examples. [Explanation of symbols]

[0083] 3 Creating a trajectory 4 Stimulation trajectory 5. Storage 6. Stimulus presentation 7 Gaze position 13 Eye Tracking 14 Filtered Data 15 Preprocessing data 17 Extracted Velocity 19 Calculated deviation 29 Cross-correlation 31 Averaging 32 Gaussian Fit 33 Features 34 Timepoint Classifier 35 Classified Fields of View 36 Histogram 39 Gaze position / stimulus position 40 2xFC layers · brightness and tracking type 41 3x bidirectional GRU layer 42 2xFC layers 43 2xFC layer 44 2xFC layers 45 Category Classifier 46 Euclidean distance 47 Space-time integral 48 Thresholdless Cluster Enhancement 49 Classified Data 50 Field of View Mapping 51 Field of View Map 52 display screen 53 Eye Tracker 54 Data Processing Systems

Claims

1. 1. A method for measuring quality of vision in a visual field of an eye, the method comprising: Displaying a stimulus to be tracked at a stimulus position in the direction of gaze relative to the eyeball; Moving the stimulus to be tracked in various directions within the visual field of the eyeball and registering the stimulus position over time; Detecting and registering the gaze position of the direction of the line of sight of the eyeball following the stimulus to be followed over time; determining and registering a deviation between the gaze position and a relevant one of the stimulus positions at which the stimulus to be tracked was displayed when the gaze position was detected, and a magnitude of the registered deviation; determining a visual field map in field portions, wherein for each of the field portions, the associated stimulus position is located relative to the gaze position such that the associated stimulus position is located in that field portion, the quality of vision is determined as a function of an estimate of the quality of vision of an associated one of the registered deviations, and for each associated one of the registered deviations, the quality of vision is estimated as a function of the magnitude of the associated one of the registered deviations and the magnitude of at least one of the registered deviations preceding and following it; during the measurement period.

2. The method of claim 1 , wherein the stimulus to be followed is moved at different speeds.

3. 3. The method of claim 1, wherein a gaze position measured while the gaze position moves at a velocity above a threshold, stops, and moves at a velocity above the threshold again is replaced with an interpolated gaze position.

4. 4. The method of claim 3, wherein positions within a predetermined range of time or number of samples before and after the gaze position measured while the gaze position moves at a speed exceeding the threshold, stops, and moves at a speed exceeding the threshold again are replaced with an interpolated gaze position.

5. 5. The method of claim 1, wherein the stimulus to be followed is the only moving stimulus being displayed.

6. 6. The method according to any one of claims 1 to 5, comprising using a recurrent neural network, in particular a recurrent neural network trained with gaze positions obtained by measuring the gaze positions of a normal eye following a displayed stimulus to be followed.

7. 7. The method of claim 6, wherein an index of the type of visual dysfunction across the visual field is determined from the registered gaze and stimulus positions using the neural network.

8. 8. The method of claim 7, wherein visual impairment is simulated by suppressing the display of the stimulus to be followed in a predetermined portion of the field.

9. During training and during use of the trained recurrent neural network, data input to the recurrent neural network is: - the maximum correlation between gaze position velocity and stimulus position velocity, - the time offset (delay) between gaze position and stimulus position, -Temporal precision of gaze position relative to stimulus location, - variance explained by a Gaussian model fitted to the correlation plot of gaze position velocity and stimulus position velocity versus time delay, - the number of occurrences of the gaze position most likely to deviate from the stimulus position; - the mean spatial offset (bias) between the gaze position and the stimulus position, - the mean deviation of the gaze position from the stimulus position, - the variance of deviation of the gaze position from the stimulus position explained by a Gaussian model fitted to a graph of the occurrence of deviations in multiple ranges, 9. The method of claim 6, wherein the data includes at least one of the following categories of data:

10. 10. The method of claim 1, wherein the stimulation location is moved continuously during the measurement period.

11. The determination of the quality of vision for each relevant one of the registered deviations is performed by integrating over the series of registered deviations, the series comprising: Minimum threshold h ０ Starting from the beginning of the series of registered deviations, which have the following magnitudes: Among the associated ones of the registered deviations, a first succession of the registered deviations having increasing magnitudes up to a magnitude h of the associated deviation; Among the registered deviations, including the relevant ones, and The magnitude h of the relevant one of the registered deviations is used to calculate the minimum threshold magnitude h of the registered deviations. ０ a second successive registered deviation having a decreasing magnitude until the last deviation in the series has a magnitude equal to or less than the magnitude of the first deviation; and The magnitude of each deviation is greater than or equal to the minimum threshold magnitude h, except for the first deviation in the series of registered deviations and the last deviation in the series of registered deviations. ０ 11. The method according to claim 1, further comprising forming a larger uninterrupted series of the registered deviations.

12. 12. The method according to claim 11, wherein, before integration of the series of registered deviations, magnitudes of the series of registered deviations that are greater than h are reduced to an upper limit value that depends on the magnitude h, the upper limit value being preferably equal to h.

13. determining the quality of vision for each field portion; A learning input includes a series of learning time points each including a stimulus position that is the position of the stimulus to be followed and a gaze position of a normal eye that follows the stimulus to be followed, and an indicator of whether the stimulus position relative to the gaze position is in a field portion where the display of the stimulus to be followed is suppressed at each learning time point; a training output including a map of field portions where presentation of the to-be-followed stimulus was suppressed during a measurement session to obtain the training input; 11. The method of claim 1, wherein the method is performed using a recurrent neural network trained to obtain a recurrent neural model having:

14. inputting a series of time points each including a stimulus position, which is a position of the stimulus to be tracked, and a gaze position of the eye to be tested that follows the stimulus to be tracked; the trained recurrent neural network preferably classifies the quality of vision at the series of time points into scotoma time points at locations within the visual field where vision is classified as functional and non-scotoma time points at locations within the visual field where vision is classified as impaired; determining, for each of the field portions, a time point at which the deviation of the stimulus position relative to the gaze position is in that field portion; generating, for each of the field portions, a visual field map showing an aggregated visual field quality value according to the estimated visual quality at the time points having the stimulus location determined to be in that field portion; The method of claim 13 , comprising, during operation:

15. 15. The method of claim 13 or 14, wherein the recurrent neural network comprises a fully connected layer in which every output unit is connected to every input unit and vice versa, and at least one gated recurrent unit with long short-term memory functionality to capture time dependencies and that processes sequential information in a recursive manner.

16. acquiring and inputting luminance and tracking data during said measurement period; at least two fully connected layers that process the luminance and tracking data into categorical data indicative of the type of scotoma; inputting the combination of time points from the gated recurrent unit and the categorical data into a softmax classifier of the recurrent neural network; The softmax classifier predicts, for each time point, whether the stimulus position at that time point is on a scotoma within the visual field; 16. The method of claim 15, further comprising:

17. 1. A system for measuring quality of vision in a visual field of an eye, the system comprising: The display and an eye tracker that tracks a gaze position of an eyeball on the display; a data processing system including a video display controller connected to the display for controlling the display to display a visual stimulus to be tracked at a stimulus position that moves on the display, and connected to the eye tracker for receiving data representing the gaze position from the eye tracker; The data processing system includes: displaying the stimulus to be followed at the stimulus position on the display; receiving data representing the gaze position from the eye tracker; moving the stimulus to be followed on the display in various directions and registering the stimulus position over time; registering the received gaze positions in chronological order; determining and registering a deviation between the gaze position and a relevant stimulus position where the stimulus to be tracked was displayed when the gaze position was detected, and a magnitude of the deviation; determining a visual field map in field portions, wherein for each of the field portions, an associated stimulus location is located relative to the gaze position such that the associated stimulus location is located in that field portion, the quality of vision is determined as a function of an estimate of the quality of vision of an associated one of the registered deviations, and for each associated one of the registered deviations, the quality of vision is estimated as a function of the magnitude of the associated one of the registered deviations and the magnitude of at least one of the registered deviations preceding and following it; during the measurement period.

18. A computer program product stored in a computer-readable form, the computer program being capable of executing, when executed on a computer: controlling a display to display a stimulus to be followed at the stimulus location; receiving data representing a gaze position on the display; controlling the display to move the stimulus to be followed in different directions on the display and to register the position of the stimulus over time; registering the received gaze positions in chronological order; determining and registering a deviation between the gaze position and a relevant stimulus position where the stimulus to be tracked was displayed when the gaze position was detected, and a magnitude of the deviation; determining a visual field map in field portions, wherein for each of the field portions, a quality of vision is determined depending on the quality of vision of associated ones of the registered deviations, relative to which the gaze position has an associated stimulus location, such that the associated stimulus location is in that field portion, and for each associated one of the registered deviations, a quality of vision is estimated depending on the magnitude of the associated one of the registered deviations and the magnitude of at least the registered deviations preceding and following it; A computer program product that causes a computer to execute the above.

Images