Stimulation of a retina with an algorithm for adjusting stimulation parameters to approximate an actual value to a target value.

Abstract

Arrangement (1) for stimulating a retina (2), comprising: - a camera (3), - a large number of electrodes (5) for interaction with the retina (2), - an electronics (6) which is set up to ▪ to determine a target value (8) for the stimulation of the retina (2) for a large number of pixels, ▪ to determine a respective set of stimulation parameters (9) for each of the pixels, ▪ to deliver to one of the electrodes (5) for each of the pixels to stimulate the retina (2), ▪ to receive action potentials (11) generated by the retina (2) for each of the pixels, ▪ to determine an actual value (13) of the stimulation of the retina (2) for each of the pixels, ▪ to adjust the determination of the set of stimulation parameters (9) for each of the pixels such that the actual value (13) approaches the target value (8), wherein at least the plurality of electrodes (5) is designed as part of a retinal implant (19) of the arrangement (1).

Description

[0001] The invention relates to an arrangement and a method for stimulating a retina.

[0002] Retinal implants are known to provide blind people with shadowy vision. Such retinal implants typically have numerous electrodes for stimulating the retina. Electrical signals are sent to the retina via these electrodes. These signals are generated using a chain of algorithms based on a signal from a camera that optically records the surroundings.

[0003] Current retinal implants do not provide satisfactory visual impressions. In particular, the visual impression achievable with current retinal implants is often described as unnatural.

[0004] EP 0 971 770 B1 discloses a learning-capable, sensorimotor encoder for a visual prosthesis, which comprises a central control unit for signal processing functions, monitoring functions, control functions and external intervention functions as well as a group of adaptive spatiotemporal filters for converting sensor signals into stimulation pulse sequences.

[0005] The object of the present invention is to achieve an improved visual impression compared to known retinal implants by stimulating a retina.

[0006] This object is achieved by the arrangement and method according to the independent claims. Further advantageous embodiments are specified in the dependent claims. The features presented in the claims and in the description can be combined with one another in any technologically expedient manner.

[0007] According to the invention, an arrangement for stimulating a retina is presented. The arrangement comprises: - a camera, - a variety of electrodes for interaction with the retina, - an electronic device which is designed to ▪ use an image evaluation algorithm to determine a target value for the stimulation of the retina for a plurality of pixels from a camera signal output by the camera, ▪ to determine a respective set of stimulation parameters for each of the pixels using a decoding algorithm based on the determined target value, ▪ for each of the pixels, based on the set of stimulation parameters, to deliver a stimulation signal to one of the electrodes for stimulating the retina, ▪ to receive action potentials generated by the retina for each of the pixels via one of the electrodes, ▪ to determine an actual value of the stimulation of the retina for each of the pixels from the received action potentials using a classification algorithm, ▪ using an optimization algorithm for each of the pixels based on the determined target value and the determined actual value to adapt the determination of the set of stimulation parameters such that the actual value approaches the target value, wherein at least the plurality of electrodes is formed as part of a retinal implant of the arrangement.

[0008] With the described arrangement, a retina can be stimulated via electrical signals. In particular, retinal ganglion cells can be stimulated. Other cell types can also be stimulated. By stimulating the retina, a visual impression can be created. The described arrangement can be used in humans or animals. For this purpose, the arrangement is inserted wholly or partially into an eye. With the described arrangement, a totally or partially blind person or animal can visually perceive their surroundings, at least in a vague way. The arrangement interacts with the retina. However, the retina is not part of the described arrangement.

[0009] The system includes a camera. The camera can optically capture the environment. This forms the basis for the visual impression created. The camera can be a conventional camera. As is typical with cameras, the camera outputs an electronic camera signal, which represents the information about what was optically captured by the camera.

[0010] The arrangement further comprises a plurality of electrodes for interacting with the retina. Each of the electrodes can stimulate a corresponding part of the retina. The electrodes are preferably arranged in an array. This allows the retina to be stimulated in a two-dimensional grid, resulting in a two-dimensional image as the visual impression. As with a display, the highest possible resolution is preferred. However, the higher the resolution, the more complex the implementation. For example, a number of electrodes in the range of 500 to 2000 has proven to be effective. For example, the electrodes can be arranged in a 34x36 array. In this case, there are 1224 electrodes.

[0011] Each of the electrodes can correspond to a pixel of the image to be generated as a visual impression. In this case, there is a one-to-one mapping between pixel and electrode. However, there is nothing to prevent multiple electrodes from being used together for a single pixel.

[0012] The system also includes electronics. These can process the camera signal into signals that stimulate the retina via the electrodes. The electronics are preferably digital.

[0013] At least the plurality of electrodes is formed as part of a retinal implant. The retinal implant is part of the arrangement. Whether the electronics are also formed entirely or partially as part of the retinal implant is irrelevant for the functioning of the described arrangement. If the electronics are not formed as part of the retinal implant, the electronics can be formed outside the retinal implant and connected to the retinal implant by a wireless connection or a wired connection. Whether the camera is also formed entirely or partially as part of the retinal implant is also irrelevant for the functioning of the described arrangement. If the camera is not formed as part of the retinal implant, the camera can be formed outside the retinal implant and connected to the retinal implant by a wireless connection or a wired connection.

[0014] The electronics are specifically designed to execute the algorithms described in more detail below. Whether these algorithms are executed on a single chip or on multiple interconnected chips is irrelevant to the functionality of the described arrangement. In the simplest case, the electronics are designed as a chip that can be placed inside the human or animal.

[0015] The part of the electronics that executes the image analysis algorithm can be referred to as the image evaluator. The part of the electronics that executes the decoding algorithm can be referred to as the decoder. The part of the electronics that executes the classification algorithm can be referred to as the classifier. The part of the electronics that executes the optimization algorithm can be referred to as the optimizer. However, it is irrelevant for the functionality of the system whether the algorithms are executed with physically separate processors or as sub-algorithms with the same processor.

[0016] The electronics are configured to use an image analysis algorithm to determine a target value for retinal stimulation for each of the multiple pixels from the camera signal output by the camera. The image analysis algorithm is performed jointly for all pixels. The image analysis algorithm receives the camera signal as input. The image analysis algorithm outputs a target value for each pixel.

[0017] The target value is a measure of the extent to which the respective pixel should be stimulated. Numerous different definitions of the target value are conceivable. In the simplest case, the target value expresses how bright the pixel should appear in the visual impression to be created. However, the nature of the retina is preferably taken into account when defining the target value. In particular, the target value can be a measure of the ratio in which the various ganglion cells of the retina should be stimulated for the respective pixel. The target value can also be expressed simply as a percentage, where 0% can represent black and 100% white, for example.

[0018] The image analysis algorithm aims to replicate the functionality of retinal image processing. This can be achieved, for example, by implementing a Difference of Gaussian (DoG) algorithm. This involves a convolution between a DoG kernel and the camera image, which encodes either so-called ON or OFF cells by changing the sign of the kernel. By subtracting two resulting frames, it is possible to identify temporal changes that influence the activation rates. The frames generated by the DoG convolution and the subtraction operator then lead to the desired activation ratios of the ON and OFF cells by forming the relationships between the generated ON and OFF values ​​pixel by pixel.

[0019] However, the detailed design of the image analysis algorithm is irrelevant to the functionality of the system. It is sufficient that the image analysis algorithm is capable of determining a value for a large number of pixels from the camera signal. This value is a measure of the desired retinal stimulation and can therefore be considered a target value. In a simpler case, the image analysis algorithm simply determines a respective brightness for each pixel and outputs this directly as a target value.

[0020] The electronics are further designed to use a decoding algorithm to determine a respective set of stimulation parameters for each of the pixels based on the target value determined for this pixel using the image evaluation algorithm.

[0021] The decoding algorithm is performed individually for each pixel. The decoding algorithm receives the target value for that pixel as input. The decoding algorithm outputs the set of stimulation parameters for that pixel as output.

[0022] The decoding algorithm can be performed for each pixel sequentially. Alternatively, the decoding algorithm can be performed for a subset of pixels or for all pixels simultaneously.

[0023] The set of stimulation parameters contains the information on how the retina should be stimulated for the corresponding pixel. The set of stimulation parameters varies over time. If the image of the surroundings captured by the camera changes, the respective set of stimulation parameters generally also changes for each pixel. This allows a visual impression to be achieved in real time.

[0024] The decoding algorithm can, for example, be implemented as a neural network. This can calculate corresponding stimulation parameters from input parameters for so-called ON and OFF activation ratios. Various layer types can be used for this purpose, such as dense (fully connected) layers, convolutional layers, or LSTM layers. This decoder network can be trained, for example, by creating a loss function, which represents an error between the actual data (classifier output) and the desired data (image processing output) as a "cross-entropy loss."

[0025] However, the detailed design of the decoding algorithm is irrelevant to the functionality of the system. It is sufficient that the decoding algorithm is capable of determining a respective set of parameters for each pixel based on the target value determined for that pixel by the image analysis algorithm. These parameters can be interpreted as stimulation parameters. In particular, it is not mandatory for the decoding algorithm to be implemented as a neural network. In a simpler case, the decoding algorithm simply converts a respective brightness value for each pixel into an amplitude for the stimulation signal.

[0026] The electronics are further configured to deliver a stimulation signal to one of the electrodes for stimulating the retina for each pixel based on the set of stimulation parameters determined for that pixel using the decoding algorithm. This can be done for each pixel sequentially. Alternatively, it can be done for some or all pixels simultaneously.

[0027] The stimulation signal is an electrical signal. The stimulation signal acts on the retina and stimulates it. This activates the retinal cells, creating a corresponding visual impression. The set of stimulation parameters characterizes the stimulation signal. A visual impression is composed of the individual pixels, which at least roughly corresponds to how the camera captures the surroundings.

[0028] With the described setup, it is not only possible to stimulate the retina as described. Furthermore, the described setup can also be used to record how well this is achieved. The described setup thus allows feedback on the stimulation of the retina. This allows the success of the stimulation to be evaluated. The result obtained can be considered when determining the stimulation parameters. The stimulation can therefore also be referred to as "adaptive closed-loop stimulation."

[0029] When the retina is stimulated, it generates action potentials. These ultimately cause the visual impression. The invention is based on the discovery that these action potentials can be measured and used as feedback.

[0030] The electronics are configured to receive action potentials generated by the retina for each pixel via one of the electrodes. This can be done for each pixel sequentially. Alternatively, it can be done for some or all pixels simultaneously.

[0031] The stimulation signal for a pixel can be delivered via the same electrodes used to record the action potentials generated by the retina for that pixel. In this case, the electrodes can also be described as bidirectional. However, the same functionality can also be achieved if the action potentials are recorded using dedicated electrodes. Based on these general considerations, the following three embodiments are described.

[0032] In a first preferred embodiment, the electronics are configured to deliver a stimulation signal for each pixel to one of the electrodes for stimulating the retina based on the set of stimulation parameters, and to receive action potentials generated by the retina from this electrode. In this embodiment, the stimulation signal for a pixel is thus delivered to the electrode from which the action potentials are also received.

[0033] In a second preferred embodiment, the electronics are configured to deliver a stimulation signal to one of the electrodes for stimulating the retina for each of the pixels based on the set of stimulation parameters, and to receive action potentials generated by the retina from another of the electrodes. In this embodiment, the stimulation signal for a pixel is therefore delivered to a first of the electrodes, while the action potentials for this pixel are received by a second electrode. The second electrode is preferably adjacent to the first electrode. There may be multiple second electrodes. For example, the stimulation signal for a pixel can be delivered to a first of the electrodes, while the action potentials for this pixel are received by multiple second electrodes surrounding the first electrode.

[0034] In the second embodiment, the stimulation signals can generally be output independently of the reception of the action potentials. However, the individual electrodes can also perform different functions sequentially. In order to generate the highest possible resolution, all electrodes are preferably used to output stimulation signals. In the second embodiment, this can be achieved, for example, by using the individual electrodes alternately as the first electrode for one of the pixels and then as the second electrode for another of the pixels.

[0035] In a third preferred embodiment, a part of the pixels are treated according to the first embodiment and another part of the pixels according to the second embodiment.

[0036] The fact that the stimulation signal for a pixel is delivered to one of the electrodes generally means that the stimulation signal for a pixel is delivered to at least one of the electrodes. It can also be multiple electrodes. The fact that the action potentials for a pixel are received via one of the electrodes generally means that the action potentials for a pixel are received via at least one of the electrodes. It can also be multiple electrodes.

[0037] As far as the stimulation of the retina by the electrodes is concerned, the electrodes together could be considered a stimulator. As far as the reception of action potentials from the retina is concerned, the electrodes together could be considered a recorder. However, since the electrodes are preferably used for both of these purposes, such a collective designation for the electrodes is omitted here. The electrodes are preferably formed together on a chip that can be inserted into the eye.

[0038] In each case, action potentials are received for each pixel. These can be used as feedback and further processed accordingly. For this purpose, the electronics are configured to use a classification algorithm to determine the actual value of retinal stimulation for each pixel from the action potentials received for that pixel. The classification algorithm can be performed for each pixel sequentially. Alternatively, the classification algorithm can be performed for some or all pixels simultaneously.

[0039] The classification algorithm is preferably implemented as a neural network. This can, for example, consist of dense (fully connected) layers and be trained separately from the decoder algorithm. Recordings of neuronal responses to pulsed light irradiation can be used for training. Cells that emit action potentials more frequently as a result of the optical irradiation are labeled as ON cells, and cells that exhibit more action potentials in the dark are labeled as OFF cells. This labeling enables so-called supervised learning, through which the weights of the neural network can be trained so that the network can correctly assign the respective cell type. The pre-trained network can then be used in the degenerated retina to correctly assign the action potentials evoked by electrical stimulation to one of the cell types.

[0040] However, the detailed design of the classification algorithm is irrelevant to the functionality of the system. It is sufficient that the classification algorithm is capable of classifying the received action potentials in such a way that a value can be obtained that is a measure of the actual excitation of the retina. This value can therefore be considered the actual value. In particular, it is not mandatory for the classification algorithm to be implemented as a neural network. In a simple case, the classification algorithm can simply output a value that is a measure of the strength of the received action potentials. This does not require a neural network.

[0041] The action potentials generated by the retina are comparatively weak and typically lie in the µV range. The electronics are preferably designed to amplify the action potentials generated by the retina. Amplification can be performed as part of the classification algorithm or prior to the classification algorithm.

[0042] Ideally, the actual value determined for a specific pixel using the classification algorithm corresponds to the target value determined for that pixel using the image analysis algorithm, which ideally applies permanently to all pixels. As is usual with an actual-target comparison, the actual value is defined analogously to the target value. The actual value and the target value each represent a value of the same parameter. This makes a comparison of these values ​​possible.

[0043] With the described setup, deviations from this ideal state can not only be detected but also corrected. For this purpose, the electronics are configured to use an optimization algorithm for each pixel to adjust the determination of the set of stimulation parameters based on the target value determined for that pixel by the image analysis algorithm and the actual value determined for that pixel by the classification algorithm.

[0044] In the event of a discrepancy between the actual and target values, the conversion of the target value into the set of stimulation parameters can be individually adjusted for each pixel. This results in an improved visual impression compared to existing retinal implants without this type of feedback.

[0045] The set of stimulation parameters is adjusted so that the actual value approximates the target value. The error between the actual and target values ​​is thus reduced. The feedback described above ensures that the retina is reliably stimulated as desired. This optimization occurs continuously and for each pixel. The result is a visual impression that would not be achievable without this feedback.

[0046] The target value determined for a pixel is a measure of how this pixel should be represented in the generated visual impression. If the action potentials recorded for this pixel indicate that the retina for this pixel has been stimulated more strongly than desired, the optimization algorithm intervenes in the determination of the set of stimulation parameters so that this pixel is stimulated less strongly. This can be achieved, for example, by the decoding algorithm outputting a lower amplitude of the stimulation signal for this pixel as part of the set of stimulation parameters. This, in turn, can be achieved by adjusting the weights used in the decoding algorithm. The decoding algorithm can therefore be influenced for this pixel so that the decoding algorithm tends to output a lower amplitude for this pixel.The same applies if the retina for this pixel has been stimulated less than desired.

[0047] The decoding algorithm converts the desired retinal stimulation into stimulation parameters. How close the actual retinal stimulation comes to the desired stimulation is fundamentally a question of calibration. The feedback described above not only allows this calibration to be continuously corrected, but also allows this calibration adjustment to be made pixel by pixel. This allows the stimulation of the individual pixels to be precisely adjusted relative to one another. A visual impression with particularly well-adjusted excitation of the pixels relative to one another is a good visual impression.

[0048] The optimization algorithm intervenes in the decoding algorithm. The decoding algorithm is preferably the only one of the described algorithms that is variable.

[0049] Various algorithms can be used as optimization algorithms, such as the Adams Optimizer. If the decoding algorithm is implemented as a neural network, the optimization algorithm can be designed to train the neural network. In this case, the optimization algorithm can also be referred to as an "optimizer" or a "training optimizer." These terms are commonly used in the context of AI. Most optimizers have in common that they minimize a predefined loss function (e.g., categorical cross-entropy) using a gradient descent method. If the decoder algorithm is implemented as a neural network, this is done by adjusting the neuron weights in the decoder algorithm until a minimum is found.

[0050] However, the detailed design of the optimization algorithm is irrelevant to the functionality of the system. It is sufficient that the optimization algorithm is capable of influencing the decoding algorithm in such a way that the actual value approaches the target value. This does not require the decoding algorithm to be a neural network, nor does it require the optimization algorithm to be a training optimizer for this neural network.

[0051] In a preferred embodiment of the arrangement, the determined target value for the stimulation of the retina expresses the ratio in which ON ganglion cells and OFF ganglion cells of the retina are to be stimulated.

[0052] Previous retinal implants have met with neither clinical nor commercial success, as the evoked visual impressions only contained phosphenes and were not qualitatively comparable to natural vision. The reason for this lies in the non-selective activation of the different ganglion cell types in the retina. Such non-selective activation does not correspond to the natural processes in the retina. In the healthy eye, different subclasses of ganglion cells encode different lighting conditions. So-called OFF ganglion cells ensure the encoding of dark edge sections, while so-called ON ganglion cells inversely encode bright edge sections. The present embodiment takes this circumstance into account. For this purpose, the image analysis algorithm determines the target value for each pixel, indicating the ratio in which the ON ganglion cells and the OFF ganglion cells of the retina are to be stimulated for that pixel.This can provide an improved visual impression.

[0053] In this embodiment, it is preferred that the actual value of the retinal stimulation determined using the classification algorithm expresses the ratio in which ON ganglion cells and OFF ganglion cells of the retina were stimulated. The actual value and the target value are thus defined analogously to one another. The described definition of the actual value is possible because the neurons of the retina emit signals with a characteristic shape. Thus, it is possible to distinguish whether a specific signal emitted by the retina was generated by an ON ganglion cell or an OFF ganglion cell. Consequently, the ratio of these signals can be determined.

[0054] In a further preferred embodiment of the arrangement, the decoding algorithm is designed as a neural network, wherein the optimization algorithm is configured to influence weights of the decoding algorithm.

[0055] It has been shown that the use of a neural network for the decoding algorithm is particularly advantageous because the retina is also a neural network. Due to the similarity of the networks, the retina can be integrated particularly well into data processing in this embodiment. The optimization algorithm can intervene in the decoding algorithm by influencing the weights of the decoding algorithm in such a way that the actual value approaches the target value. The weights of the decoding algorithm can also be referred to as neuron weights.If the received action potentials indicate that the retina in a pixel is being stimulated more strongly than desired, the decoding algorithm can be influenced via the optimization algorithm so that the decoding algorithm outputs a set of stimulation parameters for that pixel that corresponds to a weaker retinal stimulation. The same applies if the retina in a pixel is being stimulated less strongly than desired.

[0056] In a further preferred embodiment of the arrangement, the stimulation signal is a superposition of several sinusoidal signals, each of which has an individually selectable amplitude and / or an individually selectable frequency. This applies pixel by pixel. This means that for each pixel, several sinusoidal signals are superimposed to obtain the stimulation signal for that pixel.

[0057] In this embodiment, the set of stimulation parameters contains one or more stimulation parameters for each of the sinusoidal signals. For example, the set of stimulation parameters can contain an amplitude, a frequency, and a repetition rate for each of the sinusoidal signals. The individual amplitudes can range between 0 and 100% of a reference value. The amplitudes can therefore be used to determine the ratio in which the individual sinusoidal signals are superimposed.

[0058] The more sinusoidal signals are superimposed, the more realistic the visual impression. However, the more sinusoidal signals to be superimposed, the more effort required. In practice, it has proven effective to create a stimulation signal consisting of a superposition of two to six sinusoidal signals, especially four sinusoidal signals.

[0059] The present embodiment is based on the finding that the retina reacts to stimulation with a selective response. This applies in particular to stimulation in the kHz range with an amplitude in the µA range. Accordingly, in this embodiment in particular, it is preferred that the frequencies are each between 1 and 10 kHz and / or the amplitudes are in the range from 0 to 200 µA. The selective response means that the functional cell classes that encode different impressions can be activated in isolation by electrical stimulation. The resulting visual impression approximates the natural visual impression of a healthy eye.

[0060] It was also recognized that the waveform of the stimulation signal also plays a role in the activation quality. Conventional retinal implants are based on stimulation patterns with rectangular current pulses. In contrast, the sinusoidal stimulation waveforms used in the present embodiment lead to significantly more robust cell activation.

[0061] The stimulation signals for the individual electrodes are preferably generated with a signal generator that has a dedicated output for each electrode. For example, the signal generator can have 1224 output channels to achieve a resolution of 34x36. For each output channel, the signal generator can generate a mix of, for example, four freely selectable sine frequencies in the range between 1 and 10 kHz. The respective amplitude of these four sine waves can also be freely defined for each frequency, enabling highly versatile stimulation waveforms. An individual mix of frequencies can be output on each output channel, resulting in a high resolution of the evoked visual impression.

[0062] Existing retinal stimulators are unable to generate both the high frequency in the kHz range and the required sine waveform, and simultaneously use these on independent output channels for stimulation. While stimulators capable of generating arbitrary signal shapes do exist, they have very few output channels, which cannot replicate the high resolution of natural vision, or which can only deliver the output signal asynchronously via multiplexers. Due to the high data rate required, these devices are also unsuitable for implantation in humans.

[0063] In the described arrangement, a retinal stimulator is preferably used, which is divided into globally available and locally available circuit components. The waveforms with their different frequencies can be generated in the global circuit section via digital-to-analog converters, while the amplitude modulation of the individual frequencies can be performed locally via capacitively stored voltage values ​​at the pixels. The capacitively stored voltages can control the reference current of operational transimpedance amplifiers (OTAs), which enables amplitude modulation of the amplitude-homogeneous sinusoidal signals. This can reduce the required data rate and circuit complexity without compromising signal quality or synchronicity. This design overcomes the limitations described above.

[0064] In a further preferred embodiment of the arrangement, the set of stimulation parameters comprises an amplitude, a frequency and a repetition rate.

[0065] In this embodiment, the decoding algorithm determines an amplitude, a frequency, and a repetition rate for each pixel. These parameters are collectively referred to as the set of stimulation parameters for that pixel.

[0066] In this embodiment, the stimulation signal is delivered to the retina in pulsed form. The repetition rate indicates how quickly the pulses are delivered to the retina one after the other.

[0067] As a further aspect of the invention, a method for stimulating a retina with a plurality of electrodes for interacting with the retina is presented, wherein at least the plurality of electrodes are formed as part of a retinal implant, and wherein a) using an image analysis algorithm, a target value for the stimulation of the retina is determined for a large number of pixels from a camera signal output by a camera, b) a decoding algorithm is used to determine a respective set of stimulation parameters for each of the pixels based on the determined target value, c) for each of the pixels, a stimulation signal is delivered to one of the electrodes for stimulating the retina based on the set of stimulation parameters, d) for each of the pixels, action potentials generated by the retina are received via one of the electrodes, e) using a classification algorithm, an actual value of the stimulation of the retina is determined for each of the pixels from the received action potentials, f) using an optimization algorithm for each of the pixels based on the determined target value and the determined actual value, the determination of the set of stimulation parameters is adjusted in such a way that the actual value approaches the target value.

[0068] The advantages and features of the arrangement are applicable and transferable to the method, and vice versa. The arrangement is preferably configured to operate according to the method. The method is preferably carried out using the arrangement.

[0069] The invention is explained in more detail below with reference to the figure. The figure shows a particularly preferred embodiment, to which the invention is not limited, however. The figure and the proportions depicted therein are only schematic. It shows: Fig. 1: an arrangement according to the invention for stimulating a retina.

[0070] Fig. 1 shows an arrangement 1 for stimulating a retina 2. The retina 2 is in Fig. 1 is included, but is not part of assembly 1. Assembly 1 includes a camera 3. Camera 3 optically records the surroundings. Camera 3 outputs a camera signal 7 with corresponding information.

[0071] The arrangement 1 further comprises electronics 6 and a plurality of electrodes 5 for interacting with the retina 2. The electrodes 5 are part of a chip 4, which is applied to the retina 2. The electronics 6 processes the camera signal 7 into signals that are transmitted to the electrodes 5. This stimulates the retina 2 in such a way that a visual impression is created that reflects the surroundings captured by the camera.

[0072] In Fig. 1, nine electrodes 5 are shown as an example, arranged side by side for simplicity. Only one of the electrodes 5 is provided with a reference symbol. In reality, the electrodes 5 are preferably arranged in a two-dimensional array. This allows a two-dimensional image to be generated.

[0073] The plurality of electrodes 5 are formed as part of a retinal implant 19. The retinal implant 19 is part of the arrangement 1. Whether the camera 3 and the electronics 6 are also formed entirely or partially as part of the retinal implant 19 is irrelevant for the functioning of the described arrangement 1.

[0074] The electronics 6 are configured to use an image evaluation algorithm 14 to determine a target value 8 for the stimulation of the retina 2 for a plurality of pixels from the camera signal 7 output by the camera 3. The determined target value 8 for the stimulation of the retina 2 expresses the ratio in which ON ganglion cells and OFF ganglion cells of the retina 2 are to be stimulated for this pixel.

[0075] The image evaluation algorithm 14 is executed for all pixels together. The algorithms described below are executed individually for each pixel using the electronics 6. Fig. This is illustrated in Figure 1 in that an algorithm block 20 is shown for one of the pixels as an example. This algorithm block 20 interacts with two of the electrodes 5. Further identical algorithm blocks 20 are provided for the remaining electrodes 5. For the sake of clarity, however, only one of the algorithm blocks 20 is shown.

[0076] The algorithm block 20 comprises a decoding algorithm 15. This allows a respective set of stimulation parameters 9 to be determined for each of the pixels based on the determined target value 8.

[0077] The electronics 6 are configured to deliver a stimulation signal 10 to one of the electrodes 5 for stimulating the retina 2 for each of the pixels based on the set of stimulation parameters 9. In the example shown, the Fig. 1, the set of stimulation parameters 9 is converted into the stimulation signal 10 using a signal generator 12. The stimulation signal 10 is a superposition of several sinusoidal signals, each of which has an individually selectable amplitude and an individually selectable frequency. The set of stimulation parameters 9 includes an amplitude, a frequency, and a repetition rate for each of these sinusoidal signals.

[0078] The electronics 6 are further configured to receive action potentials 11 generated by the retina 2 for each of the pixels via one of the electrodes 5. In the example of the Fig. 1, this is done with an electrode 5, which is located next to the electrode 5 that receives the stimulation signal 10 for this pixel. Alternatively, however, the action potentials 11 could also be received by the same electrode 5 that also receives the stimulation signal 10 for this pixel.

[0079] The algorithm block 20 further comprises a classification algorithm 16. This allows an actual value 13 of the stimulation of the retina 2 to be determined from the received action potentials 11.

[0080] Algorithm block 20 further includes an optimization algorithm 17. This allows the determination of the set of stimulation parameters 9 to be adjusted based on the determined target value 8 and the determined actual value 13 such that the actual value 13 approximates the target value 8. For this purpose, a corresponding correction signal 18 can be sent from the optimization algorithm 17 to the decoding algorithm 15. The correction signal 18 causes a change in the weights in the decoding algorithm 15. List of reference symbols 1 arrangement 2 Retina 3 Camera 4 chips 5 electrodes 6 Electronics 7 Camera signal 8 Target value 9 Set of stimulation parameters 10 Stimulation signal 11 action potentials 12 Signal generator 13 Actual value 14 Image analysis algorithm 15 Decoding algorithm 16 Classification algorithm 17 Optimization algorithm 18 Correction signal 19 Retinal implant 20 Algorithm Block

Claims

[1] Arrangement (1) for stimulating a retina (2), comprising: - a camera (3), - a plurality of electrodes (5) for interaction with the retina (2), - an electronic device (6) which is designed to ▪ to determine a target value (8) for the stimulation of the retina (2) for a plurality of pixels using an image evaluation algorithm (14) from a camera signal (7) output by the camera (3), ▪ to determine a respective set of stimulation parameters (9) for each of the pixels using a decoding algorithm (15) based on the determined target value (8), ▪ for each of the pixels, based on the set of stimulation parameters (9), to deliver a stimulation signal (10) to one of the electrodes (5) for stimulating the retina (2), ▪ to receive action potentials (11) generated by the retina (2) for each of the pixels via one of the electrodes (5), ▪ to determine an actual value (13) of the stimulation of the retina (2) for each of the pixels from the received action potentials (11) using a classification algorithm (16), ▪ using an optimization algorithm (17) for each of the pixels based on the determined target value (8) and the determined actual value (13) to adapt the determination of the set of stimulation parameters (9) such that the actual value (13) approaches the target value (8), wherein at least the plurality of electrodes (5) is designed as part of a retinal implant (19) of the arrangement (1). [2] Arrangement (1) according to claim 1, wherein the determined target value (8) for the stimulation of the retina (2) expresses the ratio in which ON ganglion cells and OFF ganglion cells of the retina (2) are to be stimulated. [3] Arrangement (1) according to one of the preceding claims, wherein the decoding algorithm (15) is designed as a neural network, and wherein the optimization algorithm (17) is arranged to influence weights of the decoding algorithm (15). [4] Arrangement (1) according to one of the preceding claims, wherein the stimulation signal (10) is a superposition of several sinusoidal signals, each of which has an individually selectable amplitude and / or an individually selectable frequency. [5] Arrangement (1) according to one of the preceding claims, wherein the set of stimulation parameters (9) comprises an amplitude, a frequency and a repetition rate. [6] Method for stimulating a retina (2) with a plurality of electrodes (5) for interacting with the retina (2), wherein at least the plurality of electrodes (5) are formed as part of a retinal implant (19), and wherein: a) using an image evaluation algorithm (14) a target value (8) for the stimulation of the retina (2) is determined for a plurality of pixels from a camera signal (7) output by a camera (3), b) a decoding algorithm (15) is used to determine a respective set of stimulation parameters (9) for each of the pixels on the basis of the determined target value (8), c) for each of the pixels, a stimulation signal (10) is delivered to one of the electrodes (5) for stimulating the retina (2) based on the set of stimulation parameters (9), d) for each of the pixels, action potentials (11) generated by the retina (2) are received via one of the electrodes (5), e) an actual value (13) of the stimulation of the retina (2) is determined for each of the pixels from the received action potentials (11) using a classification algorithm (16), f) using an optimization algorithm (17) for each of the pixels based on the determined target value (8) and the determined actual value (13), the determination of the set of stimulation parameters (9) is adapted in such a way that the actual value (13) approaches the target value (8).

Images