Technical field

Examples relate to concepts for processing biology-related data.

Background

In many biological applications, a vast amount of data is generated. For example, images are taken from a huge amount of biological structures and stored in databases. It is very time- consuming and expensive to analyze the biological data manually.

Summary

Hence, there is a need for an improved concept for processing biology-related data.

This need may be satisfied by the subject matter of the claims.

Some embodiments relate to a system comprising one or more processors and one or more storage devices. The system is configured to obtain a plurality of high-dimensional repre sentations of a plurality of biology-related image-based input data sets of a set of images of one or more biological structures and determine a typical high-dimensional representation based on the plurality of high-dimensional representations. The typical high-dimensional representation comprises at least 3 entries each having a different value. Further, the system is configured to generate a biology-related image-based output data set of a typical image representing the set of images based on the typical high-dimensional representation by a trained machine-learning algorithm executed by the one or more processors. By using a trained machine- learning algorithm for generating a typical image for a set of images, a high-quality image containing relevant and/or important features common to the set of images may be obtainable. By allowing the high-dimensional representations to have entries with various different values (in contrast to one hot encoded representations), images with semantically similar content may have similar high-dimensional representations. Espe cially for biology related images, the semantic similarity may be reproduced very well by the high-dimensional representations. The trained machine- learning algorithm may be able to provide a typical image accurately reproducing important and/or relevant features of the set of images due to the semantic similarity of the high-dimensional representations being close to each other. The combination of applying a trained machine-learning algorithm to high-dimensional representations with various different values reproducing semantical prox imity may allow to generate typical images with significantly improved quality.

Some embodiments relate to a system comprising one or more processors and one or more storage devices. The system is configured to obtain an input high-dimensional representa tion of biology-related image-based input data of a biological structure. The input high dimensional representation comprises at least 3 entries each having a different value. Fur ther, the system is configured to generate training image data based on the input high dimensional representation by a generative network of a generative adversarial network ex ecuted by the one or more processors. Additionally, the system is configured to generate an output high-dimensional representation based on the training image data by a discriminative network of the generative adversarial network executed by the one or more processors. The output high-dimensional representation comprises at least 3 entries each having a different value. Further, the system is configured to adjust the generative network of the generative adversarial network based on a comparison of the input high-dimensional representation and the output high-dimensional representation.

By training a generative adversarial network to generate an image based on high dimensional representations of biology related images, a machine- learning algorithm very well suited for generating typical images of biological images represented by high dimensional representations may be obtained. By allowing the high-dimensional representa tions to have entries with various different values (in contrast to one hot encoded representa tions), images with semantically similar content may have similar high-dimensional repre sentations. Especially for biology related images, the semantic similarity may be reproduced very well by the high-dimensional representations. The trained generative adversarial net work may be able to provide a typical image of a biological structure accurately reproducing important and/or relevant features.

Short description of the Figures

Some examples of apparatuses and/or methods will be described in the following by way of example only, and with reference to the accompanying figures, in which

Fig. 1 is a schematic illustration of a system for generating a biology-related image-based output data set of a typical image of a biological structure;

Fig. 2 is a schematic illustration of a system for generating a typical image;

Fig. 3 is a schematic illustration of a system for training a generative adversarial network;

Fig. 4 is a schematic illustration of another system for training a generative adversarial network;

Fig. 5 is a schematic illustration of a system for processing biology-related data including a microscope;

Fig. 6 is a flow chart of a method for generating a biology-related image-based output data set of a typical image of a biological structure; and

Fig. 7 is a flow chart of a method for training a generative adversarial network.

Detailed Description

Various examples will now be described more fully with reference to the accompanying drawings in which some examples are illustrated. In the figures, the thicknesses of lines, layers and/or regions may be exaggerated for clarity. Accordingly, while further examples are capable of various modifications and alternative forms, some particular examples thereof are shown in the figures and will subsequently be described in detail. However, this detailed description does not limit further examples to the particular forms described. Further examples may cover all modifications, equivalents, and alternatives falling within the scope of the disclosure. Same or like numbers refer to like or similar elements throughout the description of the figures, which may be implemented iden tically or in modified form when compared to one another while providing for the same or a similar functionality.

It will be understood that when an element is referred to as being“connected” or“coupled” to another element, the elements may be directly connected or coupled or via one or more intervening elements. If two elements A and B are combined using an“or”, this is to be un derstood to disclose all possible combinations, i.e. only A, only B as well as A and B, if not explicitly or implicitly defined otherwise. An alternative wording for the same combinations is“at least one of A and B” or“A and/or B”. The same applies, mutatis mutandis, for com binations of more than two Elements.

The terminology used herein for the purpose of describing particular examples is not intend ed to be limiting for further examples. Whenever a singular form such as“a,”“an” and “the” is used and using only a single element is neither explicitly or implicitly defined as being mandatory, further examples may also use plural elements to implement the same functionality. Likewise, when a functionality is subsequently described as being implement ed using multiple elements, further examples may implement the same functionality using a single element or processing entity. It will be further understood that the terms“comprises,” “comprising,”“includes” and/or“including,” when used, specify the presence of the stated features, integers, steps, operations, processes, acts, elements and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, processes, acts, elements, components and/or any group thereof.

Unless otherwise defined, all terms (including technical and scientific terms) are used herein in their ordinary meaning of the art to which the examples belong. Fig. 1 shows a system for generating a biology-related image-based output data set of a typ ical image of a biological structure according to an embodiment. The system 100 compris ing one or more processors 110 coupled to one or more storage devices 120. The system 100 is configured to obtain a plurality of high-dimensional representations 111 of a plurality of biology-related image-based input data sets of a set of images of one or more biological structures and determine a typical high-dimensional representation based on the plurality of high-dimensional representations 111. The typical high-dimensional representation com prises at least 3 entries each having a different value (or at least 20 entries, at least 50 entries or at least 100 entries having values different from each other). Further, the system 100 is configured to generate a biology-related image-based output data set 113 of a typical image representing the set of images based on the typical high-dimensional representation by a trained machine-learning algorithm executed by the one or more processors 110.

The high-dimensional representations 111 and/or the typical high-dimensional representa tion may be numerical representations (e.g. comprising numerical values only). The high dimensional representations 111 and/or the typical high-dimensional representation may comprise more than 100 dimensions (or more than 300 dimensions or more than 500 dimen sions) and/or less than 10000 dimensions (or less than 3000 dimensions or less than 1000 dimensions). Each entry of the high-dimensional representations 111 and/or the typical high-dimensional representation may be a dimension of the high-dimensional representa tions (e.g. a high-dimensional representation with 100 dimensions comprises 100 entries). For example, using high dimensional representations with more than 300 dimensions and less than 1000 dimensions may enable a suitable representation for biology-related data with semantic correlation. The high-dimensional representations 111 and/or the typical high dimensional representation may be vectors. If a vector representation is used for the entries of a high-dimensional representation, an efficient comparison and/or other calculations (e.g. normalization) may be enabled, although other representations (e.g. as a matrix) may be possible as well. For example, the high-dimensional representations 111 and/or the typical high-dimensional representation may be normalized vectors. The high-dimensional repre sentations of the plurality of high-dimensional representations 111 may be generated by a trained visual recognition machine-learning algorithm, which may have been trained by a loss function, which causes the trained visual recognition machine-learning algorithm to output normalized high-dimensional representations. However, other approaches for the normalization of the high-dimensional representation may be applicable as well. For example, each high-dimensional representation of the plurality of high-dimensional rep resentations 111 and/or the typical high-dimensional representation may comprise various entries (at least three) with values unequal 0 in contrast to one hot encoded representations. By using a high-dimensional representation, which is allowed to have various entries with values unequal 0, information on a semantic relationship between the high-dimensional rep resentations can be reproduced. For example, more than 50% (or more than 70% or more than 90%) of values of the entries of each high-dimensional representation of the plurality of high-dimensional representations 111 and/or the typical high-dimensional representation may be unequal 0. Sometimes one hot encoded representations have also more than one entry unequal 0, but there is only one entry with high value and all other entries have values at noise level (e.g. lower than 10% of the one high value). In contrast, the values of more than 5 entries (or more than 20 entries or more than 50 entries) of the typical high dimensional representation may be larger than 10% (or larger than 20% or larger than 30%) of a largest absolute value of the entries of the typical high-dimensional representation, for example. Further, the values of more than 5 entries (or more than 20 entries or more than 50 entries) of each high-dimensional representation of the plurality of high-dimensional repre sentations 111 may be larger than 10% (or larger than 20% or larger than 30%) of a respec tive largest absolute value of the entries of the high-dimensional representations. For exam ple, the values of more than 5 entries (or more than 20 entries or more than 50 entries) of one high-dimensional representation of the plurality of high-dimensional representations 111 may be larger than 10% (or larger than 20% or larger than 30%) of a largest absolute value of the entries of the one high-dimensional representation. For example, each entry of a high-dimensional representation of the plurality of high-dimensional representations 111 and/or the typical high-dimensional representation may comprise a value between - 1 and 1.

Each high-dimensional representation of the plurality of high-dimensional representations 111 and/or the typical high-dimensional representation may be a hidden representation, a latent vector, an embedding, a sematic embedding and/or a token embedding and/or may be also called hidden representation, a latent vector, an embedding, a semantic embedding and/or a token embedding.

The values of one or more entries of each high-dimensional representation of the plurality of high-dimensional representations 111 and/or the typical high-dimensional representation may be proportional to a likelihood of a presence of a specific biological function or a spe cific biological activity. By using a mapping that generates high-dimensional representa tions preserving the semantical similarities of the input data sets, semantically similar high dimensional representations may have a closer distance to each other than semantically less similar high-dimensional representations. Further, if two high-dimensional representations represent input data sets with same or similar specific biological function or specific biolog ical activity one or more entries of these two high-dimensional representations may have same or similar values. Due to the preservation of the semantic, one or more entries of the high-dimensional representations may be an indication of an occurrence or presence of a specific biological function or a specific biological activity. For example, the higher a value of one or more entries of the high-dimensional representation, the higher the likelihood of a presence of a biological function or a biological activity correlated with these one or more entries may be.

The typical high-dimensional representation may be determined by an arithmetic combina tion of the plurality of high-dimensional representations 111. For example, the arithmetic combination of the plurality of high-dimensional representations 111 may be based on a linear combination of the plurality of high-dimensional representations 111. Various arith metic combinations may be applicable, but a linear combination of the plurality of high dimensional representations may be determined with low computational effort. For exam ple, the typical high-dimensional representation may be determined by calculating the arithmetic mean (e.g. or the geometric mean, the harmonic mean, the quadratic mean or the medoid) of the plurality of high-dimensional representations 111.

For example, the trained machine- learning algorithm may comprise or may be a generative adversarial network or another trained machine- learning algorithm. The biology-related image-based output data set 113 of the typical image may be generated by a trained genera tive adversarial network. For example, a trained generative network of a generative adver sarial network may generate the biology-related image-based output data set 113 based on the typical high-dimensional representation. The biology-related image-based output data set 113 may be determined by applying the trained generative network of the generative adversarial network with a trained set of parameters to the typical high-dimensional repre sentation. For example, the trained generative network may comprise more than 30 layers (or more than 100 layers or more than 200 layers) and/or less than 800 layers (or less than 600 layers or less than 400 layers). For example, a ResNet or DenseNet may be used as gen erative network. For example, more than the trained generative network may comprise more than 200 layers (e.g. ResNet 152 with about 150 layers results in about 2*150=300 layers, if this neural network is used as generative network and discriminative network) as micro scopes may generate images with high resolution (e.g. more than 1 mega pixels, for example between 4 and 5 mega pixels) and the typical image may be desired to have a similar resolu tion or the same resolution as the input images (the set of images).

The high-dimensional representations of the plurality of high-dimensional representations 111 may be received from a database (e.g. stored by the one or more storage devices). Al ternatively, the system 100 may be configured to obtain the plurality of high-dimensional representations by generating the high-dimensional representations of the plurality of high dimensional representations 111 of the plurality of biology-related image-based input data sets by a trained visual recognition machine-learning algorithm executed by the one or more processors 110. For example, each high-dimensional representation of the plurality of high dimensional representations 111 may be determined by applying at least a part (e.g. encod er) of a trained visual recognition machine- learning algorithm with a trained set of parame ters to a biology-related image-based input data set of the plurality of biology-related image- based input data sets. For example, generating the high-dimensional representations by the trained visual recognition machine- learning algorithm may mean that the high-dimensional representations are generated by an encoder of the trained visual recognition machine learning algorithm. The trained set of parameters of the trained visual recognition machine learning algorithm may be obtained during training of the visual recognition machine learning algorithm as described below.

Each biology-related image-based input data set of the plurality of biology-related image- based input data sets and/or the biology-related image-based output data set 113 may be image data (e.g. pixel data of an image) of an image of a biological structure comprising a nucleotide or a nucleotide sequence, a biological structure comprising a protein or a protein sequence, a biological molecule, a biological tissue, a biological structure with a specific behavior, and/or a biological structure with a specific biological function or a specific bio logical activity. The biological structure may be a molecule, a viroid or virus, artificial or natural membrane enclosed vesicles, a subcellular structure (like a cell organelle) a cell, a spheroid, an organoid, a three-dimensional cell culture, a biological tissue, an organ slice or part of an organ in vivo or in vitro. For example, the image of the biological structure may be an image of the location of a protein within a cell or tissue or an image of a cell or tissue with endogenous nucleotides (e.g. DNA) to which labeled nucleotide probes bind (e.g. in situ hybridization). The image data may comprise a pixel value for each pixel of an image for each color dimension of the image (e.g. three color dimensions for RGB representation). For example, depending on the imaging modality other channels may apply related to exci tation or emission wavelength, fluorescence lifetime, light polarization, stage position in three spatial dimensions, different imaging angles. Each biology-related image-based input data set of the plurality of biology-related image-based input data sets may be an XY pixel map, volumetric data (XYZ), time series data (XY+T) or combinations thereof (XYZT). Moreover, additional dimensions depending on the kind of image source may be included such as channel (e.g. spectral emission bands), excitation wavelength, stage position, logical position as in a multi-well plate or multi-positioning experiment and/or mirror and/or objec tive position as in lightsheet imaging.

For example, the plurality of biology-related image-based input data sets may be obtained from a database (e.g. stored by the one or more storage devices) or may be obtained by an imaging device (e.g. microscope, camera) during a running experiment. For example, the system 100 may comprise a microscope configured to obtain the plurality of biology-related image-based input data sets by taking images (e.g. representing the set of images) of one or more biological specimens. The set of images may comprise more than 5 images, more than 10 images, more than 20 images or more than 50 images. For example, if more images are considered for the generation of the typical image, the typical image may reproduce the rel evant and/or important features more accurately.

The biology-related image-based output data set 113 of the typical image representing the set of images may be stored (e.g. by the one or more storage devices) and/or may be output (e.g. on a screen). The typical image may be also called average image or standardized im age and may be an artificially generated image representing the set of images. Typical im age may refer to a generalized representation of the object where individual variability in the data, be it instrument noise or morphological and topological variability gets eliminated. A typical image may represent something like the average image in embedding space which contains all the features or many features a domain expert would draw into a textbook. The trained visual recognition machine- learning algorithm may also be called image recog nition model or visual model. The trained visual recognition machine- learning algorithm may be or may comprise a trained visual recognition neural network. The trained visual recognition neural network may comprise more than 20 layers (or more than 40 layers or more than 80 layers) and/or less than 400 layers (or less than 200 layers or less than 150 layers). The trained visual recognition neural network may be a convolutional neural net work or a capsule network. Using a convolutional neural network or a capsule network may provide a trained visual recognition machine- learning algorithm with high accuracy for bi ology-related image-based data. However, also other visual recognition algorithms may be applicable. For example, the trained visual recognition neural network may comprise a plu rality of convolution layers and a plurality of pooling layers. However, pooling layers may be avoided, if a capsule network is used and/or stride=2 is used instead of stride=l for the convolution, for example. The trained visual recognition neural network may use a rectified linear unit activation function. Using a rectified linear unit activation function may provide a trained visual recognition machine- learning algorithm with high accuracy for biology- related image-based input data, although other activation functions (e.g. a hard tanh activa tion function, a sigmoid activation function or a tanh activation function) may be applicable as well. For example, the trained visual recognition neural network may comprise a convo lutional neural network and/or may be a ResNet or a DenseNet of a depth depending on the size of the input images.

More details and aspects of the system 100 are mentioned in conjunction with the proposed concept and/or the one or more examples described above or below (e.g. Figs. 2-7). The system 100 may comprise one or more additional optional features corresponding to one or more aspects of the proposed concept and/or of one or more examples described above or below.

Fig. 2 shows a schematic illustration of a system for generating a typical image according to an embodiment. From a large set of (image) data showing multiple observations of the same class of object, a system according to an embodiment may create a generalized representa tion of the object which contains all the typical features which may make it uniquely distin guishable from other objects. The set of image data 210 may be obtained as the result of an image search in a database 201 (e.g. by an Image-to-Image Search using an image as search input or Text-to-Image Search using text as search input), obtained as the recorded images of a particular specimen in a running experiment using an imaging system 203 (e.g. micro scope) and/or obtained as manually selected by a user 207. From this data set a typical im age may be created by converting each image in the data set to its respective semantic em bedding, combining all the semantic embeddings from a data set into one using arithmetic and using the combined semantic embedding as a prior to a generative model to output an image.

In more detail, a second stage visual model 220 was trained to predict semantic embeddings from images as described below. Each recorded image 210 may be converted to its respec tive semantic embedding 111. The resulting set of semantic embeddings 111 is arithmetical ly combined 240, for example by a linear combination, but non-linear combinations are pos sible as well. For example, a linear combination may be the arithmetic mean of these seman tic embeddings 111. The result of the combination of the semantic embeddings is itself a latent vector 250 (e.g. semantic embedding). It may be used as a prior for a generative mod el, such as a generative adversarial network (GAN) 260 to generate a new image 113.

The new image 113 generated by the trained GAN may contain properties in a semantic sense which are part of all of the retrieved images used to create it. It can be viewed as a “typical image”, or, in the case of using the arithmetic mean for the combination, be an“av erage image”, for example. It may be noted that such a typical or average image might not be equivalent to computing the pixel- wise arithmetic mean, but may rather be a typical or average image in semantic space.

More details and aspects of the system 200 are mentioned in conjunction with the proposed concept and/or the one or more examples described above or below (e.g. Figs. 1 or 3-7). The system 200 may comprise one or more additional optional features corresponding to one or more aspects of the proposed concept and/or of one or more examples described above or below.

Fig. 3 shows a schematic illustration of a system for training a generative adversarial net work according to an embodiment. The system 300 comprises one or more processors 110 coupled to one or more storage devices 120. The system 300 is configured to obtain an input high-dimensional representation 310 of biology-related image-based input data of a biologi cal structure. The input high-dimensional representation 310 comprises at least 3 entries each having a different value (or at least 20 entries, at least 50 entries or at least 100 entries having values different from each other). Further, the system 300 is configured to generate training image data based on the input high-dimensional representation 310 by a generative network of a generative adversarial network executed by the one or more processors 110. Additionally, the system 300 is configured to generate an output high-dimensional represen tation 320 based on the training image data by a discriminative network of the generative adversarial network executed by the one or more processors 110. The output high dimensional representation 320 comprises at least 3 entries each having a different value (or at least 20 entries, at least 50 entries or at least 100 entries having values different from each other). Further, the system 300 is configured to adjust the generative network of the genera tive adversarial network based on a comparison of the input high-dimensional representation 310 and the output high-dimensional representation 320.

The input high-dimensional representation 310 and/or the output high-dimensional represen tation 320 may be numerical representations (e.g. comprising numerical values only). The input high-dimensional representation 310 and/or the output high-dimensional representa tion 320 may comprise more than 100 dimensions (or more than 300 dimensions or more than 500 dimensions) and/or less than 10000 dimensions (or less than 3000 dimensions or less than 1000 dimensions). Each entry of the input high-dimensional representation 310 and/or the output high-dimensional representation 320 may be a dimension of the high dimensional representations (e.g. a high-dimensional representation with 100 dimensions comprises 100 entries). For example, using high dimensional representations with more than 300 dimensions and less than 1000 dimensions may enable a suitable representation for bi ology-related data with semantic correlation. The input high-dimensional representation 310 and/or the output high-dimensional representation 320 may be vectors. If a vector represen tation is used for the entries of a high-dimensional representation, an efficient comparison and/or other calculations (e.g. normalization) may be enabled, although other representa tions (e.g. as a matrix) may be possible as well. For example, the input high-dimensional representation 310 and/or the output high-dimensional representation 320 may be normal ized vectors. The input high-dimensional representation 310 may be generated by a trained visual recognition machine- learning algorithm, which may have been trained by a loss func tion, which causes the trained visual recognition machine- learning algorithm to output nor malized high-dimensional representations. However, other approaches for the normalization of the high-dimensional representation may be applicable as well. For example, the input high-dimensional representation 310 and/or the output high dimensional representation 320 may comprise various entries (at least three) with values unequal 0 in contrast to one hot encoded representations. By using a high-dimensional rep resentation, which is allowed to have various entries with values unequal 0, information on a semantic relationship between the high-dimensional representations can be reproduced. For example, more than 50% (or more than 70% or more than 90%) of values of the entries of the input high-dimensional representation 310 and/or the output high-dimensional repre sentation 320 may be unequal 0. Sometimes one hot encoded representations have also more than one entry unequal 0, but there is only one entry with high value and all other entries have values at noise level (e.g. lower than 10% of the one high value). In contrast, the val ues of more than 5 entries (or more than 20 entries or more than 50 entries) of the input high-dimensional representation 310 and/or the output high-dimensional representation 320 may be larger than 10% (or larger than 20% or larger than 30%) of a largest absolute value of the entries of the input high-dimensional representation 310 and/or the output high dimensional representation 320, for example. For example, the input high-dimensional rep resentation 310 and/or the output high-dimensional representation 320 may comprise a val ue between - 1 and 1.

The input high-dimensional representation 310 and/or the output high-dimensional represen tation 320 may be a hidden representation, a latent vector, an embedding, a sematic embed ding and/or a token embedding and/or may be also called hidden representation, a latent vector, an embedding, a semantic embedding and/or a token embedding.

The values of one or more entries of the input high-dimensional representation 310 and/or the output high-dimensional representation 320 may be proportional to a likelihood of a presence of a specific biological function or a specific biological activity.

The training image data may be generated by applying the generative network of the genera tive adversarial network with a current set of parameters (e.g. neural network weights) to the input high-dimensional representation. For example, the trained generative network may comprise more than 30 layers (or more than 100 layers or more than 200 layers) and/or less than 800 layers (or less than 600 layers or less than 400 layers). Further, the output high dimensional representation may be generated by applying the discriminative network of the generative adversarial network with a current set of parameters (e.g. neural network weights) to the training image data. The current set of parameters of the generative network and/or the discriminative network may be updated during the adjustment of the generative adversarial network.

The generative network of the generative adversarial network may be trained by adjusting neural network weights of the generative network based on the comparison of the input high-dimensional representation and the output high-dimensional representation. For exam ple, the comparison of the input high-dimensional representation 310 and the output high dimensional representation 320 for the adjustment of the generative network of the genera tive adversarial network may be based on a mean squared error loss function. For example, the system 300 may calculate the mean squared error between the input high-dimensional representation 310 and the output high-dimensional representation 320 and may adjust the neural network weights of the generative network in order to reduce the mean squared error. Other embodiments may use different metrics for comparing the input high-dimensional representation 310 and the output high-dimensional representation 320, such as cosine simi larity or Earth mover’s distance. The discriminative network may be pre-trained on predict ing high-dimensional representations 320 on a large body of images as described for train ing the visual models.

The generative adversarial network may be a machine- learning algorithm comprising two competing neural networks. The generative network of the generative adversarial network may generate training images while the discriminative network of the generative adversarial network may evaluate them. By adjusting the network weights of the generative network, the generative network may learn to map from the space of the input high-dimensional rep resentations (e.g. latent space) to images of interest, while the discriminative network may try to distinguish training images produced by the generative network from true images. The generative network's training objective may be to increase the error rate of the discrimina tive network.

The training image data may be image data (e.g. pixel data of an image) of an artificial im age of the biological structure comprising a nucleotide or a nucleotide sequence, a biologi cal structure comprising a protein or a protein sequence, a biological molecule, a biological tissue, a biological structure with a specific behavior, and/or a biological structure with a specific biological function or a specific biological activity. The image data may comprise a pixel value for each pixel of an image for each color dimension of the image (e.g. three col or dimensions for RGB representation or other larger numbers of dimensions).

The input high-dimensional representation may be received from a database (e.g. stored by the one or more storage devices). Alternatively, the system 100 may be configured to obtain the input high-dimensional representation by generating the input high-dimensional repre sentation of the biology-related image-based input data by a trained visual recognition ma chine-learning algorithm executed by the one or more processors 110. For example, the in put high-dimensional representation may be determined by applying at least a part (e.g. en coder) of a trained visual recognition machine-learning algorithm with a trained set of pa rameters to the biology-related image-based input data. For example, generating the input high-dimensional representation by the trained visual recognition machine- learning algo rithm may mean that the input high-dimensional representation is generated by an encoder of the trained visual recognition machine- learning algorithm. The trained set of parameters of the trained visual recognition machine- learning algorithm may be obtained during train ing of the visual recognition machine- learning algorithm as described below.

The trained visual recognition machine- learning algorithm may also be called image recog nition model or visual model. The trained visual recognition machine- learning algorithm may be implemented as described above or below.

The biology-related image-based input data may be image data (e.g. pixel data of an image) of an image of a biological structure comprising a nucleotide or a nucleotide sequence, a biological structure comprising a protein or a protein sequence, a biological molecule, a biological tissue, a biological structure with a specific behavior, and/or a biological structure with a specific biological function or a specific biological activity. The biological structure may be a molecule, a viroid or virus, artificial or natural membrane enclosed vesicles, a subcellular structure (like a cell organelle) a cell, a spheroid, an organoid, a three- dimensional cell culture, a biological tissue, an organ slice or part of an organ in vivo or in vitro. For example, the image of the biological structure may be an image of the location of a protein within a cell or tissue or an image of a cell or tissue to which endogen nucleotides (e.g. DNA) bind itself (e.g. in situ hybridization). The image data may comprise a pixel val ue for each pixel of an image for each color dimension of the image (e.g. three color dimen sions for RGB representation). For example, the biology-related image-based input data may be obtained from a database (e.g. stored by the one or more storage devices) or may be obtained by an imaging device (e.g. microscope, camera) during a running experiment. For example, the system 100 may comprise a microscope configured to obtain the biology-related image-based input data by taking an image of the biological specimen representing the biological structure.

The biology-related image-based input data may be one biology-related image-based input data set of a plurality of biology-related image-based input data sets of a training group. The training group may comprise more than 50 images, more than 100 images, more than 200 images or more than 500 images. For example, if more images are considered for the train ing, the typical images generated by the trained generative adversarial network may repro duce the relevant and/or important features more accurately. For example, the training group of images can be selected from a cluster of their respective high-dimensional representations (e.g. in semantic embedding space) so as to make them more homogeneous to produce a typical image. This cluster can be found by unsupervised clustering, user selection or searching a large body of data. Additionally or alternatively, the selection of suitable images may be done based biological sequences, coarse-grained search terms or their respective high-dimensional representations.

More details and aspects of the system 300 are mentioned in conjunction with the proposed concept and/or the one or more examples described above or below (e.g. Figs. 1-2 or 4-7). The system 300 may comprise one or more additional optional features corresponding to one or more aspects of the proposed concept and/or of one or more examples described above or below.

Fig. 4 shows a system for training a generative adversarial network for generating a typical image from combined semantic embeddings according to an embodiment. In general, GANs may create images from noise priors, but here the prior to the generative model 420 (e.g. generative network) would be a combined semantic embedding 310. Using a semantic em bedding as prior to the generative model 420 instead of a stochastic vector or noise prior may ensure that the discriminator 440 (e.g. discriminative network) can learn to output a matching semantic embedding 320 and use a relatively simple loss function, for example mean squared error MSE loss, for error backpropagation. The generator 420 and discrimina- tor 440 may be trained together such that the generator 420 learns to predict an image 430 which the discriminator 440 uses as input. The GAN may be fully trained if the discrimina tor cannot distinguish between a generated and a recorded image anymore.

More details and aspects of the system 400 are mentioned in conjunction with the proposed concept and/or the one or more examples described above or below (e.g. Figs. 1-3 or 5-7). The system 400 may comprise one or more additional optional features corresponding to one or more aspects of the proposed concept and/or of one or more examples described above or below.

The system described in conjunction with one of the Figs. 1-4 may comprise or may be a computer device (e.g. personal computer, laptop, tablet computer or mobile phone) with the one or more processors and one or more storage devices located in the computer device or the system may be a distributed computing system (e.g. cloud computing system with the one or more processors and one or more storage devices distributed at various locations, for example, at a local client and one or more remote server farms and/or data centers). The system may comprise a data processing system that includes a system bus to couple the var ious components of the system. The system bus may provide communication links among the various components of the system and may be implemented as a single bus, as a combi nation of busses, or in any other suitable manner. An electronic assembly may be coupled to the system bus. The electronic assembly may include any circuit or combination of circuits. In one embodiment, the electronic assembly includes a processor which can be of any type. As used herein, processor may mean any type of computational circuit, such as but not lim ited to a microprocessor, a microcontroller, a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a graphics processor, a digital signal processor (DSP), multiple core processor, a field programmable gate array (FPGA) of the microscope or a microscope component (e.g. camera) or any other type of processor or processing cir cuit. Other types of circuits that may be included in electronic assembly may be a custom circuit, an application-specific integrated circuit (ASIC), or the like, such as, for example, one or more circuits (such as a communication circuit) for use in wireless devices like mo bile telephones, tablet computers, laptop computers, two-way radios, and similar electronic systems. The system includes one or more storage devices, which in turn may include one or more memory elements suitable to the particular application, such as a main memory in the form of random access memory (RAM), one or more hard drives, and/or one or more drives that handle removable media such as compact disks (CD), flash memory cards, digital video disk (DVD), and the like. The system may also include a display device, one or more speak ers, and a keyboard and/or controller, which can include a mouse, trackball, touch screen, voice-recognition device, or any other device that permits a system user to input information into and receive information from the system.

Additionally, the system may comprise a microscope connected to a computer device or a distributed computing system. The microscope may be configured to generate the biology- related image-based input data sets by taking images from one or more specimens.

The microscope may be a light microscope (e.g. diffraction limited or sub-diffraction limit microscope as, for example, a super-resolution microscope or nanoscope). The microscope may be a stand-alone microscope or a microscope system with attached components (e.g. confocal scanners, additional cameras, lasers, climate chambers, automated loading mecha nisms, liquid handling systems, optical components attached, like additional multiphoton light paths, optical tweezers and more). Other image sources may be used as well, if they can take images of objects which are related to biological sequences (e.g. proteins, nucleic acids, lipids) or other specimens, for example. For example, a microscope according to an embodiment described above or below may enable deep discovery microscopy.

More details and aspects of the system are mentioned in conjunction with the proposed con cept and/or the one or more examples described above or below (e.g. Figs. 1-7). The system may comprise one or more additional optional features corresponding to one or more as pects of the proposed concept and/or of one or more examples described above or below.

Some embodiments relate to a microscope comprising a system as described in conjunction with one or more of the Figs. 1-4. Alternatively, a microscope may be part of or connected to a system as described in conjunction with one or more of the Figs. 1-4. Fig. 5 shows a schematic illustration of a system 500 for processing data according to an embodiment. A microscope 510 configured to take images of one or more biological specimens is connected to a computer device 520 (e.g. personal computer, laptop, tablet computer or mobile phone) configured to process biology-related data. The microscope 510 and the computer device 520 may be implemented as described in conjunction with one or more of the Figs. 1-4. Fig. 6 is a flow chart of a method for generating a biology-related image-based output data set of a typical image of a biological structure according to an embodiment. The method 600 comprises obtaining 610 a plurality of high-dimensional representations of a plurality of biology-related image-based input data sets of a set of images of one or more biological structures and determining 620 a typical high-dimensional representation based on the plu rality of high-dimensional representations. The typical high-dimensional representation comprises at least 3 entries each having a different value. Further, the method 600 compris es generating 630 a biology-related image-based output data set of a typical image of the biological structure based on the typical high-dimensional representation by a trained ma chine-learning algorithm.

By using a trained machine- learning algorithm for generating a typical image for a set of images, a high-quality image containing relevant and/or important features common to the set of images may be obtainable. By allowing the high-dimensional representations to have entries with various different values (in contrast to one hot encoded representations), images with semantically similar content may have similar high-dimensional representations. Espe cially for biology related images, the semantic similarity may be reproduced very well by the high-dimensional representations. The trained machine- learning algorithm may be able to provide a typical image accurately reproducing important and/or relevant features of the set of images due to the semantic similarity of the high-dimensional representations being close to each other. The combination of applying a trained machine-learning algorithm to high-dimensional representations with various different values reproducing semantical prox imity may allow to generate typical images with significantly improved quality.

More details and aspects of method 600 are mentioned in conjunction with the proposed concept and/or the one or more examples described above or below (e.g. Figs. 1-5). The method 600 may comprise one or more additional optional features corresponding to one or more aspects of the proposed concept and/or of one or more examples described above or below.

Fig. 7 is a flow chart of a method for training a generative adversarial network according to an embodiment. The method 700 comprises obtaining 710 an input high-dimensional repre sentation of biology-related image-based input data of a biological structure. The input high dimensional representation comprises at least 3 entries each having a different value. Fur- ther, the method 700 comprises generating 720 training image data based on the input high dimensional representation by a generative network of a generative adversarial network ex ecuted by the one or more processors and generating 730 an output high-dimensional repre sentation based on the training image data by a discriminative network of the generative adversarial network executed by the one or more processors 110. The output high dimensional representation comprises at least 3 entries each having a different value. Addi tionally, the method 700 comprises adjusting 740 the generative network of the generative adversarial network based on a comparison of the input high-dimensional representation and the output high-dimensional representation.

By training a generative adversarial network to generate an image based on high dimensional representations of biology related images, a machine- learning algorithm very well suited for generating typical images of biological images represented by high dimensional representations may be obtained. By allowing the high-dimensional representa tions to have entries with various different values (in contrast to one hot encoded representa tions), images with semantically similar content may have similar high-dimensional repre sentations. Especially for biology related images, the semantic similarity may be reproduced very well by the high-dimensional representations. The trained generative adversarial net work may be able to provide a typical image of a biological structure accurately reproducing important and/or relevant features.

More details and aspects of method 700 are mentioned in conjunction with the proposed concept and/or the one or more examples described above or below (e.g. Figs. 1-6). The method 700 may comprise one or more additional optional features corresponding to one or more aspects of the proposed concept and/or of one or more examples described above or below.

Some embodiments relate to a trained generative adversarial network trained by obtaining an input high-dimensional representation of biology-related image-based input data of a biological structure, wherein the input high-dimensional representation comprises at least 3 entries each having a different value. Further, the trained generative adversarial network was trained by generating training image data based on the input high-dimensional represen tation by a generative network of a generative adversarial network executed by the one or more processors and generating an output high-dimensional representation based on the training image data by a discriminative network of the generative adversarial network exe- cuted by the one or more processors. The output high-dimensional representation comprises at least 3 entries each having a different value. Additionally, the trained generative adversar ial network was trained by adjusting the generative network of the generative adversarial network based on a comparison of the input high-dimensional representation and the output high-dimensional representation.

The trained generative adversarial network may comprise at least a generative network and/or a discriminative network. At least a part of the trained generative adversarial network may be learned parameters (e.g. neural network weights) stored by a storage device.

More details and aspects of trained generative adversarial network are mentioned in con junction with the proposed concept and/or the one or more examples described above or below (e.g. Figs. 1-7). The trained generative adversarial network may comprise one or more additional optional features corresponding to one or more aspects of the proposed con cept and/or of one or more examples described above or below.

In the following, some examples of applications and/or implementation details for one or more of the embodiments described above (e.g. in conjunction with one or more of the Figs. 1-7) are described.

According to an aspect, an image-to-typical image creation may be proposed during which a “standardized” or“typical” image is produced from a set of images showing a number of observations of the same class of object.

Images of biological objects may be subject to an immense variability in morphology and topology. Thus, biologists who interpret small data sets manually may subjectively overin terpret certain image features as being relevant for the biological function of the object. Ex cessively large datasets with millions or trillions of images, which could provide higher sta tistical relevance, cannot be curated or interpreted manually anymore. The proposed concept may close the gap of getting an image which can be interpreted by the researcher and cap tures at the same time all the statistically relevant features which occur in a large body of data or in a running experiment. The results may lead to quantifiable descriptors and generalized representations of biologi cal objects found in the specimen. Moreover, image-to-image searches based on those im ages may lead to better hits, because the search item had eliminated irrelevant variability.

According to an aspect, important features of the object may be recognized without being misled by variability in the data. Further, a generalized representation of the object may be found providing the basis for a more precise description (e.g. as usually provided by the author of a scientific publication). Additionally, effects of noise and morphological / topo logical variability may be reduced or eliminated to quantify difference between two or more objects. Further, a new image may be obtained as search query for image-to-image search for finding related objects in running experiments or public databases to get better search hits. Additionally, a downstream image analysis may be facilitated when applied to a typical image instead of just one single observation.

According to an aspect, a combination of textual and visual models may be used in micros copy (to train a machine-learning algorithm to generate high-dimensional representations) to create typical images.

The proposed concept may enable to standardize visual representations of objects found in biomedical research in such areas as cell science and/or oncology. Additionally or alterna tively, the proposed concept may enable to follow complicated morphological changes over time by this type of semantic ensemble averaging, when one object cannot be followed long enough in such areas as developmental biology, toxicology and/or pharmacology, drug vali dation or drug screens using cell based assays. Additionally or alternatively, the proposed concept may enable to develop spatio-temporal memes of complex growth patterns in or ganoids, 3D cultures of induce pluripotent stem cells and/or regenerative medicine.

Other concepts may use elastic matching. In elastic matching a number of non-rigid trans formations is done on each image in the data set to produce an image with the smallest global deviation from the entire data set. However, an assumption about the shape and to pology of the object may be needed which serve as a boundary condition. The large varia bility in morphology and topology found in microscopic images may limit the applicability of elastic matching, especially, because the objects in the specimen react dynamically to their neighbors and to environmental conditions. Alternatively, other concepts may use Eigenimages. A principal component analysis may be performed on the set of images to minimize the variability in the data set and transform the coordinate system so as to be par allel to eigenvectors of the data. This may suffer from variations in lighting conditions which appear in higher order eigenvectors. Manual curation of those may be needed to make this technique applicable. As a result“typical images” created by other techniques may tend to be blurry and contain artifacts (e.g. such as from different lighting conditions).

In contrast, the proposed concept may need no assumptions about the inner structure, mor phology or topology of the objects and may be more robust towards changes in lighting conditions and may produce plausible, high quality images.

A trained visual recognition machine- learning algorithm may be obtained by a training de scribed in the following. A system for training machine- learning algorithms for processing biology-related data may comprise one or more processors and one or more storage devices. The system may be configured to receive biology-related language-based input training da ta. Additionally, the system may be configured to generate a first high-dimensional repre sentation of the biology-related language-based input training data by a language recogni tion machine- learning algorithm executed by the one or more processors. The first high dimensional representation comprises at least three entries each having a different value. Further, the system may be configured to generate biology-related language-based output training data based on the first high-dimensional representation by the language recognition machine- learning algorithm executed by the one or more processors. In addition, the system may be configured to adjust the language recognition machine- learning algorithm based on a comparison of the biology-related language-based input training data and the biology- related language-based output training data. Additionally, the system may be configured to receive biology-related image-based input training data associated with the biology-related language-based input training data. Further, the system may be configured to generate a second high-dimensional representation of the biology-related image-based input training data by a visual recognition machine- learning algorithm executed by the one or more pro cessors. The second high-dimensional representation comprises at least three entries each having a different value. Further, the system may be configured to adjust the visual recogni tion machine-learning algorithm based on a comparison of the first high-dimensional repre sentation and the second high-dimensional representation. The biology-related language-based input training data may be a textual input being related to a biological structure, a biological function, a biological behavior or a biological activity. For example, the biology-related language-based input training data may be a nucleotide sequence, a protein sequence, a description of a biological molecule or biological structure, a description of a behavior of a biological molecule or biological structure, and/or a descrip tion of a biological function or a biological activity. The biology-related language-based input training data may be a first biology-related language-based input training data set (e.g. sequence of input characters, for example, a nucleotide sequence or a protein sequence) of a training group. The training group may comprise a plurality of biology-related language- based input training data sets.

The biology-related language-based output training data may be of the same type as the bi ology-related language-based input training data including optionally a prediction of a next element. For example, the biology-related language-based input training data may be a bio logical sequence (e.g. a nucleotide sequence or a protein sequence) and the biology-related language-based output training data may be a biological sequence (e.g. a nucleotide se quence or a protein sequence) as well. The language recognition machine- learning algo rithm may be trained so that the biology-related language-based output training data is equal to the biology-related language-based input training data including optionally a prediction of a next element of the biological sequence. In another example, the biology-related language- based input training data may be a biological class of a coarse-grained search term and the biology-related language-based output training data may be a biological class of the coarse grained search term as well.

The biology-related image-based input training data may be image training data (e.g. pixel data of a training image) of an image of a biological structure comprising a nucleotide or a nucleotide sequence, a biological structure comprising a protein or a protein sequence, a biological molecule, a biological tissue, a biological structure with a specific behavior, and/or a biological structure with a specific biological function or a specific biological activ ity. The biology-related image-based input training data may be a first biology-related im age-based input training data set of a training group. The training group may comprise a plurality of biology-related image-based input training data sets. The biology-related language-based input training data may be a biology-related language- based input training data set (e.g. sequence of input characters, for example, a nucleotide sequence or a protein sequence) of a training group. The training group may comprise a plurality of biology-related language-based input training data sets. The system may repeat generating a first high-dimensional representation for each of a plurality of biology-related language-based input training data sets of a training group. Further, the system may generate biology-related language-based output training data for each generated first high dimensional representation. The system may adjust the language recognition machine learning algorithm based on each comparison of biology-related language-based input train ing data of the plurality of biology-related language-based input training data sets of the training group with the corresponding biology-related language-based output training data. In other words, the system may be configured to repeat generating a first high-dimensional representation, generating biology-related language-based output training data , and adjust ing the language recognition machine-learning algorithm for each biology-related language- based input training data of a training group of biology-related language-based input train ing data sets. The training group may comprise enough biology-related language-based in put training data sets so that a training target (e.g. variation of an output of a loss function below a threshold) can be fulfilled.

The plurality of all first high-dimensional representations generated during training of the language recognition machine- learning algorithm may be called latent space or semantic space.

The system may repeat generating a second high-dimensional representation for each of a plurality of biology-related image-based input training data sets of a training group. Further, the system may adjust the visual recognition machine- learning algorithm based on each comparison of a first high-dimensional representation with the corresponding second high dimensional representation. In other words, the system may repeat generating a second high-dimensional representation and adjusting the visual recognition machine-learning algo rithm for each biology-related image-based input training data of a training group of biolo gy-related image-based input training data sets. The training group may comprise enough biology-related image-based input training data sets so that a training target (e.g. variation of an output of a loss function below a threshold) can be fulfilled. For example, the system 100 uses a combination of a language recognition machine learning algorithm and a visual recognition machine- learning algorithm (e.g. also called visual-semantic model). The language recognition machine- learning algorithm and/or the visual recognition machine-learning algorithm may be deep learning algorithms and/or arti ficial intelligence algorithms.

The training may converge fast and/or may provide a well-trained algorithm for biology- related data by using the cross entropy loss function for training the language recognition machine- learning algorithm, although other loss functions could be used as well.

The visual recognition machine- learning algorithm may be trained by adjusting parameters of the visual recognition machine- learning algorithm based on the comparison of a high dimensional representation generated by the language recognition machine- learning algo rithm with a high dimensional representation generated by the visual recognition machine learning algorithm of corresponding input training data. For example, network weights of a visual recognition neural network may be adjusted based on the comparison. The adjustment of the parameters (e.g. network weights) of the visual recognition machine-learning algo rithm may be done under consideration of a loss function. For example, the comparison of the first high-dimensional representation and the second high-dimensional representation for the adjustment of the visual recognition machine- learning algorithm may be based on a co sine similarity loss function. The training may converge fast and/or may provide a well- trained algorithm for biology-related data by using the cosine similarity loss function for training the visual recognition machine- learning algorithm, although other loss functions could be used as well.

For example, the visual model may learn how to represent an image in the semantic embed ding space (e.g. as a vector). So, a measure for the distance of two vectors may be used, which may represent the prediction A (the second high-dimensional representation) and the ground-truth B (the first high-dimensional representation). For example, a measure is the cosine similarity as defined in

A B

similarity eosf#) =

I A. II If 53 with the dot product of the prediction A and ground-truth B divided by the dot product of their respective magnitudes (e.g. as in L2-Norm or Euclidian norm).

More details with respect to non-training specific aspects of the system for training ma chine-learning algorithms are mentioned in conjunction with the proposed concept and/or the one or more examples described above or below (e.g. Figs. 1-7).

Embodiments may be based on using a machine- learning model or machine- learning algo rithm. Machine learning may refer to algorithms and statistical models that computer sys tems may use to perform a specific task without using explicit instructions, instead relying on models and inference. For example, in machine-learning, instead of a rule-based trans formation of data, a transformation of data may be used, that is inferred from an analysis of historical and/or training data. For example, the content of images may be analyzed using a machine- learning model or using a machine- learning algorithm. In order for the machine learning model to analyze the content of an image, the machine- learning model may be trained using training images as input and training content information as output. By train ing the machine-learning model with a large number of training images and/or training se quences (e.g. words or sentences) and associated training content information (e.g. labels or annotations), the machine- learning model“learns” to recognize the content of the images, so the content of images that are not included in the training data can be recognized using the machine- learning model. The same principle may be used for other kinds of sensor data as well: By training a machine-learning model using training sensor data and a desired output, the machine- learning model“learns” a transformation between the sensor data and the out put, which can be used to provide an output based on non-training sensor data provided to the machine- learning model.

Machine- learning models may be trained using training input data. The examples specified above use a training method called“supervised learning”. In supervised learning, the ma chine-learning model is trained using a plurality of training samples, wherein each sample may comprise a plurality of input data values, and a plurality of desired output values, i.e. each training sample is associated with a desired output value. By specifying both training samples and desired output values, the machine-learning model“learns” which output value to provide based on an input sample that is similar to the samples provided during the train ing. Apart from supervised learning, semi-supervised learning may be used. In semi- supervised learning, some of the training samples lack a corresponding desired output value. Supervised learning may be based on a supervised learning algorithm, e.g. a classification algorithm, a regression algorithm or a similarity learning algorithm. Classification algo rithms may be used when the outputs are restricted to a limited set of values, i.e. the input is classified to one of the limited set of values. Regression algorithms may be used when the outputs may have any numerical value (within a range). Similarity learning algorithms may be similar to both classification and regression algorithms, but are based on learning from examples using a similarity function that measures how similar or related two objects are. Apart from supervised or semi-supervised learning, unsupervised learning may be used to train the machine- learning model. In unsupervised learning, (only) input data might be sup plied, and an unsupervised learning algorithm may be used to find structure in the input da ta, e.g. by grouping or clustering the input data, finding commonalities in the data. Cluster ing is the assignment of input data comprising a plurality of input values into subsets (clus ters) so that input values within the same cluster are similar according to one or more (pre defined) similarity criteria, while being dissimilar to input values that are included in other clusters.

Reinforcement learning is a third group of machine- learning algorithms. In other words, reinforcement learning may be used to train the machine- learning model. In reinforcement learning, one or more software actors (called“software agents”) are trained to take actions in an environment. Based on the taken actions, a reward is calculated. Reinforcement learn ing is based on training the one or more software agents to choose the actions such, that the cumulative reward is increased, leading to software agents that become better at the task they are given (as evidenced by increasing rewards).

Furthermore, some techniques may be applied to some of the machine-learning algorithms. For example, feature learning may be used. In other words, the machine- learning model may at least partially be trained using feature learning, and/or the machine- learning algo rithm may comprise a feature learning component. Feature learning algorithms, which may be called representation learning algorithms, may preserve the information in their input, but also transform it in a way that makes it useful, often as a pre-processing step before per forming classification or predictions. Feature learning may be based on principal compo nents analysis or cluster analysis, for example. In some examples, anomaly detection (i.e. outlier detection) may be used, which is aimed at providing an identification of input values that raise suspicions by differing significantly from the majority of input or training data. In other words, the machine- learning model may at least partially be trained using anomaly detection, and/or the machine- learning algorithm may comprise an anomaly detection component.

In some examples, the machine-learning algorithm may use a decision tree as a predictive model. In other words, the machine- learning model may be based on a decision tree. In a decision tree, observations about an item (e.g. a set of input values) may be represented by the branches of the decision tree, and an output value corresponding to the item may be rep resented by the leaves of the decision tree. Decision trees may support both discrete values and continuous values as output values. If discrete values are used, the decision tree may be denoted a classification tree, if continuous values are used, the decision tree may be denoted a regression tree.

Association rules are a further technique that may be used in machine- learning algorithms. In other words, the machine- learning model may be based on one or more association rules. Association rules are created by identifying relationships between variables in large amounts of data. The machine- learning algorithm may identify and/or utilize one or more relational rules that represent the knowledge that is derived from the data. The rules may e.g. be used to store, manipulate or apply the knowledge.

Machine- learning algorithms are usually based on a machine- learning model. In other words, the term“machine-learning algorithm” may denote a set of instructions that may be used to create, train or use a machine- learning model. The term“machine- learning model” may denote a data structure and/or set of rules that represents the learned knowledge, e.g. based on the training performed by the machine-learning algorithm. In embodiments, the usage of a machine- learning algorithm may imply the usage of an underlying machine learning model (or of a plurality of underlying machine- learning models). The usage of a machine- learning model may imply that the machine- learning model and/or the data struc ture/set of rules that is the machine- learning model is trained by a machine- learning algo rithm. For example, the machine- learning model may be an artificial neural network (ANN). ANNs are systems that are inspired by biological neural networks, such as can be found in a retina or a brain. ANNs comprise a plurality of interconnected nodes and a plurality of con nections, so-called edges, between the nodes. There are usually three types of nodes, input nodes that receiving input values, hidden nodes that are (only) connected to other nodes, and output nodes that provide output values. Each node may represent an artificial neuron. Each edge may transmit information, from one node to another. The output of a node may be de fined as a (non-linear) function of the sum of its inputs. The inputs of a node may be used in the function based on a“weight” of the edge or of the node that provides the input. The weight of nodes and/or of edges may be adjusted in the learning process. In other words, the training of an artificial neural network may comprise adjusting the weights of the nodes and/or edges of the artificial neural network, i.e. to achieve a desired output for a given in put.

Alternatively, the machine-learning model may be a support vector machine, a random for est model or a gradient boosting model. Support vector machines (i.e. support vector net works) are supervised learning models with associated learning algorithms that may be used to analyze data, e.g. in classification or regression analysis. Support vector machines may be trained by providing an input with a plurality of training input values that belong to one of two categories. The support vector machine may be trained to assign a new input value to one of the two categories. Alternatively, the machine- learning model may be a Bayesian network, which is a probabilistic directed acyclic graphical model. A Bayesian network may represent a set of random variables and their conditional dependencies using a directed acy clic graph. Alternatively, the machine- learning model may be based on a genetic algorithm, which is a search algorithm and heuristic technique that mimics the process of natural selec tion.

As used herein the term "and/or" includes any and all combinations of one or more of the associated listed items and may be abbreviated as "/".

Although some aspects have been described in the context of an apparatus, it is clear that these aspects also represent a description of the corresponding method, where a block or device corresponds to a method step or a feature of a method step. Analogously, aspects described in the context of a method step also represent a description of a corresponding block or item or feature of a corresponding apparatus. Some or all of the method steps may be executed by (or using) a hardware apparatus, like for example, a processor, a micropro cessor, a programmable computer or an electronic circuit. In some embodiments, some one or more of the most important method steps may be executed by such an apparatus.

Depending on certain implementation requirements, embodiments of the invention can be implemented in hardware or in software. The implementation can be performed using a non- transitory storage medium such as a digital storage medium, for example a floppy disc, a DVD, a Blu-Ray, a CD, a ROM, a PROM, and EPROM, an EEPROM or a FLASH memory, having electronically readable control signals stored thereon, which cooperate (or are capable of cooperating) with a programmable computer system such that the respective method is performed. Therefore, the digital storage medium may be computer readable.

Some embodiments according to the invention comprise a data carrier having electronically readable control signals, which are capable of cooperating with a programmable computer system, such that one of the methods described herein is performed.

Generally, embodiments of the present invention can be implemented as a computer pro gram product with a program code, the program code being operative for performing one of the methods when the computer program product runs on a computer. The program code may, for example, be stored on a machine readable carrier. For example, the computer pro gram may be stored on a non-transitory storage medium. Some embodiments relate to a non-transitory storage medium including machine readable instructions, when executed, to implement a method according to the proposed concept or one or more examples described above.

Other embodiments comprise the computer program for performing one of the methods de scribed herein, stored on a machine readable carrier.

In other words, an embodiment of the present invention is, therefore, a computer program having a program code for performing one of the methods described herein, when the com puter program runs on a computer. A further embodiment of the present invention is, therefore, a storage medium (or a data carrier, or a computer-readable medium) comprising, stored thereon, the computer program for performing one of the methods described herein when it is performed by a processor. The data carrier, the digital storage medium or the recorded medium are typically tangible and/or non-transitionary. A further embodiment of the present invention is an apparatus as described herein comprising a processor and the storage medium.

A further embodiment of the invention is, therefore, a data stream or a sequence of signals representing the computer program for performing one of the methods described herein. The data stream or the sequence of signals may, for example, be configured to be transferred via a data communication connection, for example, via the internet.

A further embodiment comprises a processing means, for example, a computer or a pro grammable logic device, configured to, or adapted to, perform one of the methods described herein.

A further embodiment comprises a computer having installed thereon the computer program for performing one of the methods described herein.

A further embodiment according to the invention comprises an apparatus or a system con figured to transfer (for example, electronically or optically) a computer program for per forming one of the methods described herein to a receiver. The receiver may, for example, be a computer, a mobile device, a memory device or the like. The apparatus or system may, for example, comprise a file server for transferring the computer program to the receiver.

In some embodiments, a programmable logic device (for example, a field programmable gate array) may be used to perform some or all of the functionalities of the methods de scribed herein. In some embodiments, a field programmable gate array may cooperate with a microprocessor in order to perform one of the methods described herein. Generally, the methods are preferably performed by any hardware apparatus. List of reference Signs

100 system for generating a biology-related image-based output data set of a typical im age of a biological structure

110 one or more processors

111 high-dimensional representations; semantic embedding

113 biology-related image-based output data set; new image; typical image

120 one or more storage devices

200 system for generating a typical image

201 database

203 imaging system; microscope

207 user

210 set of image data; plurality of biology-related image-based input data sets

220 trained visual recognition machine-learning algorithm; visual model

240 combining of set of semantic embeddings

250 typical high-dimensional representation; latent vector

260 generative adversarial network

300 system for training a generative adversarial network

310 input high-dimensional representation

320 output high-dimensional representation

400 system for training a generative adversarial network

420 generative model; generative network of a generative adversarial network

430 predicted image; training image data

440 discriminator; discriminative network of a generative adversarial network

500 system for processing biology-related data including a microscope

510 microscope

520 computer device

600 method for generating a biology-related image-based output data set of a typical im age of a biological structure

610 obtaining a plurality of high-dimensional representations

620 determining a typical high-dimensional representation

630 generating a biology-related image-based output data set

700 method for training a generative adversarial network

710 obtaining an input high-dimensional representation generating training image data

generating an output high-dimensional representation adjusting the generative network