Processing of cell images in a folding network with context data

Abstract

The disclosure relates to the processing of images in a machine-learned convolutional network. The techniques make it possible to improve the accuracy and reliability of applications that use machine-learned convolutional networks, particularly in cell microscopy. Contextual data is provided that is indicative of scene-global cell features, and this data is then considered in the inference of the machine-learned convolutional network.

Description

TECHNICAL AREA

[0001] Several examples of the revelation concern techniques for processing an image depicting a scene with cells in a machine-learned convolutional network. In particular, contextual data is taken into account that is indicative of scene-global cell characteristics. BACKGROUND

[0002] Deep convolutional neural networks (CNNs; also known as machine-learned convolutional networks) have become state-of-the-art in microscopy for segmentation and image-to-image tasks. One example is UNets; see, for example, Falk, Thorsten, et al. “U-Net: deep learning for cell counting, detection, and morphometry.” Nature methods 16.1 (2019): 67–70.

[0003] In detail, CNNs are used in a variety of areas, such as the position and shape determination of sample carriers, cell counting and estimation of cell confluence, live / dead cell analysis, compressed sensing, contrast-to-contrast imaging and virtual staining, cell instance segmentation, transfection analysis, deconvolution, denoising, high resolution and in artifact reduction models.

[0004] However, it has been observed that such CNNs sometimes deliver inaccurate or incorrect results. SUMMARY

[0005] Therefore, it is an object of the present invention to provide techniques that achieve improved results in the processing of images (especially microscope images) using machine-learned convolutional networks.

[0006] This task is solved by the features of the independent patent claims. The features of the dependent patent claims define embodiments.

[0007] Several examples are based on the finding that when applying CNNs to image processing tasks on images depicting a scene with cells (cell images), convolution operations in a layer of the CNN only "see" a local region around a specific position in the image; this is referred to as the "receptive field" of the CNN. This means that context located at other positions in the image is ignored by the respective convolution operation. The receptive field of a model is typically smaller the "lighter" the CNN is, i.e., the fewer machine-learned parameters the CNN has. For example, the size and number of convolution layers typically limit the receptive field. An example of this is cell counting and confluence estimation using a microscope image.If the background (the region without cells) of the microscope image exhibits a periodic structure, the CNN may mistake this for a cell cluster. If a CNN processes a section of the receptive field at a particular image position that contains only this background pattern (i.e., no cells), the CNN might interpret the background pattern as cells, leading to false detections. However, if the receptive field also contains at least some actual cells, false detections in the background are often avoided because the model has a kind of "reference" within the image.

[0008] A computer-implemented method comprises several steps to obtain cell-specific result data for a scene containing cells. First, an image depicting the scene with the cells is acquired. This image, or another image depicting at least the scene with the cells, is then processed in a context module. Within this context module, context data indicative of scene-global cell characteristics is determined. This context data, along with the image, is then processed in an inference run of a machine-learned convolutional network. The convolutional network has a receptive field in its input layer that is smaller than the representation of the scene in the image. By processing the image and the context data in this network, cell-specific result data for the scene is obtained.

[0009] An electronic data processing facility is set up to carry out such a procedure.

[0010] By incorporating contextual data into its inference process, the CNN can better distinguish between actual cells and, for example, background structures. This leads to improved accuracy and reliability in CNN-based applications, such as in microscopy. Providing contextual data to the convolutional network allows for a comprehensive view of the scene and avoids false detections that could otherwise arise from the limitations of the receptive field.

[0011] The features set out above and those described below can be used not only in the corresponding explicitly set out combinations, but also in further combinations or in isolation, without leaving the scope of protection of the present invention. BRIEF DESCRIPTION OF THE FIGURES Fig. Figure 1 is a flowchart of an example procedure. Fig. Figure 2 schematically illustrates the processing of an image in a CNN according to reference implementations. Fig. Figure 3 schematically illustrates the processing of an image in a CNN according to various examples described herein. Fig. Figure 4 schematically illustrates a data processing facility set up for techniques such as those described therein. DETAILED DESCRIPTION

[0012] The present invention is explained in more detail below with reference to preferred embodiments and the drawings. In the figures, identical reference numerals denote identical or similar elements. The figures are schematic representations of various embodiments of the invention. Elements depicted in the figures are not necessarily shown to scale. Rather, the various elements depicted in the figures are represented in such a way that their function and general purpose are understandable to a person skilled in the art. Connections and couplings between functional units and elements shown in the figures can also be implemented as indirect connections or couplings. A connection or coupling can be implemented as a wired or wireless connection. Functional units can be implemented as hardware, software, or a combination of hardware and software.

[0013] Several aspects of the disclosure concern the processing of an image depicting cells in a CNN. For example, the techniques described herein may be particularly applicable to CNNs with no more than E6 (i.e., one million) machine-learned weights. Such CNNs are relatively compact and may be particularly relevant for applications where rapid output is required and / or limited computational resources are available.

[0014] A CNN is characterized by the size of its receptive field. The receptive field refers to the portion of the processed image to which a single neuron unit in the CNN can respond (a neuron unit performs a weighted summation of its inputs, applies an activation function to generate the output, and passes it on). In other words, the receptive field defines that part of the processed data structure (typically an image, for example, with multiple channels) that can directly influence the activation of a neuron. The receptive field is limited by the size and structure of the filters used, as well as by the number and type of layers in the network. In practice, a smaller receptive field limits the network's ability to capture long-range dependencies and comprehensive contextual information.This often leads to challenges in processing data with complex or hierarchical structures, which can cause performance losses in certain applications such as image segmentation.

[0015] Typically, the fewer machine-learned parameters the CNN has, the smaller the size of the receptive field. For example, the cell image size can be smaller than the size of the receptive field; the size of the receptive field can, in turn, be smaller than the image size. Thus, it is conceivable that the cell image size is at least two orders of magnitude smaller than the image size, while the receptive field, on the other hand, is up to one order of magnitude larger than the cell image size. In other words, this means that the receptive field is smaller than the image size, but can, in principle, "see" multiple cells.

[0016] Techniques for processing cell images are described below. When a CNN infers, in addition to the local image information defined by the receptive field, a global context of the entire image is also considered for each image position. This allows the CNN, for example, to recognize which cell type is present in the image and thus prevents false detections in background areas without such cells. Context data can be provided as a vector, defined in an embedded space, i.e., a latent feature space. The vector can, for example, encode the cell type. In one variant, a corresponding method involves obtaining an input image and a context image. Both the input image and the context image depict a scene containing cells. The method also includes extracting features from the context image. Context modules can be used for this purpose.Contextual data is thus obtained that is indicative of scene-global cell characteristics. Furthermore, the method includes processing the input image in an inference run of a CNN. In this way, result data describing one or more properties related to the cells is obtained for the scene. The contextual data is taken into account during image processing. Such a method is described below in connection with the flowchart from [reference missing]. Fig. 1 explained.

[0017] Fig. Figure 1 is a flowchart of an example procedure. The procedure is from Fig. 1 can be performed by an electronic data processing device. For example, the procedure could be performed by Fig. 1. A process is executed by a processor when the processor loads program code from memory and then executes it. The procedure from Fig. 1 concerns the processing of images depicting a scene with cells in an inference run of a CNN. In this way, cell-specific result data for the scene is obtained.

[0018] Box 805 will receive an image depicting a scene with cells. For example, a microscope image could be obtained. This image could be acquired using phase-contrast imaging. Fluorescence imaging could also be used, such as in transfection analysis. The image could be obtained from a microscopy system. The image could be obtained from a camera within the microscopy system. The image could be loaded from a database or storage.

[0019] Box 810 will optionally contain another image. This additional image also depicts the same scene with the cells. However, this additional image could also represent the surrounding area of ​​the scene, for example, further peripheral regions.

[0020] The additional image could, for example, be obtained from a microscopy system. The additional image could be obtained from a camera. The additional image could be loaded from a database or storage.

[0021] The image obtained in Box 805 and the image obtained in Box 810 may have been acquired using different imaging modalities. The resolution of the two images may differ. Different contrast types may have been used in the two images.

[0022] For example, the additional image could depict the scene and also its surroundings. The additional image could be an overview image. The additional image could have a lower magnification than the image. In other words, the additional image could depict the scene at a scale smaller than the scale at which the image from Box 805 depicts the scene. The additional image could optionally have a different contrast type than the image. For example, the image could be acquired using phase contrast, while the additional image is acquired using fluorescence imaging (or vice versa). The additional image could, for example, be obtained from an overview camera of a microscopy system, with the microscopy system providing the image from Box 805 via microscopic imaging, i.e., acquired with a different camera.

[0023] The image from box 805 will be referred to as the input image; and the image from box 810 as the context image.

[0024] The input image is generally larger than the receptive field of the CNN. The role of the context image is to provide auxiliary information to improve the inference of the input image. This auxiliary information describes a context that is not available in the receptive field, which is processed by at least some neurons of the CNN.

[0025] The contextual image can encompass a larger sample area than the input image (this can specifically include a "zoomed-out" version of the input image) or images taken at different magnifications (using different lenses). However, the contextual image can also encompass the same sample area as the input image, i.e., it cannot depict any surroundings. The contextual image can also be an overview image of the sample / rehearsal space.

[0026] The input and context images can be derived from different contrast sources (phase or brightfield contrast and fluorescence contrast). The context image then provides not only spatial context but also additional local information.

[0027] In Box 815, the input image from Box 805, or—if available—the context image from Box 810, is processed in a context module. This is how context data is preserved.

[0028] The context data is indicative of scene-global cell characteristics. This context data is then considered in Box 820 during a CNN inference run when the CNN processes the input image from Box 805.

[0029] Such scene-global features can therefore (at least to some extent) not be pixel-bound to individual areas of the scene, or pixel-bound to individual pixels of the input image or the context image (depending on which image is processed in Box 815). An example is providing a feature vector valid for the entire processed image. For instance, a local feature extraction followed by an aggregation step could be performed in the context module to determine such a globally valid feature vector.

[0030] A data structure for contextual data can take various forms. For example, it could be a feature vector, an image, a scalar, a list (e.g., of regions where valid cell detections are expected), or a category (e.g., the cell type present).

[0031] The contextual data can, for example, provide indicative, scene-wide distinguishing features between cells and any background structures. Background structures that look similar to cells—for example, because they are similar in size—can thus be effectively differentiated from the cells. These distinguishing features enable the CNN to avoid confusing the different structure types.

[0032] The contextual data could, for example, include classification information for the cells. This classification information could include cell type, cell size, cell species, cell state, brightness information for the cells (e.g., average brightness, minimum brightness, brightness variation, transfection brightness, etc. in the input image, to name just a few examples), cell size information, or cell density. This classification information describes a classification of the cells that is scene-global, meaning it is not limited to local areas of the input image. Such classification information can be scene-global, i.e., it can apply to cells located at various positions across the scene. In some scenarios, the classification information may include human-interpretable information elements.Such human-interpretable information elements must be distinguished from latent feature vectors, which are interpretable by a deep layer of a deep neural network using machine-learned weights (latent feature vectors). These latent feature vectors contain feature values ​​for machine-learned features. An example of human-interpretable information elements would be scaling information for the cells and / or background structures in the scene. In other words, such scaling information can be indicative of a typical image size for the cells and / or background structures.

[0033] It would also be conceivable for the context data to include localization information for the cells and / or the background structures of the scene. Such localization information could be provided, for example, in the form of point markers (such as a "center point" marker) or in the form of bounding boxes in corresponding images. A segmentation mask could also be defined. The context data could therefore include a context image that marks cells or parts of cells.

[0034] In principle, it would be conceivable for the context data to be indicative of a comparison or deviation of locally determined cell features against a scene-global reference (in this respect, the context data is still indicative of scene-global determined features). For example, the image could be divided into sub-areas (patches), and each sub-area could be assigned a corresponding feature vector. The additional context data could thus comprise several context data components intended for different patches of the image processed in Box 815. For example, the context data components could be indicative of a change in scene properties from patch to patch in the different patches of the input image or the context image. In other words, this means that variations in the appearance of the cells and / or the background structures can be described on a global scale, i.e., across the entire scene.

[0035] In principle, it would be conceivable for the context module to operate separately from the CNN. That is, the output of the context module is independent of the output of the CNN. In some examples, it would be conceivable for the context module to determine the context data based on an output from the CNN in Box 820, where the CNN output was provided in a previous inference run, i.e., in a previous iteration 821. In this way, context data and cell-specific result data can be determined iteratively over several iterations 825. Thus, iterative inference in Box 820 is possible: For example, it would be conceivable that—in a first iteration 825—Box 815 is not yet executed. The context data could then be set to "null," for example. The CNN would then generate result data in Box 820. This could, for example, include image data of the same size as the input image in Box 805.For example, cell segmentation could be provided. It would then be conceivable to subsequently aggregate the result data generated in iteration 825 of Box 820 (in the following iteration 825 of Box 815); this aggregated result data could then be used as context data. This process can be repeated as often as desired. This allows for a more precise determination of the result data in Box 820 from iteration 825 to iteration 825.

[0036] The context module is implemented as software or program code. Furthermore, the context module can be implemented in various ways. For example, the determination of context data can be achieved by using a CNN as a feature extractor. For instance, the activations of neurons in the penultimate layer or another layer can be aggregated. In this way, scene-global features are obtained. Alternatively, a vision transformer-based machine-learning network could be used: for example, by inferring a specific token; the latent feature vector of this token can then be used as a feature vector after inference.

[0037] In general, a pre-trained model with machine-learned weights can be used as the context module. These weights can be trained on a generic (i.e., non-cell-specific) or application-related dataset (e.g., based on cell images from very different imaging modalities). For example, a foundation model trained in a domain-independent manner, i.e., not just on cell images, could be used. Such a pre-trained model can provide the context data, which describes, for example, one or more of the following cell characteristics: cell type classification, cell type, cell state, focal plane, cell size estimation, maximum background area size, average cell density in the image, etc.

[0038] Another possibility involves using an autoencoder network in the context module. This allows for the extraction of context data from the bottleneck or the generation of aggregated values, for example through averaging or max projection, from any layer.

[0039] Further techniques for implementing the context module rely on non-machine-learned algorithms. For example, image-based dimensionality reduction algorithms such as Principal Component Analysis (PCA), Non-negative Matrix Factorization (NMF), or Active Appearance Models (AAMs) (on the entire image or tiles) could be used. Statistical analyses could also be performed, such as generating distributions (e.g., cell colors for living / dead cells) and finding the optimal separation parameter. Image properties in segmented / all areas (e.g., contrast, brightness, color distribution, texture features, etc.) could be determined.

[0040] If the context module includes machine-learned weights, these weights can be trained separately from the machine-learned weights of the CNN, which is subsequently used in Box 820. This means that different training processes are used for the CNN in Box 820 and for the context module in Box 815. These different training processes allow for targeted training to achieve different objectives. For the context module in Box 815, scene-global features of the cells can be specifically extracted, while in Box 820, training can be performed taking into account the size-limited receptive field.

[0041] Various techniques for implementing the context module have been described above. Such implementation variants can also be combined, for example by maintaining several submodules that retrieve different parts of the context data and / or by combining (e.g., averaging) their outputs.

[0042] In Box 820, the input image is then processed by the CNN. The CNN performs inferences, thus obtaining cell-specific result data for the scene. The CNN has a receptive field that is smaller than the image size of the scene in the input image.

[0043] Box 820 also incorporates the context data from Box 815. There are different ways in which this context data can be considered. For example, as described above, the context data could include at least one latent feature vector, which would then be processed at a corresponding hidden layer within the CNN. In other words, the latent feature vector is not inputted to the CNN's input layer, but rather concatenated with the feature map at a deeper layer between the input and output layers.

[0044] In other examples, it would also be conceivable to combine (concatenate) the context data with the input image from Box 805 at the CNN's input layer and process them together in the input layer. The context data can thus be represented as additional "color channels" of the input image: for example, the input image could have nxm pixels for each of three color channels. Then, the three nxm matrices could be extended by one or more additional channels of the same dimension based on the context data. Individual feature vectors can be provided for each pixel in the image. This can be implemented, for example, as described above in connection with the context data components that are determined for different patches in the image. A context image could be used. Such a context image could also be considered in an output layer.For example, the output of the CNN could be made plausible using the contextual image by performing a pixel-based comparison between the contextual image and the output image of the CNN. Alternatively, the same scalar value could be inserted at each pixel value.

[0045] Instead of processing the context data within the CNN, the context data can also be used to modify certain parameters of the CNN operating on the input data. This means that the context data is not used as input to the CNN, but rather for parameterizing the CNN.

[0046] Fig. Figure 2 schematically illustrates the exemplary processing of an input image 105. The input image 105 is a microscope image acquired using phase contrast. It shows cells and background structures; the background structures are present, for example, in the central area marked by the arrow. However, no cells are present in this central area. The input image 105 (compare Fig. 1: Box 805), is then displayed in a CNN 115 (compare Fig. 1: Box 820) is processed. The CNN 115 has a receptive field 107, which is significantly smaller than the size of the input image 105, but also larger than the size of the cells.

[0047] The result will be image 106, which, in the example shown, provides a center-localization of the different cells. However, other result data is also conceivable, e.g., a scalar for cell counting, etc.

[0048] In result image 106, the center points are marked with small crosses. Fig. Figure 2 shows that individual background structures are also incorrectly recognized as cells when – unlike what was previously described in connection with Fig. As described in section 1, no context data is taken into account. This is because, for example, the receptive field 107 has an extent that is smaller than the background area, so that only background structures are seen for the image area 107-1 shown in the bottom left (this corresponds to a receptive field). However, this is different for image area 107-2; there, background structures are not incorrectly recognized as cells because real cells lie within the image area 107-2 captured by the receptive field. Fig. Figure 3 shows the same processing, but with contextual data taken into account. There is no misidentification of background structures as cells.

[0049] Two specific embodiments of the invention are discussed below.

[0050] The first embodiment concerns the estimation of transfection. A cell culture is stained with a fluorescent dye. The aim is to verify which cells have correctly "accepted" the dye, that is, which cells have undergone successful transfection. For this purpose, an image pair is acquired, consisting of a phase-contrast image (the input image, see Box 805) and a corresponding fluorescent image. All cells are visible in the phase-contrast image, regardless of whether the transfection was successful or not. Conversely, only those cells for which the transfection was successful are visible in the fluorescent image. Based on the phase-contrast image, cell instance segmentation is performed. Contextual data (see Box 805) are then used to determine the transfection rate. Fig. 1: Box 815) determines the scene-global information about the brightness value exhibited by the various cells in the fluorescence image (i.e., in this case, the context data is determined based on the input image itself; Box 810 is omitted). For example, a mean brightness value could be determined as exemplary brightness information. This is based on the understanding that two distinct cells where transfection was successful may still differ significantly in brightness in the fluorescence image. Differentiating between successful and unsuccessful transfection based on a size-limited receptive field can then yield inaccurate results. Based on this context data, as well as based on the fluorescence image (as the input image (compare Fig. 1: Box 805) and based on the cell instance segmentation, a count of the transfected and non-transfected cells is then performed using a CNN.

[0051] The second embodiment concerns cell counting or confluence estimation. These are routine tasks in most cell laboratories. They determine how many cells are present in a defined area of ​​a cell culture dish and the ratio between colonized and uncolonized areas within the dish. For example, this can be used to determine when dividing the cell culture is necessary or beneficial. The viability of the cell culture can be determined as a function of time. In a reference implementation, cell centers are identified in a phase-contrast light microscopy image. The detected centers are then counted to determine the number of cells in a defined area.Furthermore, semantic segmentation is subsequently performed to determine the ratio between the areas covered by cells and the areas not covered by cells. A challenge with such a reference implementation is that there is typically a large variance in the appearance of cells in the microscopic images across different cell lines or cell types. For example, some cell types produce a sharp contrast, while others are barely visible. Since the CNN used for localization or segmentation only sees a small portion of the entire image during initial convolution operations due to its size limitations, background structures in uncovered areas and poorly visible cells can be confused.This is particularly the case when the image area captured by a receptive field contains only background structures but no cells (compare . Fig. 2) In such a case, the CNN has no information about the cell types present in the sample and therefore cannot distinguish background artifacts or background structures from faintly visible cells, or can only do so poorly. By taking the contextual data into account (compare Fig. 3) That is, if information about the appearance of the cells or, in particular, about differences in the appearance of the cells compared to the background structures is available (e.g., implicit information about the cell type), the CNN can better distinguish the background structures from cells of the present cell type - even in areas where the image area captured by the receptive field shows only background structures.

[0052] Fig.Figure 4 schematically illustrates an electronic data processing device 80. This device comprises a processor 81, a memory 82, and a communication interface 83. For example, image data can be loaded via the communication interface 83. It would be conceivable that a camera of a microscopy system could be controlled via the communication interface 83 to acquire such image data. The processor 81 is configured to load and execute program code from the memory 82. When the processor 81 executes the program code, this causes the processor to perform techniques as described herein, such as: receiving image data, for example, via the communication interface 83; executing a context module to determine context data for a scene with cells; executing a CNN to determine result data for a scene with cells, etc.

[0053] In summary, the techniques described here are based on the understanding that, despite their performance in segmentation and image-to-image tasks, CNNs have certain limitations. One disadvantage of CNNs is that they only consider a local region around a given position in an image, the so-called "receptive field." This leads to the loss of context located at other positions in the image. An example illustrating this limitation is cell counting in an input image. The background pattern in this input image has a periodic structure and can easily be mistaken for a cluster of cells. If a CNN-based cell-counting model has a portion of the image in its receptive field at a given position that contains only the background pattern, the CNN will interpret the background pattern as cells—resulting in false detections.Only when the receptive field also contains at least some real cells will false detections no longer occur in the background, since the model has a kind of "reference" in the image.

[0054] A context module is revealed that enables the CNN to provide global scene context, including cell information, alongside local image information for each image position during inference. Using this context module, a CNN-based cell counting model, for example, can identify the cell type in the image and avoid false detections in background areas. The context module allows for the determination of scene-global cell characteristics based on a contextual image. By processing the contextual image within the context module, contextual information crucial for cell detection can be extracted. This contextual data can then be used to improve the accuracy of CNNs in solving cell-specific tasks and minimize false detections.

[0055] Naturally, the features of the embodiments and aspects of the invention described above can be combined with one another. In particular, the features can be used not only in the combinations described, but also in other combinations or individually, without leaving the scope of the invention.

[0056] For example, aspects were described above where a machine-learned CNN is used to obtain cell-specific result data for the scene. However, the techniques described here, which consider context data during the inference run, can also be applied to other types of machine-learned networks, such as vision transformer-based architectures. While these are not as susceptible to the problem described above as CNNs, since the transformer layers already facilitate global interaction between sub-areas of the image, global features from a contextual image can still be extracted. This can be done, for example, by adding them as a separate token or by adding them to the embedding of existing tokens (as with position embedding).

[0057] For example, techniques have been described above for processing images depicting cells. These techniques can then also be applied to other domains. For instance, images can be processed that do not depict cells, but rather other small, repetitive structures that are easily mistaken for background structures. Examples include satellite images of residential areas, images of microplastic samples, examinations of yeast or sperm, images of mechanical test specimens, and so on. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited non-patent literature

[0000] Nature methods 16.1 (2019): 67-70

[0002]

Claims

[1] Computer-implemented method that includes: - Obtaining an image (805) depicting a scene with cells, - Processing (815) the image or another image that at least depicts the scene with the cells in a context module, to obtain context data indicative of one or more scene-global determined features of the cells, and - Processing (820) the image and context data in an inference run of a machine-learned convolutional network (115) to obtain cell-specific result data (106) for the scene, wherein the convolutional network has a receptive field in an input layer that is smaller than a mapping size of the scene in the image. [2] Method according to claim 1, wherein the further image, which is different from the image, is processed in the context module. [3] Method according to claim 2, wherein the further image depicts the scene and an environment of the scene. [4] Method according to claim 2 or 3, wherein the further image has a different contrast type than the image. [5] Method according to any one of claims 2 to 4, wherein the further image depicts the scene at a scale that is smaller than the scale at which the image depicts the scene. [6] Method according to any of the preceding claims, wherein the context data is obtained from the context module in scalar form. [7] Method according to any of the preceding claims, where the context module has machine-learned weights, where the context module is selected from the following group: Encoder branch of an autoencoder network; Vision transformer network; a Foundation model. [8] Method according to any of the preceding claims, where the context module includes machine-learned weights, where the machine-learned weights of the context module are trained separately from the machine-learned convolutional network. [9] Method according to any of the preceding claims, wherein the context module determines the context data based on an output of the machine-learned convolutional network in a preceding inference run. [10] Method according to any of the preceding claims, wherein the context module comprises an image-based dimensionality reduction algorithm. [11] Method according to one of the preceding claims, wherein the context data comprises several context data parts which are intended for different patches of the image or further image and which indicate a comparison of the one or more scene-global determined features. [12] Method according to claim 11, wherein the context data parts are indicative of a change in properties of the scene in the different patches of the image or the further image. [13] Method according to any of the preceding claims where the context data includes at least one latent feature vector, where the at least one latent feature vector is processed at a hidden layer of the machine-learned convolutional network. [14] Method according to any of the preceding claims, wherein the context data includes classification information for the cells. [15] Method according to claim 14, wherein the classification information comprises one or more of the following information elements: cell type; cell size; cell species, cell state; cell size information; cell brightness information; and / or cell density. [16] Method according to claim 14 or 15, wherein the classification information comprises human-interpretable information elements. [17] Method according to any of the preceding claims, wherein the context data includes scaling information for the cells and / or background structures in the scene. [18] Method according to any of the preceding claims, wherein the context data includes localization information for the cells and / or background structures of the scene. [19] Method according to any of the preceding claims, where the context data includes a context image that marks cells or parts of cells, where the context image is processed at an input layer or an output layer of the machine-learned convolutional network. [20] Method according to any of the preceding claims, wherein the machine-learned convolution network comprises no more than E6 machine-learned weights. [21] Method according to any of the preceding claims, where the cell size in the image (105) is at least two orders of magnitude smaller than the image size of the image (105), where the receptive field (107) is up to an order of magnitude larger than the image size of the cells. [22] Method according to any of the preceding claims, the scene features cells as well as comparably large background structures, where the context data are indicative of scene-global distinguishing features between the cells and the background structures. [23] Electronic data processing equipment comprising a processor and a memory, wherein the processor is configured to load and execute program code from memory, wherein the processor is configured to perform the following steps based on the execution of the program code: - Obtaining an image (805) depicting a scene with cells, - Processing (815) the image or another image that at least depicts the scene with the cells in a context module, to obtain context data indicative of one or more scene-global determined features of the cells, and - Processing (820) the image and context data in an inference run of a machine-learned convolutional network (115) to obtain cell-specific result data (106) for the scene, wherein the convolutional network has a receptive field in an input layer that is smaller than a mapping size of the scene in the image. [24] Electronic data processing device according to claim 23, wherein the processor is further configured to execute the method according to any one of claims 1 to 22 based on the execution of the program code.

Images