Method for detecting binding of patient sample autoantibodies to double stranded deoxyribonucleic acid

Abstract

The invention relates to a method for detecting the binding of autoantibodies of a patient sample to double-stranded deoxyribonucleic acid by fluorescence microscopy and by digital image processing using Tetranychus cells, the method comprising: providing a substrate having a plurality of Tetranychus cells; incubating the substrate with the patient sample potentially having autoantibodies; incubating the substrate with a secondary antibody, the secondary antibody being labeled with a fluorescent dye, preferably green; acquiring a fluorescence image of the substrate; identifying sub-images representing Tetranychus cells in the fluorescence image by a first pre-trained convolutional neural network; further processing at least a subset of the sub-images by a second pre-trained convolutional neural network to determine a binding measure indicative of the degree of binding of autoantibodies in the kinetoplast region of the Tetranychus cells of the sub-images; and determining a total binding measure regarding the binding of autoantibodies of the patient sample to double-stranded deoxyribonucleic acid based on the binding measures.

Description

Technical Field

[0001] This invention relates to a method and apparatus for detecting the binding of autoantibodies to double-stranded deoxyribonucleic acid (DNA) in patient samples using fluorescence microscopy and digital image processing in the presence of short-hymened worm cells. Background Technology

[0002] For example, the confirmation of autoantibodies against deoxyribonucleic acid (DNA) is crucial for diagnosing SLE (systemic lupus erythematosus). Here, it is essential to fundamentally distinguish between two types: antibodies against dsDNA and antibodies against single-stranded denatured DNA (ssDNA). Antibodies against dsDNA react with epitopes within the DNA deoxyribophosphate skeleton. In contrast, antibodies against ssDNA primarily bind to epitopes originating from purine and pyrimidine base regions. However, they can also recognize epitopes on the phosphate deoxyribophosphate skeleton. Anti-dsDNA antibodies are almost exclusively found in SLE. Their prevalence ranges from 20% to 90%, depending on the confirmatory method and disease activity. Anti-dsDNA antibodies are sometimes confirmed in patients with other autoimmune diseases and infections, and in rare cases in clinically healthy individuals. The latter will develop SLE within 5 years of the first confirmed anti-dsDNA diagnosis in 85% of cases. However, the absence of antibodies against dsDNA does not completely rule out SLE. SLE is a systemic autoimmune disease originating from connective tissue disorders. Diagnosis is guided by the 11 revised criteria of the American College of Rheumatology (ACR) in 1997. If four of the 11 criteria are met, there is an 80% to 90% certainty of diagnosis for SLE.

[0003] Indirect immunofluorescence is an in vitro assay used to identify human antibodies against dsDNA. For example, a so-called BIOCHIP smear containing *Biomyxobolus* can be used as a substrate. This substrate is then incubated with a diluted patient sample, for example. In the presence of a positive reaction, a specific antibody binds to the antigen. In a further incubation step, the bound antibody (IgG) is stained with, for example, an anti-human antibody labeled with fluorescein and can be seen under a fluorescence microscope.

[0004] Therefore, a substrate is known in the prior art in which multiple short-hymened worm cells are fixed. For example, this fixation can be achieved using ethanol.

[0005] The substrate is then incubated with a patient sample, preferably a diluted serum, which may potentially contain the autoantibodies to be detected. According to existing techniques, the cells or substrate can then be incubated with a so-called conjugate containing secondary antibodies labeled with, for example, a green fluorescent dye.

[0006] After irradiating the cultured substrate with excitation light, the fluorescence radiation emitted by the green fluorescent dye can be used as a microscopic image for fluorescence microscopy.

[0007] For example, such microscopic images are in Figure 1 It is shown as image SG.

[0008] Figure 2 The text then shows a more detailed version of the text. Figure 1 A single brevis cell CR. This sub-image TB of a brevis cell CR clearly shows staining on the basal body K, also known as mitochondria. There are additional stainings on the nucleus Z and basal body B.

[0009] To reliably detect the binding of autoantibodies to double-stranded DNA in patient samples, obtaining fluorescent images of the substrate and staining multiple dynamic matrix particles in the SG is crucial. For example... Figure 2 As shown, binding or staining may also exist for cell nucleus Z and matrix B, thus the dynamic matrix K must be reliably determined in terms of its position in the image through image processing.

[0010] By analyzing and evaluating the corresponding staining of the corresponding motile matrix of the corresponding short hymenorrhea cells in the fluorescent image SG, the average total binding of autoantibodies to double-stranded DNA in the patient sample can be determined overall. Summary of the Invention

[0011] Therefore, the objective is to provide a method for determining the binding of autoantibodies to double-stranded deoxyribonucleic acid in patient samples using digital image processing, wherein staining of different kinematic regions of different short-hymened worm cells within a fluorescent image can be reliably determined.

[0012] Therefore, a method is proposed for detecting the binding of autoantibodies to double-stranded deoxyribonucleic acid (DDNA) in patient samples using fluorescence microscopy and digital image processing with the use of *Bifidobacterium brevis* cells. The method comprises several steps. First, a substrate containing multiple *Bifidobacterium brevis* cells is provided; then, the substrate is incubated with a patient sample potentially containing autoantibodies. Furthermore, the substrate is incubated with secondary antibodies, each labeled with a fluorescent dye, preferably a green fluorescent dye. A fluorescence image of the substrate is acquired in a color channel corresponding to the fluorescent dye, preferably a green channel. Furthermore, corresponding sub-images are identified within the acquired fluorescence image. Each sub-image represents a *Bifidobacterium brevis* cell. This identification is achieved via a pre-trained first convolutional neural network. Furthermore, at least a subset of the corresponding sub-images is processed by a pre-trained second convolutional neural network to determine a corresponding binding metric, indicating the degree of binding of the autoantibody in the corresponding motile matrix region of the corresponding *Bifidobacterium brevis* cell in the corresponding sub-image. Finally, a total binding metric for the binding of autoantibodies to DDNA in the patient sample is determined based on the corresponding binding metric.

[0013] The patient sample is a liquid patient sample, preferably liquid blood or a component of liquid blood. In particular, the liquid patient sample is liquid serum. Preferably, the patient sample is diluted with a so-called washing buffer, preferably a so-called PBS Tween.

[0014] The conjugate contains a secondary antibody labeled with a fluorescent dye.

[0015] The proposed method is particularly useful for determining the presence of primary antibodies in liquid patient samples in vitro.

[0016] The substrate, particularly biological samples of animal pathogenic blood flagellates of the short-hymened type, is important. These single-celled organisms possess giant mitochondria (kinetic matrix) containing double-stranded DNA, which are essentially devoid of the remaining antigens found in the cell nucleus. Primary autoantibodies from patient samples react with the kinetic matrix against dsDNA.

[0017] Fluorescence microscopy, especially so-called indirect immunofluorescence microscopy (IIFT microscopy).

[0018] Preferably, the substrate is irradiated with excitation radiation to excite the fluorescence of the first and second fluorescent dyes. The excitation radiation is preferably blue light and therefore has a wavelength in the blue spectrum. The fluorescent dye is preferably a green fluorescent dye, especially a fluorescent dye of the fluorescein isothiocyanate (FITC) type.

[0019] To illustrate one or more possible advantages of the method according to the present invention, a more detailed explanation follows.

[0020] Figure 5 A device V1 is shown, by which the method according to the invention can be performed. Device V1 may be referred to as a fluorescence microscope. Device V1 includes a holder H for a substrate S, which has been cultured as described above. Excitation light AL of an excitation source LQ is directed to the substrate S via an optical device O. The resulting fluorescence radiation FL is then transmitted back through the optical device O and passes through a dichroic mirror SP1 and an optional optical filter F2. Preferably, the fluorescence radiation FL passes through an optical filter FG, which filters out the green channel. Camera K1 is preferably a monochrome camera, which then acquires the fluorescence radiation FL in the green channel in the presence of the optical filter FG. According to an alternative embodiment, camera K1 is a color camera, which is sufficient without using the optical filter FG and acquires a fluorescence image in the corresponding color channel, which is the green channel, via a Bayer matrix. Camera K1 provides image information BI or a fluorescence image to a computing unit R, which processes the image information BI.

[0021] Figure 1 An exemplary fluorescence image SG in the green channel is shown.

[0022] According to the present invention, sub-images, particularly those containing at least one breech-hymenorrhiza cell, are first identified based on the fluorescence image SG. In this regard, Figure 3 The corresponding sub-images marked with rectangles are shown, specifically sub-images TB1, TBA, TBB, and TBX. The specific sub-image region TB1 is indicated in particular.

[0023] In the context of this application, a sub-image region can also be referred to as a sub-image.

[0024] Such an exemplary sub-image TB, i.e. Figure 1 Sub-image TB1 in Figure 2 As shown in the image. Figure 2 As observed, due to antibody binding, up to three significantly stained regions can be observed in *Bifidobacterium brevis* cells: the K region of the kinematic matrix, the B region of the matrix, and the Z region of the nucleus. Therefore, it is essential to ensure that no significant staining is used as an analytical criterion in the fluorescence image SG, and that staining of the kinematic matrix region is used solely as the analytical criterion.

[0025] Figure 14a Further steps of the method according to the invention are shown, wherein, for a specific sub-image TB1, a matching binding metric IBM1 is determined by a second convolutional neural network CNN2. The binding metric IBM1 indicates the degree of binding of autoantibodies in the dynamic matrix region of the short-hymened worm cells shown in the sub-image.

[0026] Figure 14bFurther steps of the method according to the invention are shown in ERS, wherein the total binding measure GBM of the autoantibody in the patient sample in the substrate to be detected is determined based on different corresponding binding measures IBM1, IBM2 for different corresponding sub-image regions.

[0027] Therefore, the objective of the method according to the invention is to determine the binding of primary autoantibodies to dsDNA in the corresponding dynamic matrix region of the corresponding short hymenorrhea cell by determining the corresponding staining of the corresponding dynamic matrix region via fluorescent dye in the corresponding sub-image region.

[0028] It should be noted that, in the method according to the present invention, Figure 1 The entire fluorescence image SG is not simply fed to a single convolutional neural network for the overall identification of multiple stained moving matrix regions. Rather, this invention is explicitly different from methods that perform overall classification of the entire fluorescence image SG using a single convolutional neural network. Specifically, according to the invention, preprocessing is first performed such that sub-image regions or sub-images are identified in the fluorescence image by means of a first convolutional neural network. For example... Figure 2 As can be seen, due to antibody binding, there can be up to three significantly stained regions in *Brachymena* cells: the kinematic matrix K, the matrix B, and the nucleus Z. Images identified by the first convolutional neural network are then processed separately by a second convolutional neural network (CNN2) to determine corresponding sub-images representing the kinematic matrix regions for each sub-image, and to determine, in particular, a corresponding binding metric based on these sub-images. This binding metric indicates the degree of binding of autoantibodies in the corresponding kinematic matrix region of the corresponding *Brachymena* cell in the corresponding sub-image. In other words, these identified sub-image regions are processed separately by the second convolutional neural network (CNN2) to identify corresponding sub-images for each sub-image, and then a corresponding binding metric is determined based on the corresponding sub-image. This binding metric indicates the degree of binding of autoantibodies in the corresponding kinematic matrix region of the corresponding *Brachymena* cell in the corresponding sub-image. The total binding metric is then determined based on these respective individual binding metrics of the corresponding sub-images.

[0029] According to the invention, a pre-trained first convolutional neural network identifies sub-images in the fluorescence image. This is advantageous because the first convolutional neural network is not used for the overall task of determining the total binding metric of autoantibodies to double-stranded deoxyribonucleic acid in a patient sample, but only for the sub-task of identifying sub-images representing breech-lymphoid cells. Therefore, the first convolutional neural network only needs to be pre-trained to identify or locate relevant sub-images in the fluorescence image. According to the invention, the identified sub-images are then processed separately by a second convolutional neural network to determine a separate binding metric of the autoantibody binding to the corresponding motile matrix region of the corresponding breech-lymphoid cell in each sub-image. Thus, the second convolutional neural network can be pre-trained such that it focuses on or is designed for the task of identifying motile matrix regions in sub-images, particularly those containing a single breech-lymphoid cell. Specifically, instead of incorporating the entire image information of the fluorescence image into a single convolutional neural network, separate convolutional neural networks are used for the separate sub-tasks of identifying the aforementioned sub-images and determining the corresponding separate binding metric of the corresponding sub-image. Therefore, these separate convolutional neural networks can be designed and pre-trained to reliably perform the corresponding tasks. In contrast, the design or training of a single convolutional neural network to process the entire fluorescence image in order to determine the total binding metric must process an extremely large amount of image information to meet the purpose of determining the total binding metric. Therefore, such a single convolutional neural network may be prone to errors.

[0030] The proposed method is further advantageous because, in a fluorescence image of a preferred color channel (green channel), the first convolutional neural network directly identifies or locates the sub-image within the fluorescence image. Therefore, according to the invention, it is unnecessary to prepare the substrate using a so-called counterstaining with another fluorescent dye, such as Evans Blue, to produce red staining in the red channel; rather, this counterstaining for the red channel can be completely eliminated according to the invention. This is achieved by designing or pre-training the first convolutional neural network for the step or task of recognizing sub-images representing breech-lymphoid cells.

[0031] In other words, the method according to the invention achieves high accuracy in locating the moving substrate region or in detecting discoloration in the moving substrate region by means of the manner described herein. If... Figure 1The entire fluorescent color image SG is fed to a single convolutional neural network to detect discoloration of a single basal body or different basal body regions of different breech-hymenorrhea cells. In this case, other regions, such as the basal body region or the nucleus region, may be incorrectly identified as basal body regions, potentially leading to errors in determining the binding of primary autoantibodies to dsDNA. Furthermore, the entire fluorescent image SG contains only background regions in approximately 90% of its area. Therefore, when processing the entire fluorescent image SG with a single CNN, image pixels belonging to the background regions must also be processed, making the single CNN very resource-intensive and inefficient. Additionally, the single CNN must further identify subordinate regions representing the basal body from the entire fluorescent image SG, which is resource-intensive because the nucleus or basal body regions may be stained. Specifically, in cases where the basal body region and / or the nucleus region is stained, staining of the basal body region may not be present. Due to the different locations of breech-hymenorrhea cells within the basal body or fluorescent color image SG, such as… Figure 1 As shown, there are too many degrees of freedom to reliably classify a single image segment as a moving matrix using a single convolutional neural network.

[0032] Instead of processing the entire fluorescence image by a single CNN, according to the present invention, a second convolutional neural network CNN2 processes each sub-image individually. The sub-image's position relative to the entire fluorescence image is determined by a first convolutional neural network CNN1 based on an analysis and evaluation of the fluorescence image. This allows the second convolutional neural network CNN2 to be explicitly trained for individual sub-images, i.e., for sub-images containing a single *Briohymena* cell. The second convolutional neural network CNN2 then only needs to detect the position of the moving matrix within that sub-image. This position can preferably be determined as a so-called sub-image. The binding metric of the cell or the moving matrix can then be determined by the second convolutional neural network CNN2 with respect to each moving matrix.

[0033] In other words, according to the present invention, this can be achieved by feeding only specific, individual sub-image regions into a second convolutional neural network (CNN2): the second CNN2 is limited to analyzing such sub-images or individual illustrations of a single brevichallomys cell to identify the dynamic matrix and determine the binding quantification of the autoantibody. If the entire fluorescent color image, such as... Figure 1If an image SG is fed to a single convolutional neural network (CNN) for a classification task involving the detection of different moving bodies during the training phase, then this single CNN must be designed with a very large number of degrees of freedom, and training such a single CNN would be very costly and inefficient. Since the second convolutional neural network CNN2 according to the invention only needs to process each individual sub-image region individually in order to determine the corresponding sigmata, the second convolutional neural network CNN2 can be limited to processing images representing breech-lymphatic cells, such as... Figure 2 The image TB is shown in the middle.

[0034] Advantageous embodiments of the invention are described in detail below with reference in part to the accompanying drawings.

[0035] Preferably, the identification of corresponding sub-images in the fluorescence image is achieved by assigning corresponding image segments of the fluorescence image to corresponding image segment categories in an image segment category group through a pre-trained first convolutional neural network, wherein the image segment category group includes at least the following image segment categories: cells and background.

[0036] Preferably, the image fragment category group includes at least the following image fragment categories: cells, cell edges, and background. Preferably, the category "cells" can also be referred to as the category "cell body".

[0037] Preferably, for a given sub-image, the method includes the following steps: selecting a corresponding sub-image of the given sub-image, wherein the corresponding sub-image represents a corresponding kinematic matrix region of the given short-hymened worm cell; and further determining a corresponding binding measure based on the corresponding sub-image. The method preferably further includes: determining a total binding measure based on the corresponding binding measure.

[0038] Preferably, the method includes the following steps: determining at least one corresponding final feature map of a corresponding sub-image using a second convolutional neural network; further determining a corresponding confidence measure regarding the presence of binding of autoantibodies in a corresponding dynamic matrix region for the corresponding sub-image or the corresponding final feature map, particularly based on the corresponding final feature map; further selecting a subset of sub-images or a subset of corresponding final feature maps based on the determined confidence measure; and further processing the corresponding feature map of the selected sub-images accordingly to determine the corresponding binding measure. The method preferably further includes: determining a total binding measure based on the corresponding binding measure.

[0039] Preferably, for a corresponding sub-image from the selected subset, the method includes the steps of: selecting a corresponding sub-image of the corresponding sub-image based on a corresponding final feature map corresponding to the corresponding sub-image, wherein the corresponding sub-image represents a corresponding kinematic matrix region of the corresponding breech-hymenorrhiza cell, and further determining a corresponding binding metric based on the corresponding sub-image. The method preferably further includes: determining a total binding metric based on the corresponding binding metric.

[0040] Preferably, for a corresponding sub-image from the selected subset, the method further includes the steps of: obtaining a corresponding masking operator based on the corresponding final feature map; selecting a corresponding sub-image of the corresponding sub-image by applying the corresponding masking operator to the corresponding sub-image; and determining a corresponding binding metric based on the corresponding sub-image. The method preferably further includes: determining a total binding metric based on the corresponding binding metric.

[0041] Preferably, the method is designed such that during the processing of a sub-image, the second convolutional neural network generates a first set of synthetic feature maps based on the sub-image through at least one first convolutional layer in a first processing level; furthermore, in a second processing level, a second set of synthetic feature maps is generated based on the first set of two-dimensional feature maps through at least one second convolutional layer; and a third set of synthetic feature maps is also generated based on the second set of two-dimensional feature maps through at least one third convolutional layer, wherein the second set has fewer synthetic feature maps than the first set, and the third set has more synthetic feature maps than the second set.

[0042] Preferably, the method is designed such that the second and third convolutional layers follow each other as sub-steps of the sequential processing path, wherein, in the second processing level, there exists another processing path in parallel with the sequential processing path, in which the second convolutional neural network generates a fourth set of synthetic feature maps based on a first set of two-dimensional feature maps through at least one fourth convolutional layer, wherein, in addition, the second convolutional neural network generates a final feature map corresponding to the sub-image based on the third and fourth sets of synthetic feature maps, and furthermore, the number of convolutional layers that follow each other in the parallel processing path is less than the number of convolutional layers that follow each other in the sequential processing path.

[0043] Preferably, the method includes the following steps: acquiring a temporary first fluorescence image in a color channel using fixed preset acquisition parameters; determining whether the brightness of the temporary first fluorescence image in the color channel exceeds the maximum brightness; if the temporary first fluorescence image in the color channel does not exceed the maximum brightness, using the temporary first fluorescence image as a fluorescence image; if the temporary first fluorescence image in the color channel exceeds the maximum brightness, acquiring a temporary second fluorescence image in the color channel and using the temporary second fluorescence image in the color channel as a fluorescence image.

[0044] Furthermore, an apparatus according to the invention is proposed for detecting the binding of autoantibodies to double-stranded deoxyribonucleic acid (DDNA) in a patient sample using fluorescence microscopy and digital image processing in the presence of *Biomyxomorphus* cells. The apparatus includes: a substrate holding device having a plurality of *Biomyxomorphus* cells, and the substrate being cultured with a patient sample containing autoantibodies and secondary antibodies, each labeled with a fluorescent dye. The apparatus further includes at least one image acquisition unit for acquiring a fluorescence image of the substrate. The apparatus also includes at least one computational unit designed to identify corresponding sub-images in the fluorescence image using a pre-trained first convolutional neural network, each sub-image representing at least one *Biomyxomorphus* cell; furthermore, to process at least one subset of each corresponding sub-image individually using a pre-trained second convolutional neural network to determine a corresponding binding metric, the binding metric indicating the degree of binding of the autoantibody in a corresponding dynamic matrix region of the corresponding *Biomyxomorphus* cell in the corresponding sub-image; and furthermore, to determine the total binding metric of the patient sample's autoantibodies to DDNA based on the corresponding binding metric.

[0045] A computational unit is also proposed, which is designed to: receive a fluorescence image during digital image processing, the fluorescence image representing a substrate stained with a fluorescent dye, the substrate having multiple short-hymened worm cells; identify corresponding sub-images in the fluorescence image using a pre-trained first convolutional neural network, each sub-image representing at least one short-hymened worm cell; process at least one subset of each corresponding sub-image individually using a pre-trained second convolutional neural network to determine a corresponding binding metric, the binding metric indicating the degree of binding of autoantibodies in the corresponding kinetic matrix region of the corresponding short-hymened worm cell in the corresponding sub-image; and further determine the total binding metric of autoantibodies to double-stranded deoxyribonucleic acid in the patient sample based on the corresponding binding metric.

[0046] Furthermore, a data network device is proposed, comprising at least one data interface for receiving a fluorescence image representing a substrate stained with a fluorescent dye, the substrate having multiple short-hymened worm cells. The data network device also includes at least one computing unit designed to, during digital image processing, identify corresponding sub-images in the fluorescence image using a pre-trained first convolutional neural network, each sub-image representing at least one short-hymened worm cell; furthermore, to process at least one subset of each corresponding sub-image individually using a pre-trained second convolutional neural network to determine a corresponding binding metric, the binding metric indicating the degree of binding of autoantibodies in the corresponding mobile matrix region of the corresponding short-hymened worm cell in the corresponding sub-image; and furthermore, to determine the total binding metric of autoantibodies to double-stranded deoxyribonucleic acid in the patient sample based on the corresponding binding metric.

[0047] A method for digital image processing is also proposed, comprising the following steps: receiving a fluorescence image representing a substrate stained with a fluorescent dye, the substrate having multiple short-hymened worm cells; identifying corresponding sub-images in the fluorescence image using a pre-trained first convolutional neural network, each sub-image representing at least one short-hymened worm cell; further processing at least one subset of each corresponding sub-image individually using a pre-trained second convolutional neural network to determine a corresponding binding metric, the binding metric indicating the degree of binding of autoantibodies in the corresponding kinetic matrix region of the corresponding short-hymened worm cell in the corresponding sub-image; and further determining the total binding metric of autoantibodies to double-stranded deoxyribonucleic acid in a patient sample based on the corresponding binding metric.

[0048] Furthermore, a computer program product comprising instructions that, when executed by a computer, cause the computer to perform a method for digital image processing according to the present invention.

[0049] In addition, a data carrier signal for transmitting the proposed computer program product is also proposed. Attached Figure Description

[0050] Without limiting the general concept of the invention, the invention will now be described in detail with reference to the accompanying drawings and specific embodiments. In the drawings:

[0051] Figure 1 The image shows a fluorescence image of a matrix containing multiple short-hymened worm cells in the color channel;

[0052] Figure 2 Show Figure 1 Sub-images of fluorescent images containing short-hymened worm cells;

[0053] Figure 3A fluorescence image of a first color channel with a corresponding identified first sub-image region is shown;

[0054] Figure 4 The image shows a fluorescence image, in which the fluorescence is indicated from... Figure 3 The selected sub-image from the total set of sub-images;

[0055] Figure 5 An embodiment of the device according to the invention is shown;

[0056] Figure 6 The binary image determined for the fluorescence image is shown as binary information or a binary mask;

[0057] Figure 7a Show Figure 3 The classification results for specific local areas of the image;

[0058] Figure 7b The final classification results for the first type are shown;

[0059] Figure 7c The final classification results for the second type are shown;

[0060] Figure 8a An example sub-image is shown;

[0061] Figure 8b This shows the first final feature mapping;

[0062] Figure 8c The preferred second final feature map to be determined is shown;

[0063] Figure 9a Show Figure 8b The interpolated version of the first final feature map in the dataset;

[0064] Figure 9b Showing from Figure 9a The binary value mask operator derived from the interpolation feature map in the model;

[0065] Figure 9c Showing from Figure 8a The sub-image region selected from the sub-image regions in the sub-image region;

[0066] Figure 10 The steps for processing a fluorescence image via a first convolutional neural network are shown.

[0067] Figure 11 The detailed steps of processing via the first convolutional neural network are shown;

[0068] Figure 12a An exemplary illustration of a fluorescence image is shown;

[0069] Figure 12b An exemplary diagram of the first classification matrix or the first classification mapping is shown;

[0070] Figure 13 The steps for performing the method according to the invention are shown;

[0071] Figure 14a This illustrates the basic principle of binding metric used to determine the binding of primary autoantibodies in patient samples to the dynamic matrix region of a sub-image region;

[0072] Figure 14b The processing of multiple binding measures is shown to determine the total binding measure;

[0073] Figure 15 The steps for determining a sub-image region in a first fluorescent color image of a color channel are shown;

[0074] Figure 16a An exemplary description of a preferred implementation of the second convolutional neural network is shown;

[0075] Figure 16b This illustrates the sub-steps of the first type of convolutional layer in the second convolutional neural network;

[0076] Figure 16c This illustrates another sub-step of a convolutional layer in a second convolutional neural network;

[0077] Figure 17 The processing steps are shown for determining a confidence metric about the presence of binding between an autoantibody and the dynamic matrix region for a corresponding sub-image based on matched feature maps;

[0078] Figure 18 The steps for selecting a subset of the identified sub-images based on a confidence metric of the corresponding sub-image are shown.

[0079] Figure 19 The diagram illustrates the steps for selecting the corresponding sub-image of the corresponding sub-image and determining the corresponding fusion metric based on the corresponding sub-image by applying a binary value masking operator to the corresponding sub-image.

[0080] Figure 20 An exemplary illustration of a computing unit according to the present invention is shown;

[0081] Figure 21 An exemplary illustration of a data network apparatus according to the present invention is shown;

[0082] Figure 22 Exemplary illustrations are shown of a computer program product according to the present invention and a data carrier signal according to the present invention;

[0083] Figure 23 An example of the first processing layer of CNN2 is shown;

[0084] Figure 24 An embodiment of the first part of the second processing layer of CNN2 is shown;

[0085] Figure 25 An embodiment of the second part of the second processing layer of CNN2 is shown;

[0086] Figure 26 An example of the third processing layer of CNN2 is shown;

[0087] Figure 27 Shows the pixel values ​​of the sub-image;

[0088] Figure 28 The corresponding confidence measure values ​​for the corresponding sub-images of the fluorescence image are shown;

[0089] Figure 29 The steps of a method for acquiring a fluorescence image according to a preferred embodiment are shown; and

[0090] Figures 30 to 33 illustrate the sub-elements of an implementation of the second convolutional neural network. Detailed Implementation

[0091] Figure 13 Different steps for performing the method according to the invention are illustrated. In step S1, a substrate having a plurality of *Brachymene* cells is provided. In step S2, the substrate is cultured together with a patient sample. The patient sample potentially contains autoantibodies. In common steps S3 and S4, the substrate is cultured together with secondary antibodies, each labeled with a preferred green fluorescent dye. In step S5, a fluorescence image of the substrate is acquired in a color channel, preferably a green channel, and this color channel corresponds to the fluorescent dye. In step S6, corresponding sub-images, each representing *Brachymene* cells, are identified in the fluorescence image using a pre-trained first convolutional neural network.

[0092] In step S7, at least a subset of the corresponding sub-images is processed by a pre-trained second convolutional neural network to determine a corresponding binding metric, which indicates the degree of binding of autoantibodies in the corresponding kinetic matrix region of the corresponding short-hymened worm cell in the corresponding sub-image. In step S8, based on the corresponding binding metrics of the corresponding sub-images, a total binding metric is determined regarding the binding of autoantibodies from the patient sample to double-stranded DNA in the matrix. In step S9, the total binding metric is preferably provided and alternatively or additionally output and / or displayed.

[0093] Figure 14aThis illustrates an exemplary individual processing of a single sub-image region TB1 via a second convolutional neural network CNN2 to determine the individual binding metric IBM1. The binding metric IBM1 indicates the individual binding degree of autoantibodies in a single kinetic matrix region of a single short-hymened cell within a single sub-image region TB1.

[0094] Figure 14b This illustrates the determination of the total binding measure GBM through the step ERS (which corresponds to...). Figure 13 Step S8) is used to determine the total binding measure GBM based on the corresponding binding measures IBM1 and IBM2 of the corresponding sub-images.

[0095] Figure 5 An embodiment of the device V1 according to the invention is shown. Device V1 includes a holding device H for a substrate S. Excitation light AL from an excitation source LQ is pre-filtered via an optical filter F1 and then guided to the substrate via an optical device O using a dichroic mirror SP1. The resulting fluorescent radiation, or the resulting fluorescent light FL, then returns from the substrate via an objective lens O, through the dichroic mirror SP1, and through a closed filter F2. The optical filter F2 filters out the wavelength of the excitation radiation or excitation light AL. The fluorescent light FL is then supplied to at least one image acquisition unit in the form of a camera K1. The fluorescent light FL is filtered in a first color channel via a preferred optical filter FG, such that the fluorescent light FL represents the so-called green channel after passing through the optical filter FG. The image acquisition unit K1 is preferably a monochrome camera. In an alternative embodiment, there is no optical filter FG and the camera K1 is a color camera that uses, for example, a Bayer matrix to filter out a green image or a green channel image and acquires a fluorescent image as the preferred green image. Device V1 can provide result data ED indicating a total binding metric via a data interface DS1.

[0096] The computing unit R is designed to receive fluorescence images in the form of digital data BI. Furthermore, the computing unit R is designed to perform steps S5 to S9 of the method according to the invention.

[0097] The computing unit R according to the present invention can also be as follows: Figure 20 The implementation is shown. Here, the computing unit R receives a fluorescence image in the form of at least one data signal SI via at least one data interface DS2. During execution... Figure 13Following steps S5 to S8 of the method according to the present invention, the calculation unit R determines the total binding measure of the autoantibody in the patient sample to the double-stranded DNA. Preferably, the calculation unit R includes an output interface AS leading to the display unit AE, through which the total binding measure can be output or displayed. Preferably, the calculation unit R includes an additional data interface DS3 leading to a data network, through which the calculation unit provides the total binding measure via a data signal SI3. The data interfaces DS2 and DS3 of the calculation unit R can also be a common data interface. The data interfaces DS2 and DS3 are preferably network data interfaces.

[0098] The computing unit R can also be part of the data network device DV according to the present invention, such as Figure 21 As shown. The data network device DV includes at least one data interface DS4 for receiving a fluorescence image BI via at least one data signal SI1. The data network device DV includes a computing unit R, which is preferably connected to a storage unit MEM and the data interface DS4 via an internal data bus IDB. The computing unit R is configured as previously described. Figure 20 The design is as described. The data network device DV can be a single computer or a so-called cloud solution. The data network device DV therefore executes the method for digital image processing according to the invention via the computing unit R, in which a fluorescence image BI is received and in which the computing unit R performs... Figure 13 Steps S5 to S8 in the process.

[0099] Figure 22 A computer program product CPP according to the invention is shown, which includes instructions that cause the computer to execute the digital image processing method according to the invention when the program is executed by the computer CO.

[0100] The computer program product CPP can be provided in the form of a data carrier signal SI2 and received by the computer CO through the data interface DSX located on the computer CO. The data carrier signal SI2 thus transmits the computer program product CPP.

[0101] Figure 15 The process of the fluorescence image SG in step S6 is illustrated to obtain sub-image data TBD, which represents or indicates a sub-image. Figure 3 A fluorescence image SG is shown, which includes indicated sub-images TB1, TBA, TBB, and TBX, recorded in the fluorescence image SG by corresponding rectangles. Other sub-images, not mentioned in detail, are recorded in the fluorescence image SG by other rectangles. Figure 4 The fluorescent image SG is shown with specially selected sub-images. For example, Figure 3The subimage TBX in the dataset was not selected to determine the subset of subimages because... Figure 3 The sub-image TBX mentioned in the text is no longer in Figure 4 The instructions are in the middle. This step of selecting a subset of sub-images will be explained in detail later.

[0102] Figure 15 Step S6 shown in Figure 10 The details are elaborated below. Step S6 is performed by a first convolutional neural network CNN1, which preferably includes a convolutional neural network CNNA and a so-called post-processing PP1 as corresponding sub-steps. The sub-image data TBD obtained by means of the pre-trained first convolutional neural network CNN1 includes the identified sub-image, such as sub-image TB1 or an indication of the sub-image.

[0103] Figure 11 Further details of the first convolutional neural network, CNN1, are shown.

[0104] The convolutional neural network CNNA used as the first sub-step in the first convolutional neural network CNN1 will be described in detail later with reference to Figures 30 to 33.

[0105] The convolutional neural network (CNNA) receives a fluorescent image SG and assigns corresponding image segments or pixels of the fluorescent image SG to corresponding image segment categories. These image segment categories form a set of image segment categories, which includes at least the image segment categories of cells and background. Preferably, the set of image segment categories includes the image segment categories of cells, cell edges, and background.

[0106] Fluorescence image SG can be understood as information

[0107]

[0108] Wherein, the index m is a pixel index or image segment index, and wherein, at a resolution of 1024x1024, the parameter M produces a value of M=1048576. Those skilled in the art will understand that different values ​​for the parameter M are produced when different image resolutions are selected for the fluorescence image SG. The fluorescence image can also be scaled down.

[0109] The convolutional neural network (CNNA) then calculates the corresponding matrices or mappings MA1, MA2, MA3 for each image segment category. For a preferred set of K=3 image segment categories, a mapping or matrix MA1, MA2, MA3 can therefore be described as...

[0110]

[0111] The resolution of the mapping is preferably the same as that of the fluorescence image SG, therefore it is suitable for...

[0112]

[0113] Among them, the preferred value is M=1048567.

[0114] to this end, Figure 12a An exemplary illustration of a fluorescence image SG is shown, wherein the first pixel SG (1) and the last pixel SG (M) are indicated.

[0115] to this end, Figure 12b The mapping or matrix MA1 for the first image fragment category is shown in a corresponding manner. Then, corresponding mappings or matrices MA2, MA3 can be set for other image fragment categories.

[0116] In a single entry MA1(1), the mapping MA1 has a confidence metric that indicates that the corresponding image fragment or pixel SG(1) belongs to the first image fragment category of the mapping or matrix MA1.

[0117] Accordingly, this can be set for the mapping or matrix MA2, MA3 of the relevant second or third image segment category.

[0118] Also there Figure 11 The mappings or matrices MA1, MA2, MA3 shown in the diagram then undergo a so-called Argmax operation in the further step AM, which determines which image fragment category the pixel or image fragment is most likely to belong to for the corresponding image fragment from the fluorescence image SG.

[0119] This is accomplished by forming a classification matrix MX, where,

[0120]

[0121] Then, to determine the classification matrix MX, the Argmax operation is performed according to the following formula.

[0122]

[0123] The values ​​of the classification matrix MX, or a single value MX(m), come from the set {1,2,3}, where, for example, 1 indicates that the corresponding image fragment or pixel from the fluorescence image SG belongs to the category: background. For example, value 2 indicates that the corresponding pixel or image fragment from the fluorescence image SG belongs to the cell edge. Preferably, value 3 indicates that the corresponding pixel or image fragment from the fluorescence image belongs to the category: cell or cell body.

[0124] Then, preferably, a so-called mapping step MS1 is performed, which transforms the classification matrix MX into a binary classification matrix BMX, which can also be understood as binary image information BSG. The result of the binary image information BSG is as follows. Figure 6 As shown. This binary image information, BSG or BMX, can also be described as...

[0125]

[0126] The mapping in mapping step MS1 is preferably performed such that all image segments or pixels in matrix MX that are associated with the background, i.e., belong to category 1 and have index m, are also ultimately assigned to the background and set to a value of 0 in binary matrix BMX. All image segments or pixels assigned to cell edges in matrix MX, i.e., belong to category 2, are also ultimately assigned to the background and set to a value of 0 in binary matrix BMX. All image segments or pixels assigned to cells or cell bodies in matrix MX, i.e., belong to category 3, are ultimately assigned to cell regions and set to a value of 1 in binary matrix BMX. This is performed according to the following rules:

[0127]

[0128] Figure 7a against Figure 6 The fully specific cell ZBA and ZBB values ​​show the classification results of the mapping matrix MX before mapping step MS1. Pixels assigned to the background are shown in black. Pixels assigned to cell edges are shown in gray as ZRA. Pixels assigned to cell bodies or cells are shown in white as ZEA. This is also... Figure 7a The cell ZBB is classified by pixels or image fragments and described by information ZEB and ZRB. Due to the mapping step MS1, the cell ZBA and ZBB are advantageously separated, thus obtaining... Figure 7c The final information allows the two cells ZBA and ZBB to be separated and now identified as regions E1 and E2. If the mapping is performed incorrectly, for example, because cell edges ZRA and ZRB are assigned to cell regions or cell bodies ZEA and ZEB, it is possible that the two cells ZBA and ZBB are interpreted as a single, unified region, which... Figure 7b The region shown is the region with subregions E11 and E21.

[0129] Figure 6 The binary image BSG in the middle Figure 11Further post-processing steps in PX are performed to determine the sub-image and provide it as sub-image data TBD. The steps of post-processing PX are now discussed in more detail. First, post-processing is performed using the method described by Satoshi Suzuki et al., “Topological structural analysis of digitized binary images by border following” (Computer Vision, Graphics, and Image Processing, 30(1):32-46, 1985), to find the relevant contours in the image BSG. Then, the so-called bounding box is post-processed using the function boundingRect from the “Structural analysis and shapedescriptors” section of the OpenCV database (https: / / opencv.org). Here, the size of the bounding box is 150×150 pixels.

[0130] Preferably, it can be analyzed and evaluated. Figure 1 The quality of the fluorescence image SG in the image is determined by the following method: Figure 6 Following the contours in the binary image BSG, at least ten identified cell regions or sub-images must possess a specific morphology and a certain minimum size. The fluorescence image is further processed only if at least ten cells meet both criteria, without outputting an error in the method. If fewer than ten identified cell regions or sub-images do not meet these criteria, the method is preferably aborted and an error notification is output.

[0131] Figures 30, 31, 32, and 33 show the sub-elements CNNA1, CNNA2, CNNA3, and CNNA4, which are considered together in a corresponding order to form... Figure 10 Or, consider the convolutional neural network CNNA in Figure 11. According to Figure 30, the fluorescence image SG is received by a portion of the convolutional neural network, CNNA1. It is evident from Figures 30 to 33 that the entire convolutional neural network CNNA consists of a sequence of multiple steps or processing operations within the neural network, where different types of steps occur. The first step, INL, is the so-called input layer, where the fluorescence image SG, for example, has dimensions of 1024 × 1024, is received. Figure 1 The fluorescence image SG in the image preferably has a first dimension or resolution higher than 1024×1024, and is reduced or rescaled to a second dimension or resolution of 1024×1024 for preprocessing purposes. Preferably, the brightness value of the fluorescence image SG is scaled to a range of 0 to 1 in the scaling step SKL.

[0132] Each subsequent step is described in detail: what dimension the input parameter of the step is, and what dimension the output parameter of the step is. Here, for each individual step, the dimension of the input parameter can be determined from the first line “Input” via the second and third entries in the subsequent parentheses. Furthermore, the fourth entry in the parentheses indicates how many input parameters are received in a single step. For example, in step CS1, a two-dimensional convolution is performed, resulting in the output parameter “Output” generating 16 distinct output parameters for the input, since 16 convolutional kernels are used for a single input parameter, where each output parameter has a dimension of 512 × 512. Therefore, for each processing step, those skilled in the art can clearly deduce from the parameters given here the dimensions of the input and output parameters and the number of convolutional kernels that may be necessary.

[0133] This design can also be referenced later. Figure 23 , 24 The descriptions of the second convolutional neural network CNN2 can be found in 25 and 26.

[0134] The convolutional neural network CNNA, with its components CNNA1 to CNNA4 as shown in Figures 30 to 32, includes different types of processing steps. The processing step CONV2D is a two-dimensional convolution with a specific number of kernels. The function LR is the activation function, used as the Leaky Rectified Linear Unit (RRL) activation function. The function AP is the so-called average pooling. The function BN is the so-called batch normalization. The processing step or function ADD is the element-wise summation of multiple input parameters to generate a single output parameter. For example, as shown in Figure 31, the processing step US is the so-called upsampling. The processing step ACT in Figure 33 is the so-called activation function, which can be given by, for example, the Sigmoid function or the so-called Softmax function. At the output of the neural network CNNA or the subnetwork CNNA4, three matrices MA1, MA2, and MA3 are generated, which have been referenced... Figure 11 It was explained in detail.

[0135] Figure 16a The different processing layers P1, P2, P3, and P4 of the second convolutional neural network CNN2 are shown for determining the total binding metric GBM based on multiple sub-image regions TB1, ..., TBN, as described above. Here, CNN2 processes the corresponding sub-image regions TB1, TB2, ..., TBN with index N respectively, in order to generate corresponding individual binding metrics IBM1, IBM2, ..., IBMN for the corresponding sub-image regions TB1, TB2, ..., TBN, as referenced above. Figure 14a What is described.

[0136] Here, Figure 16aAs shown, in the process of determining the corresponding individual combination metric for each individual sub-image TB1, ..., a corresponding final feature map FFM1 and preferably another final feature map FFM2 are generated. CNN2 can be designed to set and generate a unique final feature map FFM1 for only a single channel.

[0137] therefore, Figure 16a The result of processing level P3 in the process, for the sub-image TB with index N, is the final feature map FFM1 and preferably another final feature map FFM2. Figure 8a The first final feature map FFM1 of the sub-image TB in Figure 8b As shown in the figure. The second final feature map FFM2 for sub-image TB is in Figure 8c As shown in the image.

[0138] The Convolutional Neural Network (CNN2) solves the so-called "single-label classification" problem, which determines whether moving matrix regions in a sub-image are stained. The final feature map FFM1 represents the activation of the positive decision regarding "single-label classification" in the first classification channel, i.e., the moving matrix region is stained. Preferably, the final feature map FFM2 represents the corresponding activation of the negative decision, i.e., the moving matrix is ​​not significantly stained.

[0139] according to Figure 17 Based on the first final feature map FFM1 and the preferred second final feature map FFM2, a positive confidence measure PK is determined regarding staining of the dynamic matrix region or the presence of autoantibody binding in the corresponding dynamic matrix region K of the observed sub-image TB. Preferably, a negative confidence measure NK is determined based on the first final feature map FFM1 and the preferred second final feature map FFM2 regarding staining of the dynamic matrix region or the presence of autoantibody binding in the corresponding dynamic matrix region K of the observed sub-image TB.

[0140] Preferably, a confidence metric regarding the presence of autoantibody binding in the corresponding dynamic matrix region of the corresponding sub-image TB can be determined solely based on the first final feature map, without the need to use the second final feature map FFM2. Then, for example, in step S20, the feature map FFM1 can be supplied to a so-called max pooling process, which calculates the maximum pixel value for the final feature map FFM1 as a single scalar value. Preferably, this scalar value can be used as a confidence metric. Preferably, a value can be obtained as a confidence metric based on the scalar value, for example, using a so-called sigmoid function. Preferably, a value can be obtained as a confidence metric based on the scalar value, for example, using a so-called modified linear unit activation function.

[0141] Preferably, in the corresponding step S20, two corresponding feature maps FFM1 and FFM2 are respectively supplied to the so-called max pooling, and the max pooling calculates the maximum pixel value for each of the corresponding final feature maps as a corresponding single scalar value. Based on the scalar value, the positive probability PK can be determined in the so-called Softmax function in step S21 as a confidence measure regarding the presence of autoantibody binding in the moving matrix region or a confidence measure regarding staining of the moving matrix region. The negative probability NK can also be determined by the Softmax function. The positive probability PK and the negative probability NK are preferably added together to form a sum of 1. Therefore, in this way, for the corresponding sub-image TB, by calculating the first final feature map FFM1 and preferably the second final feature map FFM2, it is possible to determine the probability based on the given information. Figure 17 Determine the corresponding confidence measure regarding the presence of autoantibody binding in the dynamic matrix region of the corresponding sub-image.

[0142] Functions that can replace the Softmax function include, for example, the Sigmoid function, the rectified linear unit activation function, or the leaky rectified linear unit activation function.

[0143] Figure 18 The steps are illustrated for selecting a subset of identified sub-images from a fluorescence image based on the corresponding positive confidence measures of the corresponding sub-images. In step S30, the corresponding positive confidence measures PK1, PK2, PK3, ..., PKN with indices 1, ..., N in the N corresponding sub-images are sorted in ascending order with respect to their values. For example, for 29 different short-hymened worm cells or 29 different sub-images, Figure 28 The corresponding positive confidence measures PK are displayed, and these positive confidence measures are plotted in ascending order of their values ​​PKW using corresponding sorting indices 1, ..., 29. Then, in step S31, the following confidence measures and their corresponding matched sub-images are selected, wherein the matched confidence measure value of the confidence measure is 50% of the highest confidence measure value PKW. Then, a subset of matched sub-images is output or indicated by outputting the corresponding sub-image indicator with reference to the data record ID. Based on the indicator from the data record ID, the corresponding sub-images and their matched combination measures can be selected to determine the total combination measure.

[0144] The total binding metric is then determined based on the binding metrics belonging to the selected sub-images. This determination of the total binding metric occurs specifically in the post-processing PP step within the fourth processing layer P4 of the second convolutional neural network, such as... Figure 16a As shown.

[0145] Figure 19The exemplary processing of a single sub-image TB1 via a second convolutional neural network CNN2 is shown again. First, the sub-image TB1 is processed through the first three processing layers P1 to P3 of the second convolutional neural network, thereby providing the final feature map FFM1 and preferably also providing a second final feature map. Then, in the selection step SEL, based on the reference... Figure 17 and 18 The described indicator data ID potentially selects subimage TB1 into a subset of subimages. If subimage TB1 has been selected into the subset, it is then further processed as post-processing PP in the fourth processing level P4. Here, after determining the positive confidence metric and after selecting a subset of subimages, the corresponding post-processing PI for the corresponding first feature map FFM1 of the selected subimage is then performed. Exemplarily, the post-processing PI for determining the individual binding metric IBM1 is shown here for a single subimage TB1. Here, the corresponding subimage is selected for subimage TB1 based on the final feature map FFM1. The subimage represents the corresponding kinematic basal region of the corresponding brevichal cell in the subimage. Such a subimage TB is exemplarily shown in Figure 8a As shown in the diagram. Then, in step KIN, Figure 8b The first feature map FFM1 that matches in terms of size and resolution is interpolated by cubic interpolation. Figure 8a Sub-images in TB are adapted to obtain Figure 9a Amplified feature map VFM in the model.

[0146] Then, in the threshold step SB, the threshold is calculated. Figure 9b The binary value mask BM is shown in the figure. Here, it is preferable to use a value that is half the maximum possible gray value intensity of the feature map as the threshold. Therefore, if the gray value of the feature map VFM is between the values ​​0 and 1, then a value of 0.5 is used as the threshold.

[0147] Then in step MS, Figure 9b The masking operator BM is applied to the subimage TB to produce Figure 9c The corresponding sub-image SUB is then extracted from the sub-image. Therefore, the proposed second convolutional neural network CNN2 is able to provide a final feature map FFM1 for each corresponding sub-image TB, which indicates the sub-image region corresponding to the moving matrix within the sub-image by virtue of its value. Thus, the second convolutional neural network can extract or select the corresponding sub-image from the sub-image, which can then be used to determine the binding metric of the autoantibody to the double-stranded DNA on the moving matrix.

[0148] according to Figure 19Then, in step BS, the binding metric IBM1 of sub-image TB1 is determined. This is achieved by observing the pixel values ​​in sub-image SUB and selecting a value that defines the 90th percentile of pixel values ​​in the sub-image. For this purpose, Figure 27 against Figure 9c The pixels of the sub-image SUB are shown as corresponding pixel values ​​PW in ascending order of their respective sorting index QN. The value QW of index QN is such that 90% of the pixel values ​​from the sub-image SUB are less than the value QW.

[0149] In the next step, the total combined measure GBM is determined based on multiple individual combined measures IBM1, IBM2, ... from the individual sub-images of the selected subset.

[0150] Figure 16a An exemplary implementation of a second convolutional neural network CNN2 with multiple processing layers P1, P2, P3, P4 is shown, which is used to process sub-images TB1, ..., TBN with index N separately.

[0151] In the first processing layer P1, the convolutional neural network CNN2 generates a first subset of two-dimensional synthetic feature maps RFM1 based on the sub-image TB1 through at least one first convolutional layer LA1 and by applying multiple two-dimensional convolutional kernels. The feature maps RFM1 do not necessarily come directly from the convolutional layer LA1, but can be generated through additional processing steps PS2, PS3, and PSC.

[0152] In convolutional layer LA1, processing takes place in step PS1, which comprises a sequence of distinct sub-steps. Step PS1 is of type step PSA, which... Figure 16b The process is described in detail below. First, a 2D convolution CONV2D is performed on the input sub-image using multiple convolution kernels. Subsequent batch normalization is performed in step BN. Next, activation follows in step ACT.

[0153] In the context of this application, the convolutional layers of the second convolutional neural network CNN2 include layers for convolving one or more feature maps using one or more convolutional kernels. Preferably, batch normalization layers and / or activation layers may then follow such convolutional layers within the convolutional layers.

[0154] Then Figure 16a In the second processing layer P2, the first set of feature maps RFM1 is generated by passing through at least one second convolutional layer LA2 and applying multiple three-dimensional convolutional kernels to generate a second set of two-dimensional synthetic feature maps RFM2.

[0155] Then, based on the second set RFM2, a third set RFM3 of two-dimensional synthetic feature maps is generated by passing through at least one third convolutional layer LA3 and applying multiple three-dimensional convolutional kernels. The third set RFM3 is directly or indirectly input into further processing at the third level P3. In the third level, based on the third set RFM3, a first final feature map FFM1 and a preferred second final feature map FFM2 are determined by passing through another convolutional layer LAX.

[0156] The second set, RFM2, has fewer feature maps than the first set, RFM1. Furthermore, the third set, RFM3, has a greater number of synthetic feature maps than the second set, RFM2. Convolution kernels can also be called convolution operators.

[0157] So-called squeezing occurs by reducing the number of feature maps in the second convolutional layer LA2. The feature maps of the first set RFM1, or their features, are projected into the subspace by the convolutional kernel because the 3D convolutional kernel responds to the feature correlations between the feature maps. Therefore, only the most dominant features of the feature maps of the first set RFM1 are preserved and projected into the feature maps of the second set RFM2. Thus, less dominant and less persuasive features are filtered out.

[0158] Due to the increase in the number of feature maps from the second set RFM2 to the third set RFM3 via the third convolutional layer LA3, the previously reduced features or information are distributed into different feature spaces and different feature maps, where the features can be combined in different ways based on the degrees of freedom of the three-dimensional convolutional kernel used in the third convolutional layer LA3. This corresponds to the so-called expansion.

[0159] In the first processing layer P1, an additional convolutional layer LA11 may follow the first convolutional layer LA1. Layer LA11 uses the feature map created in layer LA1. Preferably, layer LA11 has processing steps PS2 and PS3 arranged in parallel with each other. The processing steps PS2 and PS3 are respectively... Figure 16c The processing step PSB is of a certain type. In the sub-step CONV3D of step PSB, 3D convolution of the feature map is performed using the corresponding 3D convolution kernel. Then, so-called batch normalization is performed in the further sub-step BN. In addition, so-called activation follows in step ACT.

[0160] The feature maps generated by steps PS2 and PS3 of layer LA11 are then merged together in the merging step PSC; in other words, the feature maps are arranged together.

[0161] Preferably, in the first processing level P1, the sub-image TB1 is convolved using a two-dimensional convolution kernel in step PS4. Step PS4 is... Figure 16bThe type of the substep CONV2D in the process.

[0162] Preferably, the feature maps generated by layer LA11 and step PS4 can be correlated in such a way that the entries of the feature maps are added element-wise. Therefore, this does not cause a change in the dimension of the feature maps, but rather the elements of the feature maps from layer LA11 are added element-wise to the elements of the feature maps from step PS4.

[0163] Step PS5 from the second convolutional layer LA2 belongs to Figure 16c The steps in the PSB type.

[0164] Preferably, the feature maps from convolutional layer LA2 are processed in the third convolutional layer LA3, such that in the corresponding steps PS7 and PS8, and in step PSC, the feature maps are processed in a manner similar to that of the feature maps from convolutional layer LA11, wherein the number and dimension of the convolutional kernels used may differ from each other. Steps PS7 and PS8 belong to... Figure 16c The type of step PSB in the process. The third set of feature maps RFM3 is generated by step PSC by element-wise summing the corresponding feature maps from the corresponding steps PS7 and PS8.

[0165] In the second processing layer P2, the second convolutional layer LA2 and the third convolutional layer LA3 follow successively as sub-steps of the sequential processing path PF1. Furthermore, in the second processing layer P2, parallel to the sequential processing path PF1, there exists another processing path PF2, in which CNN2 generates a fourth set of synthetic feature maps (RFM4) based on the first set RFM1 through at least one fourth convolutional layer LA4 and by applying multiple 3D convolutional kernels. This is achieved through step PS6, which is... Figure 16c The type of the substep CONV3D in the process.

[0166] Then, through step PSS, the third set of feature maps RFM3 and the fourth set of feature maps RFM4 are used to generate the feature map set RFM5 determined by step PSS in processing level P2. The feature map set RFM5 can then be used in the third processing level P3 to generate the first final feature map FFM1 and a preferred second final feature map FFM2 through the further step LAX, which will be described in detail later.

[0167] In the PS4's other processing layer, so-called post-processing is performed, such as... Figure 19 As detailed in the text.

[0168] Therefore, CNN2 generates the final feature map FFM1 corresponding to the sub-image TB based on the third set of feature maps RFM3 and the fourth set of feature maps RFM4. Here, the number of successive convolutional layers LA4 in the parallel processing path PF2 is less than the number of successive convolutional layers LA2 and LA3 in the sequential processing path PF1. Therefore, the parallel processing path PF2 has fewer convolutional layers than the sequential path PF1. Thus, during the training of the second convolutional neural network, the so-called "vanishing gradient" problem can be avoided or reduced by recalculating the individual weights of the convolutional kernels during backpropagation.

[0169] As mentioned above Figure 16a As mentioned, CNN2 can consist of four processing layers.

[0170] In response, Figure 23 A detailed implementation of the first processing layer P1 is shown, which corresponds to Figure 16a The first processing level P1 in the process.

[0171] For each individual step, the dimension of the input parameter, in the form of a sub-image or feature map set, is specified in detail. Here, for each individual step, the dimension of one or more input parameters can be determined in the first line “Input” via the second and third entries in subsequent parentheses. For example, the dimension of sub-image data TB1 is 150×150 pixels. For data TB1, there is only one unique input parameter, indicated by the element “1” in the fourth / last entry within parentheses. Regarding their range of values, image data TB1 is preferably normalized to a range from 0 to 1.

[0172] Then, in step PS, the input parameter TB1 is processed, for example, with convolutional kernels to produce a feature map of dimension 75×75 pixels. Here, the last entry in the lower row "Output" indicates the number of feature maps generated in the synthetic feature map set. Therefore, for each processing step, those skilled in the art can clearly deduce from the parameters given here how many convolutional kernels must be applied to the input data TB1 or the input feature map to achieve a specific number of output feature maps. In the example of step PS4, these are 64 convolutional kernels. Furthermore, those skilled in the art can deduce from the given dimensions of the input feature map and the given dimensions of the output feature map how much the so-called "striding" is, i.e., the offset in the convolution process of the feature map using a convolutional kernel, so that a specific number of pixels must be implemented. In the example of step PS4, this is a striding value of 2.

[0173] Those skilled in the art Figure 23The instructions provided in the documentation clearly outline the design of the processing layer P1 for the convolutional neural network CNN2.

[0174] Figure 24 Show Figure 16a The first part P21 of the processing layer P2. Here, the structure of the sub-processing layer P21 basically corresponds to Figure 16a The processing layer is P2. Furthermore, a so-called "Dropout" step for the training phase is also included, in which individual entries of the feature map are set to the value "0" (zero) in terms of their pixel value during training, but not during the actual classification in the testing phase. Here, the dropout factor is preferably 50% of the values, so half of the pixel values ​​are set to zero. The pixel values ​​are randomly selected in such a way that their indicators are selected by a random function.

[0175] Then, in sub-processing layer P21, the feature map set RFM5 is generated, as previously described. Figure 16a As shown in the processing level P2.

[0176] Preferably, the convolutional neural network CNN2 may have an additional sub-processing layer P22, which is described above in Figure 16a Not described in the text. Here, the feature map set RFM5 is further processed to generate a modified feature map set RFM51.

[0177] Here, in Figure 25 It also includes precise instructions for those skilled in the art on how to generate the modified feature map set RFM51. Here, the so-called dropped DRO also occurs during the training phase, not the testing phase.

[0178] Figure 26 An embodiment of a third processing layer P3 for generating a first final feature map FFM1 and a preferred second final feature map FFM2 is shown. Feature map FFM1 and the preferred feature map FFM2 are indirectly generated based on a third set of feature maps RFM3 and a fourth set of feature maps RFM4. Here, the above... Figure 16aThe processing step LAX described herein has sub-steps. In the first sub-step SEPC, a so-called "depth convolution" of the input feature map RFM51 occurs, wherein each individual two-dimensional feature map from the set RFM51 is convolved using a two-dimensional convolution kernel, thereby producing a set of two-dimensional feature maps RFM511. Subsequently, in the further sub-step CONV3D, multiple two-dimensional feature maps of the set RFM511 produced by step SEPC are convolved using a three-dimensional convolution kernel of dimension 1×1×G, where G is the number of feature maps in the set, thereby determining the first final feature map FFM1. Preferably, in the further sub-step CONV3D, multiple two-dimensional feature maps of the set RFM511 produced by step SEPC are further convolved using another three-dimensional convolution kernel of dimension 1×1×G, where G is the number of feature maps in the set, thereby determining the second final feature map FFM2.

[0179] then Figure 26 The third processing level, P3, is the so-called post-processing PP, such as... Figure 16a As shown in processing level P4, and as in Figure 19 It is explained in detail in the text.

[0180] To implement one or more exemplary embodiments of the convolutional neural networks CNN1 and CNN2 presented herein, those skilled in the art can employ the so-called open-source deep learning library known as "Keras". Detailed information can be found at https: / / keras.io. Figures 23 to 26 The proposed implementation of CNN2, including processing layers P1, P21, P22, and P3, and processing layer P4 (as shown in the figure). Figure 19 (Detailed explanation follows) Created for testing purposes using the so-called open-source deep learning library "Keras".

[0181] Different fluorescent image data records were used to train convolutional neural networks CNN1 and CNN2.

[0182] The first convolutional neural network (CNN1) was trained using 34,000 complete or fluorescent images, along with corresponding ground-truth data for three image segment categories: cell bodies, cell edges, and background. For training, 95% of the 34,000 fluorescent images, or 32,300 images, were used. During training, these 32,300 fluorescent images were fed to the CNN1 and optimized using cross-entropy loss to reproduce the labeled mask. The training was conducted for 100 epochs with an initial learning rate of 1e-3. The learning rate was variable and depended on the current epoch. For subsequent testing of the CNN1, the remaining 5% of the 34,000 fluorescent images, or 1,700 images, were then used. Corresponding ground-truth data for the three image segment categories—cell bodies, cell edges, and background—was also available. Image segment classification was performed at the pixel level or with pixel precision. The classification accuracy for image segments or pixels across the three categories reached 99.6%. Another evaluation metric is the average intersection-union ratio (Miou), because it takes into account the severe imbalance between the frequencies of background pixels and unit pixels. An 80% Miou was achieved.

[0183] The first data record of the fluorescence image is one in which the patient sample used for cultivation is known to have autoantibodies and the dynamic matrix region is therefore stained in the fluorescence image. The second data record of the fluorescence image is one in which the patient sample used for cultivation is known to have no antibodies and the dynamic matrix region is therefore not stained in the fluorescence image.

[0184] The second convolutional neural network (CNN2) was trained based on 24,980 sub-images representing "positive" test cases (positive outcome or positive patient) and 15,957 sub-images representing "negative" test cases (negative outcome or negative patient). For training itself, 17,427 positive sub-images out of the 24,980 positive sub-images and 11,233 negative sub-images out of the 15,957 negative sub-images were used, each with corresponding annotations. CNN2 was trained for 100 epochs using a learning rate of 1e-3. For testing, 7,543 positive sub-images out of the 24,980 positive sub-images and 4,724 negative sub-images out of the 15,957 negative sub-images were used. In the test case, the decisions derived by CNN2 produced 6,602 true positive decisions and 4,623 true negative decisions. Additionally, 101 false positive decisions and 941 false negative decisions were produced. This corresponds to a sensitivity of 87.52% and a specificity of 97.86%.

[0185] Figure 29Method V20 is shown, in which preferred executable steps for acquiring a fluorescence image are performed.

[0186] In step S200, a temporary first fluorescence image EVFB2 is acquired by at least one fixed preset acquisition parameter EP1, which is preferably a gain parameter.

[0187] Then, in step S201, a histogram of pixel values ​​for image EVFB2 is formed, thereby determining and providing histogram data HD.

[0188] Then, in step S202, it is determined how high the number of pixels exceeds a certain saturation level in terms of brightness. For an exemplary quantization range of pixel values ​​from 0 to 255, for example, it is determined how many pixels have a pixel value or grayscale value of 255. The number of pixels with saturated brightness is provided as data AS.

[0189] Then, in step S203, it is checked whether the number of pixels AS within the saturation range exceeds a preset threshold TH. If it does not exceed the threshold (see branch "N"), a temporary first fluorescence image EVFB2 is used as the fluorescence image SG; see [link to relevant documentation]. Figure 4 If the number of saturated pixels exceeds the threshold TH (see branch "Y"), a temporary second fluorescence image ZVFB2 is acquired in step S204 using at least one fixed preset second acquisition parameter EP2. The fixed preset acquisition parameter EP2 is different from the previously fixed preset first acquisition parameter EP1. Preferably, the acquisition parameter EP2 is a gain parameter and is less than the first acquisition parameter EP1, so that the exposure of image ZVFB2 is weaker than that of image EVFB2. The acquired temporary second fluorescence image EVFB2 is then used as the fluorescence image SG; see [link to relevant documentation]. Figure 4 Preferably, indication data is also provided, which indicates that the brightness of the temporary first fluorescence image EVFB2 has exceeded the maximum brightness, and the fluorescence image SG is a temporary second fluorescence image ZVFB2 acquired under reduced brightness conditions.

[0190] The temporary second fluorescence image EVFB2 can then be used as the fluorescence image SG in the proposed method in the usual manner. Here, the convolutional neural network CNN2 preferably determines, via a confidence metric PK, whether there is staining of the moving matrix region for a minimum number of sub-image regions, preferably at least ten sub-image regions. If this is the case, the convolutional neural network CNN2 outputs the maximum brightness or a maximum brightness value, preferably 255, as the overall binding metric. If the convolutional neural network CNN2 determines that the moving matrix region is not actually stained, the fluorescence image SG is analyzed and evaluated as overall negative.

[0191] Although some aspects have been described in conjunction with the equipment, it is self-evident that these aspects also describe the corresponding methods, and thus the blocks or structural elements of the equipment should also be understood as the corresponding method steps or features of the method steps. Similarly, aspects that have been described in conjunction with or as method steps are also descriptions of the corresponding blocks, details, or features of the corresponding equipment.

[0192] Depending on the specific implementation requirements, in an exemplary embodiment of the present invention, the computing unit R or data network device can be implemented in hardware and / or software. The computing unit R mentioned herein can be implemented as at least one computing unit or can be implemented by multiple interconnected computing units. The implementation can be carried out using a digital storage medium, such as a floppy disk, DVD, Blu-ray disc, CD, ROM, PROM, EPROM, EEPROM, or flash memory, hard disk drive, or other magnetic or optical storage device, which stores electronically readable control signals that interact with or can interact with programmable hardware components to cause the execution of corresponding methods.

[0193] As computing units, programmable hardware components can be formed from processors, central processing units (CPUs), computers, computer systems, application-specific integrated circuits (ASICs), integrated circuits (ICs), single-chip systems (SOCs), programmable logic elements, or field-programmable gate arrays (FPGAs) with microprocessors.

[0194] Therefore, digital storage media can be machine-readable or computer-readable. Consequently, some embodiments include data carriers having electronically readable control signals that can interact with a programmable computer system or programmable hardware component to perform one of the methods described herein.

[0195] Generally, embodiments or portions of the present invention can be implemented as a program, firmware, computer program, or computer program product having program code or data, wherein the program code or data can function to perform one or more methods when the program is running on a processor or programmable hardware component.

Claims

1. A method for detecting the binding of autoantibodies to double-stranded deoxyribonucleic acid in a patient sample by fluorescence microscopy and digital image processing using *Bifidobacterium brevis* cells, the method comprising: - Provide a substrate (S) with multiple short hymenorrhea cells (CR). - Culture the substrate (S) together with patient samples that may have autoantibodies. - The substrate (S) is incubated together with secondary antibodies, each labeled with a fluorescent dye. - Obtain the fluorescence image (SG) of the substrate (S) in the color channel corresponding to the fluorescent dye. Its features are, - A pre-trained first convolutional neural network (CNN1) identifies corresponding sub-images (TB) in the fluorescence image (SG), each sub-image representing at least one short hymenorrhea cell (CR). - At least one subset of the corresponding sub-image (TB) is processed by a pre-trained second convolutional neural network (CNN2). The second convolutional neural network (CNN2) determines a corresponding sub-image representing the dynamic matrix region for the corresponding sub-image, in order to determine the corresponding binding metric (IBM1, IBM2) based on the corresponding sub-image. The corresponding binding metric indicates the degree of binding of autoantibodies in the corresponding dynamic matrix region (K) of the corresponding short hymenorrhea cell (CR) in the corresponding sub-image (TB). - Determine the total binding measure (GBM) of the binding of autoantibodies to double-stranded deoxyribonucleic acid in patient samples based on the corresponding binding measures (IBM1, IBM2).

2. The method of claim 1, wherein, Identifying the corresponding sub-image (TB) in the fluorescence image (SG) is achieved by assigning the corresponding image fragment of the fluorescence image to the corresponding image fragment category in the image fragment category group through a pre-trained first convolutional neural network (CNN1), wherein the image fragment category group includes at least the following image fragment categories: -cells, and -background.

3. The method of claim 1, wherein, Identifying the corresponding sub-image (TB) in the fluorescence image (SG) is achieved by assigning the corresponding image fragment of the fluorescence image to the corresponding image fragment category in the image fragment category group through a pre-trained first convolutional neural network (CNN1), wherein the image fragment category group includes at least the following image fragment categories: -Cell body, -Cell edges, and -background.

4. The method according to claim 1, wherein the method comprises: For the corresponding sub-image (TB) - Select the corresponding sub-image (SUB) of the corresponding sub-image (TB), where the corresponding sub-image (SUB) represents the corresponding kinematic matrix region (K) of the corresponding short hymenorrhea cell (CR). - Determine the corresponding binding metric (IBM1) based on the corresponding sub-image (SUB). The method also includes determining the total binding measure (GBM) based on the corresponding binding measures (IBM1, IBM2).

5. The method according to claim 1, wherein the method comprises: - The corresponding final feature map (FFM1) for the corresponding sub-image (TB) is determined by the second convolutional neural network (CNN2). - Determine the corresponding confidence measure (PKN) regarding the presence of autoantibody binding in the corresponding dynamic matrix region (K) of the corresponding sub-image (TB). - Select a subset of sub-images (TB) based on the determined confidence measure (PKN). - Correspondingly process the feature maps of the selected sub-images to determine the corresponding combination measures (IBM1, IBM2). - Determine the total binding measure (GBM) based on the corresponding binding measures (IBM1, IBM2) of the selected sub-images.

6. The method according to claim 5, wherein the method comprises: For the corresponding sub-image (TB) from the selected subset. - Select the corresponding sub-image (SUB) of the corresponding sub-image (TB) based on the corresponding final feature map (FFM1) corresponding to the corresponding sub-image (TB), wherein the corresponding sub-image (SUB) represents the corresponding kinematic matrix region (K) of the corresponding short hymenorrhea cell (CR). - Determine the corresponding binding metric (IBM1) based on the corresponding sub-image (SUB). The method also includes: - Determine the total binding measure (GBM) based on the corresponding binding measures (IBM1, IBM2).

7. The method according to claim 6, wherein the method comprises: For the corresponding sub-image (TB) from the selected subset. - Obtain the corresponding mask operator (BM) based on the corresponding final feature map (FFM1). - Select the corresponding sub-image (SUB) of the corresponding sub-image (TB) by applying the corresponding mask operator (BM) to the corresponding sub-image (TB). - Determine the corresponding binding metric (IBM1) based on the corresponding sub-image (SUB). And the method includes: - Determine the total binding measure (GBM) based on the corresponding binding measures (IBM1, IBM2).

8. The method according to claim 5, wherein, During the processing of sub-images (TB), the second convolutional neural network (CNN2) is used: - In the first processing layer (P1), a first set of synthetic feature maps (RFM1) is generated based on the sub-image (TB) through at least one first convolutional layer (LA1) and by applying multiple two-dimensional convolutional kernels. -In the second processing level (P2), - A first set of two-dimensional feature maps (RFM1) is generated by passing through at least one second convolutional layer (LA2) and by applying multiple three-dimensional convolutional kernels to produce a second set of synthetic feature maps (RFM2). Furthermore, a third set of synthetic feature maps (RFM3) is generated based on a second set of two-dimensional feature maps (RFM2) through at least one third convolutional layer (LA3) and by applying multiple three-dimensional convolutional kernels. The second set (RFM2) has fewer synthetic feature maps than the first set (RFM1), and the third set (RFM3) has more synthetic feature maps than the second set (RFM2).

9. The method according to claim 8, in, In the second processing layer (P2), the second convolutional layer (LA2) and the third convolutional layer (LA3) follow successively as sub-steps of the sequential processing path (PF1). In the second processing layer (P2), there is another processing path (PF2) in parallel with the sequential processing path (PF1). In the other processing path, the second convolutional neural network (CNN2) generates a fourth set of synthetic feature maps (RFM4) based on the first set of two-dimensional feature maps (RFM1) through at least one fourth convolutional layer (LA4). The second convolutional neural network (CNN2) generates the final feature map (FFM1) corresponding to the sub-image (TB) based on the third set (RFM3) and the fourth set (RFM4) of synthetic feature maps. Furthermore, the number of successive convolutional layers in the parallel processing path (PF2) is less than the number of successive convolutional layers in the sequential processing path (PF1).

10. The method according to claim 1, wherein the method comprises: - Acquire a temporary first fluorescence image (EVFB2) in the color channel using fixed preset acquisition parameters (EP1). - Determine whether the brightness of the temporary first fluorescence image (EVFB2) of the color channel exceeds the maximum brightness. -If the temporary first fluorescence image of the color channel does not exceed the maximum brightness, the temporary first fluorescence image (EVFB2) is used as the fluorescence image (SG). - If the temporary first fluorescence image in the color channel exceeds the maximum brightness, a temporary second fluorescence image (ZVFB2) is acquired in the color channel and the temporary second fluorescence image (ZVFB2) in the color channel is used as the fluorescence image (SG).

11. The method according to claim 1, wherein, The fluorescent dye is a green fluorescent dye.

12. The method according to claim 1, wherein, The color channel is the green channel.

13. An apparatus (V1) for detecting the binding of autoantibodies to double-stranded deoxyribonucleic acid in a patient sample by fluorescence microscopy and digital image processing using *Synthia spp.* cells, said apparatus comprising: - A holding device (H) for a substrate (S) having multiple short hymenorrhea cells (CR) and the substrate being cultured with a patient sample containing autoantibodies and also with secondary antibodies, each labeled with a fluorescent dye. - At least one image acquisition unit is used to acquire a fluorescence image (SG) of the substrate (S) in a color channel corresponding to the fluorescent dye. The device also includes at least one computing unit (R), which is designed to be - A pre-trained first convolutional neural network (CNN1) identifies corresponding sub-images (TB) in the fluorescence image (SG), each sub-image representing at least one short hymenorrhea cell (CR). - A pre-trained second convolutional neural network (CNN2) processes at least one subset of the corresponding sub-images (TB). The second convolutional neural network (CNN2) determines a corresponding sub-image representing the dynamic matrix region for the corresponding sub-image, in order to determine the corresponding binding metric (IBM1, IBM2) based on the corresponding sub-image. The corresponding binding metric indicates the degree of binding of autoantibodies in the corresponding dynamic matrix region (K) of the corresponding short hymenorrhea cell (CR) in the corresponding sub-image (TB). - And based on the corresponding binding metric (IBM1, IBM2), the total binding metric (GBM) of the autoantibodies in the patient sample to the double-stranded deoxyribonucleic acid is determined.

14. The device according to claim 13, characterized in that, The fluorescent dye is a green fluorescent dye.

15. The device according to claim 13, characterized in that, The color channel is the green channel.

16. A computing unit (R), said computing unit for digital image processing and comprising, - A receiving unit for receiving a fluorescence image (SG) representing a substrate stained with a fluorescent dye, the substrate having multiple short hymenorrhea cells (CR). - Identification units of corresponding sub-images (TB) in the fluorescence image (SG) are identified by a pre-trained first convolutional neural network (CNN1), each sub-image representing at least one short hymenorrhea cell (CR). - Processing units that process at least one subset of the corresponding sub-images (TB) using a pre-trained second convolutional neural network (CNN2), the second convolutional neural network (CNN2) determines a corresponding sub-image representing the dynamic matrix region for the corresponding sub-image, in order to determine the corresponding binding metric (IBM1, IBM2) based on the corresponding sub-image. The corresponding binding metric indicates the degree of binding of autoantibodies in the corresponding dynamic matrix region (K) of the corresponding short hymenorrhea cell (CR) in the corresponding sub-image (TB). - and a unit for determining the total binding mass (GBM) of autoantibodies binding to double-stranded deoxyribonucleic acid in patient samples based on the corresponding binding mass (IBM1, IBM2).

17. The computing unit according to claim 16, characterized in that, The fluorescent dye is a green fluorescent dye.

18. A data network device (DV), comprising: At least one data interface (DS4) is provided for receiving fluorescence images representing a substrate stained with a fluorescent dye, the substrate having multiple short-hymened worm cells. and at least one computing unit (R), said computing unit being designed to, during digital image processing, - A pre-trained first convolutional neural network (CNN1) identifies corresponding sub-images (TB) in the fluorescence image (SG), each sub-image representing at least one short hymenorrhea cell (CR). - A pre-trained second convolutional neural network (CNN2) processes at least one subset of the corresponding sub-images (TB). The second convolutional neural network (CNN2) determines a corresponding sub-image representing the dynamic matrix region for the corresponding sub-image, in order to determine the corresponding binding metric (IBM1, IBM2) based on the corresponding sub-image. The corresponding binding metric indicates the degree of binding of autoantibodies in the corresponding dynamic matrix region (K) of the corresponding short hymenorrhea cell (CR) in the corresponding sub-image (TB). - And based on the corresponding binding metric (IBM1, IBM2), the total binding metric (GBM) of the autoantibodies in the patient sample to the double-stranded deoxyribonucleic acid is determined.

19. The data network device (DV) according to claim 18, characterized in that, The fluorescent dye is a green fluorescent dye.

20. A method for digital image processing, comprising: - Receive a fluorescence image (SG) representing a substrate stained with a fluorescent dye (S), the substrate having multiple short hymenorrhea cells (CR). - A pre-trained first convolutional neural network (CNN1) identifies corresponding sub-images (TB) in the fluorescence image (SG), each sub-image representing at least one short hymenorrhea cell (CR). - A pre-trained second convolutional neural network (CNN2) processes at least one subset of the corresponding sub-images (TB). The second convolutional neural network (CNN2) determines a corresponding sub-image representing the dynamic matrix region for the corresponding sub-image, in order to determine the corresponding binding metric (IBM1, IBM2) based on the corresponding sub-image. The corresponding binding metric indicates the degree of binding of autoantibodies in the corresponding dynamic matrix region (K) of the corresponding short hymenorrhea cell (CR) in the corresponding sub-image (TB). - And based on the corresponding binding metric (IBM1, IBM2), the total binding metric (GBM) of the autoantibodies in the patient sample to the double-stranded deoxyribonucleic acid is determined.

21. The method for digital image processing according to claim 20, characterized in that, The fluorescent dye is a green fluorescent dye.

22. A computer program product (CPP) comprising instructions that, when executed by a computer, cause the computer to perform the method for digital image processing according to claim 20 or 21.

Images