Image capturing method, image analysis method, and method for training image analysis neural network

Abstract

The image capturing method of the present invention comprises capturing images with a microscope at focal plane positions (50) shifted in equal steps in an analysis space (40); and selecting an image from the captured images for further image processing. Images are captured of a precipitate (31) of a liquid filled in the analysis space (40), the analysis space being located between a transparent upper window portion (11) and a transparent lower window portion (21) of a container (30) suitable for analyzing the liquid, wherein the precipitate (31) is centrifuged on an inner flat surface of the lower window portion (21), and wherein the method comprises capturing a first depth image sequence in a first spatial region (41) of the analysis space (40), and selecting an image with the best contrast from the first depth image sequence for further image processing, and capturing a second depth image sequence in a second spatial region (42) of the analysis space (40) by taking into account the previous steps, wherein the second depth image sequence has fewer images than the first depth image sequence, and selecting an image with the best contrast from the second depth image sequence for further image processing. The present invention also relates to an image analysis method, a method for training an image analysis neural network, and an image analysis neural network based on the above method.

Description

Technical Field

[0001] The present invention relates to an image recording method, an image analysis method, a method for training an image analysis neural network, and an image analysis neural network, which are suitable for use in connection with containers for analyzing various liquids, wherein the container can be used, for example, as a cuvette for the optical analysis of urine. Background Art

[0002] For the analysis of liquids such as urine, there are many prior art container designs. For the purpose of optical analysis, flat containers or cuvettes are generally used, which are preferably capable of analyzing the liquid filled into the analysis space between the transparent window portions arranged below each other using microscopic or digital image processing methods. In order to ensure a simple manufacturing method, these containers are composed of an upper part and a lower part connected in parallel to each other. The container includes an inlet hole for filling the liquid to be analyzed, and an outlet hole through which air escapes from the container when the liquid is filled. For example, WO2008 / 050165A1 discloses a cuvette that can also be used for the purpose of the present invention.

[0003] To optically analyze a liquid in a cuvette, an image of the liquid must be captured (i.e., recorded). To address the issues caused by the uncertainty of the depth of the liquid and the height of the object being analyzed within it, a known technical solution is to capture images of a given area using different focal plane positions, so that these images contain as much relevant information as possible. These images can be used not only for image analysis but also for training image analysis neural networks.

[0004] In US 2004 / 0202368 A1, US 2013 / 0322741 A1 and US 2014 / 0348409 A1 the possibility of using different focal planes for image analysis purposes is generally mentioned.

[0005] The technical solution disclosed in US 2008 / 0082468 A1 involves recording images at equidistant focal planes and applying them to training images belonging to two groups captured at different focal planes. However, according to this document, the applied focal plane only advances in one direction from the "ideal" focal plane, and there is no mention of applying a solution around (i.e., "sandwiching") the ideal focal plane. Consequently, this document does not teach the use of images captured below and above a selected focal plane for training and image analysis.

[0006] US2015 / 0087240A1 discloses images captured at multiple focal planes, but according to this document, the purpose is to select the best-focused image. Therefore, combined analysis of multiple images captured at different focal planes is not included in this technical solution.

[0007] WO 2013 / 104938 A2 contains a general description of the possibility of analyzing images taken at different focal planes simultaneously and thereby inputting additional information for decision making. However, this document also does not discuss taking images equidistantly on both sides and applying a wider range of training.

[0008] A common disadvantage of the prior art solutions is that, in the case of distributed images in multiple spatial regions of the analysis volume, which is necessary for a comprehensive liquid analysis, they do not provide an advantageous solution for finding the focal plane that provides the best contrast and for effectively utilizing such information for other spatial regions. In the case that the focal plane that provides the best contrast has to be found at each position (i.e. in each analyzed spatial region) by scanning the entire depth range, the process becomes lengthy and overly resource-intensive. Another common disadvantage of the prior art solutions is that they do not enable image analysis, the training of image analysis neural networks and the implementation of the trained image analysis neural networks in a manner supported by efficient image capture methods. Summary of the Invention

[0009] According to the present invention, it has been recognized that if a focal plane has been found that provides the best contrast in a relevant depth range of the analysis space, this information can be exploited for capturing images in other spatial regions of the flat analysis space, and preferably a much smaller depth range must be scanned in order to find the image that provides the best contrast. It has also been recognized that some of the images obtained during the depth scan can be exploited for training the image analysis neural network and for analyzing the images, which allows for a more robust image analysis and training process.

[0010] The object of the present invention is to eliminate as far as possible the disadvantages of the prior art solutions and to provide the advantages mentioned above.The objects of the present invention have been achieved by the claimed image analysis method, training method and image analysis neural network program product. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Preferred embodiments of the present invention are described below by way of example with reference to the accompanying drawings, in which:

[0012] Figure 1A is a spatial diagram of the upper portion of an exemplary cuvette,

[0013] Figure 1B is a spatial diagram of the lower portion of an exemplary cuvette,

[0014] Figure 1C Is displayed based on Figures 1A-1B The spatial diagram of the cuvette in the assembled state,

[0015] Figure 2 is based on Figures 1A-1C A top view of the cuvette showing the spatial region of the analysis space,

[0016] Figure 3 is a schematic cross-sectional illustration of the steps of a non-adaptive image capture method,

[0017] Figure 4 is a schematic cross-sectional illustration of the steps of the adaptive image capture method, and

[0018] Figure 5 is a schematic cross-sectional illustration of the steps of a partially adaptive image capture method. DETAILED DESCRIPTION

[0019] The container used in the method according to the present invention is preferably embodied as a cuvette for urine analysis, which allows optical analysis of the liquid contained therein. Optical analysis is preferably performed under illumination using a microscope. Prior to analysis, the filled container is centrifuged, causing the urine precipitate to settle on the polished inner surface of the transparent lower window portion of the container. The container can then be used to digitally analyze the image generated by this precipitate.

[0020] according to Figures 1A-1C The container 30 includes an upper portion 10 and a lower portion 20 . Figure 1A The upper portion 10 shown is formed with a transparent upper window portion 11 capable of optical analysis. In order to improve the accuracy of the optical analysis, the upper window portion 11 has polished surfaces on both sides.

[0021] Furthermore, in the upper part 10 there is an inlet opening 12 starting from the conical recess, and an outlet opening 13 designed to evacuate air from the container when filling with liquid. Figure 1B The analysis space of the container 30 is shown to have two parts; an upper analysis space part located in the upper part 10 of the container and a lower analysis space part located in the lower part 20 of the container.

[0022] Figure 1B 3 is a spatial diagram showing the lower portion 20 of an exemplary container 30. The lower portion 20 is also equipped with a transparent lower window portion 21, which is polished on both sides. According to the present invention, the thickness of the lower window portion 21 is less than 0.6 mm to improve image clarity.

[0023] Prior to optical analysis, the container 30 is centrifuged so that an image of the sediment of the liquid contained in the analysis space 40 can be taken in a state where the sediment is centrifuged relative to the flat inner surface of the lower window portion 21. Figure 2 Between the transparent upper window portion 11 and the transparent lower window portion 21 of the container 30 shown.

[0024] The exemplary cuvette is suitable for receiving 175 μL of urine, of which at least 2.2 μL must be analyzed in human applications; images must be taken at a relatively large number of positions, for example fifteen analysis positions, i.e., more specifically, in the spatial regions 41, 42, 43, 44, 45. In veterinary applications, the number of image recording positions is generally greater than the number of positions at which images are preferably taken, for example, at seventy positions. The spatial regions 41, 42, 43, 44, 45 are preferably positioned laterally so that they are distributed to the greatest possible extent in the analysis space 40 without overlapping each other. In the case of positioning the cuvette by rotation, the analysis positions will have a fanned-out spatial arrangement, i.e., with Figure 2 Different spatial arrangements shown

[0025] Figures 3 to 5 The steps performed by various embodiments of the image capturing method according to the present invention are shown.

[0026] During the image capture method according to the present invention, images are captured in analysis space 40 using a microscope at focal plane positions 50 shifted in equal steps, and the image most suitable for further image processing is selected from the captured images. As described above, an image of a precipitate 31 of a liquid contained in analysis space 40, located between transparent upper window portion 11 and transparent lower window portion 21 of a container 30 suitable for analyzing the liquid, is captured such that precipitate 31 is centrifuged on the inner plane of lower window portion 21. This requires capturing the image with the cuvette oriented horizontally.

[0027] In a first step of the method, a first depth image sequence is recorded in a first spatial region 41 of the analysis space 40 at focal plane positions 50 shifted in equal steps in the depth direction perpendicular to the flat surface, and the image with the best contrast is selected from the first depth image sequence for further image processing. Figure 3 In the present invention, the focal plane position 50 of the image of the first spatial region 41 with the best contrast is indicated by a dashed line. In the context of the present invention, the term "focal plane position" 50 is used to refer to the position of the focal plane in the depth direction according to the present invention, ie its height position.

[0028] According to the invention, the information about the focal plane position 50 determined in this way is also used in the other spatial regions 42, 43, 44, 45. Figure 3In the second spatial region 42 of the analysis space 40, a second depth image sequence is captured at focal plane positions 50 shifted in equal steps in the depth direction. One of the intermediate focal plane positions 50 in the second depth image sequence corresponds to the focal plane position 50 with the best contrast in the first depth image sequence. The second depth image sequence has fewer images than the first depth image sequence; this is possible because the depth range to be scanned encompasses the optimal focal plane position 50 determined for the first spatial region 41. If a much longer depth image sequence were first captured in all analyzed spatial regions, the measurement would take an unacceptably long time. Therefore, the relatively short duration of the second depth image sequence is a fundamental feature of the present invention. The image with the best contrast is also selected from the second depth image sequence for further image processing. This latter image does not necessarily need to be located at the optimal focal plane position 50 determined for the first spatial region 41, as the height and composition of the sediment 31 may vary spatially and the transparent lower window portion 21 may deviate from a horizontal orientation. Therefore, the optimal contrast focal plane position 50 of the second spatial region 42 may be shifted by one or more steps from the dashed line.

[0029] Preferably, the best contrast focal plane position 50 of the first depth image sequence corresponds to the center focal plane position 50 of the second depth image sequence. Thus, in the case of downward shifting as well as in the case of upward shifting, the best contrast focal plane position 50 of the second spatial region 42 will most likely be included in the analysis.

[0030] The first depth image sequence is preferably selected so that it includes focal plane positions 50 extending over the maximum empirically observed depth extension of the sediment 31. For example, the first depth image sequence may consist of 100 steps of 2.5 μm step size or 200 steps of 1.25 μm step size, while the second (and subsequent) depth image sequences may have the same step size and may consist of 8, 16, 32 steps, or may have another number of steps that is less than the number of steps of the first depth image sequence.

[0031] An implementation can also be envisioned in which the depth step size of the focal plane position 50 of the second depth image sequence is smaller than the depth step size of the focal plane position 50 of the first depth image sequence. That is, by knowing the determined ideal position, it is possible to search for an image with optimal contrast by applying a finer step size while keeping the computational / processing requirements sufficiently low. However, in this case, after capturing the first depth image sequence, an additional image sequence with a smaller step size must also be captured in the first spatial region 41 near the ideal position, so that the best-focused image can also be determined there with similar accuracy as in the other spatial regions 42, 43, 44, 45.

[0032] like Figure 3-5As can be seen in FIG, additional depth image sequences are recorded in further spatial regions 43, 44, 45 of the evaluation space 40, and the image with the best contrast is subsequently selected from each depth image sequence for further image processing. The previously determined information about the position of the best contrast can be used to record these additional images in various ways, for example according to the three alternatives described in detail below.

[0033] according to Figure 3 In a first alternative represented by the preferred embodiment depicted in , an additional depth image sequence is also captured by applying the focal plane position of the second depth image sequence 50. This so-called non-adaptive alternative has the lowest computational requirements and therefore it provides the fastest operation.

[0034] according to Figure 4 In a second alternative, represented by the preferred embodiment shown in , all additional depth image sequences are captured by applying a focal plane position 50 which contains the best contrast focal plane position 50 of the previous depth image sequence as an intermediate focal plane position 50 and which is shifted in the depth direction with equal steps. Preferably, the central focal plane position 50 of the additional depth image sequence corresponds to the best contrast focal plane position 50 of the previous depth image sequence. The latter is Figure 4 This so-called adaptive embodiment has the advantage that it can follow the changes in the height level of the sediment 31 or the transparent lower window portion 21 and thus most likely find an image with the best contrast at all analysis positions.

[0035] According to Figure 5In a third alternative, represented by another preferred embodiment shown in , if the focal plane position 50 of the previous depth image sequence providing the best contrast reaches, approaches, or exceeds a predetermined depth range boundary, the focal plane position 50 of the first depth image sequence is used to capture the additional depth image sequence. In the event that the best contrast focal plane position of the previous depth image sequence falls outside the range defined by the focal plane position of the second depth image sequence, the additional depth image sequence is preferably captured using the focal plane position 50 of the first depth image sequence, although other empirical criteria may also be provided. This so-called partially adaptive embodiment has the advantage that, if the best contrast image is detected to be outside the "normal" range, it can be assumed that some abnormality has occurred, and therefore it is advantageous to re-execute the initial multi-step image capture sequence. Such an abnormality may, for example, be caused by certain characteristics of the urine sediment, such as the presence of large lipids, which may shift the focal plane position 50 detected as ideal to a position that is too high, while the portion of the sediment 31 providing relevant information remains within the deeper range, albeit with occasional exceptions at specific analysis locations.

[0036] It is particularly preferred if the depth step size for focal plane position 50 is 50-100% of the microscope's depth of focus (DOF). A depth step size of less than 50% of the DOF will result in more than one image meeting the optimal contrast criterion, which results in unnecessary calculations. In the case of depth step values greater than 100% of the DOF, images may not be captured within the range that provides optimal contrast. For example, with a DOF of 1.5 μm, a step size of 1.25 μm is preferably used.

[0037] In order to perform the analysis of the urine sediment, high-resolution grayscale digital images are recorded in each region of the field of view, i.e. in the individual planes in the spatial regions 41, 42, 43, 44 and 45. In the next step, a focusing algorithm known per se is applied to select the "best" image, i.e. the image with the best contrast according to a suitable function, which will constitute the input information for the nonlinear image analysis module based on artificial intelligence, which is based on a neural network.

[0038] The image capture method according to the invention has the advantage that it allows other images in addition to the best image to be input to the analysis module, preferably allowing one or more pairs of images captured at planes located at the same distance above and below the plane of the best image to be captured and used. These additional images are not as sharp as the selected image, but can contain important supplementary information about the objects in the analyzed field of view. It can happen, for example, in the case of urine sediment, red blood cells and fungi, that objects or particles that appear to be the same in the best focused image (i.e. the image with the best contrast) experience darkening / distortion / blurring in different ways in other planes located increasingly further away from the best focus plane, for example due to their different height, density / optical density / refractive index, etc. As a result, such objects can be better distinguished, which improves the recognition capabilities of the analysis module.

[0039] In a physical sense, images taken in different planes (i.e., at different focal plane positions) contain not only intensity distribution information, but also phase information, so they can even be used to generate digital phase contrast (DPC) images. For example, by generating the difference between two images taken in different planes, or by applying more complex functions, even better / more visually appealing contrast images can be provided. Although the latter function is more useful for displaying the images, images generated in this way are also worth inputting into the analysis module. Among other things, a neural network can easily "figure out" how to use the difference of the images and will learn to perform the subtraction.

[0040] It is important that when capturing a focus sequence, the focal planes of consecutive images are equidistant from each other.

[0041] Therefore, according to the present invention, at least one additional image is preferably selected from each depth image sequence for further image processing, wherein, in all depth image sequences, the focal plane position 50 of the at least one additional image is located at the same relative distance from the best contrast focal plane position 50.

[0042] Preferably, more than one additional image is selected from each depth image sequence for further image processing, wherein the additional images together with the best contrast image have respective focal plane positions 50 shifted by equal distances in the depth direction.

[0043] In a particularly preferred embodiment, at least one additional pair of images is selected from the sequence of all depth images for further image processing, the images of the pair being located at the same distance above and below the best contrast focus plane position 50 .

[0044] Therefore, based on the above, the present invention relates to an image analysis method using an image analysis module having a neural network trained for image analysis, wherein an image captured using the image capturing method according to the present invention is input to the image analysis module.

[0045] By capturing additional images in each field of view, a so-called "extended training method" can optionally be used to train the image analysis neural network. This means, for example, that in an image analysis system that analyzes three images per analysis position, i.e., field of view or spatial region, five images are captured for training, with image combinations 1-2-3, 2-3-4, and 3-4-5 also used to train the image analysis module, with image 3 being the best-contrast image. This also trains the image analysis module to within ±1 plane of uncertainty in the best-contrast image. This means that potential one-plane errors in the focusing process can be eliminated, and the analysis result will be independent of these errors, i.e., it will be invariant to shifts of ±1 plane.

[0046] For the 1+2n images to be analyzed, it is preferred to take not only 1+2n+2 images (which is the minimum number sufficient for training), but even up to 1+4n images for the training process, so that the uncertainty of the best focus plane can be trained up to half of the range applied for analysis.

[0047] In order to specify the scope to be applied for extended training, it must be taken into account that, although more images contain more information, the more images the analysis algorithm receives, the more operations it has to perform during the calculations and the longer the duration of the analysis process.

[0048] It's preferable to choose a step size between the focal planes of the captured images that's close to the DOF value. This is because if the focal planes of the captured images are significantly farther apart than the DOF, truly sharp, optimally focused images may not be captured. If a distance as large as the DOF is used, then as long as the range of the focus series is chosen appropriately, i.e., it includes optimal focus, one or two very high-quality images will always be present. Furthermore, by applying extended training, independence from the constraint of a best focus error of at least + / - one plane can be achieved. In practice, it's advisable to obtain a 3-image input and 5-image training file, provided that the distance is chosen close to the DOF value.

[0049] The best results, the highest information content, are not necessarily obtained using images that are directly adjacent to the best-focused image. When implementing the analysis module, it is worthwhile to try the 3-input plane training process using images that are located + / - 2 or even 3 planes away from the best-focused image. Additionally, the further away the focal planes of the selected images are, the less similar those images are to the best-focused image (which is advantageous because those images have different information content), but the less sharp the images will be the further away they are from the best-focused plane (which is disadvantageous because the information contained in the images is less direct; images with focal planes that are actually far away from the best-focused plane can even be completely uniform, i.e., like the background, providing very little additional information).

[0050] Theoretical approaches indicate that the phase information content is optimal if images located at the same distance above and below the best-focused image are applied rather than applying a theoretically possible asymmetric image selection.

[0051] Of course, the invention can in principle operate with as few as two images, for example the best-focused image and another image taken at another focal plane.Even in this case, the analysis module receives additional information, wherein phase information is also available.

[0052] Therefore, the present invention also relates to a method for training an image analysis neural network, wherein the training of the network is carried out using images recorded by the image recording method according to the present invention.

[0053] As described above, preferably, for each spatial region 41 , 42 , 43 , 44 , 45 , training is performed using more captured images than the number of images to be input to the image analysis neural network for each spatial region 41 , 42 , 43 , 44 , 45 for image analysis.

[0054] It is particularly preferred to carry out the training process using at least two additional images which are recorded above and below the image to be input to the image analysis neural network for the image analysis.

[0055] The training process may also be performed with as many additional images as the number of images taken above and below the best contrast focal plane position 50 to be input to the image analysis neural network for image analysis.

[0056] The invention also relates to an image analysis neural network suitable for use in an apparatus for analyzing body fluids and trained by applying the method according to the invention.The first (input) layer of the neural network structure according to the invention is adapted to receive more than one input image.

[0057] For the implementation of the present invention, it is important to capture images of the same field of view from different planes quickly, i.e., the time between the images is as short as possible, so that the object to be analyzed, for example, urine sediment or bacteria, moves to the minimum possible extent during the image sequence. This is one reason why capturing images very far from the plane of optimal focus is not recommended; focusing on planes further away from the plane of optimal focus takes longer.

[0058] It is also important to ensure that the mechanical design of the system allows a sequence of images of the same field of view to be taken, so that the images taken at different focal planes are not shifted in the x and y directions, i.e. they actually show the same field of view. This is a condition for the correct operation of the invention.

[0059] Of course, the invention is not limited to the embodiments shown by way of example, but further modifications and variations are possible within the scope of the claims.The container according to the invention can be used not only for optical analysis of urine, but also for optical analysis of other liquids.

Claims

1. An image analysis method for analyzing an image by inputting the image into an image analysis module having a neural network trained for image analysis, the image being captured by an image capturing method comprising capturing an image with a microscope and selecting an image from the captured images for image analysis, wherein an image of a precipitate (31) of a liquid filled in an analysis space (40) located between a transparent upper window portion (11) and a transparent lower window portion (21) of a container (30) suitable for analyzing the liquid is captured, wherein the precipitate (31) is centrifuged on an inner flat surface of the lower window portion (21), wherein the image analysis method comprises: - in a first spatial region (41) of the analysis space (40), taking a first sequence of depth images at focal plane positions (50) shifted in equal steps in a depth direction perpendicular to the flat surface, and selecting an image with the best contrast from the first sequence of depth images for image analysis, - taking a second depth image sequence in a second spatial region (42) of the analysis space (40) at focal plane positions (50) shifted with equal steps in the depth direction, wherein the step size of the second depth image sequence is the same as or smaller than the step size of the first depth image sequence, wherein the second depth image sequence has fewer images than the first depth image sequence, and wherein one of the intermediate focal plane positions (50) of the focal plane positions (50) of the second depth image sequence is the same as the best contrast focal plane position (50) of the first depth image sequence, and selecting the image with the best contrast from the second depth image sequence for image analysis, and - further selecting at least one additional image from each depth image sequence for image analysis, wherein the focal plane position (50) of the at least one additional image is located at the same relative distance from the best contrast focal plane position (50) in each depth image sequence.

2. The image analysis method according to claim 1, wherein: The focal plane position (50) of the first depth image sequence having the best contrast is the same as the center focal plane position (50) of the second depth image sequence.

3. The image analysis method according to claim 1 or claim 2, characterized in that: Additional depth image sequences are recorded in further spatial regions (43, 44, 45) of the evaluation space (40), and the image with the best contrast is selected from each depth image sequence for image analysis.

4. The image analysis method according to claim 3, wherein: The additional depth image sequence is captured using the focal plane position (50) of the second depth image sequence.

5. The image analysis method according to claim 3, wherein: Each additional depth image sequence is captured using a focal plane position (50) that includes the best contrast focal plane position (50) of the previous depth image sequence as an intermediate focal plane position (50) and that is shifted in equal steps in the depth direction.

6. The image analysis method according to claim 5, characterized in that The center focal plane position (50) of the additional depth image sequence is the same as the best contrast focal plane position (50) of the previous depth image sequence.

7. The image analysis method according to claim 5 or 6, characterized in that: If the focal plane position (50) of the previous depth image sequence with the best contrast reaches, approaches a predetermined level, or exceeds a boundary of a predetermined depth range, the additional depth image sequence is captured using the focal plane position (50) of the first depth image sequence.

8. The image analysis method according to claim 7, wherein: If the best contrast focal plane position (50) of the previous depth image sequence is outside the range defined by the focal plane position (50) of the second depth image sequence, the additional depth image sequence is captured using the focal plane position (50) of the first depth image sequence.

9. The image analysis method according to claim 1, wherein: More than one additional image is selected from each depth image sequence for image analysis, wherein the additional images together with the image having the best contrast have corresponding focal plane positions (50) shifted by an equal distance in the depth direction.

10. The image analysis method according to claim 9, characterized in that At least one pair of images is selected from all depth image sequences for image analysis, wherein the images constituting the pair are located above and below the best contrast focal plane position (50) at the same distance from the best contrast focal plane position (50).

11. A training method for training an image analysis neural network using images captured by an image capturing method, the image capturing method comprising capturing images with a microscope and selecting images from the captured images for training, wherein an image of a precipitate (31) of a liquid filled in an analysis space (40) located between a transparent upper window portion (11) and a transparent lower window portion (21) of a container (30) suitable for analyzing the liquid is captured, wherein the precipitate (31) is centrifuged on an inner flat surface of the lower window portion (21), wherein the training method comprises: - taking a first sequence of depth images in a first spatial region (41) of the analysis space (40) at focal plane positions (50) shifted in equal steps in a depth direction perpendicular to the flat surface, and selecting an image with the best contrast from the first sequence of depth images for training, - taking a second depth image sequence in a second spatial region (42) of the analysis space (40) at focal plane positions (50) shifted with equal steps in the depth direction, wherein the step size of the second depth image sequence is the same as or smaller than the step size of the first depth image sequence, wherein the second depth image sequence has fewer images than the first depth image sequence, and wherein one of the intermediate focal plane positions (50) of the focal plane positions (50) of the second depth image sequence is the same as the best contrast focal plane position (50) of the first depth image sequence, and selecting the image with the best contrast from the second depth image sequence for training, and - further selecting at least one additional image from each depth image sequence for training, wherein the focal plane position (50) of the at least one additional image is located at the same relative distance from the optimal contrast focal plane position (50) in each depth image sequence.

12. The training method according to claim 11, characterized in that: For each spatial region (41, 42, 43, 44, 45), more captured images than the number of images to be input to the image analysis neural network for each spatial region (41, 42, 43, 44, 45) for image analysis are applied for the training.

13. The training method according to claim 12, characterized in that: The training is performed using at least two additional images taken above and below the image to be input to the image analysis neural network for image analysis.

14. The training method according to claim 13, characterized in that: The training is performed using as many additional images above and below the optimal contrast focal plane position (50) as the number of images to be input to the image analysis neural network for image analysis.

15. An image analysis neural network program product executed by a device suitable for analyzing body fluids, comprising a computer program, characterized in that When the computer program is executed by a processor or a computer, the training method according to any one of claims 11 to 14 is implemented.

Images