Improved method for fluorescence measurement of sample

Abstract

A method of performing fluorescence measurements on a sample is disclosed, the method comprising irradiating the sample with a light source, acquiring at least one fluorescence image of the irradiated sample, and processing the fluorescence image to determine a fluorescence intensity of the sample, characterized in that the fluorescence intensity of the sample is at least one of the fluorescence intensity of the sample. Processing the fluorescence image to determine the fluorescence intensity of the sample includes extracting an artifact-free target region (ROI) from the fluorescence image using a trained model, and determining the fluorescence intensity of the sample from the extracted target region.

Description

Technical Field

[0001] The present disclosure relates to methods and systems for performing fluorescence measurements on samples, which are particularly suitable for determining the presence of biological analytes or quantifying the analytes. Background Art

[0002] Fluorescence measurements are known to be used in many in vitro diagnostic systems and other biological applications to detect the presence of analytes or to quantitatively determine their concentration. Fluorescence is the ability of a substance to absorb electromagnetic radiation and subsequently emit light at a specific wavelength. Fluorescence measurements are therefore performed by irradiating a sample contained in a reading cuvette with an excitation wavelength selected to correspond to the excitation wavelength of the target analyte and detecting and measuring the fluorescence emission induced by the sample.

[0003] To obtain reliable values for the concentration of the target analyte, the fluorescence signal must be accurately detected and processed. However, the measurement of fluorescence emission can be affected by artifacts such as dust, air bubbles, or even the cuvette walls, which may have their own fluorescence contribution.

[0004] Document WO 2014 / 102502 discloses a method for analyzing a sample, comprising detecting a first fluorescence signal from a cuvette before introducing a reaction medium into the cuvette, and detecting a second fluorescence signal from the cuvette after introducing the reaction medium, and comparing the two signals to obtain a signal corresponding only to the reaction medium.

[0005] While this method is able to eliminate the fluorescence contribution from the cuvette, it does not ensure that the obtained signal is not affected by artifacts such as dust or bubbles. Summary of the Invention

[0006] The present disclosure is intended to improve upon the prior art.

[0007] In particular, an object of the present disclosure is to improve the accuracy of the fluorescent signal acquired on a sample.

[0008] Another object of the present disclosure is to be able to correct fluorescence measurements for contributions caused by artefacts such as bubbles or dust.

[0009] Therefore, a method for performing fluorescence measurement on a sample is disclosed, the method comprising:

[0010] Illuminate the sample with a light source.

[0011] acquiring at least one fluorescent image of the illuminated sample, and

[0012] Process the fluorescence images to determine the fluorescence intensity of the sample,

[0013] Characterized in that processing the fluorescence image to determine the fluorescence intensity of the sample comprises:

[0014] Extracting artifact-free target regions from fluorescence images using the trained model, and

[0015] Determine the fluorescence intensity of the sample from the extracted target region.

[0016] In embodiments, the method further comprises acquiring a plurality of fluorescence images of the sample using respective different combinations of illumination and exposure, selecting one of the plurality of fluorescence images, and processing the selected fluorescence image to determine the fluorescence intensity of the sample.

[0017] In an embodiment, selecting one of the plurality of fluorescent images comprises:

[0018] For each acquired fluorescence image, determine the number of saturated pixels, and

[0019] Fluorescence images with a number of saturated pixels below a predetermined threshold were selected.

[0020] In an embodiment, a plurality of fluorescence images are acquired at different illumination intensities, and the selected image is the fluorescence image acquired with the highest illumination intensity and a number of saturated pixels below a predetermined threshold.

[0021] In an embodiment, determining the fluorescence intensity of the sample comprises measuring the intensity of the fluorescent signal on the selected image, and deriving the fluorescence intensity of the sample from the intensity and the illumination and exposure conditions under which the selected image was acquired.

[0022] In an embodiment, the method further comprises extracting at least one other region from the fluorescence image corresponding to at least one type of artifact.

[0023] In an embodiment, determining the fluorescence intensity of the sample from the extracted target region includes calculating the fluorescence signal intensity of the target region by extrapolating the fluorescence signal intensity of the region to a region of the fluorescence image outside the target region.

[0024] In an embodiment, extracting the target region from the fluorescence image includes performing semantic segmentation on the fluorescence image.

[0025] In an embodiment, extracting the target region comprises applying a trained classification model to the fluorescence image, the classification model being configured to classify pixels of the fluorescence image according to a plurality of categories, the plurality of categories comprising at least:

[0026] A category corresponding to the absence of artifacts, and

[0027] At least one other category corresponding to one of the following artifacts:

[0028] bubble,

[0029] shadow,

[0030] dust,

[0031] background,

[0032] And wherein the target region is formed by pixels classified as corresponding to the absence of artifacts.

[0033] In an embodiment, the method further comprises a preliminary step of training a classification model by supervised learning on a training database comprising identifying regions corresponding to artifacts for each of a plurality of training fluorescence images, wherein each trained fluorescence image is rescaled by a randomly selected factor less than or equal to 1 and cropped to a constant size.

[0034] In an embodiment, the trained model is a convolutional neural network.

[0035] In an embodiment, the trained model is a segmented multi-scale attention network comprising:

[0036] Convolutional encoder,

[0037] intermediate modules configured to process the output of the convolutional encoder at multiple different scales, and

[0038] Decoder.

[0039] According to another aspect, a system for performing fluorescence measurements on a sample, wherein the sample is contained in a cuvette, is disclosed, the system comprising:

[0040] an irradiation device configured to irradiate the cuvette with at least one determined wavelength,

[0041] a detection device configured to acquire at least one fluorescence image comprising fluorescence emissions of the sample subsequent to illumination thereof by the light source, and

[0042] a computing device configured to process the fluorescence image to determine the fluorescence intensity of the sample,

[0043] It is characterized in that the system is configured to implement the method according to the above description.

[0044] The claimed method enables fluorescence measurement from a target region extracted from a fluorescence image, wherein the target region does not contain regions containing artifacts, thereby enabling more reliable fluorescence measurement.

[0045] The extraction of the target region can be achieved by applying a trained model configured to perform semantic segmentation on the image and classify each pixel or group of pixels in the image as belonging to a region free of artifacts or a region corresponding to a given type of artifact. This allows for automatic detection of artifacts even when their location in the image changes.

[0046] Once the region of interest is extracted and fluorescence measurements are taken of the region, the fluorescence measurements can be extrapolated to the excluded portions of the fluorescence image to obtain a complete fluorescence measurement, thereby calculating the fluorescence intensity of the sample.

[0047] In an embodiment, the accuracy of fluorescence intensity measurement is further improved by selecting a fluorescence image from which a target region is extracted among a plurality of fluorescence images acquired for different illumination or acquisition conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Other characteristics and advantages of the present invention will become apparent from the following detailed description given by way of non-limiting example and with reference to the accompanying drawings, in which:

[0049] Figure 1 shows an example of a system for performing fluorescence measurements on a sample.

[0050] Figure 2 Schematically represents the main steps of a method for performing fluorescence measurements according to one embodiment,

[0051] Figure 3a and 3b represents the first example of semantic segmentation of fluorescence images,

[0052] Figure 4a and 4b The second example represents semantic segmentation of fluorescence images,

[0053] Figure 5 Schematic representation of the structure of a neural network that can be used to extract the target region,

[0054] Figure 6a represents an example of the fluorescence signal measured on the target area, Figure 6b FIG. 4 shows an example of a fluorescence signal calculated on a region of an image excluding the target region based on the fluorescence signal measured on the target region.

[0055] Figure 7a represents the target region extracted from the fluorescence image, Figure 7b Represents an image reconstructed from the fluorescence signal calculated over the entire image based on the fluorescence signal measured on the target area.

[0056] Figure 8 Shown are comparative fluorescence signals acquired from a cuvette before and after the introduction of air bubbles and with and without the application of a method according to one embodiment. DETAILED DESCRIPTION

[0057] Detailed Description of At Least One Embodiment

[0058] refer to Figure 1 , shows an example of a system for performing fluorescence measurements on a sample. The sample may come from a variety of sources, such as food, environmental, veterinary, clinical, pharmaceutical or cosmetic sources.

[0059] Examples of food-derived samples include, but are not limited to, dairy products (yogurt, cheese, etc.), meat, fish, eggs, fruits, vegetables, water, and beverages (milk, juice, soda, etc.). Of course, these food-derived samples may also be derived from sauces or more complex meals, or from unprocessed or partially processed raw materials. Food samples may also be derived from animal feed, such as oil cakes and animal meal.

[0060] As previously mentioned, a biological sample may be of environmental origin and may consist of, for example, a surface sample, a water sample, etc.

[0061] The sample may also consist of a biological sample of clinical, human or animal origin, which may correspond to a biological fluid sample (urine, whole blood or derivatives such as serum, plasma, saliva, pus, cerebrospinal fluid, etc.), a stool sample (e.g. diarrhea caused by cholera), a sample from the nose, throat, skin, wound, organ, tissue or isolated cells. The above list is obviously not exhaustive.

[0062] In general, the term "sample" refers to a portion or quantity, more specifically a small portion or quantity, taken from one or more entities for analytical purposes. The sample may have undergone pre-treatment, for example, including mixing, dilution, or even comminution stages, particularly when the starting entity is in a solid state. The sample being analyzed may contain or be suspected of containing at least one analyte representative of the presence of the microorganism or disease to be detected, characterized, or monitored.

[0063] For the purpose of performing fluorescence measurements, the sample is placed in a cuvette 4. The cuvette comprises a bottom wall 40 and a side wall 41 extending from the bottom wall to an upper edge 43. The upper edge defines a hole 44 suitable for filling and emptying the cuvette.

[0064] In an embodiment, the bottom wall may include a bottom 401 extending perpendicular to the side wall, and an inclined wall 402 extending between the bottom 401 and the side wall 41, the inclined wall forming a 45° angle with the side wall. According to a non-limiting example, the bottom wall 40 may have a truncated cone shape. The cuvette may be made of a material suitable for storing liquids and other materials required for performing biological analysis. The cuvette 4 may be made of, for example, a plastic material. In addition, the cuvette may be transparent to the wavelength used to irradiate the sample, hereinafter referred to as the irradiation wavelength. It may also be transparent to the wavelength emitted by the sample due to fluorescence. For example, the cuvette may be made of polypropylene, glass, polymetil metacrilate, polystyrene, polycarbonate and other optical plastics, depending on the irradiation and fluorescence wavelength ranges.

[0065] The system 1 for performing fluorescence measurement includes an irradiation device 10, which includes a light source 11, such as a light emitting diode (LED). The light source 11 may include any monochromatic light source corresponding to the excitation peak wavelength of the analyte, i.e., the chemical molecule used as a marker or to be measured in the sample. Alternatively, the light source 11 can generate light of multiple wavelengths, such as white light, and the irradiation device may also include a filter 12 that can select at least one target wavelength corresponding to the excitation peak of the target molecule. The irradiation device may also include an optical element 13 that is suitable for making the irradiation beam conform to a suitable shape. The optical element may, for example, include at least one optical lens, such as an aspheric lens.

[0066] The system 1 for performing fluorescence measurements also includes a detection device 20, which is configured to acquire at least one fluorescence image of the sample, the fluorescence image comprising a fluorescence signal emitted by the sample sequentially with illumination by the light source. The detection device includes a detector 21, such as a camera, a high-sensitivity CMOS, a CCD or other 2D optical sensor. The detection device may also include an optical element 23 adapted to align the light beam emitted by the sample toward the detector in response to illumination. The optical element 23 may include an optical lens, such as an aspheric lens. In addition, the detection device may also include an optical filter 22 adapted to limit detection to a target wavelength or a narrowband spectrum centered on a target wavelength, which wavelength may typically be a fluorescence wavelength emitted by the analyte.

[0067] According to one non-limiting example, the illumination device can be configured to illuminate the cuvette 4 containing the sample with incident light at an angle of 90° relative to the cuvette wall, in particular, the cuvette inclined wall 402 at an angle of 45° relative to the cuvette side wall 41. The detection device 20 can be configured to collect light that leaves the cuvette through the inclined wall 402 and at an angle of 90° relative to the wall, so that the axes of the illumination device and the detector are located at 90° relative to each other.

[0068] According to another example, the illumination device and the detector can be located on the same side of the cuvette (i.e., the illumination optical axis and the detection optical axis form a zero angle), or they can be located on two opposite sides of the cuvette (i.e., the illumination optical axis and the detection optical axis form a 180° angle), because fluorescence is radiation emitted from sample molecules at 360°, regardless of the illumination direction. In the latter case, the wall on which the illumination light is incident and the fluorescence is measured can be the side wall 41.

[0069] The system 1 for performing fluorescence measurements also includes a computing device 30, which includes at least a computer 31 and a memory 32. The computer 31 is used to control the operation of the irradiation device and the detection device, and to receive the fluorescence image acquired by the detection device 20 from the latter. To this end, the computer 31 is connected to the detection device 20 and the irradiation device 10 by wire or wireless means. The computer 31 is also used to process the acquired fluorescence image to calculate the fluorescence intensity emitted by the sample in response to the irradiation. To this end, the computer can execute code instructions stored in the memory 32 to perform the method disclosed below. The computer may include one or more processors, such as a central processing unit CPU or a graphics processing unit GPU. The memory 32 may, for example, include a magnetic hard disk, a solid-state drive, an optical disk, an electronic memory or any type of computer-readable storage medium. The memory also stores a trained model that is configured to extract a target area from a fluorescence image, as described in more detail below.

[0070] refer to Figure 2 , the main steps of the method for performing fluorescence measurements on a sample are now described.

[0071] The method includes irradiating 100 a sample using an irradiation device 10. Irradiation can be performed at a selected wavelength that corresponds to the excitation wavelength of the target molecule to be detected, quantified, or analyzed by the fluorescence measurement, i.e., a wavelength suitable for inducing fluorescence. Irradiation can be continuous or in the form of multiple flashes. Using flashes can reduce sample degradation. Irradiation is also performed under specified irradiation intensity conditions. When the light source includes an LED, the excitation intensity can be controlled by the input current driving the LED.

[0072] The method further includes acquiring 200 at least one fluorescence image of the sample illuminated by the detection device 20. Since fluorescence begins almost instantaneously, acquisition 200 can be performed while the sample is being illuminated. Image acquisition is performed using defined acquisition parameters, including detector gain and integration time, i.e., the time window over which the camera collects light for a single image.

[0073] The acquired fluorescence image comprises a plurality of pixels, each corresponding to a point on the cuvette. Each pixel is associated with an intensity, which can be represented, for example, as a grayscale value or an RGB value, corresponding to the intensity of the fluorescence signal emitted by the cuvette. The fluorescence image is then processed to determine the fluorescence intensity of the sample from the fluorescence signal in the fluorescence image.

[0074] In an embodiment, the method includes acquiring 200 a plurality of fluorescence images corresponding to different sets of parameters regarding illumination and acquisition. The set of parameters that can be changed to change the exposure can include illumination intensity, sensor gain, sensor integration time.

[0075] In an embodiment, the method includes obtaining 200 multiple fluorescence images corresponding to different working points, wherein each working point is defined by illumination intensity and exposure, and exposure is determined by the integration time and camera gain of image acquisition. The working point can be set to provide an image with brightness value scaled by a known multiple. In fact, depending on the sample type and analyte concentration, the fluorescence in the cuvette can span a very wide range (more than four orders of magnitude). Therefore, the brightness, i.e., the light level captured by the pixel when reading with a given exposure and illumination intensity, can exceed the maximum readable value of the detector. Therefore, obtaining multiple fluorescence images under different exposures and illumination intensities can select the most suitable image for brightness, thereby providing the most information. Therefore, 2 to 10 fluorescence images, preferably 2 to 5 fluorescence images, can be obtained at different exposures and illumination intensities.

[0076] According to one embodiment, the number of images and operating points can be selected to cover the aforementioned fluorescence range. Thus, according to a non-limiting example, four images can be acquired, with the brightness values multiplied by four. The inverse ratio between the brightness obtained at each operating point and the brightness obtained at the most sensitive point is called the "exposure factor."

[0077] When multiple images are acquired in steps 100 and 200, the method further includes step 300 of selecting a fluorescence image from the multiple acquired images. The selection of the fluorescence image may be based on a condition regarding the number of saturated pixels in the image. If the intensity level of a pixel is the maximum intensity value that can be acquired by the sensor, the pixel is saturated. Therefore, the selection of the fluorescence image may include:

[0078] For each acquired image, count the number of saturated pixels, and

[0079] The brightest image with the number of saturated pixels below a determined threshold is selected.

[0080] To count the number of saturated pixels, a lower threshold can be considered. For example, a pixel is considered saturated if its intensity value is above 90% of the maximum intensity value that the sensor can acquire.

[0081] The method then includes processing the fluorescence image 400 to determine the fluorescence intensity of the sample. The fluorescence image herein refers to the acquired fluorescence image when one image has been acquired, or to the selected fluorescence image when multiple images have been acquired and one is selected in step 300.

[0082] Processing 400 the fluorescence image includes extracting 410 a target region ROI free of any artifacts from the fluorescence image, and determining 420 the fluorescence intensity of the sample from the extracted target region.

[0083] "Artifact" refers to a signal that is inconsistent with the fluorescence of the analyte being measured. In this disclosure, artifacts include any of the following:

[0084] The presence of bubbles in the cuvette may be due to the presence of other types of artifacts.

[0085] Dust particles may include particles or peals or other inhomogeneous materials commonly found in applications.

[0086] Shadows, such as those caused by bubbles in a cuvette,

[0087] Background, i.e. the image area that does not correspond to the illuminated area of the cuvette; in particular, the image portion corresponding to the cuvette wall (except for the wall extending perpendicular to the direction of the detection device and passed through by the fluorescence emitted by the sample) and the image portion surrounding the cuvette wall,

[0088] Extracting 410 the target region free of artifacts is achieved by applying a trained model. In an embodiment, the trained model is configured to perform semantic segmentation on the image, i.e., labeling each pixel of the image with a corresponding category. Here, the trained model is configured to label the pixels of the image with a corresponding category of at least two of the following categories:

[0089] A class corresponding to regions free of artifacts, and

[0090] A category corresponding to an illusion.

[0091] In a preferred embodiment, the model can be configured to label pixels of an image with corresponding categories from a plurality of categories, including one category corresponding to regions free of artifacts and a plurality of other categories corresponding to various artifacts.

[0092] For example, the categories may include:

[0093] At least one class corresponding to an artifact-free region,

[0094] Corresponding to a category of bubbles,

[0095] (passed by fluorescent light)

[0096] corresponds to a class of dust particles,

[0097] corresponds to a category of shadow,

[0098] A category corresponding to the background (everything not included in the other used categories).

[0099] Defining a category for each type of artifact allows for a more precise definition of the target area.

[0100] refer to Figure 3a and 3b , shows the first example of semantic segmentation of an image. According to this example, the categories are defined as follows:

[0101] For the cuvette area free of artifacts, two classes are defined, but located at different locations, one for bubbles,

[0102] A class is defined for shadows, and

[0103] One category is defined for "background" for all parts not included in the previous categories (including walls).

[0104] Figure 3a Displays the fluorescence image acquired by the detection device. Figure 3b The segmentation map output by the trained model for this image is shown. The bubble at the bottom of the cuvette is identified in the segmentation map, along with the shadow cast by the bubble in the direction of the light source. Only regions classified as artifact-free are retained to form the region of interest (ROI).

[0105] refer to Figure 4a and 4b , shows a second example of semantic segmentation of fluorescence images. According to this example, only two classes are defined, one corresponding to artifacts and the other corresponding to artifact-free regions. Figure 4a In , the acquired image shows the walls of the cuvette. Figure 4b In this image, the target area is located in the center of the image and the walls of the cuvette are excluded.

[0106] Before executing the model on fluorescence images, the model is trained 90 by supervised learning on a training database of fluorescence images, where each image in the database is annotated according to the desired class definitions, i.e., artifacts and target regions are labeled for each image. The fluorescence images in the training database can all be of the same size. Annotation can be performed manually by an operator. Model training can be performed by the system's computing device 30 or a separate computing device (not shown).

[0107] Furthermore, the model or its training can be designed to handle images of different scales.

[0108] According to one example, during training, each image can be randomly rescaled by a factor of 0 to 1 to better train the model to handle situations where objects and elements of varying scales are present. For example, each training image can be randomly rescaled by a factor selected from the following: 1, 0.8, or 0.6. Random crops of constant size are extracted from the rescaled images and used to train the model. Thus, the images used to train the model have a constant size, but the size of the artifacts may vary.

[0109] As for the trained model, a convolutional neural network is preferred.

[0110] In an embodiment, to process images of different scales, a convolutional neural network may include:

[0111] Convolutional encoder part,

[0112] Multi-scale modules, i.e., intermediate modules configured to process the output of the convolutional encoder at multiple different scales, and

[0113] Convolutional decoder part.

[0114] In an embodiment, the neural network can be a segmented multi-scale attention network (SMANet), whose structure is as follows Figure 5 As shown in and disclosed in detail in the following article Simone Bonechi et al. “Weak supervision for generating pixel-level annotations in scene text segmentation”, Pattern Recognition Letters, Volume 138, 2020, Pages 1-7, ISSN 0167-8655. The convolutional encoder part of the network corresponds to a ResNet neural network, such as ResNet50, in which the convolution is replaced by dilated (i.e., hole) convolution to expand the receptive field of the neural network. The multi-scale module includes a pyramid of hole convolutions with different dilation rates, each followed by spatial pyramid pooling, i.e., parallel pooling of multiple different block sizes. The decoder part includes two decoder stages with skip connections between the encoder and decoder parts. This neural network developed for scene text segmentation is relevant to applications in processing fluorescent images, because images containing text have difficulties in recognizing the size and position of letters with artifacts similar to those in this case.

[0115] Once the target region is extracted at step 410 , the method comprises determining 420 an intensity signal of a sample from the target region.

[0116] Determining the intensity signal of the sample from the target region includes extracting intensity values for pixels of the extracted target region, the values corresponding to the intensity of the fluorescence signal at each pixel; and inferring the intensity of the fluorescence signal across the entire image from which the target region was extracted from the values by extrapolation or interpolation. According to one embodiment, this is performed by fitting a polynomial function to the intensity values of the pixels of the target region and deriving, from the polynomial function, approximate values for the intensity values of pixels removed from the image.

[0117] refer to Figure 6a and6b , shows an example of a polynomial approximation of a target region, displayed in three dimensions. The z-axis corresponds to the measured intensity of a pixel at a location on the image defined by coordinates (x,y). Figure 6a Displays the intensity of the fluorescent signal in the target area. Figure 6b The polynomial function fitted to the intensity is represented by a continuous surface in . The points in the figure represent the original intensity values of the pixels in the target area.

[0118] exist Figure 7a and 7b Another example is shown in two dimensions in Figure 7a Displays the target region extracted from the real fluorescence image (i.e., artifacts have been removed), Figure 7b A reconstructed fluorescence image is displayed in which the areas corresponding to the artifacts have been replaced by pixels whose values were determined according to the fitted polynomial function.

[0119] A specific experiment was designed to evaluate the effectiveness of the proposed method. Specifically, a set of fixed dilution cuvettes was collected, where the same cuvette was obtained before and after inoculation with artificial bubbles. The results were Figure 8 The figure shows a graph with the horizontal axis representing each sample measured and the vertical axis representing the measured intensity of the sample. It can be seen how this method significantly improves the measurement by reducing the influence of artifacts. Four different measurement strategies are shown and compared on this graph. Line A corresponds to the measurement performed in a cuvette without bubbles. Line B corresponds to the measurement performed after introducing bubbles into the cuvette, without implementing the above method. It can be seen that the presence of the defect leads to a large gap between the two measurements. Line C corresponds to the fluorescence intensity measurement of the sample obtained only from the target area, i.e. the values of the removed image areas are not evaluated. Line D then corresponds to the fluorescence intensity measurement of the sample, where the values for the area containing the defect are deduced from the values for the target area. It can be seen that lines A and D are essentially at the same level, so the above method allows accurate compensation of errors induced by the presence of defects in the cuvette.

[0120] Once the polynomial function is fitted to the intensity values of the target region, the fluorescence intensity of the sample can be calculated as follows:

[0121] The raw intensity values of the pixels in the target area, and

[0122] The intensity values of pixels within the artifact are approximated by fitting a polynomial function.

[0123] In an embodiment, the fluorescence intensity of the sample is calculated as the average of the intensity values of the pixels contained in the above-mentioned regions (the initially identified regions and the regions approximated by the polynomial fit).

[0124] After extracting the resulting brightness value, the value is multiplied by the exposure factor corresponding to the image acquisition as defined above to obtain the actual value of the fluorescence intensity. When multiple images are acquired and one image is selected in step 300, the value is multiplied by the exposure factor of the selected image.

Claims

1. A method for measuring fluorescence of a sample, the method comprising: The sample is illuminated (100) using a light source, acquiring (200) at least one fluorescent image of the illuminated sample, and processing (400) the fluorescence image to determine the fluorescence intensity of the sample, Characterized in that processing the fluorescence image to determine the fluorescence intensity of the sample includes: extracting (410) an artifact-free target region from the fluorescence image using the trained model, and The fluorescence intensity of the sample is determined (420) from the extracted target region.

2. The method of claim 1 , comprising acquiring a plurality of fluorescence images of the sample using respective different combinations of illumination and exposure, selecting one of the plurality of fluorescence images, and processing the selected fluorescence image to determine the fluorescence intensity of the sample.

3. The method of claim 2, wherein selecting one of the plurality of fluorescent images comprises: For each acquired fluorescence image, determine the number of saturated pixels, and Fluorescence images with a number of saturated pixels below a predetermined threshold were selected. 4 . The method according to claim 3 , wherein the plurality of fluorescence images are acquired at different illumination intensities, and the selected image is the fluorescence image acquired with the highest illumination intensity and the number of saturated pixels being below a predetermined threshold.

5. The method according to any one of claims 2 to 4, wherein determining the fluorescence intensity of the sample comprises measuring the intensity of the fluorescence signal on the selected image, and deriving the fluorescence intensity of the sample from the intensity and the illumination and exposure conditions for acquiring the selected image. 6 . The method according to claim 1 , further comprising extracting from the fluorescence image at least one further region corresponding to at least one type of artifact.

7. The method according to any one of the preceding claims, wherein determining the fluorescence intensity of the sample from the extracted target region comprises: The fluorescence signal intensity of the target region is extrapolated to a region of the fluorescence image outside the target region, thereby calculating the fluorescence signal intensity of the region. 8 . The method according to claim 1 , wherein extracting the target region from the fluorescence image comprises performing semantic segmentation on the fluorescence image.

9. The method of the preceding claim, wherein extracting the target region comprises applying a trained classification model to the fluorescence image, the classification model being configured to classify pixels of the fluorescence image according to a plurality of classes, the plurality of classes comprising at least: A category corresponding to the absence of artifacts, and At least one other category corresponding to one of the following artifacts: bubble, shadow, dust, background, And wherein the target region is formed by pixels classified as corresponding to the absence of artifacts.

10. The method according to any of the preceding claims, further comprising a preliminary step of training the classification model by supervised learning on a training database comprising, for each of a plurality of training fluorescence images, identifying regions corresponding to artifacts, wherein each trained fluorescence image is rescaled by a randomly selected factor less than or equal to 1 and cropped to a constant size.

11. The method according to any of the preceding claims, wherein the trained model is a convolutional neural network.

12. The method according to any one of the preceding claims, wherein the trained model is a segmentation multi-scale attention network, comprising: Convolutional encoder, an intermediate module configured to process the output of the convolutional encoder at a plurality of different scales, and Decoder.

13. A system (1) for performing fluorescence measurements on a sample, wherein the sample is contained in a cuvette (4), the system comprising: an irradiation device (10) configured to irradiate the cuvette with at least one determined wavelength, A detection device (20) configured to acquire at least one fluorescence image comprising fluorescence emissions of a sample subsequent to illumination thereof by a light source, and a computing device (30) configured to process the fluorescence image to determine the fluorescence intensity of the sample, It is characterized in that the system is configured to carry out the method according to one of the preceding claims.

Images