Apparatus and method for white blood cell counting in extravascular body fluids

Abstract

Apparatus and method for white blood cell counting in extravascular body fluids. The apparatus (100) comprises an ultrasound device (110) for obtaining ultrasound images (117) from a body or sample, and an image processing device (120) configured to analyse the ultrasound images (117) to determine a level of attenuation (123), modify a parameter of an excitation signal (113) of the ultrasound device (110) until the level of attenuation (123) is within a predetermined range, and then estimate a concentration of white blood cells (130) from the ultrasound images (117) using a first convolutional neural network (127) trained with images with an assigned ground truth concentration. The apparatus (100) may also analyse the ultrasound images (117) to determine an effective focal position (610) and modify the focal position of the ultrasound signal (115), by actuating a focusing unit (602) of the ultrasound device (110), until the effective focal position (610) is within a range.

Description

[0001] DESCRIPTION

[0002] APPARATUS AND METHOD FOR WHITE BLOOD CELL COUNTING IN EXTRAVASCULAR BODY FLUIDS

[0003] FIELD OF THE INVENTION

[0004] The invention belongs to the field of devices and methods for estimating the concentration of white blood cells in extravascular body fluids, such as blood plasma, urine, transcellular fluids (cerebrospinal fluid, aqueous humour), joint fluids (synovial fluid), serous body fluids (peritoneal fluid, pleural fluid, saliva).

[0005] BACKGROUND OF THE INVENTION

[0006] Patent document WO2024125831-A1 , the content of which is herein included by reference in its entirety, discloses a system and method for obtaining a concentration of white blood cells in superficial body fluids. The present invention improves the system and method therein disclosed by providing an automatic adjustment of the excitation signal of the ultrasound transducer, which allows optimizing the received signal quality and sensitivity to enable and obtain a more precise estimation of the concentration of white blood cells.

[0007] BRIEF DESCRIPTION OF THE INVENTION

[0008] The present invention refers to an apparatus and method for white blood cell counting in extravascular body fluids.

[0009] The apparatus comprises an ultrasound device and an image processing device. The ultrasound device includes a driving unit configured to excite an ultrasound transducer with an excitation signal; the ultrasound transducer, configured to acquire at least one ultrasound signal from a body or a sample of extravascular body fluid; and an ultrasound scanner configured to generate at least one ultrasound image from the at least one ultrasound signal.

[0010] The image processing device comprises a first image processing unit configured to determine a level of attenuation of the at least one ultrasound signal by analysing the at least one ultrasound image; a control unit configured to actuate the driving unit of the ultrasound device to modify at least one parameter of the excitation signal until the level of attenuation is within a predetermined range; and a second image processing unit configured to estimate a concentration of white blood cells from the at least one ultrasound image, generated when the level of attenuation is within the range, using a first convolutional neural network trained with images with an assigned ground truth concentration. The white blood cell counting may be performed according to the prior art (e.g. as described in WO2024125831 -A1).

[0011] In an embodiment, the apparatus is a stand-alone device comprising a tray configured to support the sample of extravascular body fluid, and a chamber arranged under the tray and adapted to house a coupling material and the ultrasound transducer. The apparatus may further comprise a motor configured to move the ultrasound transducer within the chamber in an axial direction to change the focal position depth_and / or in a rotational direction to collect more in-plane data and obtain faster measurements (when compared to the transducer being at a fix position). The tray may include a concave region adapted to receive a bag (such as a peritoneal dialysis drainage bag or a urine drainage bag) containing a sample of extravascular body fluid, the ultrasound transducer being arranged under the concave region. The concave region may include an off-centered vertex, the ultrasound transducer being arranged under the vertex.

[0012] In another embodiment, the apparatus may be integrated in another system, such as an automated peritoneal dialysis machine. The ultrasound transducer may be connected, for instance, to a cassette of a cycler through a coupling layer.

[0013] Another aspect of the present invention refers to a method for white blood cell counting in extravascular body fluids, the method comprising the following steps: exciting an ultrasound transducer with an excitation signal; acquiring, by the ultrasound transducer, at least one ultrasound signal from a body or a sample of extravascular body fluid; generating at least one ultrasound image from the at least one ultrasound signal; determining a level of attenuation of the at least one ultrasound signal by analysing the at least one ultrasound image; modifying at least one parameter of the excitation signal until the level of attenuation is within a predetermined range; and estimating a concentration of white blood cells from the at least one ultrasound image, generated when the level of attenuation is within the range, using a first convolutional neural network trained with images with an assigned ground truth concentration.

[0014] A further aspect of the present disclosure refers to a computer program product for white blood cell counting in extravascular body fluids, comprising computer code instructions that, when executed by a processor, causes the processor to perform the steps of the method. The computer program product may comprise a non-transitory computer-readable storage medium having recorded thereon the computer code instructions.

[0015] BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To complete the description and in order to provide for a better understanding of the invention, a set of drawings is provided. Said drawings form an integral part of the description and illustrate an embodiment of the invention, which should not be interpreted as restricting the scope of the invention, but just as an example of how the invention can be carried out. The drawings comprise the following figures:

[0017] Figure 1 depicts an apparatus for white blood cell counting in extravascular body fluids, according to an embodiment.

[0018] Figure 2 is a flow diagram of a method for white blood cell counting in extravascular body fluids, according to an embodiment.

[0019] Figure 3A depicts ultrasound images in high resolution of infected peritoneal fluid in peritoneal dialysis drainage bags with different level of attenuation. Figure 3B depicts high- resolution ultrasound images with different thickness of fontanelle tissue. Figure 3C represents high-resolution ultrasound images with a different level of attenuation caused by the hair covering the fontanelle tissue.

[0020] Figure 4A describes a way of determining a level of attenuation according to an embodiment. Figures 4B and 40 describe a way of detecting white blood cell traces in the ultrasound image. Figure 4D shows cropped areas of ultrasound images with visible white blood cell traces.

[0021] Figures 5A and 5B represent another way of determining a level of attenuation.

[0022] Figure 6 depicts an embodiment of an apparatus including the additional feature of an automatic adjustment of the effective focal position.

[0023] Figures 7A and 7B depict the correction of the effective focal position.

[0024] Figure 8A is a flow diagram of a method for white blood cell counting including an automatic adjustment of the effective focal position. Figure 8B is an embodiment of a way to determine the effective focal position in an ultrasound image.

[0025] Figures 9A and 9B represent an embodiment of a stand-alone apparatus for measuring the concentration of white blood cells.

[0026] Figure 10 depicts another embodiment of the stand-alone apparatus.

[0027] Figure 11 represents an apparatus integrated in a cycler.

[0028] Figures 12A and 12B depict quality defects in the ultrasound image detectable by the reflections of the plastic bag containing the sample.

[0029] Figure 13 is another embodiment of the method that includes previous quality checks.

[0030] Figure 14 shows an embodiment of a quality check corresponding to the case described in Figure 12A.

[0031] Figure 15 shows an embodiment of a quality check corresponding to the case described in Figure 12B.

[0032] Figures 16A and 16B show another embodiment of the quality check.

[0033] Figures 17A and 17B show yet another embodiment of the quality check.

[0034] Figures 18A to 18C show an embodiment of the quality check in which the gain of the ultrasound signal is automatically modified based on the clutter detected in the ultrasound image.

[0035] Figure 19 represents a flow diagram of a method including the main feature of the present invention (dynamic ultrasound signal configuration) and two additional and optional features: a quality check and an automatic focal position adjustment.

[0036] DETAILED DESCRIPTION OF THE INVENTION

[0037] Figure 1 illustrates the main components of an apparatus 100 for white blood cell counting in extravascular body fluids. The apparatus 100 comprises an ultrasound device 110 and an image processing device 120. The apparatus 100 may be battery-operated and may include a display (not shown in Figure 1).

[0038] The ultrasound device 110 includes a driving unit 112, an ultrasound transducer 114 and an ultrasound scanner 116. The driving unit 112 excites the ultrasound transducer 114, applying an excitation signal 113, or driving signal, with specific parameters (e.g. pulse repetition frequency, pulse amplitude, pulse duration, pulse excitation frequency). The ultrasound transducer 114 acquires at least one ultrasound signal 115 from a body or a sample of extravascular body fluid, and the ultrasound scanner 116 generates at least one ultrasound image 117 from the at least one ultrasound signal 115.

[0039] The ultrasound images 117 may be obtained from a human body, such as the head 102 of a newborn baby (in particular, the fontanelle tissue), an eye 104 (in particular, a closed eye or an eyelid), the bladder or a joint of a patient. When the ultrasound device 110 is applied to the head of a newborn baby, ultrasound images 117 of the cerebrospinal fluid located between the fontanel tissue and the cortex are generated. When applied to an eye of a patient, ultrasound images 117 of the aqueous humour are obtained. When applied to the bladder of a patient, ultrasound images 117 of urine in the bladder are obtained. When applied to a joint of a patient (e.g. knee, elbow), ultrasound images 117 of synovial fluid are obtained. Alternatively, the ultrasound device 110 may be directly applied to a sample of extravascular body fluid (e.g. serous body fluid, urine, blood plasma, saliva) previously extracted from a patient. The sample is preferably contained in a container, such as a plastic test tube 106 or a plastic bag (e.g. a peritoneal dialysis drainage bag 108, a urine drainage bag). The ultrasound device 110 does not need to contact the sample, it may instead contact the container. When the ultrasound device 110 is applied to a peritoneal dialysis drainage bag 108, ultrasound images 117 of the peritoneal fluid contained therein are generated.

[0040] The image processing device 120 includes a first image processing unit 122, a control unit 124 and a second image processing unit 126. The first image processing unit 122 analyses the at least one ultrasound image 117 and determines a level of attenuation 123 of the at least one ultrasound signal 115 based on said analysis. In an embodiment, the first image processing unit 122 determines a level of attenuation 123 for each ultrasound image 117 received. The first image processing unit 122 may process a set of ultrasound images 117 sequentially generated by the ultrasound scanner 116 and determine an overall level of attenuation 123 (e.g. an average) from said images.

[0041] The control unit 124 generates an excitation control signal 125 that actuates the driving unit 112 of the ultrasound device 110 to modify at least one parameter of the excitation signal

[0042] 113 (such as pulse amplitude, pulse duration, pulse repetition frequency, pulse excitation frequency, or a combination thereof) until the level of attenuation 123 is within a predetermined range. To that end, the image processing device 120 continuously receives and analyses ultrasound images 117 generated by the ultrasound device 110 and modifies the excitation signal 113 in a feedback loop until the level of attenuation reaches a certain range.

[0043] The second image processing unit 126 is configured to estimate or calculate a concentration of white blood cells 130 from the at least one ultrasound image 117, generated when the level of attenuation 123 is within the predetermined range, using a first convolutional neural network 127 trained with images with an assigned ground truth concentration.

[0044] In an embodiment, to determine a level of attenuation 123 of each ultrasound signal 115 the first image processing unit 122 uses the first convolutional neural network 127 (although a different convolutional network may be employed), whose weights are obtained using training images with an assigned ground truth concentration. The first image processing unit 122 and the second image processing unit 126 may be implemented in a single image processing unit (i.e. the first image processing unit 122 and the second image processing unit 126 may be the same image processing unit). The image processing device 120, that includes the control unit 124 and the image processing units (122,126), may be implemented, for instance, in one or more processors (e.g. one or more CPUs, one or more GPUs, one or more FPGAs, or a combination thereof).

[0045] The present invention also refers to a computer-implemented method for white blood cell counting in extravascular body fluids. In the embodiment shown in Figure 2 the method 200 comprises the following steps:

[0046] Exciting 210 an ultrasound transducer 110 with an excitation signal 113.

[0047] - Acquiring 220, by the ultrasound transducer 110, at least one ultrasound signal 115 from a body (e.g. the head 102 or an eye 104 of a patient) or a sample of extravascular body fluid (e.g. contained in a peritoneal dialysis drainage bag 108). Generating 230 at least one ultrasound image 117 from the at least one ultrasound signal 115.

[0048] Determining 240 a level of attenuation 123 of the at least one ultrasound signal 115 by analysing the at least one ultrasound image 117.

[0049] Comparing 245 the level of attenuation 123 with a predetermined range (or threshold) and modifying 250 at least one parameter of the excitation signal 113 if the level of attenuation 123 is outside the range. The previous steps (210,220,230,240,245) are repeated until the level of attenuation 123 is within a predetermined range (e.g. until it is lower than a threshold). If at the first comparison the level of attenuation 123 is within the established range, the excitation signal 113 is not modified and the previous steps are not repeated.

[0050] Estimating 260 a concentration of white blood cells 130 from the at least one ultrasound image 117 generated when the corresponding level of attenuation 123 is within the range. The estimation is performed using a first convolutional neural network 127 trained with images with an assigned ground truth concentration.

[0051] In case of high signal attenuation, the level of attenuation 123 detected by the first image processing unit 122 will be higher than a predetermined, desired range. Alternatively, an effective signal strength reaching the focal area may be determined; in that case, the effective signal strength will be lower than a predetermined range. Determining the level of attenuation 123 or the effective signal strength reaching the focal area are equivalent ways to express the ultrasound signal attenuation (they can be considered inversely proportional). The range may be defined by a single threshold (e.g. range=[0, threshold]) or a plurality of thresholds (e.g. range=[ thresholdl , threshold2]).

[0052] If the initial level of attenuation (i.e. the level of attenuation 123 first computed by the first image processing unit 122), obtained from the initial excitation signal 113 set by the driving unit 112, is outside the desired range, the control unit 124 actuates the driving unit 112 to modify the excitation signal 113 so that the level of attenuation 123 of the next ultrasound image 117 is closer to the predetermined range. Normally, the initial level of attenuation 123 will be higher than desired, and the excitation signal 113 will be repeatedly modified (e.g. the value of some of its parameters increased) to reduce the level of attenuation 123 until it is lower than a threshold, e.g. by increasing the pulse excitation frequency, the pulse repetition frequency, the pulse amplitude, and / or the pulse duration by a certain quantity (which, for instance, may be a fixed or proportional quantity or a variable quantity, e.g. dependent on the distance to the threshold). This way, the control unit 124 actuates the driving unit 112 in a feedback loop until a desired output (a desired level of attenuation or a desired effective signal strength reaching the focal area) is obtained.

[0053] In an embodiment, the control unit 124 may include a proportional controller, a proportional- derivative controller or a proportional-integral-derivative (PID) controller, wherein the level of attenuation is compared to a desired output (a threshold) and the difference (error) is fed to the controller to compute the excitation control signal 125 that will modify the excitation signal 113 to increase the effective signal strength reaching the focal area and, therefore, reduce the level of attenuation and the error.

[0054] In many cases, the ultrasound signal 115 is more attenuated than desired. Figures 3A, 3B and 3C depict several examples in which there is an important attenuation in the ultrasound signal strength.

[0055] When the present invention is applied to estimate a concentration of white blood cells where the ultrasound signal attenuation is not normally high, such as in the eye 104 of a patient or in a sample (e.g. peritoneal dialysis drainage bag 108), the displacement of leukocytes within the fluid areas can normally be seen in the ultrasound images 117, especially in high- resolution ultrasound images. The ultrasound images 117 are preferably high-resolution images (e.g. ultrasound images at a second spatial resolution as defined in WO2024125831 -A1). Figure 3A depicts ultrasound images 117 of a peritoneal dialysis drainage bag 108 corresponding to an ultrasound signal 115 with low signal strength (left image) and high signal strength (right image). In the left image the signal is more attenuated leading to fewer white blood cell traces being visible and, at the same time, appearing less tilted, while in the right image the signal is less attenuated leading to the opposite, more white blood cell traces are visible and appear more tilted. In the case of leukocytes contained in peritoneal dialysis drainage bags 108, the signal can be attenuated by changes in the composition of the serous body fluid contained therein due to infection, by the composition of the material used to fabricate the drainage bag or by bad coupling, among other factors. In both cases, the displacement of leukocytes, or white blood cell traces 310, can be seen at the focal area of the ultrasound image 117.

[0056] However, the white blood cell traces 310 are more clearly shown for a high signal strength, where a high number of traces can also be distinguished. In addition, the angle of tilt a of the white blood cell traces 310 with respect to a horizontal line is higher for a high signal strength. The thickness (or width) of the white blood traces 310 is also affected by the ultrasound signal strength: the lower the signal strength, the less inclined the traces appear and the thicker they are. The length of the white blood traces 310 is also affected by the level of attenuation: if there is less attenuation, more signal arrives and therefore the white blood cell traces can already be seen outside the focus area, whereas if there is more attenuation, the traces they can only be seen in the focus area and therefore they are shorter. The height of the focal area is also affected by the level of attenuation (it decreases with higher levels of attenuation). Therefore, the lower the level of attenuation 123:

[0057] The higher the angle of tilt a of the white blood traces.

[0058] The lower the thickness of the white blood traces.

[0059] The higher the length of the white blood traces.

[0060] The higher the height of the focal area.

[0061] The present invention can also be applied to estimate a concentration of white blood cells where the ultrasound signal attenuation is normally high, such as the head 102 of a newborn baby. Given a good coupling, the effective acoustic pressure reaching the area of interest depends a lot on the medium (or media) it passes through. In the case of cerebrospinal fluid of newborns, the ultrasound signal is absorbed and highly attenuated by the hair and the layers of fontanel tissue that the acoustic beam must pass through before reaching the cerebrospinal fluid. Figure 3B depicts ultrasound images 117 with different thickness of fontanelle tissue 320: a thick fontanelle (with thickness THi, left image) and a thin fontanelle (with thickness TH2, right image), wherein THI»TH2. In the case of the fontanelle tissue 320, the thicker the tissue, the higher the signal attenuation.

[0062] Figure 3C represents another example of an important reduction in the ultrasound signal strength due to hair covering the fontanelle tissue. The presence of hair leads to a higher signal attenuation and a signal refraction, which corresponds with tissue saturation (which can be detected when the tissue does not appear saturated or white at all). Left image of Figure 3C depicts a non-saturated fontanelle tissue (not white) which corresponds to a high ultrasound signal attenuation caused by the hair. However, the right image depicts a saturated fontanelle tissue 320 (white) corresponding to lower signal attenuation (caused by less quantity of hair covering the fontanelle). The higher the saturation, the lower the level of attenuation. In the cases depicted in Figures 3B and 3C, the white blood cell traces are hardly distinguishable due to the overall very high attenuation.

[0063] For the examples depicted in Figures 3A and 30, a level of attenuation 123 of the ultrasound signal(s) 115 is determined by analysing the corresponding ultrasound image(s) 117. The level of attenuation 123 will be determined depending on each particular case. Figures 4A and 5A depict different ways to determine the level of attenuation 123 for the examples depicted in Figures 3A and 30, respectively. Figure 4A is a flow diagram of the step of determining 240 a level of attenuation 123 for the cases in which the ultrasound signal attenuation is not very high (e.g. the one shown in Figure 3A). According to this embodiment, the step of determining 240 a level of attenuation 123 comprises detecting 410 white blood cell traces 310 in the at least one ultrasound image 117, calculating 420 at least one feature of the white blood cell traces, and obtaining 430 a level of attenuation 123 based on the calculated feature (or features), wherein the level of attenuation 123 can be the value of the feature itself.

[0064] The at least one feature of the white blood cell traces (310 includes the angle of tilt (a) of the white blood cell traces, the thickness of the white blood cell traces, the length of the white blood cell traces, the height of the focal area in which the white blood cell traces (310) are visible, an average level of intensity of the white blood cell traces, or a combination thereof. The embodiment of Figure 4A is particularly used to estimate the concentration of white blood cells in applications where the ultrasound signal attenuation is not very high, such as counting white blood cells in the aqueous humour in the eye 104 of a patient or in a sample of extravascular body fluid (e.g. peritoneal dialysis drainage bag 108), so that the white blood cell traces 310 can be somehow identified in the ultrasound image 117. The angle of tilt, thickness and length of the white blood cell traces, and the height of the focal area, is representative of the level of attenuation 122. For instance, the angle of tilt a depicted in Figure 3A is inversely proportional to the level of attenuation 123, i.e. the higher the angle of tilt a, the lower the level of attenuation 123 of the ultrasound signal 115. However, the angle of tilt considered may be instead the angle that the white blood cell trace 310 forms with a vertical line (angle of tilt p=90-a); in that case the angle of tilt p would be proportional to the level of attenuation 123, and the level of attenuation 123 obtained can be the angle of tilt itself. When comparing 245 the level of attenuation with a range (or a threshold), the calculated feature itself (e.g. thickness or angle of tilt) may be compared instead. In an embodiment, the level of attenuation is obtained from the feature by using a predefined constant, such that L=K f, wherein L represents the level of attenuation, K is the constant and f is a value of the feature. A combination of features may also be used to compute the level of attenuation; for instance, L=Krfi+K2-f2, wherein Ki and K2 are constants, fi and f2 are the corresponding values of a first feature (e.g. length) and a second feature (e.g. thickness). Other mathematical operations may be used to obtain the level of attenuation.

[0065] According to an embodiment depicted in Figure 4B, the step of detecting 410 one or more white blood cell traces 310 in the at least one ultrasound image 117 comprises the following steps:

[0066] Obtaining 412 an activation map of a predetermined layer of the first convolutional neural network 127, the input of the first convolutional neural network 127 being the ultrasound image 117 (or a preprocessed image obtained from the ultrasound image 117).

[0067] Binarizing 414 the activation map 416 using a binarization threshold, obtaining a binarized activation map 415.

[0068] Detecting 416 one or more white blood cell traces 310 in the binarized activation map 415. This detection may include a previous step of performing a morphological closing operation on the binarized activation map 415. The morphological closing operation is useful for closing small holes inside the traces. Structures obtained after applying this operation may be detected as white blood cell traces 310 when they contain at least a predetermined number of pixels.

[0069] Once the white blood cell traces 310 are detected, a feature (e.g. an angle of tilt) of one or more white blood cell traces 310 is calculated 420. The level of attenuation 123 if obtained from the calculated feature. For example, a single angle of tilt (corresponding to the angle of a white blood cell trace) may be calculated to obtain the level of attenuation. In another embodiment, several angles of tilt are calculated (the angles corresponding to different detected white blood cell traces), and the level of attenuation is obtained using said angles of tilt; for instance, the level of attenuation can be obtained from the average of the angles of tilt.

[0070] In an embodiment depicted in Figure 4C, for detecting 410 the white blood cell traces 310 the first convolutional neural network 127 (based for instance on a MobileNetV3 architecture or a ResNet architecture), which predicts the number or concentration of white blood cell traces present in an ultrasound image, is trained based on previously labelled data giving information about the number or concentration of white blood cell traces present in an ultrasound image and it is preferably optimized via a mean squared error cost function. In an embodiment, the input of the first convolutional neural network 127 is a Sobel filtered cropped ultrasound image 411 , depicted in the upper image of Figure 4C, which is obtained after cropping the ultrasound image 177 and applying a Sobel filter. Once the first convolutional neural network 127 is trained, activation maps (or feature maps) of the different layers of the first convolutional neural network 127 can be extracted. The areas which are activated in specific layers of the first convolutional neural network 127 correspond to the white blood cell traces 310 visible in the ultrasound images. The activation map of the first layer of the first convolutional neural network 127 is preferably used, since it faithfully represents the white blood cell traces 310 (the first convolutional neural network 127 has learned that these traces need to be searched for in order to perform the white blood cell counting).

[0071] The middle image in Figure 4C shows the activation map 413 of the first layer of the first convolutional neural network 127 used for the automatic counting of the number of cells, although different layers may be used instead. The activation map 413 is then binarized with a threshold; for example, the activation map 413 can be first scaled to the range (0,1) and then binarized using a predetermined threshold (e.g. of 0.25). The lower image of Figure 4C shows the binarized activation map 415 used for calculating a feature, in particular the angle of tilt a of the white blood cell trace 310. White blood cell traces 310 can be extracted from the binarized activation map 415 by performing a morphological closing operation. Structures obtained after applying the morphological closing operation are considered as white blood cell traces 310 when containing a minimum number of pixels (e.g. 75 pixels). Once the white blood cell traces 310 are detected, the value of the feature (thickness, tilt, length) can be calculated. For instance, the angle of tilt a can be calculated 420 with standard trigonometric calculations, using for example the line 421 connecting the starting point 417 and the ending point 419 of a white blood cell trace 310. The length can be calculated by measuring the distance, e.g. in number of pixels, between the starting point 417 and the ending point 419.

[0072] Figure 4D shows cropped areas of ultrasound images with visible white blood cell traces 310 having different cell tilts. As previously explained, in the case of lower signal strength arriving at the focal area, the cell traces appear less tilted and, with regard to cell concentration, fewer cells are visible in the focal area, whereas when a higher signal strength is arriving the visible cell number increases, and the cells appear more tilted. The upper left image corresponds to an ultrasound image with high signal attenuation, wherein the white blood cell traces 310 appear very flat (angle of tilt Qi). The upper right image has some signal attenuation, but less than the upper left image, since the white blood cell traces 310 appear more tilted (angle of tilt 02, with O2>ai). The lower left image has even less signal attenuation, the white blood cell traces 310 appearing more tilted (angle of tilt 03, with 03>02). Finally, the lower right image has almost no signal attenuation, the white blood cell traces 310 appearing very tilted (almost vertical, angle of tilt 04, with 04>03) and very thin, indicating a lot of signal strength arriving at the focal area. Figure 4E shows cropped areas of ultrasound images with visible white blood cell traces 310 having different lengths ( / ). In the upper image, wherein there is low signal attenuation, the cell traces are longer. However, in the lower image the signal attenuation is high and the white blood cell traces 310 are shorter. Therefore, a higher ultrasound signal attenuation leads to a higher length ( / ) of the white blood cell traces 310, whereas a lower ultrasound signal attenuation has the opposite effect.

[0073] Figure 4F represents the activation map 413 of the first layer of the first convolutional neural network 127 showing visible white blood cell traces 310 having different thickness (t). In the left image the high signal attenuation leads to thicker white blood cell traces 310, whereas in the right image low signal attenuation leads to thinner white blood cell traces 310 (apart from higher angles of tilt). Therefore, the thickness (t) of the white blood cell traces 310 increases with the signal attenuation (the more attenuated is the ultrasound signal, the greater is the thickness). The thickness (t) of a white blood cell trace 310 can be calculated, for instance, by computing the limits in x-direction of a white blood cell trace 310 at each y- coordinate. By subtracting the left limit in x-direction from the right limit in x-direction, the thickness of a white blood cell trace 310 at a specific y-coordinate is obtained. To obtain the final estimation of the thickness of the white blood cell trace, this subtraction is performed at each y-coordinate of the white blood cell trace, and finally an average is computed over all differences. Different ways to compute the thickness (t) of the white blood cell traces may be employed.

[0074] Another feature which can be used to determine the level of attenuation based on the white blood cell traces is the height of the area where the white blood cell traces appear, the focal area. White blood cell traces 310 are only visible in the focal area of the ultrasound beam, but the height of the focal area where the white blood cell traces are visible depends on the signal strength which arrives at the white blood cells. Figure 4G shows ultrasound images 117 with different levels of attenuation and hence different heights ( / ?) of visible white blood cells in the focal area. On the left image a higher attenuation level is shown and thus the focal area, where white blood cells are visible, is smaller. On the right image, with less signal attenuation, the focal area with visible white blood cells is larger. Therefore, the focal area with visible white blood cells is wider the less is the signal attenuation, which is directly related to the white blood cell trace length ( / ) described in Figure 4F. The height ( / ?) of the focal area with visible white blood cells can be computed by calculating the individual lengths ( / ) of the white blood cell traces 310 and then computing an average over all lengths. Another feature of the white blood cell traces that can be calculated 420, individually or in combination with other features, is the intensity of the pixels forming the white blood cell trace, since the less the signal attenuation, the greater the intensity of the pixels of the white blood cell trace. Once the white blood cell traces 310 are detected 410 in the ultrasound image 117, an average level of intensity of the pixels that form the white blood cell traces 310 may be computed, and a level of attenuation may be obtained from the calculated average level of intensity, alone or in combination with other parameter(s), such as the angle of tilt or the thickness. For instance, the level of attenuation may be defined as L=K i, (wherein L represents the level of attenuation, K is a constant and / is the average level of intensity), or by L=Kri+K2-f2 (wherein Ki and K2 are constants and f2 is the value of a second feature, such as thickness of the white blood cells).

[0075] Based on the information about the feature of the white blood cells, the ultrasound signal configuration can be adapted (e.g. using a feedback loop), such that the final level of attenuation (e.g. the effective tilt) is kept substantially the same (within a certain range), independently of the properties of the liquid.

[0076] The concentration of white blood cells 130 is estimated 260 using the first convolutional neural network 127 previously trained with training images having an assigned ground truth concentration. In the example of Figure 3A, to improve the estimation of the concentration of white blood cells 130 the displacement of leukocytes should be clearly seen in the ultrasound images 117; the clearer, the better the estimation can be. To make the displacement of leukocytes clearer in the ultrasound images 117, the right amount of signal strength emitted by the acoustic beam of the ultrasound transducer 114 should reach the focal area of interest, in particular a high signal strength must reach this area such that the level of attenuation is within a predetermined range. This can be achieved, for instance, by increasing the angle of tilt a until it is equal or higher than a predetermined angle of tilt a™. In an embodiment, the image processing device 120 actuates on the driving unit 112 of the ultrasound device 110 until the predetermined angle of tilt a™ (and / or a predetermined length, thickness, height of focal area) is reached or exceeded.

[0077] For instance, in a first iteration in which the predetermined angle of tilt is 80° (OTH=80°), the initial angle of tilt detected is 45°, which would imply a high level of attenuation of the ultrasound signal 115. The control unit 124 increases one or more parameters of the excitation signal 113, such as the pulse duration, based on the error (error=80°-45°=35°), e.g. using a proportional controller. In the second iteration, the angle of tilt detected is 70°, and the pulse duration (and / or other parameters) is increased proportionally to the error (error=80°-70°=10°). In the third iteration, the angle of tilt detected is 78°, and the pulse duration (and / or other parameters) is further increased proportionally to the error (error=80°- 78°=2°). In the fourth iteration, the angle of tilt detected is 80.1°, and the excitation signal 113 is no longer modified since the level of attenuation 123 is within the predetermined range (in this example the level of attenuation is defined by the angle of tilt a, and the predetermined range is a>80°, or 80°<a<90°).

[0078] Advantageously, the final effective level of attenuation is kept almost the same in every estimation, and the ultrasound signal attenuation can be reduced so that the displacement of leukocytes is clearly shown in the ultrasound images 117 (more white blood cell traces are visible and appear more tilted and with more length), thus allowing a better estimation of the white blood cell concentration. In addition, the effective level of attenuation would also be the same (or very similar) as the level of attenuation of the training images, which further improves the precision of the estimation performed by the first convolutional neural network 127 since the conditions of the ultrasound images 117 (level of attenuation) to be analyzed are maintained the same as those of the training images.

[0079] Figure 5A represents another embodiment of the step of determining 240 a level of attenuation 123, in this case applied for white blood cell counting in the cerebrospinal fluid of a newborn baby. In the example depicted in Figure 3C, ultrasound signal strength is highly attenuated due to a lot of hair covering the fontanelle tissue 320. This leads to higher signal attenuation and also to signal refraction. This situation can be detected when the tissue does not appear saturated at all (not white) (see not saturated vs saturated tissue in Figure 3C). The higher the saturation, the lower the level of attenuation.

[0080] In an embodiment, the step of determining 240 a level of attenuation 123 comprises:

[0081] Binarizing 510 the at least one ultrasound image 117, the pixels of the binarized image 512 having a first value (e.g. 1 or white) or a second value (e.g. 0 or black).

[0082] Detecting 520 in the binarized image 512 an upper border 522 of a fontanelle tissue 320, the pixels of the upper border 522 having the first value. In an embodiment, the upper border 522 is the upmost horizontal line of pixels having the first value and a length equal or higher than an upper threshold.

[0083] Detecting 530 in the binarized image 512 a lower border 532 of the fontanelle tissue 320, the pixels of the lower border 532 having the first value. In an embodiment, the lower border 522 is the upmost horizontal line of pixels located below the upper border 522, having the first value and a length equal or lower than a lower threshold.

[0084] Determining 540 a region 542 above the lower border 532 of the fontanelle tissue 320.

[0085] Calculating 550 a ratio 552 of the pixels within the region 542 that have the same value (e.g. the first value).

[0086] Obtaining 560 a level of attenuation 123 based on said ratio 552.

[0087] Figure 5B depicts an example of a binarized image 512 showing the upper border 522, the lower border 532 and the region 542.

[0088] According to an example for detecting the lack of tissue saturation in an automatized manner, the ultrasound image is first binarizing (e.g. with a size of 1600x730 pixels) by using a threshold value (e.g. of 170) for the binarization (value range from 0-255). Then the image is being looped over to look from the top line of the ultrasound image 117 for the first united line in the x-direction (all white pixels) with a minimum length (e.g. 450 pixels, upper threshold), obtaining the upper border 522 of the fontanelle tissue 320. A similar process is applied for detecting the lower border 532 of the fontanelle tissue 320: the first horizontal line having fewer than a certain number (e.g. 50, lower threshold) of white pixels is indicative of the lower border 532. Optionally, to make sure that the lower border 532 is correctly found, in an upper horizontal line (e.g. y-20) it is checked if the number of white pixels is greater than the lower threshold and in a lower horizontal line (e.g. y+10) it is checked that the number of white pixels is fewer than the lower threshold.

[0089] Once the lower border 532 of the fontanelle tissue 320 is detected, a window (region 542) with a certain height (e.g. a height of 200 pixels) is drawn above the lower border 532. A ratio 552 or percentage of white pixels within the region is then calculated, the ratio representing the level of saturation of the fontanelle tissue 320, which in turn represents the level of attenuation 123 of the ultrasound signal. The higher the ratio of white pixels within the region 542, the higher the saturation. Alternatively, a ratio of black pixels may be calculated, although in that case the higher the ratio of black pixels, the lower the saturation would be. A level of attenuation 123 is obtained 560 using the ratio 552. The level of attenuation 123 may be considered inversely proportional to the ratio 552, defined for instance as: L=K / r, wherein L represents the level of attenuation, K is a predefined constant and r is the ratio of white pixels (the higher the ratio, the higher the saturation and the lower the level of attenuation). If the ratio r employed is the ratio of black pixels, the higher the ratio the lower the saturation is, and the level of attenuation 123 may in that case be defined as L=K r (directly proportional).

[0090] Back to the flow diagram of Figure 2, once the level of attenuation 123 is determined 240, the level of attenuation 123 is compared 245 with a predetermined range. The ratio may be checked instead (in that case, K=1). For instance, it is checked whether more than a predetermined ratio (e.g. 70%) of all pixels within the region 552 are white pixels. In that case the fontanelle tissue 320 is saturated enough, indicating that enough signal is passing through, and the level of attenuation 123 is considered to be within the predetermined range. On the contrary case, the signal strength parameters can be adapted accordingly by modifying 250 the parameters of the excitation signal 113 until the ratio 552 is equal or higher than the predetermined threshold (e.g 70%).

[0091] This way, in case of the level of saturation being lower, or the calculated ratio 552 being higher, than a predetermined threshold, the ultrasound signal can be adapted by modifying the parameters of the excitation signal 113 to increase the signal strength in the focal area, so that the level of attenuation 123 is reduced down to a desired range.

[0092] Advantageously, the final effective level of attenuation is kept similar for every estimation, since the ultrasound signal attenuation is reduced so that the level of saturation is kept at a predetermined range, thus allowing a better estimation of the white blood cell concentration.

[0093] In addition to the automatic adjustment of the parameters of the excitation signal 113 to obtain a desired level of attenuation of the ultrasound signal 115, the invention may further include a previous automatic adjustment of the effective focal position of the ultrasound beam. The ultrasound signal needs to pass through different types of media with different acoustic propagation speeds (e.g. hair, tissue, plastic, liquid with different composition) and arrives thus more or less attenuated at the focal area. As a consequence, the effective focal position of the ultrasound beam shifts slightly up), meaning it comes closer to the ultrasonic transducer. While the theoretical focal position of the ultrasonic beam is well defined beforehand by the technical specifications of the ultrasound transducer 114, the effective focal position can change depending on the media the ultrasound beam needs to pass through.

[0094] Figure 6 depicts an embodiment of an apparatus 100 for white blood cell counting in extravascular body fluids including the additional feature of an automatic adjustment of the effective focal position. In particular, the image processing device 120 is configured to modify the focal position of the ultrasound signal via a focus control signal 606 actuating on a focusing unit 602 of the ultrasound device 110, so that the ultrasound device 110 is correctly focused.

[0095] As it is well known in the prior art, in the case of a fixed-focus single-element ultrasound transducer, the focal position can be moved by a mechanical displacement of the ultrasound transducer 114 automatically by a motor 604 of the ultrasound device 110; when the ultrasound transducer 114 is formed by an array of ultrasonic elements, the focal position can be adjusted instead electronically by configuring the excitation signal 113 for each of the elements of the array that are actuated (in that case a motor is not required). The motor 604, depicted in dashed lines, is therefore an optional element of the ultrasound device 110.

[0096] In an embodiment, the focusing unit 602 includes at least one motor 604 controlled by the image processing device 120 through the focus control signal 606. The focal position may be changed, for instance, by activating the motor 604 that moves the ultrasound transducer 114 in an axial direction ( / direction, corresponding to the direction of the ultrasound wave propagation) to change the focal position depth, such as in the invention disclosed in patent document WO2024179688-A1 .

[0097] The first image processing unit 122 determines an effective focal position 610 by analysing the at least one ultrasound image 117. The effective focal position 610 is compared with a target effective focal position (e.g. a theoretical focal position of the ultrasound device with the current configuration). If the effective focal position 610 is not within a predetermined range (e.g. closer to the target effective focal position than a fixed value), the control unit 124 actuates on the focusing unit 602 to adjust the focal position of the ultrasound signal 115, so that the effective focus 610 comes closer to the target effective focal position. Once the effective focal position 610 has been adjusted and it is within the desired range, the level of attenuation 123 of the ultrasound signal 115 is checked, and automatic adjustment of the parameters of the excitation signal 113 is performed until the level of attenuation 123 is within a predetermined range, as previously explained.

[0098] Figure 7A shows, on the left image, the theoretical focal position 710 and the effective focal position 610 of an ultrasound image 117 acquired from an infected peritoneal fluid contained in a peritoneal dialysis drainage bag 108. Due to the attenuation, the effective focal position 610 of the ultrasound beam has shifted up with respect to the theoretical focal position 710. On the right image, the effective focal area 720, centered around the effective focal position 610, is depicted.

[0099] For being able to correctly detect a possible infection by determining the number of leukocytes present in the serous body fluid, the effective focal position 610 must fall into the liquid area, and the area where the leukocyte counting models are applied must be contained within the effective focal area 720. To correct the shifting between the effective focal position 610 and the theoretical focal position 710, such that the effective focal area 720 falls into the desired area, the effective focal position 610 is calculated based on features extracted from the ultrasound image.

[0100] A crop covering the areas of a hypothetical effective focal position shift is taken out of the ultrasound image 117. Subsequently, the white blood cell traces 310 are detected 410, the starting point 417 and ending point 419 of the white blood cell traces 310 are calculated (as explained in Figure 4C), and an average starting point and an average ending point over all detected white blood cell traces 310 are computed. The middle point of the average starting and ending points indicates the effective focal position 610 of the ultrasound beam. The offset between the theoretical focal position 710 and the effective focal position 610 is calculated and the effective focal position can be shifted in such a way that the effective focal position falls exactly into the desired area. Figure 7B depicts the correction of the focal position, such that the effective focal position 610 matches the theoretical focal position 710 (left image); the corresponding effective focal area 720 is shifted according to the effective focal position 610 (right image).

[0101] Figure 8A is a flow diagram of the steps of the computer-implemented method 200 for white blood cell counting in extravascular body fluids according to an embodiment that includes an automatic adjustment of the effective focal position. The method 200 comprises the steps shown in Figure 2 and the additional steps of determining 832 an effective focal position 610 of the at least one ultrasound signal 115 by analysing the at least one ultrasound image 117, and actuating 836 a focusing unit 602 of the ultrasound device 110 to modify the focal position of the at least one ultrasound signal 115 until the effective focal position 610 is within a desired range.

[0102] The effective focal position 610 is compared 834 with a range, the range being defined by the theoretical focal position 710 of the ultrasound signal 115 based on the current configuration; for example, the range may be [fr-A, fT+A], wherein fTis the theoretical focal position 710 and A is a predefined constant. The control unit 124 may include a controller (such as a proportional controller, a proportional-derivative controller or a PID controller), wherein the effective focal position 610 is compared to a target (the theoretical focal position 710) and the difference (error) is fed to the controller to compute a focus control signal 606 that will modify the focal position of the ultrasound signal 115 to reduce the error.

[0103] The actuation 836 on the focusing unit 602 is performed if the effective focal position 610 is outside the range; otherwise, the method 200 continues (steps 240,245,250) with the automatic adjustment of the level of attenuation 123. Once the effective focal position 610 and the level of attenuation 123 are duly adjusted, a concentration of white blood cells 130 can be estimated 260 with much more precision.

[0104] According to an embodiment shown in Figure 8B, the step of determining 832 an effective focal position 610 comprises a first step of detecting 410 white blood cell traces 310 in the at least one ultrasound image 117. The white blood cell traces 310 can be detected 410, for instance, as previously explained in Figure 4B. Then an average starting point and an average ending point of all detected white blood cell traces 310 are computed 840, using the starting point 417 and the ending point 419 of the detected white blood cell traces 310. Finally, the effective focal position 610 is determined 850 as the middle point of the average starting point and the average ending point (middle point = average starting point + 1 (distance between the average ending point and the average starting point)). The depth (y-coordinate) of the middle point corresponds to the effective focal position 610.

[0105] Figure 9A represents an embodiment of a stand-alone apparatus 100 for measuring the concentration of white blood cells in samples of extravascular body fluid, such as peritoneal dialysis drainage bags or urine drainage bags. Figure 9B depicts a lateral section view showing some of the internal components.

[0106] Ultrasound systems in healthcare allow non-ionizing (safe) and real-time imaging of several body parts and organs supporting clinical diagnostic, monitoring and treatment decisions. It is a well-established and widespread imaging technique increasingly used by trained nonradiologist physicians in areas like Pediatrics.

[0107] A primary factor allowing the ubiquity of ultrasound in healthcare is its portability and cost when compared to any other imaging techniques. These factors turn ultrasound imaging as the technique with most potential for global accessibility both in industrialised and in low- and-middle income economies. However, usability and cost still remain the major limitations to enable global access of ultrasound diagnostic imaging.

[0108] The irruption of Al in healthcare and, particularly, in ultrasound empower this imaging technique with tools to guide and ease navigation, recognise structures and perform quantitative measurements. This combination of ultrasound and Al-based software has been baptised by the industry as digital ultrasound.

[0109] Digital ultrasound and all related advances rely in fast-imaging systems with embedded high computational power to support navigation and often includes robotic systems to make the imaging procedure as user-independent as possible. It is the case of 3D ultrasound breast imaging for cancer screening or vascularisation imaging with probe pressure-control systems. Intuitively, these systems are hardly scalable, versatile and profitable if they are going to be used at a global scale by personnel with a low level of expertise. Mainly, the major limitations of digital ultrasound devices that impair their widespread use both in hospital and domiciliary settings at a global scale are:

[0110] - Expertise is needed: despite the integration of Al-guided systems, multi-purpose ultrasound imaging systems still require the user to understand the anatomy, develop a set of skills to interact with the Al support and the interpretation of the images.

[0111] - Cost of components: ultrasound devices rely on fast-acquisition systems (transducer arrays, beamforming, FPGA) that are expensive and require significant computing power, power supply, cabling and connectors that set a floor cost too high to be adopted in domiciliary settings and in LMIC economies.

[0112] - Designed for imaging, not for measuring: ultrasound systems regardless of having Al-powered tools or not are designed for imaging structures, not to measure analytes or micron-sized biomarkers that are often relevant to support diagnostic decisions. To do this several factors lacking in current state of the art systems must be addressed such as: higher frequencies, higher sensitivity and spatial resolution, stability of the measurement, skills of the user, management of artifacts typical in ultrasound images and specific to the needed configuration to measure target biomarkers, etc.

[0113] - Remote support is not an option: While in industrialised countries real-time visualisation is feasible and minimum delays might be affordable, most of the applications using ultrasound require an immediate result. Needless to say, not all hospitals even less in LMICs (low-and lower-middle-income countries) are technologically prepared to afford a reliable online streaming of the ultrasound data.

[0114] The general technical challenge addressed herein is to provide the lay user with a platform technology that provides quantitative and / or qualitative measurements of relevant biomarkers that support diagnostic, treatment and monitoring decisions by an expert. In particular, the apparatus depicted in Figure 9A is a low-cost and portable ultrasound-based device fully automatised by means of Al. It is configured as a non-invasive white blood cell counter of extravascular body fluids for early detection (prevention) and monitoring of infections. The device also includes Al-driven dynamic ultrasound signal configuration to guide the user in the proper disposition of the setup and to guarantee the quality of the data in order to output a reliable measurement, altogether unburdening the user from a high level of skillfulness and training.

[0115] The apparatus may be used, for instance, to measure the concentration of white blood cells contained in peritoneal dialysis drainage bags. The main factor that limits domiciliary peritoneal dialysis (PD) survival time is, after kidney transplant, peritonitis. Generally, there are patients that are more prone to develop peritonitis and who, eventually, are permanently transferred to in-clinic hemodialysis because of the peritoneum damage inflicted by the infection. This is particularly problematic for patients with reduced mobility or living far from the clinic, which are a significant portion since on average the PD patient is 64 years old. Once in in-clinic hemodialysis other complications may arise, the quality of life of the patient drops significantly and healthcare costs are, on average, doubled. Prevention or early detection of peritonitis would enable early treatment and peritoneum filtering properties preservation, hence, extending the survival time of the patient under PD. However, currently infection is detected by the patient in the presence of fever or of abdominal pain and in either case these are signs and symptoms that take place several days after the onset of the infection.

[0116] Current ultrasound systems, yet in their portable version, are too expensive even for a daily domiciliary use and are not sensitive to low white blood cell concentrations; therefore, they are not suited for prevention of peritonitis. Besides, ultrasound systems still lack the level of usability that patients at home are used to with other products in the consumer market. Despite the potential use of probe support mechanisms, the user still would need to apply coupling gel and ascertain the coupling and positioning of the probe. These two factors alone introduce a great deal of image quality variability and may impair the measurement, leading to frustration and low rates of adherence.

[0117] The apparatus 100 of Figure 9A solves the aforementioned problems, allowing accurate white blood cell counting and, therefore, can potentially detect white blood cell count exceeding peritonitis diagnostic threshold, this factor being one of the well-established criteria associated with peritonitis diagnosis. The specific technical challenge addressed herein is, therefore, using digital ultrasound in a domiciliary setting to seamlessly alert the user (who has no expertise at all) of an elevated cell count in the drained fluid (or effluent).

[0118] The apparatus 100 includes the image processing device 120 and the ultrasound device 110. The apparatus is composed of an accessory 904 containing the ultrasound transducer 114 and a base 902 housing the electronics (ultrasound signal electronics 920, preamplifier 922, communication and power electronics 924, computing subsystem 926) of the ultrasound device 110 and the image processing device 120, a display 906 and a battery 928 (or batteries) and power supply subsystems. The ultrasound signal electronics 920 includes the driving unit 112 and the ultrasound scanner 116. The computing subsystem 926 includes the first image processing unit 122 and the second image processing unit 126. The control unit 124 may also be part of the computing subsystem 926. The ultrasound transducer 114 can be a focused single-element piezoelectric transducer, an array or PMLIT (Piezoelectric Micromachined Ultrasound Transducer). However, for overall cost minimisation single-element transducers, as the one shown in Figure 9A, are currently a preferred option. In an embodiment, the ultrasound transducer 114 may be displaced (e.g. axially) by a motor 604 to automatically change the focal position depth (as explained in Figure 8A).

[0119] Figure 10 depicts another embodiment of the apparatus 100, comprising a base 902 and an accessory 904 containing an ultrasound probe 1002 housing an ultrasound transducer 114, the ultrasound probe 1002 being placed in the target region of interest where the measurement should be performed. The ultrasound probe 1002 is connected to the electronics and power supply subsystems of the base 902. The ultrasound probe 1002 housing the transducer may rest on a compartment or a cavity practiced on the accessory 904.

[0120] The apparatus of Figure 10 can be applied on the skin of the patient (in vivo). In the case of screening for infant meningitis, the ultrasound probe 1002 would be placed on the fontanelle of the baby, whereas for uveitis screening the ultrasound probe 1002 would be placed on the eyelid.

[0121] The apparatus of Figure 9A can be applied on a sample container (in vitro, like a drainage bag), the accessory 904 includes a tray 908 having a scale ike shape used to accommodate bags containing the sample, the ultrasound transducer 114 being housed inside the tray structure. In the case of peritonitis prevention, the drainage bag containing the peritoneal fluid effluent after dialysis would be placed on top of the accessory. Likewise, urine drainage bags can be measured for prevention of urinary tract infection in catheterised patients.

[0122] The tray 908 is attached to the base 902 by fixing means included both in the base 902 and the tray 908. The connector including cabling 910 from the ultrasound transducer 114, motor 604 (if present) and power is engaged with a female-connector on the base 902. When the accessory 904 is engaged no connectors are visible to the user. The upper face of the tray 908 is designed with a tilt to form a concave shape (concave region 916) with an off- centered vertex 912 where the fluid in the bag (e.g. a peritoneal dialysis drainage bag 108) accummulates. This design strategy allows measurements in bags of different sizes and with even little amount of fluid. The ultrasound transducer 114 is placed right below the vertex 912 in a chamber 914 and can be adapted to a motorised system if in-plane, axial or a combination of both (for example, as in WO2024179688-A1), displacement of the transducer is needed to sample a larger volume or to conduct a faster measurement. A lid mounted on a sliding guide is placed on top of the chamber 914 to protect the transducer when no measurement is taking place. Before starting the measurement, the chamber 914 must be filled with water, ideally degassed deionized water, for the ultrasound signal to travel from the ultrasound transducer 114 to the bag. Gel can be used instead of water and in either case, a wipe tissue or a sponge can be used to remove them after finishing the measurement. Alternatively, semisolid PVA-based hydrogels can be used and disposed after the measurements.

[0123] On the display 906, the user will find messages that guide them during the measurement that may ask him / her to take an action, e.g. reposition the bag, or simply inform about the progress of the measurement and the result. The user can start the measurement by pressing the start button on the screen. Alternatively, a weight or proximity sensor can be used to minimize the number of steps to be taken by the user to automatically detect the sample positioned on the tray 908 and start the measurement process. In an embodiment, the base 902 is common for the in-vitro (Figure 9A) and in-vivo (Figure 10) versions and the accessory 904 (the one including the tray 908 of Figure 9A or the ultrasound probe 1002 of Figure 10) is exchanged according to the intended purpose of the apparatus. In another embodiment, the base 902 and accessory 904 is a single piece.

[0124] The apparatus 100 can be implemented as a stand-alone device (e.g. the one depicted in Figures 9A and 10) or as an integrated system, such as the one depicted in Figure 11. In automatic peritoneal dialysis, a cycler 1100 (also known as automated peritoneal dialysis machine) pumps in the dialysis fluid and pumps out the effluent through and to a drainage system. In this use case, the ultrasound transducer 114 is attached to the cassette 1102 of the cycler 1100 through which the dialysis fluid and the effluent circulate. A permanent solid coupling material (coupling layer 1104), such as a thin silicone layer, arranged between the ultrasound transducer 114 and the cassette 1102 could be used.

[0125] Every cycler is normally equipped with power supply, communication and display subsystem as it is the stand-alone version depicted in Figure 9A. As a result, the ultrasound signal electronics 920, the ultrasound transducer 114 and the computing subsystem 926 running the Al can be the only elements to be integrated in the cycler 1100. In the embodiment shown in Figure 11 , the computing subsystem 1110 of the cycler 1100 is adapted to cope with the processing requirements of the present invention (Al-based algorithms that perform the image processing and white blood cell counting), and the ultrasound signal electronics 920 is also integrated within the cycler 1100 (e.g. as a plugin), further facilitating the integration of the system. Alternatively, the ultrasound signal electronics 920 and the computing subsystem 926 that performs the image processing may be components external to the cycler 1100.

[0126] Back to Figure 9A, upon start the Al-based algorithm is initiated. In an embodiment, the Al- based algorithm includes previous quality checks to identify possible secondary reflections of the bag within the effective focal area 720 or an unsaturated bag reflection under the effective focal area 720 that affects the accuracy and reliability of the measurement. These defects in the bag reflections may be caused, among other reasons, by a bad positioning of the bag and / or a bad coupling or contact of the bag with the ultrasound transducer. In case of any of these quality checks are not passed, a message is displayed on the screen asking the user to reposition the bag and to make sure there is water or gel in the chamber to improve the coupling. Alternatively, or in addition to the display, one or more loudspeakers may be used to convey an audio message instructing the user to reposition the bag and / or check the ultrasound coupling. If these quality checks are passed, the system enters measurement mode (to measure the concentration of white blood cells) and collects signals from a static transducer position or from the region (e.g. a plane) scanned by the transducer.

[0127] Figure 12A shows an example of bad positioning of ultrasound sensor with respect to the medium in a peritoneal dialysis drainage bag 108, used for draining liquid from the abdomen in peritoneal dialysis patients. In the left ultrasound image 117 an example of bad positioning is shown, where a reflection 1202 (secondary reflection) of the bag is falling right into the effective focal area 720 where leukocytes are visible in case of an infection. The ultrasound image 117 shows another bag reflection 1204 correctly positioned outside the effective focal area 720. In the right ultrasound image 117 the ultrasound sensor is correctly positioned after a feedback message 1206 has been displayed to the user and the latter has corrected the bad positioning of the bag, so that the bag reflection 1202 inside the effective focal area 720 no longer appears. To automatically detect this defect and alert the user to adjust the position of the bag and eliminate this artifact, a deep learning model (e.g. including one or more convolutional neural networks) is trained on the cropped focal area images, previously labelled to contain the artifact or not, preferably based on a MobileNetV3 architecture (Howard et al., “Searching for MobileNetV3”. Proceedings of the IEEE International Conference on Computer Vision, 1314-1324, October 2019). The same methodology can be applied to other artifacts possibly appearing in the high-resolution ultrasound images, where user feedback and intervention is needed.

[0128] Figure 12B shows another example of bad positioning of the ultrasound sensor with respect to the measured medium, which can be detected by certain structures visible in a correct or incorrect way, which may be caused by bad contact or bad coupling. This figure depicts ultrasound images 117 of peritoneal fluid in a peritoneal dialysis drainage bag 108. In the left image an example of bad coupling is shown, where the bag reflection appearing below the effective focal area 720, which is used for applying a white blood cell count (cropped area), has a characteristic interrupted shape. In the right image the transducer is well coupled with the peritoneal dialysis drainage bag 108, leading to a saturated bag reflection (white pixels). When observing the cropped areas, in the well coupled case white blood cell traces 310 are clearly visible, whereas in the badly coupled example the white blood cell traces can be distinguished only with difficulty. A high-resolution ultrasound scan serves for the detection of the presence of leukocytes in the liquid indicative of the presence of infection. In the case of bad ultrasound sensor coupling, the bag reflection is visible in an interrupted manner, not well visible. As a result, the white blood cells, in case of their presence, are less well visible and also their appearance is affected (see cropped zoomed-in images of Figure 12B). This affects the quality of the models applied on these type of image crops to establish automatically the number of leukocytes present and thus giving information about a possible infection. The bag reflection can be detected in a region 1210 with a segmentation model, preferably based on a ll-Net architecture, and subsequently the mean value of the pixels within the region (corresponding to the segmented area) can be calculated. If the mean value lies above a pre-established value, the bag reflection is saturated enough, meaning the coupling is good enough to do the scan. In the contrary case, the user can be alerted, and the bag position can be adjusted.

[0129] Figure 13 is a flow diagram of the steps of the computer-implemented method 200 for white blood cell counting in extravascular body fluids according to an embodiment that includes one or more quality checks aimed to detect defects on the at least one ultrasound image 117 and try to correct them in the following ultrasound images 117. The method 200 comprises the steps shown in Figure 2 and the additional steps of the quality check(s), which include performing 1332 at least one quality check on the at least one ultrasound image 117, checking 1334 whether the at least one quality check is passed or not and, if the at least one quality check is not passed, executing an action 1336 aimed to improve the quality of the subsequently generated ultrasound images 117. Once the action has been executed, the process starts again in step 210.

[0130] In an embodiment, the at least one quality check is performed 1332 on the at least one ultrasound image 117 acquired from a sample of extravascular body fluid, wherein the sample of extravascular body fluid is contained in a bag, and the quality checks are aimed to determine that the reflections of the bag appearing in the at least one ultrasound image 117 are correct, as explained in Figures 12A and 12B (e.g. when there is no secondary bag reflection 1202 in the effective focal area 720 and / or when the bag reflection outside the effective focal area 720 is saturated). For these cases, the step of executing an action 1336 includes providing automatic feedback to a user (feedback message 1206) about the quality of the ultrasound image 117 and, in particular, instructing the user (e.g. by sending / emitting a visual and / or acoustic message) to reposition the sample and / or check the ultrasound coupling. The process starts again in step 210 after a certain event occurs (e.g. a predetermined time has elapsed or the user presses a button).

[0131] Figure 14 shows an embodiment of a quality check corresponding to Figure 12A. According to this embodiment, the step of performing at least one quality check 1332 comprises analysing 1432 the at least one ultrasound image 117, acquired from a bag containing a sample of extravascular body fluid, using a second convolutional neural network trained to identify reflections of the bag within the effective focal area 720 of the ultrasound image 117. The second convolutional neural network is preferably based on a MobileNetV3 architecture. The output of the second convolutional neural network is the detection, or not, of this defect (i.e. a secondary bag reflection 1202 within the effective focal area 720). If the secondary bag reflection 1202 inside the effective focal area 720 is detected, the quality check is not passed; otherwise, the quality check is passed. The step of executing an action 1336 includes instructing 1436 the user to reposition the bag.

[0132] Figure 15 shows another embodiment of the quality check, in this case corresponding to Figure 12B. According to this embodiment, the step of performing at least one quality check 1332 comprises detecting 1532, in the at least one ultrasound image 117 acquired from a bag containing a sample of extravascular body fluid, a region 1210 including a bag reflection using image segmentation with a third convolutional neural network (preferably based on a ll-Net architecture); and determining 1533 a level of saturation of the bag reflection based on the value of the pixels within the region 1210, the level of saturation being representative of a level of quality of the ultrasound coupling between the ultrasound transducer 114 and the bag (e.g. good coupling or bad coupling). In an embodiment, a mean value of the pixels within the region 1210 is calculated and compared with at least one saturation threshold in order to determine a level of saturation. For example, if the mean value exceeds a saturation threshold, the bag reflection is considered to be saturated (involving a good coupling) and the quality check is passed; otherwise, it is considered unsaturated due to a bad coupling and the quality check is not passed. The step of executing an action 1336 includes instructing 1536 the user to check the ultrasound coupling and / or reposition the bag.

[0133] Figure 16A shows another embodiment of the quality check. According to this embodiment, the step of performing at least one quality check 1332 comprises analysing 1632 the at least one ultrasound image 117, acquired from a head 102 of a newborn baby, using a fourth convolutional neural network trained to classify a level of pressure exerted on the fontanelle tissue. The step of executing an action 1336 includes instructing 1636 the user to reduce the pressure exerted on the fontanelle tissue by the ultrasound device 110. The ultrasound images 17 of Figure 16B show the infant fontanelle tissue 320 (white area) with the cerebrospinal fluid area below. In the left image, the fontanelle tissue 320 appears straight and is thus not deformed by the force exerted by the user with the ultrasound probe 1002, while in the right image the fontanelle tissue 320 is being highly pressed by the user, complicating the analysis of the cerebrospinal fluid below the fontanelle tissue 320. On the left image the fontanelle tissue 320 is well defined and tissue intensity levels are high, both representing a sign of the maintenance of the ultrasound beam and intensity integrity. On the right image, the interface between the fontanelle tissue 320 (subarachnoid layer) and the fluid is badly defined, a clear indication of the ultrasound beam de-focusing.

[0134] The quality of the scan and how the user handles the ultrasound probe 1002 is of utmost importance when wanting to detect infections via high-resolution ultrasound images 17. An example of mishandling the probe is when the ultrasound sensor is pressed too much against the tissue or the measured medium, such that it is deformed and thus does not maintain its correct properties to perform the intended measurement. This is especially important, when the tissue is very soft, such as in the case of the fontanelle tissue of a newborn baby’s head 102. A newborn baby’s fontanel tissue is scanned for the non-invasive screening of meningitis via high-resolution ultrasound images, where an increase of leukocytes in the cerebrospinal fluid takes place. Here the measurement needs to be carried out in the liquid (cerebrospinal fluid in this case) area visible in the ultrasound image. In case of too much pressing of the ultrasound probe 1002 against the fontanelle tissue 320, the liquid lying below the fontanelle tissue 320 can be moved to the side, such that not enough liquid remains for doing the measurement correctly. To avoid this and detect this problem automatically, an algorithm is designed to detect if the fontanelle tissue is being pressed (with some force) by the user of the ultrasound probe 1002. The algorithm employs a fourth convolutional neural network preferably based on the Resnet50 network architecture, which classifies an image into two classes: containing fontanel tissue in the form of a slight u-shape (pressed tissue, right image of Figure 16B) or containing fontanel tissue straightly shaped (left image of Figure 16B).

[0135] The strategy of training the model may be, for instance, based on a k-fold cross validation technique where the model is always trained on 80% of the data (80% on a patient level) and validated on the remaining 20%. In the next round of training, another combination of patients’ data is used to train the model and other 20% of data is used to validate and so on, until having had all patients once in the validation data set. This way the generalization ability of the model is checked. The input of the model is a single ultrasound image, and the output is a probability for belonging to each of the 2 classes. The predicted class is chosen based on the highest probability value. The user can thus be given automatic feedback about the quality of the ultrasound image and can adapt the pressing force exerted by the ultrasound probe 1002 on the head 102 accordingly. In case the ultrasound probe is mechanically positioned or motorized to control applied pressure, Al-outputs can guide motor movements until a flat tissue is detected by Al-models.

[0136] By depressing the ultrasound probe 1002, fontanel tissue 320 will decompress and the acoustic beam will likely be distorted, hence modifying the focal position and lowering the acoustic pressure in the focal volume. This sensitivity problem will be observed as a drop in tissue intensity levels and, therefore, the ultrasound signal configuration could be adapted as previously explained (by increasing, for instance, pulse duration, number of pulses or pulse amplitude). The focal position, which has been modified by the decompression, may also be automatically adjusted, as explained in Figure 8A.

[0137] Figure 17A shows another embodiment of the quality check. According to this embodiment, the step of performing at least one quality check 1332 comprises analysing 1732 the at least one ultrasound image 117, acquired from a head 102 of a newborn baby, using a fifth convolutional neural network trained to classify a level of quality of the ultrasound coupling between the ultrasound transducer 114 and the head 102. The step of executing an action 1336 includes instructing 1736 the user to check and improve the ultrasound coupling (e.g. applying more coupling gel, improving the contact between the ultrasound probe 1002 and the head 102, etc.).

[0138] Figure 17B shows examples of bad and good ultrasound coupling between the ultrasound transducer 114 and the medium to be measured. In the right image an example of good coupling is shown, where the structures or cells can be seen. In the left image the ultrasound probe 1002 is not well coupled with the tissue of the head 102, generating a typical pattern visible in this image. In case of bad coupling, the structures or cells are not well visible and the white blood cell counting cannot be performed. To automatically detect if the probe is well coupled with the medium to be measured, a fifth convolutional neural network, preferably based on a Resnet50 network architecture, is trained to classify images into containing structures indicating bad coupling or not. The model training strategy is similar to the one described in the example of Figure 16B. Figure 18A shows another embodiment of the quality check. According to this embodiment, the step of performing at least one quality check 1332 comprises analysing 1832 the at least one ultrasound image 117 using a sixth convolutional neural network trained to classify a level of clutter (due to tissue reverberations) present in ultrasound images. The step of executing an action 1336 includes modifying 1836 the gain or amplification of the ultrasound signal 115 based on the level of clutter. The presence of clutter in the ultrasound image 117 can confound the white blood cell counting algorithm, and it is therefore important to minimize this effect.

[0139] Figure 18B shows ultrasound images of the infant fontanelle tissue 320 with the below lying cerebrospinal fluid. Examples of different configurations of gain are shown. In the case of higher gain (32dB) a lot of clutter appears especially below the fontanelle tissue 320. In the case of a lowered gain (29dB) the clutter disappears slightly, and in the case of an even lower gain (26dB) the clutter is almost completely gone.

[0140] The gain of the ultrasound signal 115 (i.e. the amplification of the ultrasound signal 115 once it returns to the ultrasound transducer 114 after travelling through the tissue or the medium it passed through) is a parameter which can be adapted in the generation of the ultrasound image 117. Such amplification can be done at the analogue and the digital stages independently. Figure 18C shows an amplifier 1810 configured to amplify the ultrasound signal 115 acquired by the ultrasound transducer 114 with a certain gain, thereby obtaining an amplified ultrasound signal 1812 used to generate the ultrasound image 117. A higher gain means that meaningful signal is amplified, but clutter produced by signal reverberations when passing through tissue is amplified as well. In cases where too much clutter is visible, meaningful information is masked with noise. In this case the gain can be lowered automatically, by taking into account information about the noise level present in the image to obtain a clearer image. To achieve this, a sixth convolutional neural network 1820, preferably based on the ResNet50 architecture, is trained to classify images into either containing too much clutter or not, based on previous manual labelling. Based on the output of this model, if an ultrasound image 117 contains too much clutter, the gain can be lowered, and the check 1334 is run again.

[0141] The sixth convolutional neural network 1820 obtains a level of clutter 1822. The level of clutter 1822 may be, for instance, a value (e.g. a percentage) or a class (high level, moderate level, low level). It is then checked whether the level of clutter 1822 is within a predetermined range (e.g. having a value lower than a predetermined threshold, or classified as “low level”). If the level of clutter 1822 is outside a desired range, the control unit 124 modifies, using a gain control signal 1824, the gain applied to the ultrasound signal 115 by the amplifier 1810. The change in the gain may be dependent upon the level of clutter 1822 detected (e.g. the more the level of clutter 1822, the more the gain is modified). Once the ultrasound image 117 is classified as being good for further processing (i.e. the level of clutter 1822 being within a desired range), the white blood cell counting can be continued with the adapted gain.

[0142] The present invention may use one or more of the quality checks previously described in Figures 13 to 18. Once all the quality checks are passed, measurement of white blood cells is performed. Because the level of proteins in the fluid as well as the composition of the material used to fabricate the peritoneal dialysis drainage bag 108 can affect while blood cell backscattered signal properties and signal strength (Figure 3), Al-based dynamic ultrasound signal configuration (as explained for instance in Figures 2 and 4A) can compensate such signal changes to guarantee that the Al model outputting a measurement estimate is fed with data that maintain the properties of the training dataset.

[0143] The aforementioned quality checks described in Figures 12 to 18 may be combined not also with the dynamic ultrasound signal configuration of Figures 1 to 5, but also with the automatic focal position adjustment described in Figures 6 to 8. Figure 19 represents a flow diagram of a method 200 for white blood cell counting in extravascular body fluids including these three features: one or more quality checks, an automatic focal position adjustment and a dynamic ultrasound signal configuration. In this embodiment, the steps of the quality checks (steps 1332, 1334 and 1336) are performed before the steps of the automatic focal position adjustment (steps 832, 834 and 836). Although this is a preferred embodiment, the order of execution may be inverted (i.e. the automatic focal position adjustment may be carried out before the quality checks).

Claims

CLAIMS1 . An apparatus for white blood cell counting in extravascular body fluids, the apparatus (100) comprising: an ultrasound device (110) including: a driving unit (112) configured to excite an ultrasound transducer (114) with an excitation signal (113); the ultrasound transducer (114), configured to acquire at least one ultrasound signal (115) from a body or a sample of extravascular body fluid; and an ultrasound scanner (116) configured to generate at least one ultrasound image (117) from the at least one ultrasound signal (115); and an image processing device (120) comprising: a first image processing unit (122) configured to determine a level of attenuation (123) of the at least one ultrasound signal (115) by analysing the at least one ultrasound image (117); a control unit (124) configured to actuate the driving unit (112) of the ultrasound device (110) to modify at least one parameter of the excitation signal (113) until the level of attenuation (123) is within a predetermined range; and a second image processing unit (126) configured to estimate a concentration of white blood cells (130) from the at least one ultrasound image (117), generated when the level of attenuation (123) is within the range, using a first convolutional neural network (127) trained with images with an assigned ground truth concentration.

2. The apparatus of claim 1 , wherein the first image processing unit (122) is configured to determine the level of attenuation (123) by: detecting white blood cell traces (310) in the at least one ultrasound image (117); calculating at least one feature of the white blood cell traces (310); and obtaining a level of attenuation (123) from the at least one feature.

3. The apparatus of claim 2, wherein the first image processing unit (122) is configured to detect white blood cell traces (310) by: obtaining an activation map (413) of a predetermined layer of the first convolutional neural network (127); binarizing the activation map (413) using a binarization threshold; and detecting one or more white blood cell traces (310) in the binarized activation map(415).

4. The apparatus of any of claims 2 to 3, wherein the least one feature of the white blood cell traces (310) includes any of the following: angle of tilt (a) of the white blood cell traces (310), thickness (t) of the white blood cell traces (310), length ( / ) of the white blood cell traces (310), height ( / ?) of the focal area in which the white blood cell traces (310) are visible, an average level of intensity of the white blood cell traces (310), or a combination thereof.

5. The apparatus of claim 1 , wherein the first image processing unit (122) is configured to determine the level of attenuation (123) by: binarizing the at least one ultrasound image (117); detecting in the binarized image (512) an upper border (522) of a fontanelle tissue (320); detecting in the binarized image (512) a lower border (532) of the fontanelle tissue (320); determining a region (542) above the lower border (532) of the fontanelle tissue (320); calculating a ratio (552) of the pixels within the region (542) having the same value; and obtaining a level of attenuation (123) based on said ratio (552).

6. The apparatus of any preceding claim, wherein: the ultrasound device (110) comprises a focusing unit (602) configured to adjust the focal position of the at least one ultrasound signal (115); the first image processing unit (122) is further configured to determine an effective focal position (610) of the at least one ultrasound signal (115) by analysing the at least one ultrasound image (117); and the control unit (124) is further configured to actuate the focusing unit (602) of the ultrasound device (110) to modify the focal position of the at least one ultrasound signal (115) until the effective focal position (610) is within a range.

7. The apparatus of claim 6, wherein the first image processing unit (122) is configured to determine the effective focal position (610) by:detecting white blood cell traces (310) in the at least one ultrasound image (117); computing an average starting point and an average ending point of the detected white blood cell traces (310); and determining the effective focal position (610) as the middle point of the average starting point and the average ending point.

8. The apparatus of any preceding claim, wherein the at least one parameter of the driving signal comprises pulse amplitude, pulse duration, pulse repetition frequency, pulse excitation frequency, or a combination thereof.

9. The apparatus of any preceding claim, comprising: a tray (908) configured to support the sample_of extravascular body fluid; and a chamber (914) arranged under the tray (908) and adapted to house a coupling material and the ultrasound transducer (114).

10. The apparatus of claim 9, further comprising a motor (606) configured to move the ultrasound transducer (114) within the chamber (914) in an axial direction to change the focal position depth, in an in-plane rotation, or a combination thereof.11 . The apparatus of any of claims 9 to 10, wherein the tray includes a concave region (916) adapted to receive a bag containing a sample of extravascular body fluid, the ultrasound transducer (114) being arranged under the concave region (916).

12. The apparatus of claim 11 , wherein the concave region (916) includes an off-centered vertex (912), the ultrasound transducer (114) being arranged under the vertex (912).

13. The apparatus of any of claims 11 to 12, wherein the bag is a peritoneal dialysis drainage bag (108) or a urine drainage bag.

14. The apparatus of any of claims 1 to 8, wherein the ultrasound transducer (114) is connected to a cassette (1102) of a cycler (1100) for automated peritoneal dialysis through a coupling layer (1104).

15. The apparatus of any preceding claim, wherein the first image processing unit (122) is configured to perform at least one quality check on the at least one ultrasound image (117); and wherein the control unit (124) is configured to execute an action if the at least onequality check is not passed.

16. The apparatus of claim 15, the first image processing unit (122) is configured to perform at least one quality check by analysing the at least one ultrasound image, acquired from a bag containing a sample of extravascular body fluid, using a second convolutional neural network trained to identify reflections (1202) of the bag within the effective focal area (720) of the ultrasound image (117); and wherein the control unit (124) is configured to execute the action of instructing a user to reposition the bag.

17. The apparatus of claim 15, wherein the first image processing unit (122) is configured to perform at least one quality check by: detecting, in the at least one ultrasound image (117) acquired from a bag containing a sample of extravascular body fluid, a region (1210) including a bag reflection using image segmentation with a third convolutional neural network; and determining a level of saturation of the bag reflection based on the value of the pixels within the region (1210), the level of saturation being representative of a level of quality of the ultrasound coupling between the ultrasound transducer (114) and the bag; and wherein the control unit (124) is configured to execute the action of instructing a user to check the ultrasound coupling and / or reposition the bag.

18. The apparatus of claim 15, wherein the first image processing unit (122) is configured to perform at least one quality check by analysing the at least one ultrasound image (117), acquired from a head (102) of a newborn baby, using a fourth convolutional neural network trained to classify a level of pressure exerted on the fontanelle tissue; and wherein the control unit (124) is configured to execute the action of instructing a user to reduce the pressure exerted on the fontanelle tissue by the ultrasound device (110).

19. The apparatus of claim 15, wherein the first image processing unit (122) is configured to perform at least one quality check by analysing the at least one ultrasound image (117), acquired from a head (102) of a newborn baby, using a fifth convolutional neural network trained to classify a level of quality of the ultrasound coupling between the ultrasound transducer (114) and the head (102); and wherein the control unit (124) is configured to execute the action of instructing a user to check the ultrasound coupling.

20. The apparatus of claim 15, wherein the first image processing unit (122) is configured to perform at least one quality check by analysing the at least one ultrasound image (117) using a sixth convolutional neural network (1820) trained to classify a level of clutter (1822) present in ultrasound images; and wherein the control unit (124) is configured to execute the action of modifying the gain of the ultrasound signal (115) based on the level of clutter (1822).

21. A method for white blood cell counting in extravascular body fluids, comprising: exciting (210) an ultrasound transducer (110) with an excitation signal (113); acquiring (220), by the ultrasound transducer (110), at least one ultrasound signal (115) from a body or a sample of extravascular body fluid; generating (230) at least one ultrasound image (117) from the at least one ultrasound signal (115); determining (240) a level of attenuation (123) of the at least one ultrasound signal (115) by analysing the at least one ultrasound image (117); modifying (250) at least one parameter of the excitation signal (113) until the level of attenuation (123) is within a predetermined range; and estimating (260) a concentration of white blood cells (130) from the at least one ultrasound image (117), generated when the level of attenuation (123) is within the range, using a first convolutional neural network (127) trained with images with an assigned ground truth concentration.

22. The method of claim 21 , wherein determining (240) a level of attenuation (123) comprises: detecting (410) white blood cell traces (310) in the at least one ultrasound image (117); calculating (420) at least one feature of the white blood cell traces (310); and obtaining (430) a level of attenuation (123) from the at least one feature.

23. The method of claim 22, wherein detecting (410) white blood cell traces (310) in the at least one ultrasound image (117) comprises: obtaining (412) an activation map (413) of a predetermined layer of the first convolutional neural network (127); binarizing (414) the activation map (413) using a binarization threshold; and detecting (416) one or more white blood cell traces (310) in the binarized activation map (415).

24. The method of any of claims 22 to 23, wherein the least one feature of the white blood cell traces (310) includes any of the following: angle of tilt (a) of the white blood cell traces (310), thickness (t) of the white blood cell traces (310), length ( / ) of the white blood cell traces (310), height ( / ?) of the focal area in which the white blood cell traces (310) are visible, an average level of intensity of the white blood cell traces (310), or a combination thereof.

25. The method of claim 21, wherein determining (240) a level of attenuation (123) comprises: binarizing (510) the at least one ultrasound image (117); detecting (520) in the binarized image (512) an upper border (522) of a fontanelle tissue (320); detecting (530) in the binarized image (512) a lower border (532) of the fontanelle tissue (320); determining (540) a region (542) above the lower border (532) of the fontanelle tissue (320); calculating (550) a ratio (552) of the pixels within the region (542) having the same value; and obtaining (560) a level of attenuation (123) based on said ratio (552).

26. The method of any of claims 21 to 25, further comprising: determining (832) an effective focal position (610) of the at least one ultrasound signal (115) by analysing the at least one ultrasound image (117); and actuating (836) a focusing unit (602) of the ultrasound device (110) to modify the focal position of the at least one ultrasound signal (115) until the effective focal position (610) is within a predetermined range.

27. The method of claim 26, wherein determining (832) an effective focal position (610) comprises: detecting (410) white blood cell traces (310) in the at least one ultrasound image (117); computing (840) an average starting point and an average ending point of the detected white blood cell traces (310); anddetermining (850) the effective focal position (610) as the middle point of the average starting point and the average ending point.

28. The method of any of claims 21 to 27, wherein the at least one parameter of the driving signal comprises pulse amplitude, pulse duration, pulse repetition frequency, pulse excitation frequency, or a combination thereof.

29. The method of any of claims 21 to 28, further comprising: performing at least one quality check (1332) on the at least one ultrasound image (117); and if the at least one quality check is not passed, executing an action (1336).

30. The method of claim 29, wherein performing at least one quality check (1332) comprises analysing (1432) the at least one ultrasound image (117), acquired from a bag containing a sample of extravascular body fluid, using a second convolutional neural network trained to identify reflections (1202) of the bag within the effective focal area (720) of the ultrasound image (117); and wherein executing an action (1336) includes instructing (1436) a user to reposition the bag.

31. The method of claim 29, wherein performing at least one quality check (1332) comprises: detecting (1532), in the at least one ultrasound image (117) acquired from a bag containing a sample of extravascular body fluid, a region (1210) including a bag reflection using image segmentation with a third convolutional neural network; and determining (1533) a level of saturation of the bag reflection based on the value of the pixels within the region (1210), the level of saturation being representative of a level of quality of the ultrasound coupling between the ultrasound transducer (114) and the bag; and wherein executing an action (1336) includes instructing (1536) a user to check the ultrasound coupling and / or reposition the bag.

32. The method of claim 29, wherein performing at least one quality check (1332) comprises analysing (1632) the at least one ultrasound image (117), acquired from a head (102) of a newborn baby, using a fourth convolutional neural network trained to classify a level of pressure exerted on the fontanelle tissue; and wherein executing an action (1336) includes instructing (1636) a user to reduce thepressure exerted on the fontanelle tissue by the ultrasound device (110).

33. The method of claim 29, wherein performing at least one quality check (1332) comprises analysing (1732) the at least one ultrasound image (117), acquired from a head (102) of a newborn baby, using a fifth convolutional neural network trained to classify a level of quality of the ultrasound coupling between the ultrasound transducer (114) and the head (102); and wherein executing an action (1336) includes instructing (1736) a user to check the ultrasound coupling.

34. The method of claim 29, wherein performing at least one quality check (1332) comprises analysing (1832) the at least one ultrasound image (117) using a sixth convolutional neural network (1820) trained to classify a level of clutter (1822) present in ultrasound images; and wherein executing an action (1336) includes modifying (1836) the gain of the ultrasound signal (115) based on the level of clutter (1822).

35. A non-transitory computer-readable storage medium for white blood cell counting in extravascular body fluids, comprising computer code instructions that, when executed by a processor, causes the processor to perform the method of any of claims 21 to 34.

Images