Improved system for the analysis of tubes adapted to contain a biological sample

Abstract

A system for the analysis of tubes adapted to contain biological samples is herein disclosed, the system having at least one support for a tube adapted to contain a biological sample, at least one image detector configured to acquire at least one image of the tube arranged on the support, and a processing unit adapted to process the images of the tube acquired by the image detector, wherein the processing unit is configured to perform a processing procedure based on a neural network which is trained based on exemplary images of tubes, wherein, in said exemplary images, the tubes comprise a label associated thereto. The neural network is programmed to perform a classification relating to the tube contained in the input image, wherein the processing unit is configured to output an output value vector comprising a plurality of output values, each of said output values being related to a certain class of the classification(s) and representing the activation level of a corresponding neuron of an output of the neural network.

Description

[0001] Title: Improved system for the analysis of tubes adapted to contain a biological sample

[0002] DESCRIPTION

[0003] Field of application

[0004] The present invention refers to a system for the analysis of tubes adapted to contain a biological sample, in particular a blood sample, for example for the measurement of the erythrocyte sedimentation rate of said blood sample. The following description refers to this field of application with the only purpose to simplify its exposition.

[0005] Prior art

[0006] The measurement of the erythrocyte sedimentation rate (ESR) is a very common laboratory test to rapidly identify inflammatory processes. In particular, said test assesses the rate at which the erythrocytes contained in a blood sample sediment on the bottom of a tube.

[0007] As is well known in this technical field, the reference method for calculating ESR is the Westergren method, according to which the blood to be analyzed, diluted with sodium citrate, is placed into a graduated tube, and the formed sediment is measured after one hour. Thus, such a calculation method involves the use of dedicated tubes and predetermined time schedules.

[0008] There are apparatuses capable of measuring ESR even on standard tubes (for example, the normal blood-count tubes), in which, through opticalabsorption measurement of the blood sample in the tube, it is possible to obtain measurement values that are very consistent with the reference values of the above-mentioned Westergren method.

[0009] In this case, an optical reading of the blood sample contained in the tube is performed, said reading being influenced by the type of the tube used and by the arrangement of the identification label thereon, which considerably affects the absorption of the radiation incident on the tube, and thus in general the measurement performed.

[0010] In the above-mentioned apparatus type, it is thus extremely important to have preliminary information (i.e., before performing the above- mentioned optical reading) about the tube and / or about the identification label associated thereto, and the known solutions are not able to ensure providing said preliminary information.

[0011] The technical problem of the present invention is to provide a system having structural and functional features capable of overcoming the limitations and the drawbacks reported in connection to the prior art, in particular a system capable of efficiently providing preliminary information about the tube type and / or about the label associated thereto.

[0012] Summary of the invention

[0013] The solution idea underlying the present invention is to provide a system for the analysis of tubes wherein an image detector is configured to acquire images of said tubes, said images being then suitably processed to extract information about the tube, in particular about the tube type and about the label associated thereto (for example, about the label type, which may also comprise the extension and the placement of said label relative to the body of the tube and possibly in relation to the way it is attached) through machine-learning techniques (in particular, deep learning), in particular through implementation of a neural network (particularly a convolutional neural network). Even more in particular, the acquired images are provided as input to the above-mentioned convolutional neural network, which is suitably trained with exemplary images of tubes and is structured to perform an operation of classification of the tube. For example, said operation of classification can comprise the identification of the network neurons with the highest activation (i.e., with the highest probability of belonging to one of the expected classes), thereby outputting a value vector representative of the classification regarding the tube type and possibly also regarding the label type, which may also comprise the way in which the label is associated to said tube.

[0014] Based on said solution idea, the above-mentioned technical problem is solved by a system for the analysis of tubes adapted to contain biological samples, comprising at least one support for a tube adapted to contain a biological sample, at least one image detector configured to acquire at least one image of the tube, and a processing unit adapted to process the images of the tube acquired by the image detector, wherein said processing unit is configured to perform a processing procedure based on a neural network (in particular, a convolutional neural network) which is trained based on exemplary images of tubes, wherein, in said exemplary images, said tubes comprise a label associated thereto, wherein the neural network is programmed to perform a classification relating to the tube contained in the at least one input image, i.e., provided as an input to the neural network (which indicates a general analysis of the tube, which may also comprise the associated label, as will be described below), wherein the processing unit is further configured to output an output value vector structured as (comprising) a plurality of output values, each of said output values being related to a certain class of tubes of the classification(s) and representing the activation level of a corresponding neuron of an output of the neural network and thus the corresponding probability that the image’s content belongs to a class of said classification, thereby providing as output the class with the highest probability.

[0015] More in particular, the invention comprises the following additional and optional features, taken individually or in combination if needed. Said features are defined, for example, in the dependent claims, individually or in combination with each other.

[0016] According to an aspect of the present invention, the processing unit may be configured to perform a pre-processing of the acquired image (for example, to crop the ROI or to perform normalization operations or the like) . According to an aspect of the present invention, the neural network may comprise at least one input layer, wherein the processing unit is configured to provide the image to said input layer, the image being possibly pre-processed, at least one convolution layer adapted to extract features from the input image, and a fully connected network configured, based on the information from the previous layers, to perform said classification regarding the tube contained in the input image, said fully connected network comprising at least one first network layer or input network layer and a second network layer or output network layer, wherein the output value vector is representative of the output network layer of the fully connected network, said values representing the activation level of corresponding neurons of the output network layer of the fully connected network.

[0017] According to an aspect of the present invention, the processing unit may be configured to perform the classification through identification of the neurons of the output network layer of the fully connected network which correspond to the highest activation level.

[0018] According to an aspect of the present invention, the output value vector may be structured into (comprise) a first plurality of output values, each being related to a certain class of types of tubes, so as to classify the tube based on the type of the tube per se, and a second plurality of output values, each being related to a certain piece of information about the label attached to said tube (class of label type), so as to classify the tubes also based on the label (for example, the label type, which may also comprise the extension and / or the placement of the label and / or how the label is attached to the tube).

[0019] According to an aspect of the present invention, the fully connected network may comprise only the first network layer and the output network layer, wherein the first network layer comprises one hundred artificial neurons and the output network layer comprises at least nine artificial neurons (although any number thereof may be provided) . In an embodiment, at least six of said at least nine artificial neurons are related to different classes of tubes, and at least three of said nine artificial neurons are related to different information about the label on the tube. Clearly, the neuron number (in particular, the neurons of the last layer and thus the values of the output vector) may vary based on different desired classifications, for example if it is desired to classify more or less types of tubes.

[0020] According to an aspect of the present invention, the processing unit may be configured to identify the presence of the label by using information contained in at least one library, for example a zxing and / or zbar library, contained in a memory unit of said processing unit.

[0021] According to an aspect of the present invention, the processing unit may be configured to perform a pre-processing of the acquired images before providing them as input to the network.

[0022] According to an aspect of the present invention, the pre-processing may occur through definition of a region of interest (ROI) of the acquired image.

[0023] According to an aspect of the present invention, in the pre-processing, the processing unit may be configured to crop the acquired image so as to select the region of interest in which the tube is located, and to normalize said image according to a standardization procedure.

[0024] According to an aspect of the present invention, the neural network may be a feed forward network.

[0025] According to an aspect of the present invention, the size of a convolution kernel applied in the convolution layer may be 3x3. Clearly, other suitable sizes may also be used.

[0026] According to an aspect of the present invention, the support of the tube may be a rotating support which is moved by moving means and configured to allow a rotation of said tube around its own longitudinal axis during the acquisition of the images by means of the image detector.

[0027] According to an aspect of the present invention, the image detector may be configured to acquire a plurality of images during the rotation of the tubes so as to acquire images of different portions of said tubes.

[0028] According to an aspect of the present invention, the processing unit may be configured to perform an average operation among the value vectors calculated for each image of said plurality of acquired images in order to perform the classification with respect to the type of tube.

[0029] According to an aspect of the present invention, the processing unit may be configured to verify a number of times in which a certain state of the label appears in said acquired images, performing the corresponding classification through the majority voting technique.

[0030] According to an aspect of the present invention, the processing unit may be programmed so as to operate based on a state machine at least for managing the reading of configuration files, for the control of a lighting system adapted to lighten the tube, for the control of the moving means of said tube, and for the identification of desired acquisition instants according to the movement of said moving means.

[0031] According to an aspect of the present invention, the neural network may comprise at least one pooling layer configured to subsample the information contained in the image by executing a maximum or average operation among the pixel values of said image.

[0032] According to an aspect of the present invention, the system may comprise at least one detection unit configured to perform optical absorption measurements on a blood sample contained in the tube.

[0033] According to an aspect of the present invention, the support may be in the form of a tube gripper, which is configured to pick up the tubes from a housing area and to move said tubes towards an area of analysis thereof. According to an aspect of the present invention, the processing unit may be configured to control the detection unit at least based on the first plurality of output values of the output value vector, and to control the operation of the tube gripper at least based on the second plurality of output values of the output value vector

[0034] According to an aspect of the present invention, the system may further comprise an agitating element (or shaking element) configured to shake (or agitate) said tube after the processing of the images, and a chain structure which is movable and defines a closed path for said tube, said chain structure comprising a plurality of measurement supports for a corresponding plurality of tubes, which are movable integrally with said chain structure.

[0035] According to an aspect of the present invention, the agitating element may comprise guides in engagement with portions of the chain structure, which is structured in portions that are connected to each other and are configured to rotate around an axis which is parallel to the advancement direction of the tubes, said agitating element comprising motorized means configured to move said guides and therefore to consequently cause the rotation of the portion of the chain structure engaged therewith.

[0036] The present invention also refers to a method for the analysis of tubes adapted to contain biological samples, comprising the steps of:

[0037] - acquiring at least one image of a tube by an image detector; and

[0038] - processing, through a processing unit, the images of the tube acquired by the image detector, wherein said processing step comprises at least the steps of

[0039] - possibly and preferably performing a pre-processing of the acquired images; and

[0040] - performing, using the at least one acquired image as an input, a processing procedure based on a neural network (in particular, a convolutional neural network) which is trained based on exemplary images of tubes to perform a classification, wherein, in said exemplary images, said tubes comprise a label associated thereto, the method further comprising a step of outputting an output value vector comprising (structured in) a plurality of output values, each of said output values being related to a certain class (in particular class of tubes) and representing the activation level of a corresponding neuron of an output of the convolutional neural network.

[0041] According to an aspect of the present invention, the output value vector may comprise (structured in) a first plurality of output values, each being related to a certain class of types of tubes, and a second plurality of output values, each being related to a certain piece of information about the label attached to said tube.

[0042] According to an aspect of the present invention, the neural network may comprise at least one input layer, wherein the processing unit provides the image to the input layer, the image being possibly pre-processed, at least one convolution layer which extracts features from the input image, and a fully connected network which performs, based on the information from the previous layers, the classification regarding the tube contained in the input image, said fully connected network comprising at least one first network layer or input network layer and a second network layer or output network layer, wherein the values of the output value vector represent the activation level of corresponding neurons of the output network layer of the fully connected network.

[0043] The present invention also refers to a computer program product for the control of a system for the analysis of tubes adapted to contain biological samples, said computer program product comprising code portions which, when executed by a processing unit of the system, are adapted to allow the execution of the above method. The features and advantages of the system and the method according to the invention will become apparent from the following description of an embodiment thereof, given by way of non-limiting example with reference to the accompanying drawings.

[0044] Brief description of the drawings

[0045] In the drawings:

[0046] - Figure 1 is a schematic representation of a system according to the present invention, in particular of a set of components for the preliminary analysis of tubes;

[0047] - Figure 2 shows a support for a tube, said support being in the form of a tube gripper, according to embodiments of the present invention;

[0048] - Figure 3 is a general outline of a convolutional neural network implemented according to embodiments of the present invention;

[0049] - Figure 4 is an example of the convolutional neural network implemented to perform a classification according to embodiments of the present invention;

[0050] - Figure 5 is a general outline of the system in accordance with embodiments of the present invention, the system also comprising an optical-detection unit for the analysis of blood samples contained in the tube;

[0051] - Figure 6 is a perspective view of the optical-detection unit in accordance with an embodiment of the present invention;

[0052] - Figure 7 is a top view of components of the system according to an exemplary embodiment of the present invention;

[0053] - Figure 8 is a perspective view of an agitating element for shaking tubes in accordance with an embodiment of the present invention; and - Figures 9A and 9B are examples of reading curves obtained with the system according to the present invention.

[0054] Detailed description

[0055] With reference to those figures, a system for the analysis of tubes adapted to contain a biological sample according to the present invention is globally and schematically indicated with the reference number 1.

[0056] It is worth noting that the figures represent schematic views and are not always drawn to scale, but instead they are drawn so as to emphasize the important features of the invention. Further, in the figures, the different elements are shown in a schematic way, their shape being variable depending on the desired application. It is further worth noting that in the figures identical reference numerals refer to identical elements in shape or function. Finally, particular features described in relation to an embodiment illustrated in a figure are also applicable to other embodiments illustrated in the other figures.

[0057] It is also noted that, unless the opposite is expressly indicated, the described process / processing steps may also be inverted if necessary.

[0058] The present invention provides a system for the analysis of tubes (identified with reference P) containing blood samples, and a measurement of the erythrocyte sedimentation rate (ESR, but not limited thereto) is then performed in relation to said blood samples. The tubes P are not limited to a particular type and may also be normal blood-count tubes (it should be noted that the inventive aspects described herein are not limited to the above-mentioned type of blood-count tube and are potentially applicable to any type of tube) .

[0059] In its most general form, the present invention provides a system and a related method capable of recognizing the type of tube P containing the blood sample to be analyzed and also to obtain information about the label (identified with reference L, and thus, for example, information regarding the extension and the placement of the label on the tube P, and possibly also regarding the way said label L is attached to the body of the tube P).

[0060] In order to allow the execution of the operations of the present invention, the system 1 comprises a processing unit or control unit (identified with reference C), including suitable memory units MEM and suitably programmed and designated for managing and automatically controlling the system, and for processing and analyzing data of measurement. The processing unit C may be, for example, a computerized unit integrated in the system 1 , or even external to the system 1 and operatively connected thereto. Moreover, the processing unit C may be a single unit or may comprise a plurality of local and / or remote units, possibly communicating with each other and each being designated for performing specific operations. The processing unit C is thus apt to control the system 1 to obtain the processing procedure which will be described below. In any case, the present invention is not, in any way, limited by the architecture used for the control unit C, which may be in general any suitable computerized unit, comprising one or more unit(s) depending on the needs and / or circumstances.

[0061] Further, the term “system 1” means a general apparatus for analysis, provided with a suitable casing and containing a plurality of components cooperating with each other to obtain the desired processing, calculation and subsequent measurement operations. However, said apparatus is not limited to a particular type. In any case, the present invention will be illustrated below in relation to a specific example wherein the tubes P, after their preliminary analysis, are moved by a chain structure along various reading stations, although, as mentioned above, the teaching herein described is not limited to this embodiment and can also be applied to many other types of apparatuses having a different configuration.

[0062] In particular, the present invention relates to an innovative method for processing images which is implemented in an apparatus of the above- mentioned type (but is not limited thereto) through the execution of the steps of a software which is inside the processing unit C (which may be a CPU, such as for example a mini-carrier IMX8 board). For example, the analysis module may be made by C++ language through a modular structure. The main modules may be related to the implementation of a serial communication protocol for driving the hardware componentry of the system 1, to a state machine for managing the acquisition sequence and synchronization of the steps, and clearly to the implementation of a system for acquisition, analysis and classification of images, as will be detailed below.

[0063] With reference to Figure 1, the system 1, in its most general form, comprises a support 2 for housing at least one tube P containing a blood sample. For example, in a preferred embodiment, the support 2 of the tube P is a rotating support that can be rotated by suitable moving means 2m’ and it is thus configured to allow a rotation of said tube P around its own longitudinal axis H-H, in particular a rotation of 360°.

[0064] The system 1 comprises at least one image detector 3 configured to acquire at least one image (identified with the reference Img) of the tube P which is located in the support 2. In particular, the image detector 3 is configured to acquire a plurality of images of the tube P, which, during the acquisition procedure, rotates according to the above-mentioned rotation movement.

[0065] Alternatively, less preferred embodiments are possible, wherein the tube P does not move and the image detector is moved, or wherein there are several (i.e., more than one) image detectors for acquiring images of said tube P from different views.

[0066] In any case, the image detector 3 is configured to acquire one or more images Img, for example eight images (but not limited to this number, even just one image being sufficient in some cases), each image being related to a different portion of the tube P (and, thus, to a different view thereof), said images Img being then suitably processed in the way described below. The system 1 also comprises a lighting system 4 adapted to lighten the tube P during the acquisition of the images Img. In an embodiment, the lighting system 4 comprises one or more LED, which are kept on during the acquisition to best lighten the area of analysis.

[0067] In an embodiment of the present invention, illustrated in Figure 2, the support 2 is in the form of a gripper (also indicated below with reference 2), which is configured to pick up the tubes P from a housing area (for example, from a corresponding rack) and to move said tubes P towards an area of analysis thereof.

[0068] A previously mentioned, in a preferred embodiment of the present invention, in order to obtain different views, the tube P is rotated by the rotation of the gripper 2, which is thus a rotating gripper under the action of the moving means 2m’. As illustrated in Figure 2, the gripper 2 comprises a load-bearing structure 2’ and it is put in rotation by the above-mentioned moving means 2m’. There are also further moving means 2m”, for example adapted to allow the above-mentioned movement of the tube P from the housing area towards the area of analysis and vice versa.

[0069] Clearly, the above embodiment is only a non-limiting example and it does not limit the scope of the present invention in any way, wherein any suitable support for the tubes may be used.

[0070] Moreover, as mentioned above, in an embodiment, the processing unit C is programmed so as to operate based on a state machine at least for managing the reading of configuration files for the configuration of the system 1, for the control of the lighting system 4, for the control of the moving means 2m’ and 2m” of the tube P, as well as for the identification of desired acquisition instants as a function to the movement of said moving means, in particular as a function of the steps of the motor of the moving means 2m’ that cause the rotation of said tube P (i.e., by defining the moment in which the images Img should be acquired during the rotation of the tube P under the action of the moving means 2m’). In an exemplary embodiment, the system 1 thus implements the acquisition of a plurality of images Img, as prescribed in a configuration file (for example, an external file). The standard operating mode provides switching on all the LEDs of the lighting system 4 to obtain the maximum lighting of the tube P and acquiring eight (but not-limited to this number) images thereof while said tube makes a full rotation around its own axis H-H, in order to display and make a view of all its sides and thus to obtain different views. When the tube P reaches one of the rotation angles at which the acquisition should take place, the acquisition module is activated, which then performs the acquisition and the processing of the single image (which may be done asynchronously), thereby obtaining the desired classification (in the way described below). At the end of the rotation of the tube P, the single classifications, namely the results of each independent image, are integrated with each other, so as to provide the most suitable result.

[0071] According to the present invention, the image detector 3 is thus configured to acquire multiple images Img of the tube P so as to obtain preliminary information about it, i.e., before said tube P is analyzed through optical detection units dedicated to the measurement of the erythrocyte sedimentation rate of the blood sample contained therein. As mentioned above, said images Img are in particular shot when the tube P is held by the gripper 2, which is adapted to rotate the tube P around its own longitudinal axis H-H. In particular, the processing unit C is in operative communication with said image detector 3 and it is configured to process the images Img of the tube P acquired by it to obtain information regarding said tube P and / or preliminary information regarding the label L (for example, its extension / position / placement on the body of the tube and / or the way it is attached to the body of the tube P), and to accordingly control the successive detection units. For example, based on the information obtained by processing the images acquired through the image detector 3, it is possible to set some measurement values and thus to accordingly control the system, for example by defining which tubes to examine and which not, as well as other important settings , as will be detailed below.

[0072] Firstly, the processing unit C is configured to perform a pre-processing of the acquired images Img. In particular, the processing unit C is configured to perform the pre-processing of the images Img through definition of a region of interest of the acquired image Img; said image Img then is cropped so as to select the region of interest in which the tube P is located, and said image is then normalized according to a standardization procedure. Specifically, in the above-mentioned standardization procedure, an image (indicated as J in the following expression) is calculated through normalization of the pixels I of the image Img; the normalization is performed according to the following formula: wherein mean(I) indicates the mean of the pixels of the image and wherein std(I) indicates the standard deviation of said pixels.

[0073] Thus, this enables a preliminary normalization of the image Img, which will be then provided as an input to a convolutional neural network, as will be detailed below. In this way, the processing model implemented in accordance with the present invention takes as input the images of the samples acquired through the instrument, said images being suitably pre-processed, and then proceeds to the processing.

[0074] In an embodiment, the processing unit C is configured to identify the presence of the label L by using information contained in at least one library stored therein, for example a zxing and / or zbar library.

[0075] More in particular, in this embodiment, the processing unit C is configured to decode, once the presence of the label has been identified, the barcode which is on said label. This decoding can be also performed using algorithms contained in external libraries, such as, for example, the zxing and / or zbar library (in an embodiment, this library recognizes the presence of the label if there is a valid barcode printed thereon) .

[0076] Obviously, the reading of the barcode is independent of the specific implemented neural network which will be detailed below in the present description, however, it may be included in the same package stored in the processing unit C, comprising various modules for the execution of respective operations. As will be described below, the neural network, very advantageously, possibly allows to identify if the label is well attached (when the system recognizes a part of the body of the tube), if it is detached (because the system recognizes parts of raised label), or if the label is wrapped (because the system does not detect the body of the tube).

[0077] In other words, processing the acquired images Img firstly comprises the possibility of defining a region of interest to crop said images, so as to compensate for possible mechanical discrepancies between different apparatuses, and to eliminate areas are not related to the tube P. Afterward, detection of the label L is performed, said label carrying a barcode which the processing unit C uses for the recognition of the label, as mentioned before. To decode the barcode, the program executed by the processing unit C uses one of the two preexisting libraries (namely the above-mentioned zxing and zbar libraries, preferably the zbar library) . To further improve the detection, said operation is performed in two steps: first, the whole image Img is processed to verify if the used library is capable of identifying the barcode of the label L. If no barcode is identified, the second step comprises analyzing the image Img to identify the area in which there may be the barcode, through techniques based on image gradient. Once the candidate area has been identified, the image gradient provides the direction and / or the orientation of the bars of the barcode, so that the image Img can be digitally rotated and said barcode can be accurately displayed, for example when the label L is applied not in parallel to the axis of the tube P. The so-corrected image is analyzed again through the selected library.

[0078] Summing up, one or more images (for example, eight) are then acquired by rotating the tube and cropping the regions of interest (ROI, in particular two ROIs) which are used for coding the barcode and for the classification of the type of tube, which will be described below. For each cropped portion, the library (i.e., the zbar library) is apt to decode a possible barcode and a neural network (described below) is apt to yield the probability that the image contains a tube of one of the training types. The above information provide, through a probabilistic majority voting, the decoding of the barcode and the type of tube. As a consequence, barcode decoding is a software module that operates independently.

[0079] Moreover, in an embodiment, as regards the barcode of the labels L, the code considered correct is the code detected in the greatest number of acquisitions, provided that it has been detected in at least two of them. As a matter of fact, during rotation, the barcode is typically visible in four photographs (in the case eight photographs were shot, or, more in general, in 50% of the acquired photos, or any suitable percentage), thus providing excellent percentages of identification of the above-mentioned barcode.

[0080] As mentioned above, the pre-processed image is used as a base for the subsequent processing, and the description will further illustrate below the innovative system and method used to obtain the classifier that provides the desired classification. For the implementation of said classifier, deep-learning techniques are used, said deep learning being a branch of artificial intelligence based on artificial-neural-networks algorithms. Artificial neural networks comprise layers of “artificial neurons” which, similarly to the biological neurons, receive signals, process them and can send signals to the neurons connected to them. Said algorithms are capable of learning by examples, without the need for a step of selection and extraction of features from the input data, which is automatically performed by the architecture, as will be detailed below.

[0081] In particular, advantageously according to the present invention, the processing unit C is configured to perform a processing procedure based on a neural network, in particular for example a convolution neural network (indicated with the acronym CNN, whose general structure is outlined in the non-limiting example of Figure 3, which shows the main blocks of this model), which is particularly suitable for the processing of images.

[0082] The convolutional neural network CNN is trained based on exemplary images of tubes P, wherein said tubes P comprise a label L associated thereto. In particular, in said exemplary images, different types of tubes P and different states of the label L are displayed, thereby training the network to automatically recognize the different cases that may occur in reality. This training allows to define the weights of the various neurons of the network in order to obtain the desired classification.

[0083] The implemented and trained convolutional neural network CNN comprises, in the first place, an input layer LI. The processing unit C is then configured to provide the above-discussed pre-processed image Img to said input layer LI. Further, this input layer LI is the input layer to which the exemplary images are provided and its dimensionality is equal to the image size.

[0084] Subsequently to the input layer LI, there are the so-called hidden layers (herein indicated in general with reference L2), whose name is due to the fact that they are invisible from outside the network, which interfaces only through the input layer LI and the output layer. As will be detailed hereinafter, said hidden layers L2 comprise convolution layers and pooling layers which are configured to extract features from the image Img. In these layers, the image is processed with the mechanism of the so-called “sliding window”, in which a matrix (indicated in the art as filter or kernel) is made to slid on the image by performing a specific operation, in particular a scalar product between the pixel matrix of the image and the applied kernel.

[0085] The architecture underlying the present invention is thus a convolutional neural network CNN of the feed-forward type (wherein the output of a layer is used as input for the next layer), whose peculiarity is the presence of particular layers, called convolution layers and pooling layers, whose structure is inspired by biological processes, since the connectivity between neurons resembles the organization of the human visual cortex.

[0086] In particular, there are convolution layers (indicated with reference L2c) adapted to extract features from the input image for the subsequent classification, in particular allowing to obtain the so-called “feature map”. In particular, in these convolution layers L2c, a convolution operation between input data and the kernel is performed, wherein said operation allows to obtain as output the above-mentioned features of the image (namely, it allows to obtain the above-mentioned feature map, i.e., particular representations of the input image in which features of interest may be inferred) .

[0087] Said features are thus defined by the kernel matrices with which the convolution operations are performed. In general, the weights of the kernel matrix are initialized with small causal values, for example, but not limited to, based on the He method. After the first initialization, the weights are optimized based on the error; in particular, the training procedure optimizes the weights of the network so as to autonomously identify the most significant features to perform the desired classification.

[0088] In an exemplary embodiment, the size or dimension of the convolution kernel applied in the convolution layers L2c is 3x3.

[0089] As indicated above, the convolutional neural network CNN also comprises pooling layers (indicated with reference L2p), said pooling layers L2p being configured to subsample the information contained in the image, preferably by executing a maximum operation (“max pooling”) or alternatively an average operation (“average pooling”) among the pixel values of said image. In the case of max pooling, the maximum value (or in any case the greatest value for each feature map) is calculated, and the result is subsampled or collected together with the other values so as to generate a matrix containing the most present values. In an embodiment, the size or dimension selected for the pooling kernel applied in the pooling layers L2p is 2x2, without being limited to this value.

[0090] As indicated in Figure 4, which is a graphic representation of the implemented model, the layers adapted to extract the feature maps comprise a series of blocks, in each of which the above-mentioned convolution layers L2c and pooling layers L2p alternate with the so-called batch-normalization layers (indicated with reference L2n) which are adapted to normalize the input to the successive layers.

[0091] All these hidden layers, block by block, produce as output an increasing number of feature maps.

[0092] Summing up the above, the set of hidden layers L2 is characterized by the presence of several convolution, normalization and pooling layers, wherein the convolution layer allows to extract, through the use of filters, particular features of the images for the desired classification. The filter is a kernel (for example, 3x3) which is convoluted with the input and allows to obtain the above-mentioned feature map, and at the end of the operation it is thus possible to have several feature maps of the same acquired image.

[0093] Moreover, as indicated in Figure 4, there is also a flatten layer L2f which reduces the size of the output of the previous layers, creating a data vector (i.e., a one-dimensional array).

[0094] The following block comprises, in an embodiment, a fully connected network (identified with reference L3) configured, based on the information from the previous layers (i.e., based on the extracted features), to perform the classification regarding the tube P contained in the input image and regarding the label L associated thereto. This fully connected network L3 comprises at least one first network layer L3’ (also called input network layer L3’) and a second network layer L3” (also called output network layer L3”) which is adapted to produce said classification) . The fully connected network L3 is a feed-forward network in which each neuron in a layer is connected to all the neurons of the successive layer, the layers of this network enabling to produce the final classification.

[0095] In general, “classification of the type of tube” refers to the commercial type of tube, and possibly, as in the example of the figures and described herein, it is also possible to verify the type of attached label (which also possibly involves the way the label is attached to the tube) . “Information regarding the label” thus means the type of label (which, besides the type in the strict sense, may also comprise the way the label is attached to the tube, such as the placement on the body of the tube).

[0096] More in particular, in an exemplary, non-limiting embodiment, the fully connected network L3 comprises only the first network layer L3’ and the output network layer L3”, namely, it comprises two layers of artificial neurons. Even more in particular, the first network layer L3’ comprises one hundred artificial neurons and the output network layer L3” comprises nine artificial neurons, six of which related to different classes of tubes, and the remaining three related to different information about the label L attached to the body of the tube P; namely, said nine neurons correspond substantially to the desired classifications. In other words, the fully connected network L3 comprises at least two layers of neurons, the first having a number of units equal to one hundred and the second one having a number of units equal to the number of classes to be identified, e.g. nine, wherein the first six outputs are designated for the classification of the tube type, whereas the last three are designated for the identification of the label’s status (i.e., the placement / the way the label is attached and / or other information). Clearly, the number of neurons may vary based on the needs and / or circumstances, in particular based on the desired classification.

[0097] In an embodiment, the above-mentioned convolutional neural network CNN was thus trained with various exemplary images of tubes, with different brand types, and different states of the label L attached to the tube, so as to obtain the suitable weights to implement the desired classification.

[0098] Finally, an output value vector V, which is representative of the output network layer L3” of the fully connected network L3 (i.e., of the last network layer) and thus representative of the desired classification, is outputted.

[0099] In general, the output value vector V comprises (is structured into) a plurality of output values, each of said output values being related to a certain class of the classification(s) and representing the activation level of a corresponding neuron of an output of the convolutional neural network CNN, and thus the corresponding probability that the image’s content belongs to a class of said classification, thereby outputting the class having the highest probability.

[0100] In particular, in an embodiment, the output value vector V comprises a first plurality of output values VI, each being related to a certain class of type of tubes, and a second plurality of output values V2, each being related to a certain piece of information about the label L on said tube P (for example, the type - label type - which may also comprise the extension and the placement of the label L relative to the body of the tube and / or the way it is attached, and thus, in general, the state of the label L).

[0101] The values of the output value vector V represent the activation level of a corresponding neuron of the output network layer L3” of the fully connected network L3. In other words, in accordance with the present invention, the output value vector V is produced as an output of the network, said vector representing the activation for each neuron of the last network layer.

[0102] Thus, in one embodiment, the classification is obtained by identifying the neurons corresponding to the highest activation, each neuron having its own activation percentage, and the neuron with the highest activation represents the output class.

[0103] Based on the output value vector V, it is thus possible to provide a classification result, indicating in particular the type of tube P (through the first plurality of output values VI, namely, the first six values corresponding to the first six neurons of the output network layer L3”) and information about the label L applied thereto (through the second plurality of output values V2, namely, the last three values corresponding to the last three neurons of the output network layer L3”). Clearly, as mentioned above, any number of output value is possible depending on the applications.

[0104] In other words, the output value vector V is part of the output layer in which the output is provided. According to a point of view, the output network layer L3” corresponds with the output value vector V, which represents the output of the network. In general, it is the output of the network.

[0105] In an example, the processing unit C is thus configured to perform the desired classification through identification of the neurons of the output network layer L3” of the fully connected network L3 which correspond to the highest activation level, thereby creating, for each acquired image, the above-mentioned output value vector V.

[0106] As known in the art, the activation level of the neurons of the network is given through an activation function, which is not limited to a particular type. Possibly, different activation functions may be used for different layers, such as, for example, the function RELU for intermediate states and the function SoftMax for the last layer (which takes as an output the activation maximum). Clearly, other solutions are also contemplated and fall within the scope of the present invention.

[0107] The processing unit C is thus configured to perform the desired classification through identification of the neurons of the output network layer L3” of the fully connected network L3 which correspond to the highest activation level, thereby creating, for each acquired image, the above-mentioned output value vector V.

[0108] As indicated above, a plurality of images is acquired for a single tube P, for example eight images, so that the processing unit C is programmed to integrate all the acquired data, starting from said plurality of images; namely, the final classification is produced considering the predictions obtained on each angle shot.

[0109] In particular, according to the present invention, the processing unit C is configured to perform an average operation among the output value vectors V calculated for each image in order to perform the classification of the type of tube P (e.g., in order to obtain the first six values of the averaged output value vector). More in particular, in a non-limiting example of the present invention, the so-implemented classifier discriminates primarily the type (or brand) of the tube P, attributing one on the following classes to it:

[0110] - STANDARD;

[0111] - GREINER (BIO - ONE);

[0112] - SARSTEDT S;

[0113] - Greiner Minicollect;

[0114] - Kima Microtest; and

[0115] - BD Microtainer MAP.

[0116] Clearly, the classification is not limited to the above-mentioned classes but may be also extended to other classes, in addition or as an alternative thereto; for example, the classifier may be trained with additional classes for pediatric tubes, as well as other tube types not limited to the above- mentioned six. For this reasons, the value number of the output value vector may be different from the one herein shown as an example. In this case, as mentioned above, the final classification is calculated by averaging all the output value vectors and considering the class related to the brand type that obtained the highest average value. An average operation among the results of the classification of the various acquired images is then performed (namely, the average of the various tube exposures), selecting the tube type showing the best average probability. The average operation allows to improve the results of the classification.

[0117] Moreover, in an embodiment, in connection with the label L applied on the tube P, the processing unit C is configured to verify, through the majority voting technique, the number n of times in which a certain state (for example, a certain placement) of the label L appears among all the acquired images. More in particular, in a particular embodiment, for the discrimination of the tube based on the application of the label, the following classes are defined:

[0118] - well-adhering label (namely, visibility of both the label L and the body of the tube, and therefore the optimal condition);

[0119] - raised label (namely, presence of a not correctly adhering label, regardless of the visibility of the body); and

[0120] - wrapped label (visibility only of the label and totally covered body of the tube).

[0121] Also in this case, further classes may be identified, in addition and as an alternative, and therefore the number of the vector values may be different from those herein shown as an example. In any case, this enables the discrimination of the tube based on label application.

[0122] In general, the number of output values of the output value vector V (and thus of the neurons of the output layer) is thus only indicative and nonlimiting, and other classes could be also discriminated depending on specific needs and / or applications.

[0123] In the above example, in this way, the number of votes taken by each of the above-mentioned three classes is calculated from the single predictions. In an embodiment, the identification of raised labels, which could cause system malfunctioning and thus interrupt the subsequent analysis, is prioritized. Initially, therefore, the number of classifications associated to the class “raised” is verified: in an example, if at least three images fall within this class, the sample is classified as “raised label” and thus it will not be inserted into the chain for the successive analysis. Subsequently, it is verified that the tube’s body is visible from at least three angle shots (i.e., three different acquired view), namely, the label is well adhering and does not cover the whole surface of the sample (in this case the output of the system is “label ok”) . Finally, in the case the model does not identify the presence of a raised label nor the presence of the body of the tube, the sample is classified as belonging to the class “wrapped label”. In other words, as regards the label, it is taken into account that the configuration varies with the rotation of said tube, thus defining the above-mentioned criterion, based on the minimum number of exposures or view (in the case considered, with a threshold of three exposures, but obviously any other number may be adopted). The abovedescribed sequence of classification is the preferred sequence of classification for the label L, but obviously other sequences are not excluded.

[0124] In any case, all this provides very useful information for the successive analysis of the biological sample contained in the tube P: for example, based on how the label L is positioned, one may choose to process the tube P or to place it back into the rack (as in the case of “raised” label), as well as, based on this information, the parameters of the successive optical-absorption measurement of the biological sample are adjusted, as will be detailed below.

[0125] Obviously, the number and the type of classes are not limited to the above-mentioned examples and other classifications may be implemented through the procedure of the present invention, namely, by applying the above-described convolutional neural network CNN. In this case, the number of elements of the output vector may vary.

[0126] In an embodiment of the present invention, the processing unit C executes two threads in parallel for managing the classification procedure. Each thread handles the management of a single image Img, from its acquisition up to completion of the analysis procedure. The implementation of two parallel threads allows to parallelize the processing of two images, if the time required for the processing is longer than the interval between the two photographs. Experimentally, it has been observed that the processing time is indeed very close to the above- mentioned interval, and therefore two sessions or threads are sufficient to ensure an appropriate safety margin.

[0127] As mentioned above, advantageously according to the present invention, the performed classification allows to efficiently control the components of the system 1 that are dedicated to the successive measurement on the blood sample contained in the tube P, for example for the measurement of the erythrocyte sedimentation rate. Figure 5 shows a general outline of these additional components, wherein the components dedicated to the above-mentioned classification are globally indicated with reference 1 ’.

[0128] The following pages thus illustrate additional components of the system

[0129] I that are adapted to enable measurements on the blood sample, according to a non-limiting example of the present invention.

[0130] In an embodiment, the tubes P are picked up by the gripper 2 and are arranged on measurement supports 10. In particular, the system 1 comprises a plurality of measurement supports 10 for housing a corresponding plurality of tubes P, said measurement supports 10 being arranged in a chain structure that will be described below.

[0131] The system 1 further comprises an agitating element or shaking element

[0132] I I configured to agitate or shake the tube P and thus to allow the successive evaluation of the sedimentation process. The agitating element 11 is not limited to a particular configuration and substantially depends on the type of support used for housing and possibly moving the tubes. An example in which the agitating element 11 cooperates with a movable chain structure on which the tubes P are arranged will be illustrated below. However, said example does not limit the protection scope to said configuration. It is indeed observed that, when the tubes are arranged on other types of support, such as for example flat circular supports (circular plates), the agitating elements are obviously different and adapted to the specific case.

[0133] There is also at least one optical-detection unit (identified with reference 12 and hereinafter also called reading unit or opto-electronic unit) which is configured to perform optical measurements, in particular opticalabsorption measurements, on the blood sample contained in the tube P.

[0134] In particular, the detection unit 12 comprises at least one emitter 12’ and a corresponding detector 12” arranged so as to irradiate with electromagnetic radiation the tube P and to collect the radiation after it has crossed said tube P. The emitter 12’ is preferably a LED configured to emit substantially white light, so that the presence of labels on the tube P (and other external factors) do not affect the measurement. Obviously, the emitter 12’ is not limited to the above-mentioned type. For example, it may be any light source configured to emit radiation in the infrared, or more in general at any suitable wavelength, and the detector 12” may thus be selected accordingly.

[0135] The detection unit 12 thus allows, through absorption measurements on the blood sample in the tube P, to obtain reading curves (indicated with reference R) that are used as starting point for calculating the quantities of interest, such as for example the erythrocyte sedimentation rate (ESR) .

[0136] The system 1 also comprises suitable moving means 12m that are configured to cause a relative movement between the detection unit 12 and the tube P during the optical-absorption measurement, so as to irradiate said tube P in different portions, and thus so as to obtain a plurality of discrete measurement points that make up the reading curves R. In particular, the moving means 12m are configured to cause a steplike movement of the detection unit 12 (for example, a raising and lowering movement relative to the longitudinal axis H-H of the tube P). As will be detailed below, the reading curves obtained are representative of the light intensity detected as a function of the reading steps (said steps being possibly convertible into time instants) .

[0137] As illustrated in Figure 6, which shows a non-limiting example of the reading unit 12, the moving means 12m may comprise a carriage moved by the action of a suitable motor unit 12mu (comprising its own control driver board, which is also identified with reference 12mu). The detection unit 12 (in particular both the emitter 12’ and the detector 12”) is then arranged on the carriage 12m, enabling its movement in a direction that is substantially parallel to the longitudinal axis H-H of the tube P, and thus allowing the acquisition of the plurality of measurement points along said longitudinal axis H-H. In this example, the motor unit 12mu moves a worm screw 12v, which in turn causes the movement of the carriage 12m. Moreover, in the example of Figure 6, the emitter 12’ and the detector 12” are both moved by the carriage 12m, said carriage being suitably shaped to allow housing of the tube (not illustrated in Figure 6) in a substantially central position thereof. All the above-mentioned components are supported by a support 12s, which thus acts as a loadbearing structure of the detection unit 12.

[0138] However, the present invention is not, in any way, limited by the configuration of the detection unit 12, ant it is thus possible to adopt any other suitable configuration, for example in relation to the movement of the detectors / emitters or their arrangement.

[0139] Referring now to Figure 7, in a particular embodiment of the present invention, as previously mentioned, the system 1 comprises a chain structure 15 on which the measurement supports 10 for the tubes P are formed. The chain structure 15 is movable and defines a closed path for the tubes P, said closed path lying substantially in a horizontal plane, for example parallel to the surface on which the system 1 is arranged. The chain structure 15 comprises a plurality of portions (for example said measurement supports 10) that are connected to each other, each portion providing a support for a respective tube P. In an embodiment, the portions of the chain 15 are connected with each other through a ball joint and can rotate with respect to each other.

[0140] In this way, the measurement supports 10 for the tubes P are comprised in the chain structure 15, which is movable and defines the closed path of analysis of said tubes P, said chain structure 15 comprising a plurality of measurement supports 10 for a corresponding plurality of tubes P, which are thus e movable integrally with said chain structure.

[0141] In a particular embodiment, the system 1 comprises at least two detection units, preferably four detection units, arranged along the chain structure 15 so that each of said four detection units is configured to analyze a tube P moved by the chain structure 15 at a certain time instant.

[0142] In particular, a first detection unit 12a acquires a first reading curve immediately after shaking of the tubes (thus performing a reference reading), whereas a second reading unit 12b, arranged in a different position along the chain 15, carries out a measurement after the tube P has traveled for a certain sedimentation time, in particular after twenty minutes. There may optionally be also other two detection units, arranged at intermediates points of the chain structure 15, so as to also perform measurements at intermediate time instants (for example, at twelve and seventeen minutes).

[0143] Before being arranged on the chain structure 15, the images Img of the tube P are acquired through the image detector 3, as indicated above.

[0144] Summing up, in the embodiment of Figure 7, the analysis module M of the system 1 comprises the chain structure 15, which may have, for example, eighty-nine meshes (supports) into which the tubes P are inserted, said meshes being free to rotate at their junction point. The chain 15 rotates clockwise inside the analysis module due to the action of two drive wheels 15t moved by a motor unit 15m, thereby transferring the tubes P to a mixing unit and subsequently to opto-electronic units.

[0145] The movement speed of the chain structure 15 is set so as to enable the samples to stabilize for twenty minutes before the last reading is performed. As mentioned, the reading units in the analysis module are preferably four: the first reader is immediately after the agitating element, the optional second reader is located so that the samples are read after twelve minutes, the optional third reader is in the position corresponding to a time of analysis of seventeen minutes, and the fourth reader is located near the position where the sample exits at a reading time of twenty minutes.

[0146] As mentioned, before performing the reading of the blood samples, the tubes are shaken by the agitating element or mixing element or shaking element, herein indicated with reference 11. As illustrated in Figure 8, the agitating element 11 may comprise guides 1 lg in engagement with portions (for example, lateral tracks) of the chain structure 15, which - as seen above - is structured in various portions that are connected to each other and are configured to rotate around an axis Y-Y which is parallel to the advancement direction of the tubes P. There are also suitable motorized means 11m configured to move said guides 11g by means of a suitable gear system and to consequently cause the rotation of the portion(s) of the chain structure 15 engaged therewith. Thus, a system based on guides and guide blocks (skates) that allows shaking of the tubes P is obtained. As illustrated in the non-limiting example of Figure 8, the guides 11g are formed in two flanges I lf, for example circular flanges, which moves integrally due to the action of the motorized means 11m.

[0147] Clearly, as repeatedly mentioned, the present invention is not, in any way, limited by the structure of the components of the system 1 and many other embodiments are possible. For example, in an embodiment not illustrated in the figures, the measurement support 10 and the agitating element 11 may also be a single component, which is suitable for both supporting and shaking the tubes P, of any suitable shape.

[0148] As discussed above, reading curves (reference curves) are acquired immediately after shaking of the tubes and then after a certain sedimentation time (up to twenty minutes), said reading curves containing information about the sedimentation process of the blood sample in the tube P. In general, the reading curve R has, in the transition from plasma to sediment, an evident variation of the intensity of the measured radiation, in particular, the radiation absorption is higher at the sediment, resulting in a decrease of the detected light intensity. There is also a plateau extending to the tube bottom, as illustrated in the example of Figures 9A and 9B, which show examples of reading curves R acquired through a detection unit 12 according to an embodiment of the present invention, in different instants, namely, immediately after shaking (Figure 9A) and after a certain sedimentation time (Figure 9B), their analysis allowing to assess ESR.

[0149] Suitably, based on the performed classification, the processing unit C is configured to control the above-mentioned detection unit 12 (for example, the intensity of the radiation emitted by the LEDs) based at least on the first plurality of output values VI, and to control the gripper 2 of the tube P based at least on the second plurality of output values V2.

[0150] Possibly, the results of the classification may also be provided to the user, for example on a display 16 of the system 1 or in any other suitable way.

[0151] In the light of the above, it is clear that the present invention also relates to a method for the analysis of tubes adapted to contain biological samples, said method comprising all the above-mentioned steps of the processing procedure, as well as to a computer program product for the analysis of tubes adapted to contain biological samples, said computer program product comprising code portions adapted to allow the execution of the above processing procedure.

[0152] In conclusion, the present invention thus allows to brilliantly overcome the technical problem, providing the above system and method and solving all the drawbacks of the prior art.

[0153] Advantageously, the system according to the present invention is provided with an analysis and processing module (for example, a suitable software stored in the processing unit) for the identification of the tubes and the evaluation of the way of attachment / placement of the label on said tubes, by extracting significant parameters from the acquired images.

[0154] An innovative analysis and processing module is thus provided, said module performing several functions, such as, for example, drive the LEDs for lighting the tube, controlling the gripper for picking up the tubes and thus the corresponding tube movement, acquiring the corresponding images and analyzing said images by deep learning. The outputs of the implemented convolutional neural network are represented through a value vector representing the various neural units (namely, the probability that an image is classified in a certain type of available tubes and labels), said value vector being ideally divided into two parts, which are related to the two performed classifications. For each image, the various output units are activates, simultaneously indicating the two performed classifications.

[0155] Thus, very useful preliminary information is provided, since the identification of the tube type and / or the label type (which, for example, also involves verifying the label’s placement) allows to adjust the measurement parameters for the successive analysis of the erythrocyte sedimentation rate of the blood sample contained in the tube.

[0156] For example, recognizing the tube type allows to preliminarily evaluate structural features of the tube, such as, for example, the type of plastic and the volume of the contained blood, and therefore it allows to select the most suitable parameter set to perform the subsequent analysis.

[0157] Similarly, recognizing the label’s placement on the body of the tube provides information about how to set the intensity of the LEDs of the optoelectronic detection units during the analysis of the blood samples. To correctly detect the blood level of a tube in which the whole surface of the body is covered by the label, the emitted radiation intensity is higher than the intensity used for a tube in which a single label that leaves some body portions free is applied. Information is also provided in the case a label is poorly adhering to the tube body, so that such a tube is not processed but only placed again into the rack, since poorly adhering labels might stick into the meshed (or portions) forming the analysis chain and thus cause mechanical malfunctioning of the instrument.

[0158] According to the present invention, the probability of recognizing the type of tube and of the correct identification of the label’s placement / arrangement are extremely high, thereby showing the goodness of the used method.

[0159] Therefore, the present invention not only recognizes the tube, but discriminates among various classes of tubes.

[0160] A comprehensive analysis is performed and the various tube classes are discriminated, obtaining the form factor of the tubes (which also affects the matrix used in the analysis of the sample).

[0161] Therefore, according to the present invention, the analysis matrix of the sample may vary based on the tube: in other words, the processing unit, in the analysis of the sample, is configured to apply an analysis matrix to selected from a plurality of analysis matrixes based on the performed classification.

[0162] As seen before, also the experimental parameters of the system may be set and varied based on the classification.

[0163] Obviously, a person skilled in the art, in order to meet particular needs and specifications, may carry out several changes and modifications to the system and the method described above, all included in the protection scope of the invention as defined by the following claims.

Claims

CLAIMS1. A system (1) for the analysis of tubes adapted to contain biological samples, comprising:- at least one support (2) for a tube (P) adapted to contain a biological sample;- at least one image detector (3) configured to acquire at least one image (Img) of the tube (P); and- a processing unit (C) adapted to process the images (Img) of the tube (P) acquired by the image detector (3), wherein said processing unit (C) is configured to perform a processing procedure based on a neural network (CNN) which is trained based on exemplary images of tubes (P), wherein, in said exemplary images, said tubes (P) comprise a label (L) associated thereto, wherein the neural network (CNN) is programmed to perform a classification relating to the tube (P) contained in the at least one image (Img) provided as an input to said neural network (CNN), and wherein the processing unit (C) is further configured to output an output value vector (V) comprising a plurality of output values, each of said output values being related to a certain class and representing the activation level of a corresponding neuron of an output of the neural network (CNN).

2. The system (1) according to claim 1, wherein the neural network (CNN) comprises:- an input layer (LI), wherein the processing unit (C) is configured to provide the image to said input layer (LI);- at least one convolution layer (L2c) adapted to extract features from the input image; and- a fully connected network (L3) configured, based on the information from the previous layers, to perform said classification regarding the tube (P) contained in the input image, said fully connected network (L3) comprising at least one first network layer or input network layer (L3’) and a second network layer or output network layer (L3”), wherein the output value vector (V) is representative of the output network layer (L3”) of the fully connected network (L3), said values representing the activation level of corresponding neurons of said output network layer (L3”) of the fully connected network (L3).

3. The system (1) according to claim 2, wherein the processing unit (C) is configured to perform said classification through identification of the neurons of the output network layer (L3”) of the fully connected network (L3) which correspond to the highest activation level.

4. The system (1) according to claim 2 or 3, wherein the fully connected network (L3) comprises only the first network layer (L3’) and the output network layer (L3”), wherein the first network layer (L3’) comprises one hundred artificial neurons and the output network layer (L3”) comprises at least nine artificial neurons.

5. The system ( 1) according to any one of claims 2 to 4, wherein the size of a convolution kernel applied in the convolution layer (L2c) is 3x3.

6. The system (1) according to any one of the previous claims, wherein the processing unit (C) is configured to identify the presence of the label using information contained in at least one library, for example a zxing and / or zbar library, contained in a memory unit (MEM) of said processing unit (C).

7. The system (1) according to any one of the previous claims, wherein the processing unit (C) is configured to perform a pre-processing of the acquired images (Img) before providing them as input to the neural network, in particular through definition of a region of interest of the acquired image.

8. The system (1) according to claim 7, wherein, in said pre-processing, the processing unit (C) is configured to crop the acquired image (Img) so as to select the region of interest in which the tube (P) is located, and to normalize said image (Img) according to a standardization procedure.

9. The system (1) according to any one of the previous claims, wherein the neural network (CNN) is a feed forward network.

10. The system (1) according to any one of the previous claims, wherein the support (2) of the tube (P) is a rotating support which is moved by moving means (2m’) and configured to allow a rotation of said tube (P) around a longitudinal axis (H-H) during the acquisition of the images by means of the image detector (3).

11. The system ( 1) according to claim 10, wherein the image detector (3) is configured to acquire a plurality of images during the rotation of the tubes (P) so as to acquire images of different portions of said tubes (P), and wherein the processing unit (C) is configured to:- perform an average operation among the value vectors (V) calculated for each image of said plurality of acquired images in order to perform the classification with respect to a type of tube (P); and- verify a number of times in which a certain state of the label (L) appears in said acquired images, performing a classification through the majority voting technique.

12. The system (1) according to claim 10 or 11, wherein the processing unit (C) is programmed so as to operate based on a state machine at least for managing the reading of configuration files, for the control of a lighting system (15) adapted to lighten the tube (P), for the control of the moving means (2m j of said tube (P), and for the identification of desired acquisition instants according to the movement of said moving means (2m j.

13. The system (1) according to any one of the previous claims, whereinthe neural network (CNN) comprises at least one pooling layer (L2p) configured to subsample the information contained in the image by executing a maximum or average operation among the pixel values of said image.

14. The system (1) according to any one of the previous claims, comprising at least one detection unit (12) configured to perform optical absorption measurements on a blood sample contained in the tube (P).

15. The system (1) according to any one of the previous claims, wherein the support (2) is in the form of a tube gripper, which is configured to pick up the tubes (P) from a housing area and to move said tubes (P) towards an area of analysis thereof.

16. The system (1) according to any one of the previous claims, further comprising an agitating element (11) configured to shake said tube (P) after the processing of the images, and a chain structure (15) which is movable and defines a closed path for said tube (P), said chain structure (15) comprising a plurality of measurement supports (10) for a corresponding plurality of tubes, which are movable integrally with said chain structure (15), and wherein the agitating element (11) comprises guides (11g) in engagement with portions of the chain structure (15), which is structured in portions that are connected to each other and are configured to rotate around an axis (Y-Y) which is parallel to the advancement direction of the tubes (P), said agitating element (11) comprising motorized means (11m) configured to move said guides (11g) and to consequently cause the rotation of the portion of the chain structure (15) engaged therewith.

17. The system (1) according to any one of the previous claims, wherein the output value vector (V) comprises a first plurality of output values (VI), each being related to a certain class of types of tubes, and a second plurality of output values (V2), each being related to a certain piece of information about the label (L) attached to said tube (P).

18. The system (1) according to claims 4 and 17, wherein at least six of said at least nine artificial neurons are related to different classes of tubes, and at least three of said nine artificial neurons are related to different information about the label (L) on the tube (P).

19. The system (1) according to claims 14, 15, and 17, wherein the processing unit (C) is configured to:- control the detection unit (12) at least based on the first plurality of output values (VI) of the output value vector (V); and- control the operation of the gripper (2) of the tube (P) at least based on the second plurality of output values (V2) of the output value vector (V).

20. A method for the analysis of tubes adapted to contain biological samples, comprising the steps of:- acquiring at least one image (Img) of a tube (P) by an image detector (3); and- processing, through a processing unit (C), the at least one image (Img) of the tube (P) acquired by the image detector (3), wherein said processing step comprises at least the steps of performing, using said at least one image (Img) as input, a processing procedure based on a neural network (CNN) which is trained based on exemplary images of tubes (P), wherein, in said exemplary images, said tubes (P) comprise a label (L) associated thereto, the method further comprising the step of outputting an output value vector (V) comprising a plurality of output values, each of said output values being related to a certain class and representing the activation level of a corresponding neuron of an output of the neural network (CNN).

21. The method according to claim 20, wherein the neural network (CNN) comprises at least:- one input layer (LI), wherein the processing unit (C)provides the image to said input layer (LI);- at least one convolution layer (L2c) which extracts features from the input image; and- a fully connected network (L3) which performs, based on the information from the previous layers, a classification in relation to the tube (P) contained in the input image, said fully connected network (L3) comprising at least one first network layer or input network layer (L3’) and a second network layer or output network layer (L3”), wherein the values of the output value vector (V) represent the activation level of corresponding neurons of the output network layer (L3”) of the fully connected network (L3).

22. The method according to claim 20 or 21, wherein the output value vector (V) comprises a first plurality of output values (VI), each being related to a certain class of types of tubes, and a second plurality of output values (V2), each being related to a certain piece of information about the label (L) attached to said tube (P).

23. A computer program product for the control of a system (1) for the analysis of tubes adapted to contain biological samples, said computer program product comprising code portions which, when executed by a processing unit (C) of the system ( 1), are adapted to allow the execution of the method according to any one of claims 20 to 22.

Images