A super multiplex PCR detection system and method for cell-free DNA in blood.

Abstract

This application relates to the field of nucleic acid detection technology and discloses a super multiplex PCR detection system and method for cell-free DNA in blood. The multiplex PCR detection system includes a detection experimental body, a delivery module, a sample preparation module, a reaction solution addition module, a mixing module, an oil phase liquid addition module, a drop preparation module, a dropwise sample preparation module, a sealing and shaking module, a digital PCR instrument, a transfer module, a disposal chamber, and a transfer robot. The dropwise sample preparation module includes a multi-axis robotic arm housed within the detection experimental body. The free end of the multi-axis robotic arm is equipped with a syringe and an injection nozzle mounted on the syringe. The opening edge of the injection nozzle forms a shearing blade. A trachea is connected to the syringe, and a pressure sensor and a pressure regulating valve are installed on the trachea. The pressure sensor is connected to a first controller via wired or wireless means, and the first controller controls the pressure sensor. This application can ensure the accuracy of experimental results in large-scale PCR detection.

Description

Technical Field

[0001] This application relates to the field of nucleic acid detection technology, and in particular to a super multiplex PCR detection system and method for cell-free DNA in blood. Background Technology

[0002] Multiplex PCR, also known as multiplex primer PCR or multiplex PCR, is a PCR reaction in which two or more pairs of primers are added to the same PCR reaction system to amplify multiple nucleic acid fragments at the same time. In medical testing, multiplex PCR is often used to detect viruses and diseases, and is particularly suitable for the detection of the current novel coronavirus.

[0003] The specific procedure for PCR testing is as follows: A small amount of secretions is collected from the subject's throat or nose using a cotton swab; sometimes, sampling from the deep respiratory tract is also necessary. After sample collection, PCR reaction solution is added to the sample, which is then sealed and thoroughly stirred before being placed in a digital PCR instrument for amplification and detection. The resulting images are then acquired and the test results analyzed. During the procedure, the PCR reaction solution must be added dropwise to the oil phase; otherwise, the accuracy of the test results will be reduced. While manual operation can meet the requirements, it is unsuitable for PCR testing of large batches of samples. Summary of the Invention

[0004] To ensure the accuracy of experimental results in high-volume PCR detection, this application provides a super multiplex PCR detection system and method for cell-free DNA in blood.

[0005] The ultra-multiplex PCR detection system for cell-free DNA in blood provided in this application adopts the following technical solution:

[0006] A super multiplex PCR detection system for cell-free DNA in blood includes a detection experimental body for providing a clean detection experimental environment;

[0007] The delivery module, located inside the main body of the testing experiment, is used to deliver the test samples into the main body of the testing experiment;

[0008] The sample preparation module, located within the main body of the detection experiment, is used to store the PCR reaction solution;

[0009] The reaction solution addition module is used to add PCR reaction solution to the experimental samples on the delivery module.

[0010] The mixing module is used to mix the PCR reaction solution with the experimental samples.

[0011] The oil phase liquid addition module is used to add oil phase liquid to the culture plate;

[0012] A droplet-making module for injecting a mixture of PCR reaction solution and experimental sample into an oil phase liquid and preparing a droplet-like mixture;

[0013] A drop-by-drop sample preparation module for placing the droplet-like mixture into a culture plate drop by drop;

[0014] A plate-sealing and shaking module for sealing the culture plate and shaking the droplet-like PCR pre-reaction solution evenly after sealing;

[0015] A digital PCR instrument for amplifying and fluorescently detecting the target DNA in the droplet-like PCR pre-reaction solution;

[0016] A transfer module is arranged inside the detection experiment main body, for transferring the experimental sample on the conveying module to the plate-sealing and shaking module, and after the droplet-like PCR pre-reaction solution is sealed and shaken, sending the droplet-like PCR pre-reaction solution into the digital PCR instrument;

[0017] A waste disposal cabin is arranged inside the detection experiment main body, for storing waste culture plates;

[0018] A transfer manipulator is arranged inside the detection experiment main body, for grasping the waste culture plate and throwing it into the waste disposal cabin;

[0019] Among them, the drop-by-drop sample preparation module includes a multi-axis robotic arm arranged inside the detection experiment main body. A syringe is arranged at the free end of the multi-axis robotic arm, and an injection gun head is arranged on the syringe. A shearing edge is formed at the opening edge of the injection gun head. A trachea is connected to the syringe, and a pressure sensor and a pressure regulating valve are installed on the trachea. The pressure sensor is connected to a first controller in a wired or wireless manner, and the first controller is control-connected to the above-mentioned pressure sensor;

[0020] Among them, the first controller controls the pressure magnitude output by the pressure regulating valve through a pre-trained air pressure regulation model. The pre-trained air pressure regulation model is F = a×(b - 1 / 3 πw^2 x) + c; where F is the pressure magnitude output by the pressure regulating valve, the value range of a is between 1.05 and 1.5, x is the measurement scale value of the liquid level height of the solution formed in the gun head when the injection gun head aspirates the solution, w is 1 / 2 of the diameter of the circular liquid surface of the solution formed in the gun head when the injection gun head aspirates the solution, the value of b is not less than the total volume value of the solution that can be accommodated when the injection gun head is used, c is an offset amount, 0 < c < y, and the value of y is not greater than the calculated value of (b - 1 / 3 πw^2x).

[0021] By adopting the above technical solution, this application utilizes a delivery device to send experimental samples into the main detection chamber. A reaction solution addition module adds PCR reaction solution to the experimental samples to prepare droplet-shaped PCR pre-reaction solution. A transfer module then transfers the droplet-shaped PCR pre-reaction solution to a sealing and shaking module for sealing and agitation. The solution is then sent to a digital PCR instrument for amplification and detection. Finally, a transfer robot picks up the experimental samples and discards them into a disposal chamber, thus achieving fully automated mechanical multiplex PCR detection and meeting the needs of large-scale sample detection. Specifically, when extracting the PCR reaction solution, the PCR reaction solution is cut into droplets by the shearing blade on the syringe tip and stored in the syringe. A PCR droplet data model is established and stored in the first controller. The air pressure in the trachea is adjusted to achieve droplet-by-drop addition of the PCR reaction solution to the oil phase, thereby ensuring the accuracy of the experimental results.

[0022] Preferably, the free end of the multi-axis robotic arm is equipped with multiple image acquisition devices around the injection nozzle tip. These image acquisition devices are connected to a secondary data image analyzer via wired or wireless means. The secondary data image analyzer is used to determine whether the PCR reaction solution flows out of the injection nozzle tip in a droplet manner. The secondary data image analyzer is electrically connected to a first alarm. The secondary data image analyzer stores a liquid drop intelligent recognition model for determining whether the injection nozzle tip is adding solution in a droplet manner. The liquid drop intelligent recognition model includes a convolutional neural network model.

[0023] By adopting the above technical solution, convolutional neural network visual recognition is used to identify whether the PCR reaction solution is added drop by drop to the oil phase, further ensuring the accuracy of the detection operation.

[0024] Preferably, the secondary data image analyzer is also used to analyze whether the injection nozzle tip is contaminated, and the secondary data image analyzer is also connected to a second alarm; wherein, the secondary data image analyzer also stores an injection nozzle tip availability model for judging whether the injection nozzle tip is contaminated, and the injection nozzle tip availability model includes a convolutional neural network model.

[0025] By adopting the above technical solution and utilizing a convolutional neural network model, visual identification of whether the injection nozzle tip is contaminated can be achieved, thus avoiding the impact of contamination on the accuracy of the detection results.

[0026] Preferably, it further includes: a master data image analyzer, which is connected to the digital PCR instrument via wired or wireless means, for receiving experimental images from the digital PCR instrument and analyzing the experimental results; wherein, the master data image analyzer stores a convolutional neural network model.

[0027] By adopting the above technical solution and utilizing a convolutional neural network model, the experimental images of the digital PCR instrument are visually recognized, and the test results are automatically generated.

[0028] Preferably, the convolutional neural network model stored in the master data image analyzer analyzes fluorescence detection image data in the clinical sample test result image.

[0029] By adopting the above technical solution, a convolutional neural network model is constructed using the fluorescence detection image data in the clinical sample test result image, which saves a lot of manpower and avoids the unsustainability of human eyes recognizing a large amount of experimental data, thus ensuring the sustainability of a large number of sample tests.

[0030] The ultramultiplex PCR detection method for cell-free DNA in blood provided in this application adopts the following technical solution:

[0031] A method for detecting cell-free DNA in blood using ultramultiplex PCR, the method being performed using a device including a digital PCR instrument and / or a computer, comprising the following steps:

[0032] S1, Target DNA extraction and primer-probe mixture preparation;

[0033] S2, take the prepared primer-probe mixture;

[0034] S3, add the prepared primer-probe mixture to the target DNA sample and shake well to obtain a pre-reaction solution, and then prepare the pre-reaction solution into a droplet-shaped pre-reaction solution;

[0035] S4, dPCR amplification and fluorescence detection of the target DNA sample were performed using a digital PCR instrument;

[0036] S5, acquire fluorescence detection image data of the target DNA sample;

[0037] S6, Construct an intelligent recognition model for fluorescence detection image data of target DNA samples;

[0038] S7 intelligently identifies the genetic information of the target DNA;

[0039] S8 intelligently outputs recognition results;

[0040] The ultra-multiplex PCR detection method is automated through a designated device, which includes an injection pipette tip, a drop-making module, an image acquisition device for the injection pipette tip, and a microbial cleaning module A. The image acquisition device for the injection pipette tip is used to collect image data information from the injection pipette tip; the microbial cleaning module A is used to inactivate microorganisms in the active space of the injection pipette tip.

[0041] By adopting the above technical solution, a microbial cleaning module A is set up next to the injection nozzle to inactivate microorganisms in the active space of the injection nozzle, thereby ensuring the accuracy of batch experimental testing.

[0042] Preferably, step S3 further includes: using an injection nozzle to draw up the pre-reaction liquid, the injection nozzle being provided with a shearing blade, cutting the pre-reaction liquid into droplets and adding it to the oil phase substance to obtain droplet pre-reaction liquid.

[0043] By adopting the above technical solution, the pre-reaction liquid is cut into droplets using a shearing blade, and then an oil phase substance is added to obtain the droplet-shaped pre-reaction liquid.

[0044] Preferably, in step S1, primers are designed based on the target DNA; a nucleotide sequence is added to one end of the designed primer: GTACCATCTGTAGACTCACTATAGGAAGAGATGTCAACTCGTGCACGAGTTGACATCTCTTCTCTCCGAGCCGGTCGAAATATTGGAGGAAGCTCGAGCTGGAGGAAAAGTGAGTCTACAGATGGTAC.

[0045] By adopting the above technical solution, increasing the nucleotide sequence can improve the accuracy of PCR detection and avoid non-specific amplification during batch detection.

[0046] Preferably, the pretreatment method before primer amplification is as follows: the synthesized nucleotide sequence is added to buffer B and heated to 68°C for 7 minutes, then cooled to 30°C and incubated for 30 minutes to obtain the full-length nucleotide sequence for detection; wherein, the components of buffer B are: 300mM NaCl, 5mM MgCl2, 20mM Tris (pH 7.6).

[0047] By adopting the above technical solution, increasing the nucleotide sequence can improve the accuracy of PCR detection and avoid non-specific amplification during batch detection.

[0048] Preferably, it also includes: using microbial cleaning module B to inactivate and clean microorganisms in the detection operation environment.

[0049] By adopting the above technical solutions, the cleanliness of the experimental operating environment can be improved.

[0050] In summary, this application includes at least one of the following beneficial technical effects:

[0051] 1. This application sets up a shearing blade so that after the PCR reaction solution is drawn out, it is cut into droplets by the shearing blade on the syringe tip and then stored in the syringe. In conjunction with the PCR drop data model and pressure regulating valve, the air pressure in the trachea is adjusted to achieve the drop-by-drop addition of the PCR reaction solution into the oil phase substance, thereby ensuring the accuracy of the experimental results.

[0052] 2. Utilize a convolutional neural network model to visually identify the dropping status of the PCR reaction solution, ensuring the standardization of experimental operations. At the same time, identify the contamination status of the syringe tip to ensure that the experimental results are not affected by contamination.

[0053] 3. By setting up microbial cleaning modules A and B to kill microorganisms, it is beneficial to maintain the cleanliness of the testing environment and ensure the accuracy of experimental testing. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the overall structure of a super multiplex PCR detection system for cell-free DNA in blood, according to an embodiment of this application.

[0055] Figure 2 This is a schematic diagram of the internal structure of a super multiplex PCR detection system for cell-free DNA in blood according to an embodiment of this application (the main structure of the detection experiment is hidden).

[0056] Figure 3 yes Figure 2 Enlarged view of part A in the middle.

[0057] Explanation of reference numerals in the attached diagrams: 1. Main body of the detection experiment; 2. Transport module; 3. Sample preparation module; 4. Reaction solution addition module; 5. Mixing module; 6. Oil phase liquid addition module; 7. Shaking machine; 8. Aluminum foil sealing machine; 9. Sealing plate shaking module; 10. Digital PCR instrument; 11. Transfer module; 12. Disposal chamber; 13. Transfer robot; 14. Multi-axis robotic arm; 15. Syringe; 16. Injection nozzle; 17. Shearing blade; 18. Trachea; 19. Pressure sensor; 20. Pressure regulating valve; 21. First control. 21. Device; 22. Image acquisition device; 23. Secondary data image analyzer; 24. First alarm; 25. Second alarm; 26. Main data image analyzer; 27. Injection device; 28. Extraction nozzle; 29. ​​Negative pressure tube; 30. Microbial cleaning module B; 31. Drawer; 32. Container; 33. Lifting arm; 34. Sliding platform; 35. Liquid pump; 36. Liquid extraction tube; 37. Lifting rod; 38. Sliding seat; 39. Infusion pump; 40. Infusion tube; 41. Storage tank; 42. Microbial cleaning module A. Detailed Implementation

[0058] The following is in conjunction with the appendix Figure 1-3 This application will be described in further detail.

[0059] This application discloses a super multiplex PCR detection system and method for cell-free DNA in blood.

[0060] Reference Figure 1 and Figure 2A multiplex PCR detection system for cell-free DNA in blood includes a detection experimental body 1, which provides a clean detection experimental environment. The detection experimental body 1 consists of an instrument rack and an instrument housing. A transport module 2, which is a conveyor belt, is installed inside the detection experimental body 1 to transport experimental samples. One end of the conveyor belt extends from inside the detection experimental body 1 to the outside. A sample preparation module 3 is also installed inside the detection experimental body 1. The sample preparation module 3 includes a drawer 31 mounted on the detection experimental body 1 and a container 32 placed inside the drawer 31. During use, PCR reaction solution is added to the container 32 by pulling out the drawer 31.

[0061] Reference Figure 2 The detection experimental body 1 is equipped with the following modules arranged sequentially along the conveying direction of the conveying module 2: a reaction solution adding module 4 for adding PCR reaction solution to the experimental sample; a mixing module 5 for shaking the PCR reaction solution and the experimental sample mixture; an oil phase adding module 6 for adding oil phase liquid to the culture plate; a drop making module for injecting the PCR reaction solution and the experimental sample mixture into the oil phase liquid and preparing a drop-shaped mixture; a drop-by-drop sample making module for placing the drop-shaped mixture drop by drop into the culture plate; and a sealing and shaking module 9 for sealing the culture plate and shaking the drop-shaped PCR pre-reaction solution after sealing.

[0062] Reference Figure 2 The main body of the detection experiment 1 is equipped with a reaction solution addition module 4 between the sample preparation module 3 and the delivery module 2. The reaction solution addition module 4 extracts the PCR reaction solution from the sample preparation module 3 and adds it to the experimental sample on the delivery module 2. The reaction solution addition module 4 includes a lifting arm 33 installed on the delivery module 2. A sliding platform 34 that slides horizontally along the length of the conveyor belt is installed on the lifting arm 33. A liquid extraction pump 35 is installed on the sliding platform 34. A liquid extraction tube 36 is connected to the liquid extraction pump 35. One end of the liquid extraction tube 36 extends into the container 32 of the sample preparation module 3, and the other end is located directly above the delivery module 2. In this embodiment, a 96-well plate is used as the culture plate for the experimental sample. The experimental sample to be tested is placed only in one row of the 96-well plate. Therefore, the PCR reaction solution can be added to the wells of the 96-well plate containing the experimental sample by moving the outlet end of the liquid extraction pump 35 and the liquid extraction tube 36 driven by the sliding platform 34.

[0063] The delivery module 2 is equipped with a mixing module 5 for shaking the PCR reaction solution and the experimental sample mixture. The mixing module 5 is a shaker. The mixing module 5 is in contact with the conveyor belt and drives the 96-well plate on the conveyor belt to move, thereby shaking the PCR reaction solution and the experimental sample evenly.

[0064] Reference Figure 2The oil phase liquid addition module 6 includes a lifting rod 37 installed on the delivery module 2. The free end of the lifting rod 37 is horizontally slidably equipped with a sliding seat 38. An infusion pump 39 is installed on the sliding seat 38, and an infusion tube 40 is installed on the infusion pump 39. The inlet end of the infusion tube 40 is connected to a storage tank 41 for holding the oil phase liquid, and the other end is located above the delivery module 2. The sliding seat 38 drives the infusion pump 39 and the outlet end of the infusion tube 40 to move, adding the oil phase liquid into the wells of another row in the 96-well plate. The movement trajectories of the infusion pump 39 and the suction pump 35 are parallel but not collinear, to avoid adding the oil phase liquid into the PCR reaction solution and the experimental sample mixture.

[0065] Reference Figure 2 and Figure 3 The drop-making module includes a multi-axis robotic arm 14 installed within the main body of the detection experiment 1. The multi-axis robotic arm 14 can achieve lifting and lowering motion at its free end, as well as horizontal movement. A syringe 15 and an injection nozzle 16 are mounted on the free end of the multi-axis robotic arm 14. The injection nozzle 16 has a beveled tip at its extraction end, and the edge of the beveled tip is configured as a shearing blade 17 for cutting the mixture into droplets. An air tube 18 is connected to the syringe 15, and a pressure sensor 19 and a pressure regulating valve 20 are installed on the air tube 18. The pressure sensor 19... A first controller 21 is connected via wired or wireless means. The first controller 21 controls the air pressure sensor 19 connected to it. The first controller 21 controls the pressure output of the pressure regulating valve 20 through a pre-trained air pressure regulation model. According to the model data, the air pressure in the air pipe 18 is adjusted through the pressure regulating valve 20. The mixture is added to the oil phase liquid in droplets. The density of the oil phase liquid is greater than that of the mixture. When the volume of the mixture is injected into a drop, it is sheared by the shearing blade 17 under the action of buoyancy. At the same time, it is wrapped by the oil phase liquid to form a droplet that floats on the surface of the oil phase liquid.

[0066] Among them, the pre-trained barometric pressure regulation model is F represents the output pressure of the pressure regulating valve, and the value of a ranges from 1.1 to 1.3. The scale value for measuring the liquid level height of the solution formed in the syringe tip (read by monitoring the liquid level height on the syringe tip via image acquisition device 16), w is half the diameter of the circular liquid surface formed in the syringe tip, b is not less than the total volume of solution that the syringe tip can hold (default setting is the total volume of solution that the syringe tip can hold), and c is the bias (adjustable according to the application scenario of the instrument, 0...). c y, the value of y is not greater than The value of w is obtained by comparing the value of x obtained by the image acquisition device 16 with the data pre-stored in the database to obtain the corresponding value, thereby calculating the size of F.

[0067] Reference Figure 2 and Figure 3 The free end of the multi-axis robotic arm 14 is equipped with a circumferential array of multiple image acquisition devices 22 around the injection nozzle 16. In this embodiment, the circumferential array is equipped with four cameras. The four image acquisition devices 22 are connected to a secondary data image analyzer 23 via wired or wireless means. The secondary data image analyzer 23 stores a model for determining whether the injection nozzle 16 is adding solution in a droplet manner: a liquid droplet intelligent recognition model. The secondary data image analyzer 23 is electrically connected to a first alarm 24 (see...). Figure 1 If it is determined that the syringe tip 16 has failed to titrate the PCR reaction solution, the first alarm 24 will be activated to remind the experimenter to check and maintain it. The auxiliary data image analyzer 23 also stores a model for determining whether the syringe tip 16 is contaminated: the syringe tip 16 availability model. The auxiliary data image analyzer 23 is also electrically connected to a second alarm 25 (see Figure 1 If the syringe tip 16 is determined to be contaminated, the second alarm 25 is activated to remind the experimenter to replace the syringe tip 16. Both the first alarm 24 and the second alarm 25 are located in the main experimental unit 1, allowing experimenters to promptly observe any abnormalities. The liquid dispensing intelligent recognition model and the syringe tip 16 availability model can share a single neural network model or use their own independent neural network models.

[0068] The droplet-by-drop sample preparation module includes an injection device 27 fixed to the free end of a multi-axis robotic arm 14. The injection device 27 is connected to a negative pressure tube 29 and an extraction pipette tip 28. In use, the droplet-by-drop sample preparation module first aspirates droplets of the mixture floating on the surface of the oil phase, and then places one or more droplets into a clean well of a 96-well plate. This experimental operation facilitates subsequent quantification of experimental results. To avoid droplet breakage during extraction, the surface roughness Ra of the extraction pipette tip 28 does not exceed 0.9. Both the extraction pipette tip 28 and the injection pipette tip 16 are connected to the free end of the multi-axis robotic arm 14 via a rotating and retractable mechanism to prevent interference and contamination during operation.

[0069] The training method for the intelligent liquid dispensing recognition model or the syringe tip availability model includes the following steps:

[0070] Step S1: Read the liquid-carrying image data of the syringe tip or the solution-dropping image data of the syringe tip acquired by the image acquisition device and perform preprocessing; the specific operation of step S1 includes: reading at least 5000 liquid-carrying image data of the syringe tip or the solution-dropping image data of the syringe tip.

[0071] S2. Select the gun head portion of the image data and refine the shape of the gun head portion image;

[0072] The specific operation of step S2 includes: first performing Gaussian filtering to remove noise from the liquid-carrying image data of the syringe tip or the image data of the solution being added to the syringe tip read in process S1, so as to obtain preprocessed enhanced image data.

[0073] S3. Divide the gun head portion image shape extracted in step S2 into p groups on average, extract the data features of these gun head portion image shapes using a convolutional neural network, and then normalize them. The specific operations of step S3 include:

[0074] Step 1: Select 5000 image data sets after preprocessing in Step 2; Step 2: Experts extract the shape portions of the injection nozzle and non-injection nozzle images, and then train an automatic segmentation model using a convolutional neural network; Here, the convolutional neural network for the injection nozzle liquid-carrying image data or the injection nozzle solution-dropping image data consists of a network structure with 12 convolutional layers and 2 downsampling layers. The kernel sizes are: 12x12 for the first layer, 5x5 for the second and third layers, and 3x3 for the remaining layers. The stride is 2 for the first two convolutional layers and 1 for the rest. The downsampling layers are all 3x3 with a stride of 2.

[0075] The specific method for training an automatic segmentation model using a convolutional neural network is as follows:

[0076] (1) Features are automatically learned and extracted through the convolutional layers and downsampling layers of the convolutional neural network. The specific steps are as follows:

[0077] Step A: In a convolutional layer, the feature maps from the previous layer are convolved by a learnable convolutional kernel, and then passed through an activation function to obtain the output feature map; each output is the value of a convolutional kernel convolving one input or combining the values ​​of multiple convolutional inputs (here we choose to combine the values ​​of multiple input and output maps): Where * represents the convolution operator; l Indicates the number of floors; i express l -1st floor i One neuron node; j express l The first layer j One neuron node; M j This represents the set of selected input maps; It is the output; It means l The output of layer -1, as l Input at level 1; f It's the activation function; here we take... sigmoid function f(x) = As an activation function; e represents the Euler number 2.718281828, e x It's an exponential function; k is the convolution operator; b is the bias; each output map is given an additional bias b, but for a specific output map, the convolution kernel for each input map is different. This step also requires gradient calculation to update the sensitivity, which represents how much the error changes with how much b changes. ;in, l Indicates the number of floors; j express l The first layer j 1 neuron node; * indicates multiplication of each element; δ This represents the sensitivity of the output neuron, i.e., the rate of change of the bias b. s l = W l x l ; W As weight; b For bias; f It's the activation function; here we take... sigmoid function f(x) = As an activation function; e represents the Euler number 2.718281828, e x It is an exponential function, f'(x) is the derivative of f(x). If f takes the sigmoid function, then f'(x) = (1-f(x))*f(x); This represents the weights shared across all layers; `up(.)` represents an upsampling operation. If the downsampling factor is n, the upsampling operation copies each pixel n times horizontally and vertically, thus restoring its original size; then... l The gradient of bias b is quickly calculated by summing the values ​​of all nodes in the sensitivity map of the layer. ;in, l Indicates the number of floors; j express l The first layer j One neuron node; b Indicates bias; δ The bias indicates the sensitivity of the output neuron. b The rate of change; u , v This indicates the output maps ( u , v Position; E is the error function, here E= C represents the dimension of the label. If it's a binary classification problem, the label can be denoted as y. h ∈{0,1}, where C=1, which can also be denoted as yh ∈{(0, 1), (1, 0)}, where C = 2; the The h-th dimension represents the label corresponding to the nth sample; This represents the h-th output of the network corresponding to the n-th sample.

[0078] Finally, the backpropagation (BP) algorithm is used to calculate the weights of the convolution kernel: Where W is the weighting parameter; E is the error function, and C represents the dimension of the label. If it is a binary classification problem, the label can be denoted as y. h ∈{0,1}, where C = 1, which can also be denoted as y h ∈{(0,1), (1,0)}, where C = 2; the The h-th dimension represents the label corresponding to the nth sample; This represents the h-th output of the network corresponding to the n-th sample; η is the learning rate, i.e., the step size; since the weights of many connections are shared, for a given weight, it is necessary to calculate the gradient of all connections related to that weight at that point, and then sum these gradients: ;in, l Indicates the number of floors; i express l The first layer i One neuron node; j express l The first layer j One neuron node; b Let represent the bias, δ represent the sensitivity of the output neuron, i.e., the rate of change of the bias b; u, v represent the (u, v) positions of the output maps; E is the error function, where E = C represents the dimension of the label. If it is a binary classification problem, the label can be denoted as y. h ∈{0,1}, where C = 1, which can also be denoted as y h ∈{(0,1),(1,0)}, then C=2; This represents the h-th dimension of the label corresponding to the nth sample; This represents the h-th output of the network corresponding to the n-th sample; It is a convolution kernel; yes The elements in the convolution are compared with... The element-wise multiplication of the patch means that all regions in the image with the same size as the convolution kernel are multiplied. The value at position (u,v) of the output convolution map is the result of the patch at position (u,v) in the previous layer multiplied by the convolution kernel. The result of element-wise multiplication.

[0079] Step B: The downsampling layer has N input maps, so it has N output maps, but each output map is smaller. Therefore: ;in, f It's the activation function; here we take... sigmoid function f(x) = As the activation function, e represents the Euler number 2.718281828. x It is an exponential function; This represents the weights shared across layers; `down(.)` represents a downsampling function; it sums all pixels in different nxn blocks of the input image, thus reducing the output image by a factor of n in both dimensions (here, each element of the input image data is assigned a 3x3x3 block, and the sum of all elements within it is used as the value of that element in the output image, thereby reducing the output image by a factor of 3 in all dimensions); each output map corresponds to its own weight parameter β (multiplicative bias) and an additive bias b; the parameters β and b are updated using gradient descent.

[0080] ;

[0081] ;

[0082] ;

[0083] Wherein, conv2 is a two-dimensional convolution operator; rot180 is a rotation of 180 degrees; This refers to performing a full convolution; the aforementioned l Indicates the number of layers; the i express l The first layer i The number of neuron nodes; j express l The first layer j The number of neuron nodes; b Indicates bias; the δ The bias indicates the sensitivity of the output neuron. b The rate of change; the u , v This indicates the output maps ( u , v Position; E is the error function, i.e., E = The C represents the dimension of the label. If it is a binary classification problem, the label can be denoted as y. h ∈{0,1}, where C = 1, which can also be denoted as y h ∈{(0,1),(1,0)}, where C=2; the The h-th dimension represents the label corresponding to the nth sample; This represents the h-th output of the network corresponding to the n-th sample; β is a weight parameter (generally taking values ​​in [0,1]); down(.) represents a downsampling function; It is the first l +1 layer of convolutional kernels; the yes l -1 layer output of the j-th neuron node; the s l =W 1 x l-1 +b l Where W is the weight parameter and b is the bias. It is s 1 The j-th component.

[0084] Step C: The convolutional neural network automatically learns the combination of feature maps, then the j-th feature map combination is:

[0085] ;

[0086] st =1, and 0≤ ≤1;

[0087] Wherein, the symbol * represents the convolution operator; the l Indicates the number of layers; the i express l The first layer i The number of neuron nodes; j express l The first layer j The number of neuron nodes; f It's the activation function; here we'll choose the sigmoid function. f(x) = As an activation function e This represents the Euler number, 2.718281828. e x It is an exponential function; the aforementioned It is the first l -1 layer output of the first i Each component; the aforementioned This represents the number of input maps; the... It is a convolution kernel; the stated It is a bias; the stated express l The output map of layer -1 is used as l When the layer is input, l -1 layer obtains the first j The first of the output maps i The weights or contributions of each input map.

[0088] (2) Using the features extracted in (1) and combined with softmax, the target region of the image is automatically identified, and the automatic segmentation model is determined. Specifically, the softmax recognition process is to output a probability value for a given sample. This probability value represents the probability that the sample belongs to a certain category. The loss function is:

[0089] ;

[0090] Among them, the m Indicates shared ownership m One sample; the c This means that these samples can be divided into a total of c Class; the It is a matrix, where each row represents the parameters corresponding to a category, namely the weight and bias; 1{·} is an indicator function, meaning that the function returns 1 when the value within the curly braces is true, and 0 otherwise; λ is the parameter balancing the fidelity term (first term) and the regularization term (second term), where λ is a positive number (adjusted according to experimental results); J(θ) refers to the system's loss function; e represents the Euler number 2.718281828, e x It refers to the exponential function; T represents the transpose operator in matrix calculations; lg represents the natural logarithm, i.e., the logarithm with Euler's number as the base; n represents the dimension of the weights and bias parameters; x (i) It is the i-th dimension of the input vector; y (i) It is the i-th dimension of each sample label; then the gradient is used to solve:

[0091] ;

[0092] in, The m represents a total of m samples; It is a matrix, where each row represents the parameters corresponding to a category, namely the weight and bias; 1{·} is an indicator function, which returns 1 when the value in the curly braces is true, and 0 otherwise; λ is the parameter balancing the fidelity term (first term) and the regularization term (second term), where λ is a positive number (adjusted according to experimental results); J(θ) refers to the system's loss function; It is the derivative of J(θ); the e represents the Euler number 2.718281828, e x It refers to the exponential function; T represents the transpose operator in matrix calculation; In represents the natural logarithm, i.e., the logarithm with Euler's number as the base; x(i) is the i-th dimension of the input vector; y(i) is the i-th dimension of each sample label.

[0093] This uses a novel Softmax classifier, specifically a binary Softmax classifier. For the fluorescence detection image data of a target DNA sample, the probability given by the softmax classifier can be used to obtain a probability map that distinguishes the target region from the non-target region in the liquid-laden image data or the solution-dropping image data of the syringe tip. Based on this map, the output result of the liquid-laden image data or the solution-dropping image data of the syringe tip can be obtained.

[0094] (3) Use convolutional neural networks to automatically segment the liquid-carrying image data of the injection nozzle or the image data of the solution dripping from the injection nozzle, and refine the segmented three-dimensional structure shape, that is, fill the holes and remove non-target areas by using erosion and dilation morphological operators.

[0095] Step 3: Using the model obtained in Step 2, automatically segment all the liquid-carrying image data of the syringe tip or the image data of the syringe tip adding solution (i.e., 5000 image data) to obtain the target area data of the liquid-carrying image data of the syringe tip or the image data of the syringe tip adding solution.

[0096] S4. Select p-1 sets of data from step S3 as the training set, and use the remaining set for testing. Train a model using a convolutional neural network for testing. The specific operations of step S4 include: dividing the target region data automatically segmented in process three into p groups, normalizing the data (i.e., after automatically segmenting the target region, extracting the features of the target region, and performing a linear transformation on these features to map the result values ​​to [0,1]); when performing step S4, all images in the training set (i.e., p-1 sets of data) need to be read in first to train an intelligent system based on a deep convolutional neural network for automatic recognition, and then the remaining set of data needs to be read in to test the system.

[0097] When using this system to automatically identify liquid-laden or solution-dispensing images of syringe tips, simply read the liquid-laden or solution-dispensing images of the syringe tips to be detected.

[0098] S5. Repeat step S4, perform p cross-validations, and obtain the optimal parameters of the model trained by the convolutional neural network. Finally, determine the intelligent recognition system based on the deep convolutional neural network for automatically recognizing liquid-filled images of syringe tips or images of syringe tips dripping solution. The specific operations of step S5 include: training the recognition model using the convolutional neural network, extracting features from all target region data (the specific process is the same as the feature extraction process in process S3 automatic segmentation, except that the object here is only liquid-filled images of syringe tips or images of syringe tips dripping solution. The network structure has three fewer convolutional layers and three more fully connected layers than in automatic segmentation, with 64, 64, and 1 neurons respectively; the convolutional kernel sizes are: 14x14 for the first layer, 5x5 for the second and third layers, and 3x3 for the remaining layers; the stride is 2 for the first three convolutional layers and 1 for the rest; the downsampling layer size is 3x3 and the stride is 2; while the automatic segmentation part extracts features from both non-target and target regions of the liquid-filled images of syringe tips or images of syringe tips dripping solution).

[0099] This embodiment employs a novel Softmax classifier, specifically a binary Softmax classifier, to find the optimal value of a loss function, i.e., optimize J(θ). The number of classes p in the Softmax classifier is equal to 2 (i.e., the target area and non-target area of ​​the image). The gradient descent method is used to obtain the probability of the accuracy of the label data information belonging to the liquid-filled image data of the syringe tip or the image data of the syringe tip adding solution (the difference between the liquid adding intelligent recognition model and the syringe tip availability model lies in the difference of the label data information of the technical features of the target area). The specific process is the same as the automatic segmentation process in process three (except that here a classification label is predicted based on these probabilities, which is to identify a liquid-filled image data of the syringe tip or the image data of the syringe tip adding solution).

[0100] The new process involves repeating the experiment in step S5. For each set of p data, p-1 sets are selected for training, and the remaining sets are used for testing. This process yields the optimal parameters for the recognition model, resulting in a system based on a deep convolutional neural network for automatically recognizing liquid-laden images of syringe tips or images of syringe tips dispensing solutions. By inputting the liquid-laden images of syringe tips or images of syringe tips dispensing solutions into this intelligent recognition system, the system can output the corresponding results.

[0101] Reference Figure 2 and Figure 3The main body 1 of the detection experiment also includes a digital PCR instrument 10 and a sealing plate shaking module 9 for sealing experimental samples and shaking the samples after sealing. The sealing plate shaking module 9 includes a shaking and mixing machine 7 installed in the main body 1 and an aluminum foil sealing machine 8 fixed above the shaking and mixing machine 7. The digital PCR instrument 10 is connected to a main data image analyzer 26 via wired or wireless means. The main data image analyzer 26 receives experimental images from the digital PCR instrument 10 and analyzes the experimental results. A transfer module 11 is set between the digital PCR instrument 10, the sealing plate shaking module 9 and the transport module 2 in the main body 1. The transfer module 11 is a multi-axis robot. The transfer module 11 first transfers the experimental samples on the transport module 2 to the sealing plate shaking module 9. After the experimental samples are sealed and shaken, the experimental samples are then sent into the digital PCR instrument 10, where the digital PCR instrument 10 completes the amplification and detection.

[0102] Table 1. PCR reaction solution

[0103] Components Final concentration Added amount PCRMix / 10μL Upstream primer (10 μM) 0.4μM 0.8μL Downstream primer (10 μM) 0.4μM 0.8μL Mutant probe (10 μM) 0.2μM 0.4μL Wild-type probe (10μM) 0.2μM 0.4μL Template DNA 1ng / μL 2μL <![CDATA[ddH2O]]> / 5.6μL

[0104] The sample preparation module includes PCR reaction solution. The PCR reaction solution was prepared according to Table 1 above. PCRMix was purchased from NEB. PCRMix does not contain cytosine triphosphate deoxynucleotide (dCTP) and contains 0.1% Triton-X-100, 1U of thermostable pyrophosphatase, and 5μg / μL BSA. Following the order of ddH2O, PCRmix, probe, primer, and template DNA, the above sample was added to a 0.2mL PCR tube at 20μL additions as per the reaction system in Table 1. The mixture was gently vortexed for 20s, and the solution was collected at the bottom of the tube by short-term centrifugation. The designed primers were then modified by adding the nucleotide sequence “GTACCATCTGTAGACTCACTATAGGAAGAGATGTCAACTCGTGCACGAGTTGACATCTCTTCTCTCCGAGCCGGTCGAAATATTGGAGGAAGCTCGAGCTGGAGGAAAAGTGAGTCTACAGATGGTAC” to one end, thus obtaining the primer sequence for amplification. The primer sequences for amplification were sent to Sangon Biotech (Shanghai) Co., Ltd. for synthesis. The synthesized nucleotides needed to be pretreated. The pretreatment method was as follows: the synthesized nucleotide sequence was added to buffer B and heated to 68°C for 7 minutes, then cooled to 30°C and incubated for 30 minutes to obtain the full-length nucleotide sequence for detection. The components of buffer B were: 300 mM NaCl, 5 mM MgCl2, and 20 mM Tris (pH 7.6).

[0105] After digital PCR amplification, a clinical sample detection result image (i.e., cluster analysis of gene loci) is obtained, with the vertical axis representing the FAM fluorescence channel and the horizontal axis representing the HEX fluorescence channel. This system supports efficient resolution of six fluorescence signals and can combine two fluorescence channels to form three two-dimensional planar data images. The effective fluorescence points in the three two-dimensional planar data images are identified using a data image analyzer, and the results are analyzed. A convolutional neural network model can detect target signals in different reaction systems, corresponding to the detection results of multiple gene loci. The method for constructing an intelligent recognition model for fluorescence detection image data in the clinical sample detection result image includes the following steps:

[0106] S1. Read the fluorescence detection image data of the target DNA sample and perform preprocessing;

[0107] S2. Select an image and use a convolutional neural network to automatically learn and segment the region of interest (AOI) image, and refine the shape of the AOI image;

[0108] S3. Divide the AOI image shapes extracted in step S2 into p groups on average, extract the data features of these AOI image shapes using a convolutional neural network, and normalize them.

[0109] S4. Select p-1 sets of data from step S3 as the training set, and use the remaining set as the test set. Train the model using a convolutional neural network and then test it.

[0110] S5. Repeat step S4 and perform p cross-validations to obtain the optimal parameters of the recognition model, and finally determine the intelligent recognition system based on the fluorescence detection image data of the target DNA sample that automatically identifies the target DNA sample.

[0111] The process S1 specifically involves: reading at least 5000 fluorescence detection image data of target DNA samples after digital PCR amplification; during step S4, all images in the training set (i.e., p-1 sets of data) need to be read first to train an intelligent system for automatic recognition based on a deep convolutional neural network, and then the remaining 1 set of data is read to test the system. When using this system to automatically identify the fluorescence detection image data of target DNA samples, only the fluorescence detection image data of the target DNA sample after digital PCR amplification to be detected needs to be read.

[0112] Specifically, process S2 involves performing Gaussian filtering to denoise the fluorescence detection image data of the target DNA sample read in process one, thereby obtaining preprocessed enhanced image data.

[0113] The process S3 specifically comprises: Step 1: Selecting 5000 image data samples preprocessed in Step 2; Step 2: Experts extract the AOI (Area of ​​Interest) and non-AOI (Non-AOI) portions, and then train an automatic segmentation model using a convolutional neural network; Here, the convolutional neural network for the fluorescence detection image data of the target DNA sample consists of a network structure with 13 convolutional layers and 2 downsampling layers. The kernel sizes are: 13x13 for the first layer, 5x5 for the second and third layers, and 3x3 for the remaining layers. The stride is 2 for the first two convolutional layers and 1 for the rest. The downsampling layers are all 3x3 with a stride of 2.

[0114] The specific method for training an automatic segmentation model using a convolutional neural network is as follows:

[0115] (1) Features are automatically learned and extracted through the convolutional layers and downsampling layers of the convolutional neural network. The specific steps are as follows:

[0116] Step A: In a convolutional layer, the feature maps from the previous layer are convolved by a learnable convolutional kernel, and then passed through an activation function to obtain the output feature map; each output is the value of a convolutional kernel convolving one input or combining the values ​​of multiple convolutional inputs (here we choose to combine the values ​​of multiple input and output maps): Where * represents the convolution operator; l Indicates the number of floors; i express l -1st floor i One neuron node; j express l The first layer j One neuron node; M j This represents the set of selected input maps; It is the output; It means l The output of layer -1, as l Input at level 1; f It's the activation function; here we take... sigmoid function f(x) = As an activation function; e represents the Euler number 2.718281828, e x It's an exponential function; k is the convolution operator; b is the bias; each output map is given an additional bias b, but for a specific output map, the convolution kernel for each input map is different. This step also requires gradient calculation to update the sensitivity, which represents how much the error changes with how much b changes. ;in, l Indicates the number of floors; j expressl The first layer j 1 neuron node; * indicates multiplication of each element; δ This represents the sensitivity of the output neuron, i.e., the rate of change of the bias b. s l = W l x l ; W As weight; b For bias; f It's the activation function; here we take... sigmoid function f(x) = As an activation function; e represents the Euler number 2.718281828, e x It is an exponential function, f'(x) is the derivative of f(x). If f takes the sigmoid function, then f'(x) = (1-f(x))*f(x); This represents the weights shared across all layers; `up(.)` represents an upsampling operation. If the downsampling factor is n, the upsampling operation copies each pixel n times horizontally and vertically, thus restoring its original size; then... l The gradient of bias b is quickly calculated by summing the values ​​of all nodes in the sensitivity map of the layer. ;in, l Indicates the number of floors; j express l The first layer j One neuron node; b Indicates bias; δ The bias indicates the sensitivity of the output neuron. b The rate of change; u , v This indicates the output maps ( u , v Position; E is the error function, here E= C represents the dimension of the label. If it's a binary classification problem, the label can be denoted as y. h ∈{0,1}, where C=1, which can also be denoted as y h ∈{(0, 1), (1, 0)}, where C = 2; the The h-th dimension represents the label corresponding to the nth sample; This represents the h-th output of the network corresponding to the n-th sample.

[0117] Finally, the backpropagation (BP) algorithm is used to calculate the weights of the convolution kernel: Where W is the weighting parameter; E is the error function, and C represents the dimension of the label. If it is a binary classification problem, the label can be denoted as y. h ∈{0,1}, where C = 1, which can also be denoted as y h ∈{(0,1), (1,0)}, where C = 2; the The h-th dimension represents the label corresponding to the nth sample; This represents the h-th output of the network corresponding to the n-th sample; η is the learning rate, i.e., the step size; since the weights of many connections are shared, for a given weight, it is necessary to calculate the gradient of all connections related to that weight at that point, and then sum these gradients: ;in, l Indicates the number of floors; i express l The first layer i One neuron node; j express l The first layer j One neuron node; b Let represent the bias, δ represent the sensitivity of the output neuron, i.e., the rate of change of the bias b; u, v represent the (u, v) positions of the output maps; E is the error function, where E = C represents the dimension of the label. If it is a binary classification problem, the label can be denoted as y. h ∈{0,1}, where C = 1, which can also be denoted as y h ∈{(0,1),(1,0)}, then C=2; This represents the h-th dimension of the label corresponding to the nth sample; This represents the h-th output of the network corresponding to the n-th sample; It is a convolution kernel; yes The elements in the convolution are compared with... The element-wise multiplication of the patch means that all regions in the image with the same size as the convolution kernel are multiplied. The value at position (u,v) of the output convolution map is the result of the patch at position (u,v) in the previous layer multiplied by the convolution kernel. The result of element-wise multiplication.

[0118] Step B: The downsampling layer has N input maps, so it has N output maps, but each output map is smaller. Therefore: ;in, f It's the activation function; here we take... sigmoid function f(x) = As the activation function, e represents the Euler number 2.718281828. x It is an exponential function; This represents the weights shared across layers; `down(.)` represents a downsampling function; it sums all pixels in different nxn blocks of the input image, thus reducing the output image by a factor of n in both dimensions (here, each element of the input image data is assigned a 3x3x3 block, and the sum of all elements within it is used as the value of that element in the output image, thereby reducing the output image by a factor of 3 in all dimensions); each output map corresponds to its own weight parameter β (multiplicative bias) and an additive bias b; the parameters β and b are updated using gradient descent.

[0119] ;

[0120] ;

[0121] ;

[0122] Wherein, conv2 is a two-dimensional convolution operator; rot180 is a rotation of 180 degrees; This refers to performing a full convolution; the aforementioned l Indicates the number of layers; the i express l The first layer i The number of neuron nodes; j express l The first layer j The number of neuron nodes; b Indicates bias; the δ The bias indicates the sensitivity of the output neuron. b The rate of change; the u , v This indicates the output maps ( u , v Position; E is the error function, i.e., E = The C represents the dimension of the label. If it is a binary classification problem, the label can be denoted as y. h ∈{0,1}, where C = 1, which can also be denoted as y h ∈{(0,1),(1,0)}, where C=2; the The h-th dimension represents the label corresponding to the nth sample; This represents the h-th output of the network corresponding to the n-th sample; β is a weight parameter (generally taking values ​​in [0,1]); down(.) represents a downsampling function; It is the first l +1 layer of convolutional kernels; the yes l -1 layer output of the j-th neuron node; the s l =W1 x l-1 +b l Where W is the weight parameter and b is the bias. It is s 1 The j-th component.

[0123] Step C: The convolutional neural network automatically learns the combination of feature maps, then the j-th feature map combination is:

[0124] ;

[0125] st =1, and 0≤ ≤1;

[0126] Wherein, the symbol * represents the convolution operator; the l Indicates the number of layers; the i express l The first layer i The number of neuron nodes; j express l The first layer j The number of neuron nodes; f It's the activation function; here we'll choose the sigmoid function. f(x) = As an activation function e This represents the Euler number, 2.718281828. e x It is an exponential function; the aforementioned It is the first l -1 layer output of the first i Each component; the aforementioned This represents the number of input maps; the... It is a convolution kernel; the stated It is a bias; the stated express l The output map of layer -1 is used as l When the layer is input, l -1 layer obtains the first j The first of the output maps i The weights or contributions of each input map.

[0127] (2) Using the features extracted in (1) and combined with softmax, the target region of the image is automatically identified, and the automatic segmentation model is determined. Specifically, the softmax recognition process is to output a probability value for a given sample. This probability value represents the probability that the sample belongs to a certain category. The loss function is:

[0128] ;

[0129] Among them, the m Indicates shared ownership m One sample; the c This means that these samples can be divided into a total of c Class; the It is a matrix, where each row represents the parameters corresponding to a category, namely the weight and bias; 1{·} is an indicator function, meaning that the function returns 1 when the value within the curly braces is true, and 0 otherwise; λ is the parameter balancing the fidelity term (first term) and the regularization term (second term), where λ is a positive number (adjusted according to experimental results); J(θ) refers to the system's loss function; e represents the Euler number 2.718281828, e x It refers to the exponential function; T represents the transpose operator in matrix calculations; lg represents the natural logarithm, i.e., the logarithm with Euler's number as the base; n represents the dimension of the weights and bias parameters; x (i) It is the i-th dimension of the input vector; y (i) It is the i-th dimension of each sample label; then the gradient is used to solve:

[0130] ;

[0131] in, The m represents a total of m samples; It is a matrix, where each row represents the parameters corresponding to a category, namely the weight and bias; 1{·} is an indicator function, which returns 1 when the value in the curly braces is true, and 0 otherwise; λ is the parameter balancing the fidelity term (first term) and the regularization term (second term), where λ is a positive number (adjusted according to experimental results); J(θ) refers to the system's loss function; It is the derivative of J(θ); the e represents the Euler number 2.718281828, e x It refers to the exponential function; T represents the transpose operator in matrix calculation; In represents the natural logarithm, i.e., the logarithm with Euler's number as the base; x(i) is the i-th dimension of the input vector; y(i) is the i-th dimension of each sample label.

[0132] This uses a novel Softmax classifier, specifically a binary Softmax classifier. For a fluorescence detection image of a target DNA sample, the probability given by the softmax classifier can be used to obtain a probability map that distinguishes the target region from the non-target region in the fluorescence detection image of the target DNA sample. Based on this map, the output result of the fluorescence detection image of the target DNA sample can be obtained.

[0133] (3) The fluorescence detection image data of the target DNA sample is automatically segmented using a convolutional neural network, and the three-dimensional structure shape of the segmented sample is refined by filling holes and removing non-target areas through erosion and dilation morphological operators.

[0134] Step 3: Using the model obtained in Step 2, automatically segment the fluorescence detection image data of all target DNA samples (i.e., 5000 image data) to obtain AOI, which is the target region data of the fluorescence detection image data of the target DNA sample.

[0135] The process S4 specifically involves: dividing the AOI automatically segmented in process three into p groups on average, normalizing the data, that is, after automatically segmenting the target region, extracting the features of the target region, performing a linear transformation on these features, and mapping the result value to [0,1].

[0136] The process S5 specifically involves: training a recognition model using a convolutional neural network to extract features from all AOIs (the specific process is the same as the feature extraction process in process three, which is automatic segmentation; however, the object here is only the fluorescence detection image data of the target DNA sample. The network structure has three fewer convolutional layers and three more fully connected layers than in automatic segmentation, with 64, 64, and 1 neurons respectively; the convolutional kernel sizes are: 14x14 for the first layer, 5x5 for the second and third layers, and 3x3 for the remaining layers; the stride is 2 for the first three convolutional layers and 1 for the rest; the downsampling layer size is 3x3 and the stride is 2; while the automatic segmentation part extracts features from both the non-target and target regions of the fluorescence detection image data of the target DNA sample simultaneously).

[0137] This embodiment employs a novel Softmax classifier, specifically a binary Softmax classifier, to find the optimal value of a loss function, i.e., optimize J(θ). The number of classes p in the Softmax classifier is equal to 2 (i.e., the target region and non-target region of the image). The probability of the accuracy of the digital PCR detection of genetic information of the target DNA sample can be obtained through the gradient descent method. The specific process is the same as the automatic segmentation process in process three (except that here a classification label is predicted based on these probabilities, which is to identify the image data of the fluorescence detection image data of a target DNA sample).

[0138] The new process involves repeating the experiment in step S5. For each set of p data, p-1 sets are selected for training, and the remaining sets are used for testing. This process yields the optimal parameters for the recognition model, resulting in a fluorescence detection image data recognition system based on a deep convolutional neural network for automatically identifying target DNA samples. By inputting the fluorescence detection image data of the target DNA sample to be identified into this intelligent recognition system, the system can output the fluorescence detection image data of the target DNA sample.

[0139] Reference Figure 2 and Figure 3 The experimental unit is also equipped with a disposal compartment 12 for storing waste culture plates. The disposal compartment 12 is designed as a drawer 31 structure to facilitate the cleaning of waste culture plates. A transfer robot 13 is installed between the digital PCR instrument 10 and the disposal compartment 12 to grab the waste culture plates and drop them into the disposal compartment 12. The equipment structure of the transfer robot 13 is the same as that of the transfer module 11.

[0140] A microbial cleaning module A42 for inactivating microorganisms in the activity space of the injection nozzle 16 is installed on the multi-axis robotic arm 14 next to the injection nozzle 16. In this embodiment, the microbial cleaning module A42 is an ultraviolet disinfection lamp. Multiple microbial cleaning modules B30 for inactivating and cleaning microorganisms in the detection experimental body 1 are installed in the detection experimental body 1. Among them, the microbial cleaning module B30 is installed above the conveying module 2 to ensure the cleanliness of the conveying module 2. The disposal chamber 12 is also equipped with a microbial cleaning module B30 for inactivating microorganisms in the disposal chamber 12. The microbial cleaning module B30 is an ultraviolet disinfection lamp or an ozone / hydrogen peroxide fumigator. After completing a round of detection experiments, a round of treatment helps to ensure the accuracy of the next experiment.

[0141] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A super multiplex PCR detection system for cell-free DNA in blood, characterized in that, It includes the main body of the detection experiment (1), which is used to provide a clean detection experimental environment; The delivery module (2) is located inside the detection experimental body (1) and is used to deliver the experimental sample to the detection experimental body (1); The sample preparation module (3) is located inside the detection experimental body (1) and is used to store the PCR reaction solution; The reaction solution addition module (4) is used to add PCR reaction solution to the experimental sample on the delivery module (2); The mixing module (5) is used to mix the PCR reaction solution and the experimental sample. Oil phase liquid addition module (6) is used to add oil phase liquid to the culture plate; The drop-making module is used to inject the mixture of PCR reaction solution and experimental sample into the oil phase and prepare a drop-shaped mixture. The dropwise sample preparation module is used to place dropwise mixtures into the culture plate one drop at a time; The sealing and shaking module (9) is used to seal the culture plate and shake the droplet PCR pre-reaction solution after sealing. A digital PCR instrument (10) is used for amplification and fluorescence detection of the target DNA in the droplet PCR pre-reaction solution; The transfer module (11) is set inside the detection experimental body (1) and is used to transfer the experimental sample on the delivery module (2) to the sealing and shaking module (9). After sealing and shaking the droplet PCR pre-reaction solution, the droplet PCR pre-reaction solution is sent into the digital PCR instrument (10). The disposal compartment (12) is located inside the main body of the detection experiment (1) and is used to store discarded culture plates; A transfer robot (13) is installed inside the main body of the detection experiment (1) to grab the waste culture plate and throw it into the disposal chamber (12); The drop-making module includes a multi-axis robotic arm (14) installed in the main body of the detection experiment (1). The free end of the multi-axis robotic arm (14) is provided with a syringe (15) and an injection nozzle (16) installed on the syringe (15). The opening edge of the injection nozzle (16) forms a shearing blade (17). An air tube (18) is connected to the syringe (15). A pressure sensor (19) and a pressure regulating valve (20) are installed on the air tube (18). The pressure sensor (19) is connected to a first controller (21) via wired or wireless means. The first controller (21) controls the connection of the pressure sensor (19). Among them, the first controller (21) controls the pressure output by the pressure regulating valve (20) through a pre-trained air pressure regulation model, and the pre-trained air pressure regulation model is F = a×(b - πw 2 χ) + c; where F is the pressure output by the pressure regulating valve (20), the value range of a is between 1.05 and 1.5, x is the measurement scale value of the liquid level height of the solution formed in the syringe tip (16) when sucking the solution, w is 1 / 2 of the diameter of the circular liquid surface of the solution formed in the syringe tip (16) when sucking the solution, the value of b is not less than the total volume value of the solution that can be accommodated when the syringe tip (16) is used, c is the offset, 0 < c < y, and the value of y is not greater than the calculated value of (b - πw 2 χ).

2. The ultra-multiplex PCR detection system for cell-free DNA in blood according to claim 1, characterized in that, The free end of the multi-axis robotic arm (14) is provided with multiple image acquisition devices (22) around the injection nozzle (16). The multiple image acquisition devices (22) are connected to a secondary data image analyzer (23) via wired or wireless means. The secondary data image analyzer (23) is used to determine whether the PCR reaction solution flows out of the injection nozzle (16) in a droplet manner. The secondary data image analyzer (23) is electrically connected to a first alarm (24). The secondary data image analyzer (23) stores a liquid drop intelligent recognition model for determining whether the injection nozzle (16) is adding solution in a droplet manner. The liquid drop intelligent recognition model includes a convolutional neural network model.

3. The ultra-multiplex PCR detection system for cell-free DNA in blood according to claim 2, characterized in that, The secondary data image analyzer (23) is also used to analyze whether the injection nozzle (16) is contaminated. The secondary data image analyzer (23) is also connected to a second alarm (25). The secondary data image analyzer (23) also stores an injection nozzle (16) availability model for judging whether the injection nozzle (16) is contaminated. The injection nozzle (16) availability model includes a convolutional neural network model.

4. The ultra-multiplex PCR detection system for cell-free DNA in blood according to claim 1, characterized in that, Also includes: The main data image analyzer (26) is connected to the digital PCR instrument (10) via wired or wireless means to receive experimental images from the digital PCR instrument (10) and analyze the experimental results; wherein, the main data image analyzer (26) stores a convolutional neural network model.

5. A super-multiplex PCR detection system for cell-free DNA in blood according to claim 4, characterized in that, The convolutional neural network model stored in the master data image analyzer (26) analyzes the fluorescence detection image data in the clinical sample detection result image.

Images