Quality control automation in polymerase chain reaction analysis
By using machine learning models and image processing techniques to automatically identify and reject faulty separators in dPCR, the problem of inaccurate quantification caused by systematic faults and artifacts in separators is solved, thus improving the accuracy of dPCR sample results.
Patent Information
- Application Number
- CN202480039436.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-04-20
- Filing Date
- 2024-04-17
- Publication Date
- 2026-01-13
AI Technical Summary
In digital polymerase chain reaction (dPCR), systematic faults in the separator region and false external artifacts can lead to inaccurate target concentration quantification, affecting data quality.
The system automatically identifies and rejects regions with systematic faults or false external artifacts using machine learning models. It employs computer-implemented systems and methods to classify regions and combines image processing techniques to improve data accuracy.
It improves the accuracy of dPCR sample results, reduces errors introduced by systematic failures and artifacts, and ensures the accuracy of target concentration quantification.
Smart Images

Figure CN121336247A_ABST
Abstract
Description
[0001] Cross-references to related applications
[0002] This application claims the benefit of U.S. Provisional Application No. 63 / 460,879, filed April 20, 2023. The subject matter of this application is also related to the following jointly assigned provisional application numbers: US63 / 460,882; US63 / 460,884; and US63 / 460,885, all of which were filed on [date missing]. Submitted on April 20, 2023. The contents of these applications are incorporated herein by reference in their entirety. Technical Field
[0003] This disclosure generally relates to the monitoring, measurement, and / or analysis of biological and biochemical reactions, and more specifically to the method of the present invention for automating data quality control during polymerase chain reaction analysis for quantifying target concentrations in biological samples. Background Technology
[0004] Polymerase chain reaction (PCR) is a technique used to quantify biomolecules in a sample. Typically, PCR amplifies nucleic acids using DNA polymerases, which are responsible for forming new copies of DNA. Based on the theory that this amplification grows exponentially, PCR can be used to amplify specific DNA fragments (e.g., nucleic acid molecules or nucleotide sequences) millions or billions of times, producing enough copies to allow analysis based on optical signals associated with one or more different color channels, each corresponding to one or more fluorescent labels associated with a “dye,” whose detectable signal increases as the target in the sample is amplified.
[0005] “Simulated” quantification (e.g., in simulated PCR) relies on extrapolation measurements based on a measurement pattern. For example, a target analyte can be quantified by comparing the number and amount of amplification cycles of the final PCR product with those of a reference sample. However, this type of quantification can be complicated by uncertainty and inaccuracy. The detection efficiency in the test sample may differ from that in the reference sample. For instance, in PCR, the initial amplification cycles may not be exponential, and PCR amplification may plateau after an indeterminate number of cycles. In particular, when low initial concentrations of the target analyte are not amplified to detectable levels, they may be completely missed.
[0006] Digital quantification methods (e.g., digital PCR or dPCR) are biotechnological improvements over analog methods, offering more robust absolute quantification of analytes with greater accuracy and precision compared to analog methods. Digital quantification excels at detecting and quantifying the concentrations of rare, hard-to-detect targets, providing more precise sample quantification or analysis (e.g., nucleotide sequences), and measuring low-fold changes in analyte concentration. Therefore, digital quantification has numerous applications in basic research, clinical diagnostics, and environmental testing. For example, digital PCR has been applied to pathogen detection and cancer surveillance, copy number variation analysis, single-gene expression analysis, rare sequence detection, gene expression profiling and single-cell analysis, detection of DNA contaminants in bioprocessing, validation of gene editing, and detection of specific methylation changes in DNA as a cancer biomarker.
[0007] Compared to analog measurements that rely on extrapolating certain measurements based on measurement patterns (e.g., exponential amplification cycles), digital quantification methods can quantitatively and discretely measure an analyte. Digital quantification can be performed on biological samples containing or suspected of containing a target analyte of interest, such as cells, tissues, or samples like hair, biological fluids like blood, urine, saliva, etc., cell clusters like microbial colonies, or organisms, cells, microorganisms, bacteria, viruses, proteins, antibodies, or nucleic acids like DNA or RNA molecules. Target analytes include “original” analytes initially present in the biological sample as well as any “synthetic” analytes that indicate the presence of the original analyte that may be added to or generated during detection, including PCR amplicon, antigen-antibody complexes, etc. Digital quantification (e.g., digital PCR) begins with a sample containing a relatively small amount of the target analyte, such as a polynucleotide or nucleotide sequence template DNA (or RNA). The sample is divided into a large number of smaller test samples, which will ideally contain one or no target analyte, allowing for individual detection reactions to be performed separately in each partition. A suitable separating region is an individual target that is sufficiently distanced from other targets to allow for individual detection or quantification. These targets may be fluidly isolated from each other or non-fluidically isolated. The separating region may or may not include separating barriers, such as walls or membranes, or liquid or semi-solid media immiscible with the sample. Exemplary separating regions include individually separated targets, such as those deposited on a substrate, such as glass slides, tubes, open or closed pores, droplets, vesicles, chambers, or beads, or any representation of a separate signal obtained from the target that is distinguishable against background or noise (e.g., a bright spot on a dark background in a digital or analog image). In digital PCR methods, when a sample is thermally cycled using a PCR device, samples containing target concentrations are amplified and produce a positive detection signal, while samples not containing target concentrations are not amplified and do not produce a detection signal. After multiple PCR amplification cycles, the sample is imaged and subjected to fluorescence analysis to quantify the target concentration in the sample.
[0008] In digital PCR, biological samples can be affected by systematic failures or spurious external artifacts present in the separators of the array that hold the samples. Such failures can be caused by physical or chemical processes occurring during the digital PCR process. Errors may be introduced, for example, due to spurious external artifacts such as debris or dust within or on the array separators, when measuring target concentrations. Air bubbles or gases may also be present in some separators of the array, which can introduce errors. Systematic failures throughout the array can also occur in the separators, such as delamination, leakage, or filling problems within the separators.
[0009] Furthermore, optical saturation errors can occur when the signal from the imaging channel exceeds the operating range of the detection sensor on the PCR device. Consequently, the digital image of the separator may appear overly bright or oversaturated. Observed bridging across the separator array can also lead to data quality errors for specific biological samples. Summary of the Invention
[0010] This article discloses specific methods for improving the accuracy of target concentration quantification in biological samples using digital quantification, including methods for automating quality control by determining which separators are accepted for analysis. Identifying separators with systematic faults or spurious external artifacts can present various technical challenges that can adversely affect the goal of obtaining useful test results. Automated quality control methods for labeling such separators can help determine the severity and type of faults and improve the accuracy of dPCR sample results.
[0011] This document describes various computer-implemented systems, methods, and articles of art for classifying separators that accept or reject one or more target concentrations in biological samples for analysis using analyte detection (e.g., dPCR) devices, and for training machine learning models for classifying separators when analyzing one or more biological samples using analyte detection (e.g., dPCR) devices.
[0012] In one embodiment, a method is provided to automate quality control when using an analyte detection (e.g., PCR) device configured to analyze an array of compartments (e.g., an array of about 1,000 to 5,000, 10,000 to 50,000, about 100,000, or even significantly more than 100,000 compartments) to quantify one or more target concentrations in a biological sample.
[0013] In one embodiment, a method and system are disclosed for computerized image processing of digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separators so that they are not used to calculate target concentration results. Image data representing multiple separators set in a container is obtained, including background image data captured prior to the amplification cycle of the dPCR assay and endpoint image data captured after amplification. Individual separator image data of the multiple separators are extracted from the image data. The separator background image data and separator endpoint image data are preprocessed to obtain processed image data, which is then input into one or more machine learning models to generate classification results. The one or more classification results are used to determine whether to reject image data corresponding to separators in the multiple separators so that they are not used to calculate target concentration results.
[0014] In one embodiment, a method is provided for training a machine learning model for analyzing one or more biological samples using an analyte detection (e.g., PCR) device. The method is performed by one or more computing devices and includes: obtaining a first plurality of images identified as rejection region images, and obtaining a second plurality of images identified as non-rejection region images. The method further includes: generating one or more datasets using the first plurality of images and the second plurality of images, and determining a set of parameters for the machine learning model by training the machine learning model using at least one of the one or more datasets. The trained machine learning model is configured based on the set of parameters to analyze one or more target concentrations in the one or more biological samples.
[0015] The various objects, features, aspects and advantages of the subject matter of the invention will become more apparent from the following description and the accompanying drawings, in which similar reference numerals denote similar parts. Attached Figure Description
[0016] This patent or application document contains at least one color drawing. A color copy of the published drawing of this patent or patent application document shall be provided after payment of fees and approval by this Office.
[0017] Figure 1 shows a block diagram of the PCR device and the array of separators arranged in a microfluidic array plate.
[0018] Figure 2 shows a flowchart of an exemplary workflow utilizing microfluidic array plate (MAP) technology.
[0019] Figure 3 shows a flowchart of an exemplary data source analysis pipeline.
[0020] Figure 4 shows several exemplary pre-PCR and post-PCR images representing one or more separator arrays arranged in a microfluidic array plate.
[0021] Figure 5A shows an image of the segmented array and a template image used to determine the angular position of the array.
[0022] Figure 5B shows an exemplary visual indication of the matching between the template image and the edge.
[0023] Figure 6A shows the array of segmented regions represented in the image.
[0024] Figure 6B shows a magnified image of the corner region of the array in Figure 6A.
[0025] Figure 7 shows four examples of the segmented area array, one representing a normal sample and three representing various quality control issues.
[0026] Figure 8 shows an example of a normal segmented region image and an example of a segmented region image showing data quality issues.
[0027] Figure 9 shows an example of a graph displaying dPCR fluorescence data from different channels, showing normal fluorescence data and data affected by quality control.
[0028] Figure 10A is a flowchart illustrating a method for quality control workflow of a partitioned array according to various implementation schemes.
[0029] Figure 10B is a flowchart illustrating a method for performing blank image error checking according to various embodiments.
[0030] Figure 11 is a flowchart illustrating a method for classifying images of individual microchamber sites in an array according to various embodiments.
[0031] Figure 12 is a flowchart illustrating a method for generating a processed image according to various implementation schemes.
[0032] Figure 13 is a flowchart illustrating a method for classifying segmented region images according to various implementation schemes.
[0033] Figure 14 is a flowchart illustrating methods for training machine learning models according to various implementation schemes.
[0034] Figure 15 shows a block diagram of a computer system that can be used to implement one or more aspects of various implementation schemes.
[0035] Figure 16A is a flowchart illustrating a method for generating a synthetic array image according to various embodiments.
[0036] Figure 16B is a flowchart illustrating a method for calculating the color of a microchamber or partition according to various embodiments.
[0037] Although the invention has been described with reference to the accompanying drawings, these drawings are intended to be illustrative, and other embodiments are consistent with the spirit of the invention and are within its scope. Detailed Implementation
[0038] To provide a more thorough understanding of the invention, the following description sets forth many specific details, such as particular configurations, parameters, examples, etc. However, it should be understood that such description is not intended to be a limitation on the scope of the invention, but rather to provide a better description of exemplary embodiments.
[0039] As used herein, the term "biological sample" means a sample or solution containing any type of biochemical substance or component and / or any target molecule of interest to a user, manufacturer, or distributor of any embodiment of the invention described or implied herein, as well as any sample or solution containing an associated chemical substance or compound for performing a bioassay, experiment, or test. These biochemical substances, components, or target molecules may include, but are not limited to, DNA sequences (including cell-free DNA), RNA sequences, genes, oligonucleotides, molecules, proteins, biomarkers, cells (e.g., circulating tumor cells), or any other suitable target biomolecule. A biological sample may contain one or more of the following: at least one target nucleic acid sequence, at least one primer, at least one buffer, at least one nucleotide, at least one enzyme, at least one detergent, at least one blocking agent, or at least one dye, label, and / or probe suitable for detecting a target or reference nucleic acid sequence. In various embodiments, such biological components may be used in conjunction with one or more PCR methods and systems in applications such as fetal diagnostics, multiplex dPCR, viral detection and quantitative standards, genotyping, sequencing assays, experiments or protocols, sequencing validation, mutation detection, detection of genetically modified organisms, detection of rare alleles, and / or copy number changes.
[0040] In various embodiments, the devices, instruments, systems, and methods described herein can be used to detect one or more types of biological components of interest. These biological components of interest can be any suitable biological target, including but not limited to DNA sequences (including cell-free DNA), RNA sequences, genes, oligonucleotides, molecules, proteins, biomarkers, cells (e.g., circulating tumor cells), or any other suitable target biomolecule.
[0041] In various implementation schemes, such biological components can be used in conjunction with a variety of digital PCR methods or systems, such as multiplex digital PCR, viral detection and quantification standards, genotyping, sequencing validation, mutation detection, detection of genetically modified organisms, fetal diagnosis, detection of rare alleles, and copy number variation.
[0042] Embodiments of this disclosure generally relate to apparatus, instruments, systems, and methods for automating quality control while measuring or quantifying biological reactions in large sample compartments. The compartments may comprise small sample volumes contained within droplets or microfluidic structures, including microchambers, micropores, through-holes, or any combination thereof.
[0043] The term “minimum-maximum normalization” used in the specification and / or claims means that the maximum value in the data is mapped to a fixed, selected maximum value, and the minimum value in the data is mapped to a fixed, selected minimum value.
[0044] The term “rescaling” or any variation thereof, as used in the specification and / or claims, refers to the process of modifying the size of a range of pixel values in an image.
[0045] The term “distortion” or any variation thereof, as used in the specification and / or claims, refers to the process of modifying an image by applying an affine transformation to match common points between a given image and a reference image.
[0046] While digital quantification, such as PCR, is generally applicable, it should be recognized that any other suitable quantification method may be used according to the various embodiments described herein. Suitable PCR methods include, but are not limited to, digital PCR, allele-specific PCR, asymmetric PCR, ligation-mediated PCR, multiplex PCR, nested PCR, qPCR, genome walking, and bridging PCR.
[0047] As used herein, thermal cycling may include the use of, for example, thermal cyclers, isothermal amplification, thermal rule, infrared-mediated thermal cycling, or hemolytic enzyme-dependent amplification.
[0048] Depending on the implementation scheme, the detection of the target may be, but is not limited to, for example, single or combined fluorescence detection, cation or anion detection, pH detection, voltage detection or current detection.
[0049] The various implementation schemes described herein are particularly well-suited for digital PCR (dPCR). In digital PCR, a solution containing a relatively small amount of the target analyte (e.g., a polynucleotide or nucleotide sequence) can be subdivided into a large number of small test samples, such that each sample typically contains one molecule of the target analyte (e.g., a nucleotide sequence) or none at all. When the samples are subsequently thermally cycled in a PCR protocol, procedure, or experiment, samples containing the target are amplified and produce a positive detection signal, while samples not containing the target are not amplified and do not produce a detection signal. Using Poisson statistics, the number of targets in the original solution can be correlated with the number of samples that produce a positive detection signal.
[0050] It should be understood that the disclosed techniques offer numerous advantageous technical effects, including an automated method for quality control during the quantification of one or more target concentrations in biological samples using analyte detection (e.g., PCR) equipment. The techniques described herein employ logic to automate various processes, including those currently performed manually. Furthermore, the disclosed techniques are designed to support data accuracy and allow for the processing of data algorithms and complex permutations at scales and speeds unattainable manually.
[0051] It should also be understood that the following instructions are not intended as a broad overview, and therefore, concepts have been simplified for clarity and brevity.
[0052] Figure 1 shows a block diagram of a PCR device 100 and a microfluidic array plate 110 having a separator array 120 disposed therein. In one embodiment, the PCR device 100 is a digital PCR. As described above, digital PCR (dPCR) uses a solution containing a relatively small amount of target analyte (e.g., polynucleotide or nucleotide sequence template DNA (or RNA), fluorescence quencher probes, primers, and a PCR master mix containing optimal concentrations of DNA polymerase and reaction buffer). This solution is divided into a large number of small test samples, for example, tens of thousands of microchambers disposed within the microfluidic array plate 110. As shown in Figure 1, the microfluidic array plate 110 comprises an array of microchambers formed by multiple rows and columns of microchambers (e.g., nanoliter-sized microchambers). The microchambers may have nucleic acid-binding surfaces for better binding to targets in the test samples. Small test samples are disposed in each microchamber of the array, thereby forming the separator array 120. Thus, separator 122 comprises microchambers and small test samples disposed therein. Typically, some separators of array 120 include test samples with one or more targets (e.g., target analytes, such as nucleotide sequences), while some separators do not. In Figure 1, for example, separator 122 includes a target, while separator 124 does not. The separator array 120 is then thermally cycled using PCR device 100.
[0053] Figure 2 illustrates a flowchart of an exemplary workflow 200 utilizing a microfluidic array plate (MAP) technology that can be used in combination with some embodiments of the present invention. In step 202, a sample is prepared, for example, using a bulk reaction mixture. In step 204, the sample is loaded (e.g., pipetted) onto a microfluidic array plate (MAP) that dispenses or digitizes the sample into thousands of small, independent reactions, each reaction in a microchamber, for example, such that each microchamber contains one or zero copies of the target. The original concentration is then calculated using statistical methods based on the number of positive and negative microchambers. In one embodiment of the invention, the MAP comprises the MAP16 digital PCR consumable employing microinjection molding plate technology, with 16 samples (16 arrays or chips) per plate, approximately 20,000 microchambers per reaction, and analysis performed on a loading volume greater than 95%. In some embodiments, the MAP comprises a sample plate comprising 32 arrays (or chips). For a sample plate with 32 digital PCR arrays (or chips), each sample plate may include approximately 10,240 microchambers. In step 206, the digital PCR thermal cycling process is run as described above. In one embodiment, up to four optical channels can be multiplexed to enable the measurement of multiple targets per sample, thereby saving time and reagents. In one embodiment of the invention, parameters such as plate layout, optical channels, and thermal conditions can be easily modified using control software to allow the user to quickly begin generating data. Specific protocols can also be saved and reloaded to streamline workflows. In step 208, the results are displayed to the user, for example, allowing the user to visualize the multiplexed data or calculate the concentration of replicates. In one embodiment, the results can be exported for downstream analysis or compiled into reports.
[0054] In some implementations, improving the accuracy of target concentration calculations includes more accurately identifying and rejecting defective microcavities in MAP technology array plates. Automated quality control can increase the throughput of dPCR systems, unaffected by dyes, channels, and assay runs, and can also help users identify and report problems during dPCR runs.
[0055] Figure 3 illustrates a flowchart of an exemplary data source analysis pipeline 300. The boxes shown in Figure 3 represent processes and / or methods that can be implemented and / or performed by one or more computing elements (e.g., one or more computing devices) described below. The computing elements can be standalone computing elements, networked computing elements, distributed computing elements, and / or embedded computing elements. For example, the computing elements can be integrated with PCR device 100, a field-standalone computing device, a cloud-based computing device, or a combination thereof. In one embodiment, images 302 of multiple separated regions in the array are captured after background cycling and before PCR. Image 302 shows uniform fluorescence, and no intensity changes caused by PCR are observed.
[0056] Image 304 is a post-PCR fluorescence image of multiple compartments in the array shown in image 302 after amplification of one or more targets in a biological sample via thermal cycling. Therefore, each bright spot shown in image 304 represents a compartment in the array shown in image 302 after amplification. It should be understood that the brightness of the dots in image 304 shown in Figure 3 is for illustrative purposes and can vary in any way. For example, some compartments in the array shown in image 304 may appear weak or dark (e.g., because the compartment does not contain a target). In some embodiments, image 304 is associated with a specific fluorescence channel of a PCR device with multiple capabilities.
[0057] Figure 4 illustrates multiple exemplary pre-PCR images 402 and post-PCR images 404 representing a separator array 120 disposed in a microfluidic array plate 110. In Figure 4, image 402 is a pre-PCR image of the separator array 120 before performing PCR amplification of one or more targets in a biological sample. Image 404 is a post-PCR image of the separator array 120 after performing amplification of one or more targets. In one embodiment, images 402 and 404 are fluorescence images provided by multiple fluorescence channels of the PCR device 100.
[0058] PCR device 100 may have multiple capabilities and is therefore capable of simultaneously quantifying multiple targets. Multiple capabilities are achieved by having multiple fluorescence channels in PCR device 100. In some embodiments, the multiple fluorescence channels use different types of dyes to generate signals with different spectral wavelengths. Different dyes can bind to different targets and produce signals with different fluorescence colors or spectra. In Figure 4, five exemplary types of dyes are used in the different fluorescence channels of the PCR device. The five different dyes may include a carboxyfluorescein (FAM)-based dye that produces a blue fluorescence signal, a hexachlorofluorescein (HEX)-based dye that produces a green fluorescence signal, a 6-carboxy-X-rhodamine (ROX)-based dye that produces a red fluorescence signal, a tetramethylrhodamine (TAMRA)-based dye that produces a red or yellow fluorescence signal, and a TYE™-based dye that produces a dark red fluorescence signal. In some embodiments, the ROX-based dye is used as a control reference dye.
[0059] In some implementations, the preprocessing step 306 in the analysis pipeline 300 may include a process for determining the angular positions of the separator array. Referring to Figure 5A, image 500 represents separator array 502. For illustrative purposes, image 500 shows only a portion of array 502. Although the following description uses a post-PCR image as an example, it should be understood that pre-PCR images can also be used to determine the angular positions of the separator array.
[0060] As part of preprocessing step 306, the computing device uses image 500 to determine the position associated with the corner of array 502. The corner position of the segmentation array is then used to determine the expected position of the segmentation in array 502. The observation position is then determined based on the expected position, and the observation position is used to obtain images of each segmentation. Next, the images of each segmentation are classified, and the classification results are used to perform the automated quality control classification analysis described herein on each segmentation and to calculate the target concentration. The segmentation position determination process and the classification process are described in more detail below.
[0061] Referring to Figure 5A, array 502 is a two-dimensional array with four corners. The corner position determination process can begin at any of the four corners. Figure 5A uses the upper left corner of array 502 as an illustration. To determine the position of the upper left corner of array 502, a corner region 504 is first selected to indicate that the upper left corner is being analyzed. The corner region 504 can be annotated using, for example, a square or rectangle of predetermined size.
[0062] In one implementation, the size of corner region 504 is predetermined such that it is neither too small nor too large. An excessively small corner region may not include enough separators to correlate with the template image to correctly locate edges. For example, as shown in Figure 5A, template image 506 represents seven separators arranged along the horizontal edge of the separator template array. Correspondingly, the size of corner region 504 can be predetermined such that 20-40 separators arranged along the horizontal edge are included in corner region 504 for analysis. Template image 506 is described in more detail below.
[0063] Excessively large corner regions can also affect the relevance of the template image due to distortions, such as those in the segmentation array image. Image distortion can lead to a deterioration in the correlation between the template image and the segmentation array image, making it difficult to find the edges of the segmentation array. Therefore, appropriately sized corner regions should include a sufficient number of segmentations based on the template image, but not so large as to cause correlation difficulties. It should be understood that the appropriate size of the corner regions can be predetermined by a computing device and / or by user input, based on the size of the segmentation array and the size of the template image.
[0064] In one embodiment shown in Figure 5A, corner region 504 is annotated and displayed. For example, the annotation could be a colored box overlaying the array of segmented regions to indicate the selected corner region. In Figure 5A, corner region 504 is annotated by a colored box (e.g., a green box) overlaying image 500 to indicate that the upper left corner is selected for analysis.
[0065] As described above, the template image is used to determine angular positions. The template image includes multiple portions that can be used to locate the edges of the array. Figure 5A shows a template image 506 associated with the upper edge 512 of the array 502 and a template image 508 associated with the left edge 514 of the array 502. Template image 506 includes two portions. The first portion of template image 506 includes multiple bright spots representing multiple dividing zones, which are predetermined to be set along the upper edge. These predetermined dividing zones of template image 506 form a line pattern, for example, as shown in Figure 5A. The second portion of template image 506 represents areas predetermined to have no dividing zones. In template image 506, this second portion is adjacent to the first portion and is dark. Therefore, template image 506 is an image representing a known pattern of dividing zones set along a known upper edge.
[0066] Figure 5A also shows another template image 508, which represents a known pattern of dividing zones set along a known left edge. In Figure 5A, template image 508 also comprises two parts. The first part of template image 508 includes multiple bright spots representing multiple dividing zones, which are predetermined to be set along the left edge. These predetermined dividing zones form an interlaced pattern, for example, as shown in Figure 5A. This interlaced pattern can also be viewed as including two interlaced lines. A portion of the dividing zones represented in template image 508 forms a first vertical line, and another portion of the dividing zones forms a second vertical line. The first and second vertical lines are offset from each other in the horizontal direction. For example, as shown in Figure 5A, the dividing zone forming the first line may be positioned further to the left than the dividing zone forming the second line. Moreover, the dividing zones of the first and second lines interlaced in alternating rows.
[0067] Like template image 506, template image 508 also has a second portion that represents a region predetermined to not have dividing areas. In Figure 5A, this portion is shown as a dark area immediately adjacent to the first portion, where the dividing areas form an interlaced line pattern. Unlike template image 506, where the second portion is oriented horizontally, the second portion of template image 508 is oriented vertically. The two portions of template image 508 thus represent a known pattern of dividing areas set along a known left edge. Although Figure 5A shows template images for the top and left edges, it should be understood that similar template images can be configured and obtained for the bottom and right edges.
[0068] The corner of a two-dimensional array is formed by two edges. For example, in Figure 5A, the upper left corner of array 502 is formed by the upper edge 512 and the left edge 514. Therefore, if the positions of the two edges are determined, the position of the corresponding corner can be determined. In one embodiment, two template images are used to determine the positions of the two edges, and then the position of the corner formed by these two edges. For example, in Figure 5A, template images 506 and 508 are obtained and used to determine the positions of the upper edge 512 and the left edge 514 forming the upper left corner, respectively. Using template image 506 as an example, to determine the position of the upper edge 512, template image 506 is moved and / or rotated to be correlated with the corner region 504, thereby finding the upper edge 512. If at least a portion of template image 506 matches the upper edge 512, the position of the upper edge 512 can be determined. For example, by moving and / or rotating the template image 506, the upper edge 512 is found if at least four or five of the total seven dividing regions shown in the template image 506 match at least four or five dividing regions shown in the corner region 504 of the image 500, and if the dark portion shown in the template image 506 also matches the dark portion above the upper edge 512 in the corner region 504. In some embodiments, if an initial match is found, the template image 506 is further moved horizontally to see if additional or continuing matches can be found. If so, the probability of finding the upper edge 512 is high. Once the upper edge 512 is found, its position can be easily determined, for example, by measuring the number of pixels from the upper edge 512 to the upper edge of the image 500.
[0069] Similarly, to determine the position of the left edge 514 of array 502, template image 508 is moved and / or rotated to correlate with the left edge 514. The position of the left edge 514 is determined if at least a portion of template image 508 matches the left edge 514. For example, by moving and / or rotating template image 508, the left edge 454 is found if at least four or five of the total seven partitions shown in template image 508 match four or five partitions in corner region 504, and if the dark portion shown in template image 508 also matches the dark portion to the left of the left edge 514 in corner region 504. In some embodiments, if an initial match is found, template image 508 is further moved vertically to see if additional or continuing matches can be found. If so, the probability of finding the left edge 514 is high. Once the left edge 514 is found, its location can be easily determined, for example, by measuring the number of pixels from the left edge 514 to the left edge of the image 500.
[0070] When a match is found between the template image and the edge, a visual indication can be displayed on image 500 and in other ways. Figure 5B illustrates such an exemplary visual indication using the upper edge as an example. In Figure 5B, the brightness of certain dividing areas set along the upper edge 512 is increased to indicate a possible match between the template image 506 and the upper edge 512.
[0071] Referring to Figure 5A, after determining the positions of the top edge 512 and the left edge 514, the position of the top-left corner of array 502 can be determined. In one embodiment, the position of the top-left corner is represented by horizontal and vertical coordinates in units of pixels from the left and top edges, respectively. Therefore, the position of the top-left corner of array 502 shown in Figure 5A can be represented by coordinate pairs such as (135, 133). Although Figure 5A uses the top-left corner as an example, it should be understood that the positions of other corners of array 502 can be determined in a similar manner. Furthermore, when additional images of the same array 502 or different arrays are obtained, for example, from different fluorescence channels, the corner positions of the same or different arrays can be determined similarly using the above-described technique.
[0072] Referring back to Figure 3, the angular positions determined in preprocessing step 306 are provided for extracting microchamber site images 308, where images of individual compartments in the compartment array are determined and extracted. As described above, the compartment array can be arranged in multiple rows and columns. Figure 6A shows such an array 602 as represented in image 600. Figure 6B shows a magnified image representing the upper left corner region of array 502. As shown in Figures 5A and 5B, array 502 includes multiple compartments (e.g., approximately 20,000 compartments) arranged in multiple rows and columns. These compartments are highlighted in Figures 5A and 5B. Ideally, these compartments are equally spaced from each other in the underlying microfluidic array plate, forming a regular repeating pattern. Furthermore, the images of the individual compartments will ideally show that the compartments are arranged in multiple rows and columns with equal spacing. If the compartments themselves and their images are ideal, the positions of the compartments can be easily determined from the angular positions of the compartment array through simple calculations.
[0073] Referring to Figures 6A and 6B, the partitions in array 602 are arranged in multiple rows and columns. Therefore, array 602 has four corners. The positions of the four corners of array 602 can be determined using the techniques described above. The position of a corner can be represented by the number of pixels in a pair of X and Y coordinates (i.e., the number of pixels measured in the horizontal and vertical directions from the relevant edge). For example, the position of the upper left corner of array 602 can be represented as a pair of coordinates measured in pixels, such as (135, 132), where "135" is the distance in pixels from the left edge of image 500, and "132" is the distance in pixels from the top edge of image 500. The positions of the upper right, lower left, and lower right corners of array 502 can be similarly represented as (2073, 127), (130, 2049), and (2085, 2044), respectively. In the example above, the position of the top left corner of image 500 is set as the base coordinate point (0, 0), and the position of the angle is measured relative to the base coordinate point. It should be understood that any other base coordinate point or system can be used.
[0074] After determining the angular positions of the separator array, the expected positions of the separators in the array can be calculated. The expected positions of the separators are calculated based on the design and / or manufacturing specifications of the microfluidic array plate. For example, the design and / or manufacturing specifications may provide the row spacing between any two adjacent separators in the same row. Similarly, the column spacing between any two adjacent rows may also be provided by the specifications. Furthermore, in some embodiments, two adjacent separator rows may be offset horizontally, and this offset is also known according to the design and / or manufacturing specifications. An example of such offset is shown in Figure 6B, where two adjacent rows may be offset by approximately the width of one separator. Using one or more of the known row spacing, column spacing, and offset values provided by the design and / or manufacturing specifications, the expected positions of the separators can be calculated based on the determined angular positions.
[0075] Referring back to Figure 3, in process 310, the fluorescence signals represented in the segmented region image are summarized. In some embodiments, the size of the segmented region is approximately 16 pixels x 16 pixels. In other embodiments, the size of the segmented region is 9 pixels x 11 pixels.
[0076] In some implementations, to summarize the fluorescence signals represented in a segmented region image for a specific segment, a 16x16 grid or frame is applied at the observation location of the segment. Next, several brightest pixels in the corresponding 16x16 grid are ignored. Then, the fluorescence signals of the next set of brightest pixels are averaged to obtain a summary value of the signal in the specific segment represented by the segmented region image. In some implementations, several darkest pixels are also ignored when calculating the summary value of the signal. The fluorescence signals in a segment represented by an image of a non-positive segment (e.g., a negative segment) can also be summarized in a similar or different manner (e.g., simply by averaging all pixels). It should be understood that other methods for summarizing the signals of segmented regions can also be used. For example, summarizing the signal may include, but is not limited to, integration, summation, averaging, weighted averaging, etc.
[0077] Referring back to Figure 3, in some implementations, a background image subtraction process may also be performed during the fluorescence data summarization step 310 to improve the quality of the segmented region image before it is provided to the classification process based on the machine learning model, as described below with reference to Figures 11 through 14. This may occur during the user threshold determination step 312.
[0078] Referring back to Figure 3, images of each segmented region can be obtained using the observation locations of the segmented regions determined in step 308 (e.g., by cropping the image at or near the observation location or by reproducing the original image of the segmented region array). During the fluorescence data summarization step 310, these images of each segmented region are provided to the classification process in step 312, with or without background image subtraction. In some embodiments, the classification process in step 312 is a machine learning model-based process that can identify images of class-rejection and class-non-rejection segmented regions, as well as images of class-positive and class-non-positive segmented regions.
[0079] Referring again to Figure 3, in step 314 of estimating target concentration, the presence or quantification of the target concentration is determined. In some embodiments, to detect the presence or quantify the target concentration, the total number of separatory regions, the number of positive separatory region images (minus separatory region images rejected in previous processes), and the total number of accepted separatory region images (including positive and negative separatory region images minus separatory region images rejected in previous processes) are obtained. Using Poisson statistics, the presence or quantification of the target concentration can be determined within a defined confidence interval. For example, the number of targets in the original solution can be correlated with the number of samples generating a positive detection signal, as represented by positive separatory region images. By performing one or more of the above procedures (e.g., classification based on machine learning models, observation location determination, spectral compensation, separatory region rejection by thresholding, rejection of long consecutive positive separatory regions), the target concentration can be quantified more accurately than with conventional analog PCR equipment.
[0080] In addition, in or after step 314, a visual report and chart of the results of target detection or concentration quantification can be generated and displayed to the user.
[0081] In some implementations, quality control is a function that helps identify sample failures or spurious external artifacts. To improve system throughput, automated quality control becomes an important aspect, helping to identify and report problems promptly and informing customers of potential failures during dPCR runs. It is crucial to warn users of any quality control issues so that they receive adequate warning about the potential adverse impact on the reliability of dPCR results.
[0082] Sometimes, samples are lost due to systematic failures or other errors. Figure 7 shows four examples of separator arrays, one representing a normal sample 702 and three representing various quality control issues. Array 702, representing a normal sample, shows uniform fluorescence overall. The quality control issues shown in Figure 7 include separator array 704, which exhibits failure due to the presence of excess gas, as shown by some non-uniformity in fluorescence within part of array 704. Separator array 706 illustrates a quality control issue in an array that exhibits failure due to filling problems, such as incomplete or non-uniform filling of separators or microchambers. This shows some non-uniformity in fluorescence, particularly on the left side of the array, which may be caused by incomplete filling or gaps in separators on the side of the array where the fluid dPCR solution did not reach. Array 708 exhibits a more catastrophic failure, with leakage occurring within the array. This could be caused by delamination, where defects or separation may exist in the injection-molded microfluidic device, which includes an array of partition zones, or where the thermoplastic film may separate from the microfluidic device. This thermoplastic film covers and seals (e.g., hermetically seals) the partition zones or microchambers and connects the siphon orifice and microchannel. The thermoplastic film can be at least partially permeable when a pressure differential is applied across it. Other types of failures that may occur include bubbles, debris, corner-finding failures, saturation, bridging, and dust. Such quality control problems are identified and tagged in the embodiments described herein, including severity level and failure type.
[0083] Figure 8 shows an example of a normal (e.g., positive) separator image 802 extracted from a raw sample image of a microchamber site. Positive separator images represent positive separators corresponding to the presence of one or more target concentrations within the separator. Positive separators include a target (e.g., a DNA molecule) that has been amplified by thermal cycling, making it detectable via a fluorescence signal. Therefore, positive separators contain a large number of target PCR amplicones of the raw DNA. Positive separator images are subsequently used to quantify the concentration of target PCR amplicones in a biological sample. The images show uniform fluorescence with a brighter region near the center of the separator. Negative separator images (not shown) correspond to separators where no target is present. If a particular separator does not have a target (e.g., a DNA target molecule, such as a PCR amplicon), thermal cycling in PCR will not produce target amplicones. Therefore, due to the lack of fluorescence signal, the fluorescence image of a negative separator appears mostly or completely dark. Negative separator images can be classified as non-positive separator images.
[0084] Figure 8 also shows an example 804 of a separator image showing a data quality problem. Non-positive separator images also include images associated with defective microcavities in the separator, images associated with contaminated separators, images associated with defective filling of microcavities in the separator, defective separator images, and any post-PCR separator images other than positive separator images.
[0085] Images associated with defective separators are also examples of images of non-positive separators. For example, defective separators may include design and / or fabrication defects in the microchambers of the separator, defects in the microfluidic array plate (such as reflective surfaces of the plate), delamination, leaks, gases, defective filling of the microchambers of the separator (e.g., filling problems), and any other defects in the separator or the microfluidic array plate. Such defective separators may or may not include targets (e.g., DNA target molecules, such as nucleic acids originally present in the biological sample or PCR amplicon generated from such nucleic acids). While thermal cycling in PCR may produce target amplicons of the target, defective separators may affect fluorescence signal detection and image generation. Images generated based on defective separators may not be usable for accurately quantifying the target concentration in the analyzed biological sample. Some of these images may generate unique signatures, and examples of such are shown in 804 of Figure 8. For example, gases, trapped air, or bubbles may appear as rings in a portion of the separator image. Filling problems may manifest as uneven filling or a brightening of the separator image. These images can be classified as non-positive separator images. Separator images that exhibit characteristics of quality control issues can be used to train a machine learning classifier, which can then be used to quickly identify and reject defective separator images, thereby improving the quality of dPCR target concentration estimation.
[0086] Images associated with contaminated separators are additional examples of images of non-positive separators. For example, separators may be contaminated with dust, fibers, particles, parasite DNA, etc. Contaminated separators may or may not include the target. While thermal cycling in PCR may generate amplicones of the target, contaminated separators may affect fluorescence signal detection and image generation. Images generated based on contaminated separators may not be usable for accurately quantifying the target concentration in the analyzed biological sample.
[0087] Images themselves may also be defective. Defective images can also be classified as non-positive separation area images. Imaging defects can include, for example, defects associated with the imaging system, such as dust on the camera lens, damaged images, image distortion, etc. These images may not be usable for accurately quantifying the target concentration in the biological sample being analyzed. These images can also be classified as non-positive separation area images.
[0088] However, many factors can affect the septa in the microfluidic array plate and the images of the septa, resulting in deviations from the ideal arrangement and / or ideal image of the septa. For example, some microcavities in the microfluidic array plate and / or septa may be defective, causing the spacing between some septa to be smaller or larger than the designed or desired spacing. Furthermore, contamination of the septa and / or the microfluidic array plate (e.g., dust, fibers, surface contamination) can affect the images of the septa, making them not an accurate representation of the septa arrangement. The imaging system can also introduce errors. As an example, when capturing images of the septa, the camera position may be slightly incorrect, so the microfluidic array plate and / or septa may not always be centered in the image. The microfluidic array plate and / or septa may be slightly rotated relative to the camera, causing the multiple rows and columns of the septa to not be perfectly horizontal and vertical in the image. Additionally, the camera lens may introduce some distortion, causing the rows and columns of the septa to not be perfectly straight, but rather they may have some slight curvature depending on their position in the image.
[0089] Although the images of negative separators, defective separators, contaminated separators, and defective images above are described as examples of images classified as either rejection or non-positive separators, it should be understood that any other separator image that is not a positive separator image can also be classified as a non-positive separator image.
[0090] Figure 9 shows examples of plots depicting dPCR fluorescence data for different channels of normal fluorescence data 902 and examples of plots depicting dPCR fluorescence data for different channels of data 904 affected by quality control. In the plots for each fluorescence channel in 902 and 904, the horizontal axis represents the number of separators (e.g., a total of 20,000 separators), and the vertical axis represents the fluorescence signal intensity. In the plots of 902 and 904, the fluorescence signal is plotted based on the detected signal. The plot of 902 depicting normal data shows a fairly uniform pattern of fluorescence signal intensity for target concentration and negative separators containing no target molecules. As mentioned above, many fluorescence signals in 904, representing class rejection or class non-positive, may be inaccurate due to various factors such as separator defects, image artifacts, contamination, etc. These inaccuracies manifest as more uneven random fluorescence values in the 904 plot.
[0091] Figure 10A is a flowchart illustrating a method for a quality control workflow 1000 for a partitioned array according to various embodiments. For this sample-level data quality check, three levels can be considered for data quality verification: (1) Pass: The current sample does not encounter any quality control issues; (2) Warning: The current sample is affected by a quality control issue that can (possibly) be corrected by adjusting a user-set threshold. Whether to use the results is determined by the customer; and (3) Fail: The current sample fails due to a serious quality control issue that may affect the results (e.g., concentration estimates). If the sample quality control status is Fail, no results are reported to the user. In one embodiment of the workflow, each unit (partitioned array or chip containing partitions) operates independently of each other. Therefore, this quality control workflow assumes independence between arrays. This means that a quality control issue in a given array will not affect other arrays. For example, currently, if a particular array leaks, information from that array will not be used to perform quality control assessments on adjacent arrays.
[0092] In the implementation shown in Figure 10A, frame 1000 will report the fault type and warning as follows. Specifically, for warnings, the cause is described by one or more of five categories, including saturation 1010, bridging 1012, positive distribution 1013, excessive quality control rejection 1014, or array overload 1015. If any of warning categories 1010, 1012, 1013, 1014, or 1015 are met, the system will issue warnings in steps 1020, 1018, 1015, 1016, and 1019 respectively, and continue the workflow until the sample array passes in step 1022. For failures, the cause may include one or more of the following: (a) Blank image error 1003, which includes images with no signal or out-of-focus images, and requires the user to close or delete the problematic channel from the analysis at 1005; (b) Image registration error 1004, which may cause the sample to fail and exit the quality control check at 1006; and (c) Sample failure 1006, which will cause the sample to fail and exit the quality control check at 1008.
[0093] In some implementations, the root cause in the sample failure category may be one of many possible systemic failures, including leaks, stratification, gas, or other systemic problems, but may not provide a root cause granularity that clearly identifies the root cause of the sample. In other implementations, the root cause of a sample failure may be provided based on information from a machine learning classifier disclosed herein or other information.
[0094] Data quality checks can follow the priority order shown in Figure 10A, meaning quality control can be triggered by a single problem, rather than by different problems simultaneously or at different times. The first quality control problem encountered will trigger the corresponding quality control flag. For saturation, if a sample fails and reaches saturation, the sample is reported as failed for quality control purposes, but the reason may be reported as saturation. The workflow of the analysis pipeline follows the logical order in which problems may occur. For example, blank image error checks 1003 and / or image registration error checks 1004 always precede other steps in workflow 1000 because if other aspects fail due to image registration or receive blank images, the user should be informed of the first point of failure. If a sample fails in step 1006, no further warnings are needed for that sample. For warnings, saturation is first detected in step 1010, then a bridging check is performed in step 1012, and finally an excessive quality control channel rejection check is performed in step 1014. Saturation checks are performed before bridging checks because saturation can lead to bridging, but the reverse is not true. Too many quality control thresholds rejected inspection 1014 Independent of other warnings.
[0095] The following describes further example implementation details of data quality control frameworks for some implementation schemes. Blank image errors can occur during digital quantification methods, including digital PCR, under various conditions. In some implementations, these conditions include failure to load any sample, intentional running without assays or templates, or incorrect software configuration on the client or system side. Blank image errors can trigger image failure indicators, requiring images to be processed differently. According to some implementations, as shown at 1031 in Figure 10B, a blank image check 1003 can include determining whether there are any location errors in the image. In some implementations, location error measures the goodness of fit of rows and columns on the sample image for estimating the location of sample compartments (e.g., microchamber sites). A location error can be identified when the pipeline fails to accurately estimate the location of sample compartment (e.g., microchamber sites). For example, the distance between two adjacent microchambers along rows and columns on the sample image can be determined, and the difference can be identified and compared to the median variation of that distance on the same sample image. If the difference is too large and / or too numerous, a location error may be identified. If the image has no positional errors, as shown in 1032, the blank image error check is considered normal. In response to an image with positional errors, as shown in 1033, the blank score is compared to a predetermined threshold. According to the embodiment, the blank score is obtained by comparing the intensity of the original image with the intensity of the transformed image. The transformed image includes an image transformed by amplifying the shape of the sample compartment (e.g., microchamber) to emphasize high values (i.e., max pooling (maximum value of a small subset of pixels)) of the image convolved with the sample compartment shape / mean of the original image. In response to a blank score not greater than the threshold, blank image errors and / or defocusing errors are marked in box 1035. As shown in 1037, the channel containing blank image errors and / or defocusing errors is then closed to eliminate the channel's influence on the analysis. In response to a blank score greater than the threshold, an image registration error check is performed in 1039.
[0096] According to the implementation plan, the image registration error check 1004 measures the goodness of fit of the rows and columns on the sample image for estimating the location of the compartmentalized areas (microchamber sites). An error is flagged when the pipe cannot accurately estimate the location of the microchamber sites. For example, the distance between two adjacent microchambers along a row and column on the sample image can be determined, and the difference is identified and compared to the median variation of that distance on the same sample image. If the difference is too large and / or too numerous, an image registration error may occur. Causes of image registration errors include corner-finding failure or sample not being loaded.
[0097] In some implementations, sample failure check 1008 is performed by the site classifier process described in Figures 11 and 12 below. The unit of measurement is the number of sites rejected by the site classifier. When the number of sites rejected by the site classifier exceeds an analytical threshold (e.g., 6.1% of the total number of microchambers in the sample), it may cause the sample to fail. The site classifier discussed in Figure 11 below is a sophisticated algorithm that captures most of the serious failures discussed above, including delamination, bubbles, gas, leaks, and filling problems.
[0098] In some implementations, a saturation check 1010 is performed when a site is rejected by the optical saturation detection algorithm. Optical saturation is observed when the signal from the imaging channel exceeds the operating range of the detection sensor on the dPCR instrument, for example, causing the digital image to appear overly bright or saturated. In one implementation, it is detected on the raw microchamber site image at the end of the cycle and manifests as extremely high intensity values on the pixels belonging to the microchamber region in the digital site image. A small region belonging to the microchamber signal is used to detect saturation in each microchamber. In one implementation, the channel is saturated if 8% of the total microchamber sites in the sample are saturated. Saturation typically manifests as false brightness in the FAM channel, but it can also manifest in the ROX channel due to JUN2ROX spectral crosstalk.
[0099] A bridging check 1012 is characterized by the presence of a continuous length of positive (e.g., a bright microchamber) within a negative region (e.g., a sparsely lit (bright) area). In one implementation, a bridging check warning is issued if 2.5% of the total microchambers in the sample are flagged as bridging errors, allowing the user to check if the bridging warning threshold has been incorrectly set. Positive distribution errors 1013 can occur when positive septa (reaction sites) are poorly distributed in the sample plate, leading to inconsistent data interpretation. Images of septated samples may show variations and / or differences in target analyte / molecule distribution. For example, uneven distribution of target analytes / molecules within the sample may result in some septa containing many target molecules while others contain very few. These variations and / or differences can introduce uncertainty into the data analysis process, potentially affecting the accuracy and reliability of experimental results. In digital PCR, samples may be detected at concentrations outside the range of Poisson statistics. This could be due to poor dilution, for example.
[0100] In some embodiments, the positive distribution error at 1013 is determined by evaluating the uniformity of the distribution of positive sample separators across nine regions of a chip or sample plate array. These nine regions may include the upper left, upper center, upper right, middle left, middle center, middle right, lower left, lower center, and lower right regions. In some embodiments, the nine regions contain substantially the same amount of sample separators (e.g., microchambers or droplets). According to embodiments, a positive distribution error is indicated by a coefficient of variation (CV) greater than 0.3 for the positive counts across the nine regions. The coefficient of variation is calculated as CV = σ / µ, where σ is the standard deviation of the positive counts across the nine regions, and µ is the average of the positive counts across the nine regions. In some embodiments, the positive distribution error check at 1013 is protected from unreliable CV values caused by low positive counts. For example, any target with an average of less than 50 positives will be designated as "NAN" instead of a positive distribution error and / or an unreliable CV value. In the implementation, samples with a number of positive septa (e.g., microchambers) greater than 90% of the total number of positive septa (e.g., microchambers) but less than 450 are designated as having unreliable coefficients of variation. According to the implementation, a QC (quality control) warning for a sample is triggered when the number of positive septa is less than 90% of the total sample septa count but greater than 450, and the coefficient of variation for all regions is greater than 0.3. In the implementation, unreliable coefficients of variation may be caused by partial bridging in the array, background issues, and / or incorrect thresholds.
[0101] According to the implementation scheme, the number of sites outside the quality control threshold range is used to measure the quality control threshold rejection check 1014. In one implementation scheme, if 5% of the total microchambers in the sample are outside the quality control threshold range, an alert is triggered, and the user can check whether the lower limit of the quality control threshold is set too high or the upper limit of the quality control threshold is set too low, resulting in an excessive number of rejections.
[0102] According to the implementation scheme, an overload array error 1017 is triggered when the analysis detects essentially no negative sample separators or no negative sample separators at all. Based on Poisson statistics, the average number of molecules per sample separator is obtained by the following formula: λ = -ln(z / n), where z is the number of negative separators and n is the total number of separators. Therefore, an overload array with no negative separators can affect Poisson-based calculations and data analysis. In response to the overload array error at 1017, the concentration of the sample operation needs to be corrected, and / or the threshold of the overload dye needs to be adjusted.
[0103] If the sample image of the segmented array passes all checks of the quality control framework 1000 as shown in Figure 10A, no quality control problem is detected, and the image passes in step 1022. In some embodiments, the sample image may exit the quality control framework check 1000 upon the first warning, and if a warning has already been issued, the framework does not check for subsequent warnings.
[0104] Using a trained machine learning model, many rejection and non-positive signals can be removed or filtered out by correctly classifying the separator images into non-positive and class-rejection (and therefore rejected) separator images. Therefore, the sample failure inspection classification process 1008 based on the machine learning model described in Figures 10A and 11 can perform defect detection at the micro-cavity level by evaluating micro-cavity site images. In some implementations, machine learning and data analysis can be utilized to significantly improve the accuracy of identifying positive separator images.
[0105] Machine learning models can be trained using images pre-identified as positive or negative / rejection segmentation regions. For example, the normal segmentation region images and the segmentation region images affected by quality control in Figure 8 can both be used as training images for training machine learning models, for example, as positive or negative / rejection segmentation region images and negative / rejection segmentation region images, respectively.
[0106] Positive and non-positive separatory region images can be obtained from past PCR tests. For example, a set of post-PCR images can be classified and annotated (manually or automatically) as non-rejection and rejection separatory region images. In some implementations, training a machine learning model may require many annotated images to provide a sufficient number and variety of positive and non-positive, as well as rejection and non-rejection separatory region images. Therefore, images obtained solely from past PCR analyses may be sufficient or insufficient for training a machine learning model.
[0107] If the images obtained from past PCR analyses are insufficient, additional training images may be required. To obtain additional images for training a machine learning model, in some implementations, a set of non-positive separatory region (NPMR) images can be generated based on another set of NPMR images. For example, a first set of images is obtained from past PCR tests. Images in the first set are pre-identified and annotated as NPMR images. Images in the first set can be modified to obtain a second set of images. Such image modifications may include, but are not limited to, rotating, editing, cropping, distorting, mirroring, brightening, darkening, changing the contrast, changing the color, and changing the pattern of the first set of images. Modified images in the second set can also be pre-identified and annotated as NPMR images. Therefore, the second set can be included in the set of NPMR images used to train the machine learning model, thereby increasing the number and variety of available training images.
[0108] Similarly, in some implementations, a set of images representing non-positive separators can be generated based on a set of images representing positive separators. For example, a third set of images is obtained from past PCR tests. Images in the third set are pre-identified and annotated as images representing positive separators. The images representing positive separators in the third set can be modified to obtain a fourth set of images, which can be images representing non-positive separators. Such image modifications may include, but are not limited to, rotating, editing, cropping, distorting, mirroring, brightening, darkening the first set, changing its contrast, changing its color, and changing its pattern, or more of these. The modified images in the fourth set can be pre-identified and annotated as images representing non-positive separators. Therefore, the fourth set of images can be included in the set of images representing non-positive separators used to train machine learning models, thereby increasing the number and variety of available training images.
[0109] In the example above, a set of non-positive separator images and / or a set of positive separator images are modified and used to generate another set of non-positive separator images. Similarly, a set of non-positive separator images and / or a set of positive separator images can be modified and used to generate another set of positive separator images. In some implementations, the modified non-positive separator images and / or positive separator images are also referred to as seed images. By modifying the seed images, many other images can be generated to expand the number of training images. Modification of the seed images can also be configured to generate different types of non-positive separator images and / or positive separator images, thereby significantly improving the variety of training images for the machine learning model.
[0110] In some implementations, a set of images representing positive separators and a set of images representing non-positive separators, and / or a set of images representing rejections and non-rejections, are combined to generate one or more datasets for training a machine learning model. The datasets may include training datasets, validation datasets, and test datasets. For example, one-third of the images may be used as the training dataset; one-third as the validation dataset; and one-third as the test dataset. The datasets may include image files and / or features extracted from the image files.
[0111] Using one or more datasets and one or more computing devices (e.g., servers), a machine learning model is iteratively trained to determine a set of parameters. This set of parameters may include, for example, the weights and features of the machine learning model. Training can be performed using, for example, stochastic gradient descent and / or its extensions and variants, such as Adaptive Moment Estimation (Adam), Implicit Update (ISGD), Momentum Method, Average Stochastic Gradient Descent, Adaptive Gradient (AdaGrad), and Root Mean Square Propagation (RMSProp).
[0112] According to various implementations, the machine learning model for image classification can be, for example, a convolutional neural network (CNN). A CNN includes an input layer, one or more hidden layers, and an output layer. The hidden layers include layers that perform convolutions, such as one or more convolutional layers. CNNs may also include local and / or global pooling layers as well as convolutional layers. Pooling layers are used to reduce the dimensionality of the data. While CNNs are used as an example of a machine learning model for image classification, it should be understood that extensions and variations of CNNs and / or other types of neural networks for image classification can also be used. In one implementation, a CNN may include a 19-layer deep VGG19 image classification convolutional neural network to modify or optimize the network for classifying the smaller pixel size of microcavity site images.
[0113] After training, the trained machine learning model is configured with the set of parameters determined from the training. The trained machine learning model can then be used to classify images as positive or non-positive segmentation region images, or as rejection or non-rejection segmentation region images. As described above, the observation locations of the segmentation regions represented in the post-PCR images are determined. Based on the observation locations, images of each segmentation region can be obtained (e.g., by cropping or reproducing the image at or near the observation location). These individual segmentation region images are then fed to the trained machine learning model. Optionally, image subtraction is performed to improve the quality of the post-PCR images before feeding the segmentation region images to the trained machine learning model. The trained machine learning model then classifies each of the individual segmentation region images as a positive or non-positive segmentation region image, or as a rejection or non-rejection segmentation region image. In some embodiments, the trained machine learning model determines the probability that a particular segmentation region image is a positive segmentation region image or the probability that a particular segmentation region image is a rejection segmentation region image. Based on probability, an image is classified as a positive or non-positive segmentation region in a machine learning model, or as a rejection or non-rejection segmentation region in another machine learning model.
[0114] Some implementations use a combination of two machine learning classifiers, which may be architecturally similar but trained differently. These classifiers are executed on differently processed inputs from images with the same background and endpoint to perform separate classification tasks. The classification results are then logically mapped to ultimately determine whether to reject data from specific sample separator regions. Specifically, one trained machine learning model classifies separator region images as positive or non-positive, while another trained machine learning model classifies them as reject or non-reject. In a particular example, the inputs to each model are preprocessed differently using relevant source images, as further described below.
[0115] Figure 11 is a flowchart illustrating a method 1100 for classifying images of individual microchamber sites in an array according to various embodiments. In step 1104, a processed image is obtained, and in step 1120, an original endpoint image is obtained, as described below with respect to Figure 12. In step 1106, microchamber site images are extracted from the processed image, and in step 1122, corresponding microchamber site images are extracted from the original endpoint image.
[0116] To distinguish {POS-like, non-POS-like, where POS = positive} sites, each microchamber image is normalized individually, unlike the {REJ-like, non-REJ-like, where REJ = rejection} normalization. In step 1106, microchamber site images (C) are extracted from the processed images.处理 = C 终点 – C 扭曲背景 For example, C 处理 = C40 – C15). The image size is 16 pixels × 16 pixels. After extracting the microchamber image from the processed image, the values in the processed image are in the 16-bit range of [-65535, 65535]. In step 1108, this range is non-linearly mapped to [-1, 1].
[0117] In step 1110, the 90th percentile value of all pixels in the microchamber image is calculated. In some embodiments, the image of a single microchamber contains 256 pixels. The quantile values are calculated by sorting the positive and negative values. The absolute value of the 90th percentile value is then calculated, and each pixel in the image is divided by that value. Finally, values outside the range limit of [-1, 1] are reduced to these range limits. In step 1112, these values are run through a classifier, which publishes a positive or non-positive classification. In one embodiment, if the classifier publishes a positive classification, a rejection or non-rejection classification is not published to improve computational efficiency.
[0118] In step 1122, from the original endpoint image (C 终点 For example, C40) is used to extract microchamber images. The values in the original image are in the range [0, 65535]. In step 1124, this range is linearly mapped to [0, 1]. This mapping is performed as follows. For a given channel (FAM, VIC, etc.), the values from C40 are found to be within the range [0, 65535]. 扭曲背景 and C 终点 The maximum and minimum pixel values are obtained from all microchamber images. In one implementation, minimum-maximum normalization is used for each (C) channel. 扭曲背景 C 终点 Each pixel value in the image pair is linearly rescaled to the range [0, 1].
[0119] In step 1126, a distorted background image is generated by subtracting the processed image from the original endpoint image, as discussed below with respect to Figure 12. Since the input is scaled, the distorted background image is scaled to [0, 1].
[0120] In step 1128, a scaled, distorted background image and a scaled endpoint image for each microchamber of each channel (e.g., FAM, VIC) are fed to array-level minimum-maximum normalization step 1128. The minimum and maximum values of the scaled, distorted background image and scaled endpoint image for a given channel are determined across the entire segmentation array, and these values are used to normalize (e.g., linearly scale) the pixel values in the scaled, distorted background image and scaled endpoint image; in one embodiment, minimum-maximum normalization is used. In step 1130, these two images are concatenated and fed into a machine learning-based classifier (e.g., each segmentation is represented by, for example, a 16x16x2 array). Note that other types of normalization known in the art may be used in steps 1130 and 1122.
[0121] In step 1116, the outputs from the class positive and class non-positive classifiers 1112 and the class rejection and class non-rejection classifiers 1116 are logically evaluated. If the output of classifier 1112 is class positive and the output of classifier 1130 is class rejection, the segmentation region is rejected; otherwise, the segmentation region is not rejected.
[0122] Figure 12 is a flowchart illustrating a method 1200 for generating a processed image according to various embodiments. Before PCR, a raw background image is obtained from the background cycle in 1210. After PCR, a raw endpoint image is also obtained in 1220. A distortion operation is performed in 1230 to align the separators of the background image and the raw endpoint image, thereby obtaining a distorted background image 1240. In one embodiment, homography can be used to perform the distortion. In a subtraction operation 1250, the distorted background image 1240 is subtracted from the raw endpoint image 1220 to form a processed image 1260, which is stored together with the raw endpoint image for analysis. In one embodiment, the site locations of the background cycle image are not calculated, and the raw background cycle image can be discarded after image distortion. The distorted background image can be used as a substitute background image for subsequent analysis. The distorted background image can be obtained by subtracting the processed image (pixel-by-pixel) from the raw endpoint cycle image.
[0123] Figure 13 illustrates various methods 1300 for automating quality control by determining whether to reject separatory region image data to prevent its use in calculating dPCR-based target concentration results. Further details of these methods 1300 are discussed throughout the specification, particularly with respect to Figures 11 and 12. In step 1320, image data representing multiple separatory regions, including background image data and endpoint image data, are obtained. In step 1340, separatory region background image data and separatory region endpoint image data are extracted from the multiple separatory regions. In step 1360, the separatory region background image data and endpoint image data are preprocessed to obtain processed image data for input into a machine learning model. In step 1380, the multiple separatory region images are classified into class rejection and class non-rejection separatory region images, and class positive separatory region images or class non-positive separatory region images by one or more trained machine learning models. As described with respect to Figure 14, the trained machine learning model is trained using one or more images identified as class rejection, class non-rejection, class positive, and class non-positive. In one implementation, classifying the multiple separatory images includes, for each separatory image in the multiple separatory images, determining the probability that the separatory image is a class of rejection separatory image and a class of non-positive separatory image, as described with respect to Figure 11. Based on this probability, in step 1390, the separatory image data is rejected so that it is not used to calculate the target concentration result based on dPCR.
[0124] Referring to Figure 14, method 1400 is provided. Method 1400 is used to train a first machine learning model that analyzes one or more biological samples to determine whether a segmentation region image is considered class rejection or class non-rejection. A second machine learning model can also be trained using a similar approach to method 1400, which analyzes one or more biological samples to determine whether a segmentation region image is considered class positive or class non-positive. Method 1400 is performed by one or more computing devices.
[0125] In some embodiments, method 1400 begins at step 1420, during which first plurality of images (e.g., image 804 in Figure 8) are obtained, identified as images of class-rejection separators. The class-rejection separator image is associated with a class-rejection separator that may have one or more unique signatures of separator images with data quality issues. The class-positive separator image is associated with a class-positive separator that may exhibit fluorescence characteristics corresponding to one or more target concentrations present in the separator. The positive separator includes a target (e.g., a DNA molecule) that has been amplified by thermal cycling, making it detectable via a fluorescence signal. Therefore, the positive separator includes a large number of target PCR amplicones of the original DNA. The positive separator image is then used to quantify the concentration of target PCR amplicones in a biological sample.
[0126] In step 1440, a second plurality of images (e.g., normal or negative separator images shown in 802 of FIG. 8) are obtained as images identified as quasi-non-rejection separator images. The second plurality of images may include one or more images modified from one or more other images identified as quasi-non-positive separator images. The quasi-non-positive separator image may include at least one negative separator image associated with a negative separator corresponding to the absence of target concentration in the negative separator. The quasi-non-positive separator image may also include at least one image associated with a defective microcavity in the separator. The quasi-non-positive separator image may also include at least one image associated with a contaminated separator. The non-positive separator image may also include at least one image associated with defective filling of the microcavities in the separator. The non-positive separator image may also include at least one defective image of the separator.
[0127] In some implementations, obtaining the second plurality of images (step 1440) includes obtaining a first subset of the second plurality of images identified as non-positive delimited region images. The first subset of the second plurality of images is modified to obtain a second subset of the second plurality of images, which is identified as non-positive delimited region images. Step 1440 also includes including the second subset in the second plurality of images. Modifications to the first subset of the second plurality of images include rotating, editing, cropping, distorting, mirroring, brightening, darkening, changing the contrast, changing the color, and changing the pattern of one or more of the following:
[0128] In some implementations, obtaining the second plurality of images (step 1440) includes obtaining a first subset of the first plurality of images identified as positive delimited region images. Modifying the first subset of the first plurality of images to obtain a third subset of the second plurality of images, which is identified as non-positive delimited region images, further includes incorporating the third subset into the second plurality of images. Modifications to the first subset of the first plurality of images include rotating, editing, cropping, distorting, mirroring, brightening, darkening, changing the contrast, changing the color, and changing the pattern of one or more of the following:
[0129] In step 1460, one or more datasets are generated using a first plurality of images (e.g., the image shown in 804 of FIG. 8) and a second plurality of images (e.g., the image shown in 802 of FIG. 8, and other images as described above). In one embodiment, the one or more datasets include a training dataset, a validation dataset, and a test dataset.
[0130] In step 1480, a set of parameters for the machine learning model is determined. This set of parameters is determined by iteratively training the machine learning model using one or more datasets. Based on the results of the iterative training, this set of parameters for the machine learning model is determined. In one embodiment, the machine learning model includes a convolutional neural network (CNN) as described above. In one embodiment, the convolutional neural network may be a variant of the VGG19 image classifier. In another embodiment, the CNN includes 19 layers. In yet another embodiment, the CNN includes 19 layers, including 16 convolutional layers and 3 fully connected layers.
[0131] According to the implementation scheme, a composite image is generated after the QC process as a visual representation of one or more chips on the sample board, as shown in Figures 10 to 14. The composite image can be generated based on the QC results and the QC process, as shown in Figures 10 to 14. Figure 14As shown in Figure 16A, the process 1600 for generating the synthetic image includes determining at 1602 whether a channel exists. If so, it determines at 1604 whether an image exists. In response to the presence of an image, process 1600 includes obtaining cell-level rejection information at 1606, including rejection based on sample failure and rejection information based on sample quality. As shown in Figure 16A, for each channel in the cell, process 1600 may include calculating the position of a sample compartment (e.g., a microchamber) in the image coordinate space at 1608. The calculation at 1608 includes calculating the position of each sample compartment (e.g., a microchamber) after calculating the array angle. This process tiles a known number of sample compartments on the image. These tiles can then be adjusted in a larger rolling block using a template matching method to avoid minor changes in their positions, thereby detecting peaks corresponding to the positions of the sample compartments (e.g., microchambers) on the image. According to an embodiment, process 1600 includes performing minimum-maximum normalization on the processed image at 1610 using positive and negative sites to generate normalized image data. Process 1600 may also include rescaling the normalized image data to a range [0.5 to 1.0] at 1612 to generate rescaled normalized image data. The rescaling at 1612 is configured to prevent color desaturation in dark intensities. In an embodiment, as shown at 1614, process 1600 may include calculating the color of a septum (e.g., a microchamber) based on the type (positive or negative) and intensity of each septum (e.g., a microchamber) using the rescaled normalized image data. According to some embodiments, as shown at 1616, process 1600 includes generating an image based on calculated attributes. Calculated attributes may include the color of the site, the label of the site (whether the septum is positive, negative, or rejected), the location of the site, or any combination thereof. As shown at 1618, process 1600 includes adjusting image parameters of the generated image to produce an adjusted image. Image parameters may include brightness, contrast, image resolution, overall appearance of the image, or any combination thereof. At 1618, adjustment may include increasing the brightness and / or contrast of the generated image. Process 1600 may include returning the adjusted image as a composite image of one or more chips of a sample board, as shown at 1620.
[0132] According to an implementation, as shown in Figure 16B, step 1614, calculating the color of the separator regions of one or more chips in a sample plate, may include obtaining all positive sites at 1632. Information may include a label indicating whether a site is positive (i.e., whether it is above a threshold used to distinguish between positive and negative sites). Step 1614 may include obtaining a site summary for each separator region (e.g., a microchamber), as shown at 1634. In some implementations, the site summary at 1634 may include the median or average of pixels belonging to the sample separator region, or any combination thereof. Step 1614 may include rescaling the site summary to an α value, as shown at 1636. In an implementation, the α value may be obtained by performing minimum-maximum normalization on the site summaries of positive and negative sample separator regions, so that the minimum site summary is mapped to 0.5 and the maximum site summary is mapped to 1.
[0133] Step 1614 (as shown at 1638 in Figure 16B) may include generating the color of the segmented region (e.g., microchamber) of the image based on the type of site (segmentation) (including positive, negative, and rejection). As shown at 1640, for positive sites, the color is generated by mixing an α value with the color of the corresponding dye channel. As shown at 1642, for negative sites, the color is generated by mixing an α value with a negative color. In some embodiments, the negative color is selected based on some user interface test. The mixing at 1640 and 1642 involves mapping the red, green, and blue color values to the HSL (hue, saturation, and lightness) color space. In the HLS color space, the hue and saturation values are preserved, while the lightness value is modified using the following calculation: min(1.0, α * lightness of the microchamber primary color). The primary color may correspond to a fixed color value for the positive sample segmentation region. Each filter channel supported by the instrument is different. Exemplary hexadecimal color values for positive sites include: JUN dye deep red "0xC00000", ABY dye yellow "0xFFC000", VIC dye green "0x18CF2F", FAM dye blue "0x4472C4", ROX dye light red "0xFF8080", and other dye colors "0xC0C0C0". As shown at 1644, for rejected sites, a fixed color is applied. The fixed color may include a rejection-on color "0x595959" and a rejection-off color "0x000000".
[0134] The systems, apparatus, and methods described herein can be implemented using digital circuit systems or one or more computers utilizing well-known computer processors, memory units, storage devices, computer software, and other components. Typically, a computer includes a processor for executing instructions and one or more memories for storing instructions and data. A computer may also include or be coupled to one or more mass storage devices, such as one or more disks, internal hard disks and removable disks, magneto-optical disks, optical disks, etc.
[0135] The systems, apparatus, and methods described herein can be implemented using computers operating in a client-server relationship. Typically, in such a system, the client computer is located remotely from the server computer and interacts via a network. The client-server relationship can be defined and controlled by computer programs running on the respective client and server computers. Examples of client computers may include desktop computers, workstations, portable computers, cellular smartphones, tablets, or other types of computing devices.
[0136] The systems, apparatuses, and methods described herein can be implemented using computer program products tangibly embodied in an information carrier (e.g., a non-transitory machine-readable storage device) for execution by a programmable processor; and the steps of the methods described herein, including one or more of the steps of the methods in Figures 2, 3, and 10 through 14, can be implemented using one or more computer programs executed by such processors. A computer program is a set of computer program instructions that can be used directly or indirectly in a computer to perform an activity or produce a result. A computer program can be written in any form of programming language, including compiled or interpreted languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.
[0137] Figure 15 illustrates a high-level block diagram of an exemplary device that can be used to implement the systems, apparatuses, and methods described herein. Device 1500 includes a processor 1510 operatively coupled to persistent storage device 1520 and main memory device 1530. Processor 1510 controls the overall operation of device 1500 by executing computer program instructions defining these operations. The computer program instructions may be stored in persistent storage device 1520 or other computer-readable medium and loaded into main memory device 1530 when execution of the computer program instructions is desired. For example, processor 1510 may include one or more components of PCR device 100. Therefore, the method steps of Figures 2, 3, and 10 through 14 may be defined by computer program instructions stored in main memory device 1530 and / or persistent storage device 1520 and controlled by processor 1510 executing the computer program instructions. For example, computer program instructions can be implemented as computer-executable code programmed by those skilled in the art to execute the algorithm defined by the method steps of Figures 2, 3, and 10-14. Therefore, by executing the computer program instructions, processor 1510 executes the algorithm defined by the method steps of Figures 2, 3, and 10-14. Device 1500 also includes one or more network interfaces 1580 for communicating with other devices via a network. Device 1500 may also include one or more input / output devices 1590 that enable a user to interact with device 1500 (e.g., a display, keyboard, mouse, speaker, buttons, etc.).
[0138] Processor 1510 may include general-purpose and special-purpose microprocessors, and may be the sole processor of device 1500 or one of multiple processors. Processor 1510 may include one or more central processing units (CPUs) and one or more graphics processing units (GPUs), which may, for example, operate independently of one or more CPUs and / or perform multitasking with one or more CPUs to accelerate processing (e.g., for the various image processing applications described herein). Processor 1510, persistent storage device 1520, and / or main memory device 1530 may include one or more application-specific integrated circuits (ASICs) and / or one or more field-programmable gate arrays (FPGAs), supplemented by one or more ASICs and / or one or more FPGAs, or incorporated into one or more ASICs and / or one or more FPGAs.
[0139] Persistent storage device 1520 and main memory device 1530 each include a tangible, non-transitory, computer-readable storage medium. Persistent storage device 1520 and main memory device 1530 may each include high-speed random access memory, such as dynamic random access memory (DRAM), static random access memory (SRAM), double data rate synchronous dynamic random access memory (DDR RAM), or other random access solid-state memory devices, and may include non-volatile memory, such as one or more disk storage devices such as internal hard disks and removable disks, magneto-optical disk storage devices, optical disk storage devices, flash memory devices, semiconductor memory devices such as erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), optical disc read-only memory (CD-ROM), digital versatile optical disc read-only memory (DVD-ROM), or other non-volatile solid-state memory devices.
[0140] Input / output device 1590 may include peripheral devices such as printers, scanners, displays, etc. For example, input / output device 1590 may include display devices for displaying information to a user (such as cathode ray tube (CRT), plasma, or liquid crystal display (LCD) monitors), keyboards, and pointing devices (such as mice or trackballs through which a user can provide input to device 1500).
[0141] Any or all of the functions of the systems and devices discussed herein may be executed by processor 1510 and / or incorporated into devices such as PCR device 100. Furthermore, PCR device 100 and / or device 1500 may utilize one or more neural networks or other deep learning techniques executed by processor 1510 or other systems or devices discussed herein.
[0142] Those skilled in the art will recognize that specific implementations of actual computers or computer systems may have other structures and may contain other components, and Figure 15 is a high-level representation of some of the components of such a computer for illustrative purposes.
[0143] Throughout the specification and claims, unless the context clearly specifies otherwise, the following terms shall have the meaning explicitly and relating herein:
[0144] Although the phrase "in one embodiment" as used herein may refer to the same embodiment, it does not necessarily refer to the same embodiment. Therefore, as described below, various embodiments of the invention can be readily combined without departing from the scope or spirit of the invention.
[0145] As used herein, the term “or” is an inclusive “or” operator and is equivalent to the term “and / or”, unless the context clearly specifies otherwise.
[0146] The term "based on" is not exclusive and may be based on additional factors not described unless the context clearly specifies otherwise.
[0147] As used herein, and unless the context otherwise requires, the term “coupled to” is intended to include both direct coupling (where two elements coupled to each other are in contact with each other) and indirect coupling (where at least one additional element is located between the two elements). Therefore, the terms “coupled to” and “coupled with” are used synonymously. In the context of a networked environment in which two or more components or devices are able to exchange data, the terms “coupled to” and “coupled with” are also used to mean “communically coupled” to, possibly via one or more intermediate devices.
[0148] Furthermore, throughout the specification, the meanings of “a” and “the” include plural references, and the meaning of “in” includes both “in” and “on”.
[0149] While some of the embodiments given herein constitute a single combination of the elements of the invention, it should be understood that the subject matter of the invention is considered to include all possible combinations of the disclosed elements. Therefore, if one embodiment includes elements A, B, and C, and another embodiment includes elements B and D, the subject matter of the invention is also considered to include other remaining combinations of A, B, C, or D, even if not explicitly discussed herein. Furthermore, the transitional term "comprising" means having a portion or component, or such portions or components. As used herein, the transitional term "comprising" is inclusive or open-ended and does not exclude additional, unlisted elements or method steps.
[0150] The following disclosure may refer to servers, services, interfaces, engines, modules, clients, peers, portals, platforms, or other systems formed by computing devices. It should be understood that these terms are considered to refer to one or more computing devices having at least one processor (e.g., ASIC, FPGA, DSP, x86, ARM, ColdFire, GPU, multi-core processor, etc.) configured to execute software instructions stored on computer-readable tangible, non-transitory media (e.g., hard disk drives, solid-state drives, RAM, flash memory, ROM, etc.). For example, a server may include one or more computers that operate as a web server, database server, or other type of computer server in a manner that performs the described roles, responsibilities, or functions. It should also be understood that the disclosed computer-based algorithms, processes, methods, or other types of instruction sets may be embodied as a computer program product comprising a non-transitory tangible computer-readable medium storing instructions that cause a processor to perform the disclosed steps. Various servers, systems, databases, or interfaces may exchange data using standardized protocols or algorithms, possibly based on HTTP, HTTPS, AES, public-private key exchange, web service APIs, or other electronic information exchange methods. Data exchange can take place on packet-switched networks, circuit-switched networks, the Internet, LANs, WANs, VPNs, or other types of networks.
[0151] As used in the description herein and throughout the claims thereafter, when a system, engine, server, device, module or other computing element is described as being configured to perform functions on data in memory, the meaning of “configured to” or “programmed to” is defined as one or more processors or cores of the computing element being programmed by a set of software instructions stored in the memory of the computing element to perform that set of functions on target data or data objects stored in memory.
[0152] It should be noted that any language targeting computers should be understood to include any suitable combination of computing devices or network platforms, including servers, interfaces, systems, databases, agents, peers, engines, controllers, modules, or other types of computing devices operating individually or collectively. It should be understood that computing devices include processors configured to execute software instructions stored on tangible, non-transitory computer-readable storage media (e.g., hard disk drives, FPGAs, PLAs, solid-state drives, RAM, flash memory, ROM, etc.). These software instructions configure or program the computing device to provide roles, responsibilities, or other functions, as discussed below with respect to the disclosed devices. Furthermore, the disclosed technology may be embodied as a computer program product including a non-transitory computer-readable medium storing software instructions that cause a processor to perform the disclosed steps associated with the implementation of computer-based algorithms, processes, methods, or other instructions. In some embodiments, various servers, systems, databases, or interfaces use standardized protocols or algorithms, possibly based on HTTP, HTTPS, AES, public-private key exchange, web service APIs, known financial transaction protocols, or other electronic information exchange methods, to exchange data. Data exchange between devices can take place on packet-switched networks, the Internet, LANs, WANs, VPNs, or other types of packet-switched networks; circuit-switched networks; cellular-switched networks; or other types of networks.
[0153] In the context of this specification, at least the following embodiments are described. Embodiment 1 is a method for computerized image processing of digital polymerase chain reaction (dPCR) analysis of a biological sample to automatically reject data from individual separators so as not to be used in calculating dPCR-based target concentration results, the method comprising: obtaining image data representing a plurality of separators disposed in a container, the image data including background image data corresponding to a background image captured prior to an amplification cycle of the dPCR assay of the biological sample, and endpoint image data corresponding to an endpoint image captured after an endpoint cycle of the dPCR assay of the biological sample; extracting individual separator image data from the obtained image data, including separator background image data and separator endpoint image data; preprocessing the separator background image data and the separator endpoint image data to obtain processed image data for input into one or more machine learning models; generating one or more classification results for one separator among the plurality of separators using the one or more machine learning models; and using the one or more classification results to determine whether to reject image data corresponding to the separator among the plurality of separators so as not to be used in calculating dPCR-based target concentration results.
[0154] Implementation Scheme 2 is the method according to Implementation Scheme 1, wherein the one or more classification results include class-positive classification and class-non-positive classification. Implementation Scheme 3 is the method according to any one of Implementation Schemes 1 and 2, wherein the one or more classification results further include class-rejection classification and class-non-rejection classification. Implementation Scheme 4 is the method according to Implementation Scheme 3, wherein if the segmentation region is classified as class-non-positive and class-reject, the image data corresponding to the segmentation region is rejected. Implementation Scheme 5 is the method according to any one of Implementation Schemes 1 to 4, wherein extracting the image data of each segmentation region of the plurality of segmentation regions further includes determining one or more corner positions of the plurality of segmentation regions. Implementation Scheme 6 is the method according to Implementation Scheme 5, wherein extracting the image data of each segmentation region of the plurality of segmentation regions further includes determining the position of each segmentation region in the segmentation region array. Implementation Scheme 7 is the method according to any one of Implementation Schemes 1 to 6, wherein extracting the image data of each segmentation region of the plurality of segmentation regions further includes determining a summary of microchamber sites from the image data of each segmentation region of each of the plurality of segmentation regions. Implementation Scheme 8 is the method according to Implementation Scheme 7, further including performing a quality control process on the image data. Implementation scheme 9 is the method according to implementation scheme 8, wherein the quality control process includes marking image registration errors before generating one or more classification results in response to the inability to determine one or more corner locations, individual partition locations, or microchamber site aggregations.
[0155] Implementation 10 is the method according to implementation 8, further comprising: performing a blank image error check before checking for image registration errors. Implementation 11 is the method according to implementation 10, wherein the blank image error check comprises: determining whether an image has any positional errors; in response to the image having any positional errors, comparing a blank score of the image with a predetermined threshold; and in response to the blank score being greater than the predetermined threshold, designating the image as having a blank image error or a defocus error. Implementation 12 is the method according to any one of implementations 10 and 11, further comprising, in response to the blank image error or defocus error, closing the channel for acquiring the image.
[0156] Implementation 13 is the method according to any one of Implementations 8 to 12, wherein the quality control process further includes generating one or more warnings in response to one or more non-faulty quality control errors. Implementation 14 is the method according to Implementation 13, wherein the one or more non-faulty quality control errors include saturation errors, bridging errors, positive distribution errors, excessive quality control rejection errors, overload array errors, or any combination thereof. Implementation 15 is the method according to Implementation 14, wherein the positive distribution is determined by evaluating the uniformity of the positive distribution on two or more regions of a chip or sample board array. Implementation 16 is the method according to any one of Implementations 1 to 15, wherein preprocessing the partition background image data and the partition endpoint image data includes warping a background image corresponding to the background image data with an endpoint image corresponding to the endpoint image data, such that the respective partition image data corresponding to one or more of the plurality of partitions are aligned to form a warped background image. Implementation 17 is the method according to Implementation 16, wherein the warping is performed using a homography transformation. Implementation scheme 18 is the method according to any one of implementation schemes 16 and 17, wherein preprocessing the segmented region background image data and the segmented region endpoint image data further includes subtracting the segmented region image data corresponding to one or more of the multiple segmented regions in the endpoint image from the segmented region image data corresponding to one or more of the multiple segmented regions in the distorted background image to form processed image data. Implementation scheme 19 is the method according to implementation scheme 18, wherein a first machine learning model is used on the processed image data to generate one or more classification results corresponding to one or more of the segmented regions in the processed image. Implementation scheme 20 is the method according to implementation scheme 19, wherein the one or more classification results of the first machine learning model include positive class classification and non-positive class classification.
[0157] Implementation scheme 21 is the method according to implementation scheme 20, further comprising performing maximum absolute value rescaling to the range [-1, 1] on the processed image data before generating the one or more classification results. Implementation scheme 22 is the method according to implementation scheme 21, further comprising performing quantile-based normalization on the rescaled processed image data before classification. Implementation scheme 23 is the method according to any one of implementation schemes 16 to 22, wherein a second machine learning model is used on the distorted background image data and the endpoint image data to generate one or more classification results corresponding to one or more segmented regions in the distorted background image. Implementation scheme 24 is the method according to implementation scheme 23, wherein the one or more classification results of the second machine learning model include class rejection and class non-rejection classifications. Implementation scheme 25 is the method according to implementation scheme 24, further comprising rescaling the distorted background image data and the endpoint image data to the range [0, 1] before generating the one or more classification results. Implementation 26 is the method according to implementation 25, further comprising performing minimum-maximum normalization on the distorted background image data and the endpoint image data before generating the one or more classification results. Implementation 27 is the method according to any one of implementations 19 to 26, wherein one or more of the first or second machine learning model includes a convolutional neural network. Implementation 28 is the method according to any one of implementations 1 to 27, wherein the image data is associated with a first fluorescence channel of a dPCR device, the dPCR device including a plurality of fluorescence channels comprising the first fluorescence channel, the plurality of fluorescence channels being associated with different spectral wavelengths. Implementation 29 is the method according to implementation 28, wherein one of the plurality of fluorescence channels includes a channel using a 6-carboxy-X-rhodamine (ROX)-based dye. Implementation 30 is the method according to any one of implementations 1 to 29, wherein the container includes a microfluidic array plate. Implementation 31 is the method according to any one of implementations 1 to 30, further comprising generating a synthetic image of an array comprising a plurality of segmented regions. Implementation scheme 32 is the method according to implementation scheme 31, wherein the synthetic image is generated based on the one or more classification results.
[0158] Implementation 33 is a non-transitory computer-readable medium including a memory storing one or more instructions, which, when executed by one or more processors of at least one computing device, perform computerized image processing for digital polymerase chain reaction (dPCR) analysis of a biological sample to automatically reject data from individual compartments so as not to be used in calculating dPCR-based target concentration results. The processing includes: obtaining image data representing a plurality of compartments disposed in a container, the image data including background image data corresponding to a background image captured prior to an amplification cycle of the dPCR assay of the biological sample, and endpoint image data corresponding to an endpoint image captured after an endpoint cycle of the dPCR assay of the biological sample; extracting individual compartment image data from the obtained image data, including compartment background image data and compartment endpoint image data; preprocessing the compartment background image data and the compartment endpoint image data to obtain processed image data for input into one or more machine learning models; generating one or more classification results for one compartment of the plurality of compartments using the one or more machine learning models; and using the one or more classification results to determine whether to reject image data corresponding to the compartment in the plurality of compartments so as not to be used in calculating dPCR-based target concentration results. The target concentration results.
[0159] Implementation 34 is a system for computerized image processing of digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual compartments so as not to be used in calculating dPCR-based target concentration results. The system includes: one or more processors of at least one computing device; and a memory storing one or more instructions, which, when executed by the one or more processors, cause the one or more processors to perform processing, the processing including: obtaining image data representing a plurality of compartments disposed in a container, the image data including background image data corresponding to a background image captured prior to an amplification cycle of the dPCR assay of the biological sample, and background image data corresponding to the dPCR assay of the biological sample. The process involves: capturing endpoint image data corresponding to the endpoint image after the endpoint cycle; extracting individual segmentation image data from the acquired image data, including segmentation background image data and segmentation endpoint image data; preprocessing the segmentation background image data and the segmentation endpoint image data to obtain processed image data for input into one or more machine learning models; generating one or more classification results for one of the multiple segmentation regions using the one or more machine learning models; and using the one or more classification results to determine whether to reject image data corresponding to the segmentation region in the multiple segmentation regions so that it is not used to calculate dPCR-based target concentration results.
[0160] Implementation 35 is a method for training a machine learning model to perform computerized image processing on digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separatory regions so that they are not used to calculate dPCR-based target concentration results. The method is performed by one or more computing devices and includes: obtaining a first plurality of images identified as images of positive separatory regions; obtaining a second plurality of images identified as images of non-positive separatory regions; generating one or more datasets using the first plurality of images and the second plurality of images; and having the one or more computing devices train the machine learning model using at least one of the one or more datasets to determine a set of parameters for the machine learning model, wherein the trained machine learning model is configured based on the set of parameters to classify individual separatory regions as non-positive separatory regions based on one or more quality control issues, the one or more quality control issues including one or more of the following: bubbles, gas, debris, stratification, leakage, filling problems, saturation, bridging, and dust.
[0161] Embodiment 36 is the method according to Embodiment 35, wherein the positive separator image is associated with a quasi-positive separator, wherein the quasi-positive separator corresponds to the presence of one or more target concentrations in the positive separator. Embodiment 37 is the method according to any one of Embodiments 35 to 36, wherein the quasi-non-positive separator image includes at least one negative separator image associated with a negative separator, wherein the negative separator corresponds to the absence of a target concentration in the negative separator. Embodiment 38 is the method according to any one of Embodiments 35 to 37, wherein the quasi-non-positive separator image includes at least one image associated with a defective microcavity in the separator. Embodiment 39 is the method according to any one of Embodiments 35 to 38, wherein the quasi-non-positive separator image includes at least one image associated with a contaminated separator. Embodiment 40 is the method according to any one of Embodiments 35 to 39, wherein the quasi-non-positive separator image includes at least one image associated with defect filling of a microcavity in the separator. Embodiment 41 is the method according to any one of Embodiments 35 to 40, wherein the quasi-non-positive separator image includes at least one defective image of the separator. Implementation scheme 42 is a method according to any one of embodiments 35 to 41, wherein the one or more datasets include a training dataset, a validation dataset, and a test dataset. Implementation scheme 43 is a method according to any one of embodiments 35 to 42, wherein determining the set of parameters of the machine learning model includes: iteratively training the machine learning model using the one or more datasets; and determining the set of parameters of the machine learning model based on the results of the iterative training. Implementation scheme 44 is a method according to any one of embodiments 35 to 43, wherein the machine learning model includes a convolutional neural network (CNN).
[0162] Implementation 45 is a method for training a machine learning model to perform computerized image processing on digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separatory regions so that they are not used to calculate dPCR-based target concentration results. The method is performed by one or more computing devices and includes: obtaining a first plurality of images identified as images of class-rejection separatory regions; obtaining a second plurality of images identified as images of class-non-rejection separatory regions; generating one or more datasets using the first plurality of images and the second plurality of images; and having the one or more computing devices train the machine learning model using at least one of the one or more datasets to determine a set of parameters for the machine learning model, wherein the trained machine learning model is configured based on the set of parameters to reject individual separatory regions based on one or more quality control issues, the one or more quality control issues including one or more of the following: bubbles, gas, debris, stratification, leakage, filling problems, saturation, bridging, and dust.
[0163] Implementation scheme 46 is the method according to implementation scheme 45, wherein the one or more datasets include a training dataset, a validation dataset, and a test dataset. Implementation scheme 47 is the method according to any one of implementation schemes 45 and 46, wherein determining the set of parameters of the machine learning model includes: iteratively training the machine learning model using the one or more datasets; and determining the set of parameters of the machine learning model based on the results of the iterative training. Implementation scheme 48 is the method according to any one of implementation schemes 45 to 47, wherein the machine learning model includes a convolutional neural network. Implementation scheme 49 is the method according to any one of implementation schemes 45 to 48, further comprising: obtaining a pre-dPCR image representing the plurality of segmented regions prior to amplification of one or more targets in the biological sample; obtaining a post-dPCR image representing the plurality of segmented regions after amplification of the one or more targets in the biological sample; and performing image subtraction using the pre-dPCR image and the post-dPCR image to obtain the plurality of segmented region images provided to the trained machine learning model.
[0164] Implementation 50 is a non-transitory computer-readable medium comprising a memory storing one or more instructions, which, when executed by one or more processors of at least one computing device, perform processing to train a machine learning model for computerized image processing of digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual segmentation regions so as not to be used in calculating dPCR-based target concentration results. The processing includes: obtaining a first plurality of images identified as images of class-rejection segmentation regions; obtaining a second plurality of images identified as images of class-non-rejection segmentation regions; generating one or more datasets using the first plurality of images and the second plurality of images; and having the one or more computing devices train the machine learning model using at least one of the one or more datasets to determine a set of parameters for the machine learning model, wherein the trained machine learning model is configured based on the set of parameters to reject individual segmentation regions based on one or more quality control issues, the one or more quality control issues including one or more of the following: bubbles, gas, debris, stratification, leakage, filling problems, saturation, bridging, and dust.
[0165] Implementation 51 is a system for computerized image processing of digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separators so that they are not used to calculate dPCR-based target concentration results. The system includes: one or more processors of at least one computing device; and a memory storing one or more instructions, which, when executed by the one or more processors, cause the one or more processors to perform processing, the processing including: obtaining a first plurality of images identified as images of class-rejection separators; obtaining a second plurality of images identified as images of class-non-rejection separators; generating one or more datasets using the first plurality of images and the second plurality of images; and the one or more computing devices training the machine learning model using at least one of the one or more datasets to determine a set of parameters for the machine learning model, wherein the trained machine learning model is configured based on the set of parameters to reject individual separators based on one or more quality control issues, the one or more quality control issues including one or more of the following: bubbles, gas, debris, stratification, leakage, filling problems, saturation, bridging, and dust.
[0166] The foregoing description should be understood as illustrative and exemplary in every respect, not restrictive, and the scope of the invention disclosed herein is not determined by the description, but by the claims interpreted according to the full scope permitted by patent law. It should be understood that the embodiments shown and described herein are merely illustrative of the principles of the invention, and various modifications can be made by those skilled in the art without departing from the scope and spirit of the invention. Various other combinations of features can be implemented by those skilled in the art without departing from the scope and spirit of the invention.
Claims
1. A method for computerized image processing of digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separatory regions so as not to be used in calculating dPCR-based target concentration results, the method comprising: Image data representing multiple compartments set in a container is obtained, the image data including background image data corresponding to a background image captured before the amplification cycle of the dPCR assay of the biological sample and endpoint image data corresponding to an endpoint image captured after the endpoint cycle of the dPCR assay of the biological sample. Extract the image data of each segmentation region from the obtained image data, including the background image data of the segmentation region and the endpoint image data of the segmentation region; The background image data of the segmentation zone and the endpoint image data of the segmentation zone are preprocessed to obtain processed image data, which is then input into one or more machine learning models. Use the one or more machine learning models to generate one or more classification results for the partitions in the plurality of partitions; as well as The one or more classification results are used to determine whether to reject image data corresponding to the segment in the plurality of segmentation regions so that it is not used to calculate the dPCR-based target concentration result.
2. The method according to claim 1, wherein the one or more classification results include positive class classification and non-positive class classification.
3. The method according to any one of claims 1 and 2, wherein the one or more classification results further include class rejection classification and class non-rejection classification.
4. The method of claim 3, wherein if the segmentation region is classified as class non-positive and class rejection, the image data corresponding to the segmentation region is rejected.
5. The method according to any one of claims 1 to 4, wherein extracting the image data of each of the plurality of segmented regions further comprises determining one or more corner positions of the plurality of segmented regions.
6. The method of claim 5, wherein extracting the image data of each segment of the plurality of segmented regions further comprises determining the location of each segment in the segmented region array.
7. The method according to any one of claims 1 to 6, wherein extracting the image data of each segment of the plurality of segmented regions further comprises determining a summary of microchamber sites from the image data of each segment of the plurality of segmented regions.
8. The method of claim 7, further comprising: A quality control process is performed on the image data.
9. The method of claim 8, wherein the quality control process comprises: Image registration errors are flagged before generating the one or more classification results if one or more corner locations, individual compartment locations, or microchamber sites cannot be determined.
10. The method of claim 8, further comprising: Perform blank image error checks before checking for image registration errors.
11. The method of claim 10, wherein the blank image error checking comprises: Determine if there are any positional errors in the image; In response to any positional error in the image, the blank score of the image is compared with a predetermined threshold; as well as In response to the blank score being greater than the predetermined threshold, the image is designated as having a blank image error or a defocus error.
12. The method according to any one of claims 10 and 11, further comprising: In response to the blank image error or out-of-focus error, the channel for acquiring the image is closed.
13. The method according to any one of claims 8 to 12, wherein the quality control process further comprises: One or more warnings are generated in response to one or more non-fault quality control errors.
14. The method of claim 13, wherein one or more non-faulty quality control errors include saturation errors, bridging errors, positive distribution errors, excessive quality control rejection errors, overload array errors, or any combination thereof.
15. The method of claim 14, wherein the positive distribution is determined by evaluating the uniformity of the positive distribution across two or more regions of the chip or sample plate array.
16. The method according to any one of claims 1 to 15, wherein preprocessing the background image data of the segmented region and the endpoint image data of the segmented region comprises: The background image corresponding to the background image data is distorted using the endpoint image corresponding to the endpoint image data, so that the image data of each segmentation corresponding to one or more of the plurality of segmentation segments are aligned to form a distorted background image.
17. The method of claim 16, wherein the distortion is performed using a homography transformation.
18. The method according to any one of claims 16 and 17, wherein preprocessing the background image data of the segmented region and the endpoint image data of the segmented region further comprises: The processed image data is obtained by subtracting the segmented image data corresponding to one or more of the multiple segmented regions in the endpoint image from the segmented image data corresponding to one or more of the multiple segmented regions in the distorted background image.
19. The method of claim 18, wherein the processed image data is processed using a first machine learning model to generate one or more classification results corresponding to one or more of the segmented regions in the processed image.
20. The method of claim 19, wherein the one or more classification results of the first machine learning model include positive class classification and non-positive class classification.
21. The method of claim 20, further comprising: Before generating the one or more classification results, the processed image data is rescaled to the range [-1, 1] using maximum absolute value.
22. The method of claim 21, further comprising: Quantile-based normalization is performed on the rescaled image data prior to classification.
23. The method of any one of claims 16 to 22, wherein a second machine learning model is used on the distorted background image data and the endpoint image data to generate one or more classification results corresponding to one or more of the segmented regions in the distorted background image.
24. The method of claim 23, wherein the one or more classification results of the second machine learning model include class rejection classification and class non-rejection classification.
25. The method of claim 24, further comprising: Before generating the one or more classification results, the distorted background image data and the endpoint image data are rescaled to the range [0, 1].
26. The method of claim 25, further comprising: Before generating the one or more classification results, minimum-maximum normalization is performed on the distorted background image data and the endpoint image data.
27. The method according to any one of claims 19 to 26, wherein one or more of the first machine learning model or the second machine learning model comprises a convolutional neural network.
28. The method of any one of claims 1 to 27, wherein the image data is associated with a first fluorescence channel of a dPCR device, the dPCR device comprising a plurality of fluorescence channels, the plurality of fluorescence channels including the first fluorescence channel, the plurality of fluorescence channels being associated with different spectral wavelengths.
29. The method of claim 28, wherein one of the plurality of fluorescence channels comprises a channel using a 6-carboxy-X-rhodamine (ROX)-based dye.
30. The method according to any one of claims 1 to 29, wherein the container comprises a microfluidic array plate.
31. The method according to any one of claims 1 to 30, further comprising: A composite image of an array is generated, the array comprising multiple segmented regions.
32. The method of claim 31, wherein the synthetic image is generated based on the one or more classification results.
33. A non-transitory computer-readable medium comprising a memory storing one or more instructions, which, when executed by one or more processors of at least one computing device, perform computerized image processing for digital polymerase chain reaction (dPCR) analysis of a biological sample to automatically reject data from individual separatory regions so as not to be used in calculating dPCR-based target concentration results, comprising: Image data representing multiple compartments set in a container is obtained, the image data including background image data corresponding to a background image captured before the amplification cycle of the dPCR assay of the biological sample and endpoint image data corresponding to an endpoint image captured after the endpoint cycle of the dPCR assay of the biological sample. Extract the image data of each segmentation region from the obtained image data, including the background image data of the segmentation region and the endpoint image data of the segmentation region; The background image data of the segmentation zone and the endpoint image data of the segmentation zone are preprocessed to obtain processed image data, which is then input into one or more machine learning models. Use the one or more machine learning models to generate one or more classification results for the partitions in the plurality of partitions; as well as The one or more classification results are used to determine whether to reject image data corresponding to the segment in the plurality of segmentation regions so that it is not used to calculate the dPCR-based target concentration result.
34. A system for computerized image processing of digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separatory regions so that they are not used to calculate dPCR-based target concentration results, said system comprising: One or more processors of at least one computing device; as well as A memory storing one or more instructions, which, when executed by the one or more processors, cause the one or more processors to perform processing, the processing including: Image data representing multiple compartments set in a container is obtained, the image data including background image data corresponding to a background image captured before the amplification cycle of the dPCR assay of the biological sample and endpoint image data corresponding to an endpoint image captured after the endpoint cycle of the dPCR assay of the biological sample. Extract the image data of each segmentation region from the obtained image data, including the background image data of the segmentation region and the endpoint image data of the segmentation region; The background image data of the segmentation zone and the endpoint image data of the segmentation zone are preprocessed to obtain processed image data, which is then input into one or more machine learning models. Use the one or more machine learning models to generate one or more classification results for the segmentation regions among the plurality of segmentation regions; and The one or more classification results are used to determine whether to reject image data corresponding to the segment in the plurality of segmentation regions so that it is not used to calculate the dPCR-based target concentration result.
35. A method for training a machine learning model to perform computerized image processing on digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separatory regions so as not to be used in calculating dPCR-based target concentration results, said method being performed by one or more computing devices and comprising: Obtain the first plurality of images identified as positive segmentation regions; Obtain a second set of images identified as non-positive segmentation regions; One or more datasets are generated using the first plurality of images and the second plurality of images; as well as The one or more computing devices determine a set of parameters for the machine learning model by training the model using at least one of the one or more datasets. The machine learning model, trained based on the set of parameters, classifies each segmentation zone into non-positive segmentation zones based on one or more quality control issues, including one or more of the following: bubbles, gas, debris, stratification, leakage, filling problems, saturation, bridging, and dust.
36. The method of claim 35, wherein the positive segmentation image is associated with a positive-like segmentation region, wherein the positive-like segmentation region corresponds to the presence of one or more target concentrations in the positive segmentation region.
37. The method of any one of claims 35 to 36, wherein the non-positive separator image includes at least one negative separator image associated with a negative separator, wherein the negative separator corresponds to the absence of target concentration in the negative separator.
38. The method according to any one of claims 35 to 37, wherein the image of the non-positive partition region includes at least one image associated with a defective microcavity of the partition region.
39. The method according to any one of claims 35 to 38, wherein the image of the non-positive separation region includes at least one image associated with the contaminated separation region.
40. The method according to any one of claims 35 to 39, wherein the image of the non-positive partition region includes at least one image associated with defect filling of the microcavity of the partition region.
41. The method according to any one of claims 35 to 40, wherein the non-positive segmentation image includes at least one defective image of the segmentation.
42. The method according to any one of claims 35 to 41, wherein the one or more datasets include a training dataset, a validation dataset, and a test dataset.
43. The method according to any one of claims 35 to 42, wherein determining the set of parameters of the machine learning model comprises: The machine learning model is trained iteratively using the one or more datasets; as well as The set of parameters of the machine learning model is determined based on the results of the iterative training.
44. The method according to any one of claims 35 to 43, wherein the machine learning model comprises a convolutional neural network (CNN).
45. A method for training a machine learning model to perform computerized image processing on digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separatory regions so as not to be used in calculating dPCR-based target concentration results, said method being performed by one or more computing devices and comprising: Obtain the first plurality of images identified as class rejection delimitation regions; Obtain a second set of images that are identified as non-rejection segmentation regions; One or more datasets are generated using the first plurality of images and the second plurality of images; as well as The one or more computing devices determine a set of parameters for the machine learning model by training the model using at least one of the one or more datasets. The machine learning model, trained based on the set of parameters, is used to reject individual partitions based on one or more quality control issues, including one or more of the following: bubbles, gas, debris, stratification, leakage, filling problems, saturation, bridging, and dust.
46. The method of claim 45, wherein the one or more datasets include a training dataset, a validation dataset, and a test dataset.
47. The method of any one of claims 45 and 46, wherein determining the set of parameters of the machine learning model comprises: The machine learning model is trained iteratively using the one or more datasets; as well as The set of parameters of the machine learning model is determined based on the results of the iterative training.
48. The method according to any one of claims 45 to 47, wherein the machine learning model comprises a convolutional neural network.
49. The method according to any one of claims 45 to 48, further comprising: Obtain pre-dPCR images of the plurality of compartmentalized regions representing one or more targets in the biological sample prior to amplification; Obtain post-dPCR images of the plurality of septa representing the amplification of one or more targets in the biological sample; Image subtraction is performed using the pre-dPCR image and the post-dPCR image to obtain the plurality of segmentation region images provided to the trained machine learning model.
50. A non-transitory computer-readable medium comprising a memory storing one or more instructions, which, when executed by one or more processors of at least one computing device, perform processing to train a machine learning model for computerized image processing of digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual partitions so as not to be used in calculating dPCR-based target concentration results, said processing comprising: Obtain the first plurality of images identified as class rejection delimitation regions; Obtain a second set of images that are identified as non-rejection segmentation regions; One or more datasets are generated using the first plurality of images and the second plurality of images; as well as The one or more computing devices determine a set of parameters for the machine learning model by training the model using at least one of the one or more datasets. The machine learning model, trained based on the set of parameters, is used to reject individual partitions based on one or more quality control issues, including one or more of the following: bubbles, gas, debris, stratification, leakage, filling problems, saturation, bridging, and dust.
51. A system for training a machine learning model for computerized image processing of digital polymerase chain reaction (dPCR) analysis of biological samples to automatically reject data from individual separatory regions so as not to be used in calculating dPCR-based target concentration results, the system comprising: One or more processors of at least one computing device; as well as A memory storing one or more instructions, which, when executed by the one or more processors, cause the one or more processors to perform processing, the processing including: Obtain the first plurality of images identified as class rejection delimitation regions; Obtain a second set of images that are identified as non-rejection segmentation regions; One or more datasets are generated using the first plurality of images and the second plurality of images; and The one or more computing devices determine a set of parameters for the machine learning model by training the model using at least one of the one or more datasets. The machine learning model, trained based on the set of parameters, is used to reject individual partitions based on one or more quality control issues, including one or more of the following: bubbles, gas, debris, stratification, leakage, filling problems, saturation, bridging, and dust.
