Efficient and robust high-speed neural network for cell image classification

Abstract

This paper describes an efficient and robust high-speed neural network for cell image classification. The neural network for cell image classification utilizes the difference between softmax scores to determine whether a cell is ambiguous. If a cell is ambiguous, it is classified as a pseudo-class; otherwise, it is classified into the class corresponding to the highest score. This neural network for cell image classification enables high speed, high accuracy, and high recall.

Description

[0001] Cross-references to related applications

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 116088, filed November 19, 2020, entitled “EFFICIENT AND ROBUST HIGH-SPEED NEURAL NETWORKS FOR CELL IMAGE CLASSIFICATION”, which is incorporated herein by reference in its entirety for all purposes. Technical Field

[0003] This invention relates to cell sorting. More specifically, this invention relates to image-based cell sorting. Background Technology

[0004] Cell image analysis plays an increasingly important role in biological and medical research, but existing methods cannot be used for real-time, high-speed applications such as cell sorting. Deep neural networks are often used to achieve high-accuracy classification; however, their complexity limits their use in real-time applications that require extremely high speeds. Summary of the Invention

[0005] This paper describes an efficient and robust high-speed neural network for cell image classification. The neural network for cell image classification utilizes the difference between softmax scores to determine whether a cell is definite. If a cell is definite, it is classified as a pseudo-class; otherwise, it is classified into the class corresponding to the highest score. This neural network for cell image classification enables high speed, high accuracy, and high recall.

[0006] Existing methods are based on pressurization and manual feature design. Deep learning-based methods (e.g., DeepFlow, ResNet) have also been proposed. None of these existing methods meet the speed requirements of typical applications, including cell sorting based on image activation.

[0007] In one aspect, a method includes processing cell images using a neural network, including generating multiple soft-max scores, each soft-max score corresponding to a class; comparing the highest soft-max score among the multiple soft-max scores with the second highest soft-max score to determine a difference; classifying the cell image as a pseudo-class when the difference is below a threshold; and classifying the cell image as the class corresponding to the highest soft-max score when the difference is equal to or above the threshold. The neural network includes two convolutional layers and one fully connected layer. The soft-max scores are between 0 and 1, inclusive. The threshold is set based on historical information to generate high precision and high recall. The threshold is determined by machine learning and artificial intelligence. The threshold is also manually determined by the user. The method also includes acquiring cell images.

[0008] On the other hand, an apparatus includes non-transitory memory for storing an application and a processor coupled to the memory. The application is used to: process cell images using a neural network, including generating multiple soft-max scores, each soft-max score corresponding to a class; comparing the highest soft-max score among the multiple soft-max scores with a second-highest soft-max score to determine a difference; classifying the cell image as a pseudo-class when the difference is below a threshold and classifying the cell image as the class corresponding to the highest soft-max score when the difference is equal to or above the threshold; the processor is configured to process the application. The neural network includes two convolutional layers and one fully connected layer. The soft-max score is between 0 and 1, including 0 and 1. The threshold is set based on historical information to generate high precision and high recall. The threshold is determined by machine learning and artificial intelligence. The threshold is also manually determined by the user. The apparatus is also used to acquire cell images.

[0009] On the other hand, a system includes a first device and a second device configured to acquire cell images. The second device is configured to: generate multiple soft-max scores, each soft-max score corresponding to a class; compare the highest soft-max score among the multiple soft-max scores with the second highest soft-max score to determine a difference; classify the cell image as a pseudo-class when the difference is below a threshold; and classify the cell image as the class corresponding to the highest soft-max score when the difference is equal to or above the threshold. The neural network includes two convolutional layers and one fully connected layer. The soft-max score is between 0 and 1, inclusive. The threshold is set based on historical information to generate high precision and high recall. The threshold is determined by machine learning and artificial intelligence. The threshold is manually determined by the user. Attached Figure Description

[0010] Figure 1The illustration shows a flowchart of a method for implementing a neural network for cell image classification according to some embodiments.

[0011] Figure 2 The diagram illustrates a neural network architecture according to some embodiments.

[0012] Figure 3 The illustration shows a flowchart of a process for trading precision for recall based on soft-max scores, according to some embodiments.

[0013] Figure 4 A diagram illustrating exemplary classifications according to some embodiments is shown.

[0014] Figure 5 A block diagram of an exemplary computing device according to some embodiments is shown, the computing device being configured to implement a neural network for cell image classification.

[0015] Figure 6 A schematic diagram illustrates the overall configuration of a biosample analyzer according to some embodiments. Detailed Implementation

[0016] A shallow double convolutional layer neural network architecture suitable for ultra-high-speed real-time cell image classification has been designed and established. Furthermore, this algorithm minimizes the common trade-off between precision and recall to achieve both high precision and high recall.

[0017] Neural networks for cell image classification address many unmet needs in cell image classification applications, primarily two key requirements: high accuracy and high speed. Existing classical machine learning-based methods (e.g., boosted, SVMs with manually designed features) and deep learning-based methods fail to meet speed requirements. The method described in this paper achieves both high speed and high accuracy thanks to a shallow neural network architecture designed using an architecture search technique and a novel precision-recall tradeoff module.

[0018] Neural networks capable of cell image classification can be used in image-based flow cytometry systems, offering high throughput and accuracy. They can also be applied to other cell image analysis systems, such as cell image organization and retrieval.

[0019] Figure 1 The diagram illustrates a flowchart of a method for implementing a neural network for cell image classification according to some embodiments. In step 100, a cell image is acquired. The cell image can be acquired in any manner, such as using a system including a microscope and a camera.

[0020] In step 102, a neural network is used to process the cell image. The neural network can be any type of neural network, such as one with two convolutional layers and one fully connected layer. The neural network generates soft-max scores, which are the cell classification scores. For example, a soft-max score is a number from 0 to 1, where for a class, a number closer to 1 means the cell is more likely to be in that class, and for a subclass, a number closer to 0 means the cell is less likely to be in that class. Multiple soft-max scores can be generated for a single cell (e.g., one score per class). For example, the neural network analyzes the cells and generates soft-max scores for all 5 classes. For example, when classifying white blood cells, there are 5 classes / categories (e.g., neutrophils, eosinophils, basophils, monocytes, lymphocytes). The scores from a single exemplary image for the different classes sum to 1. For example, [0.5, 0.2, 0.1, 0.1, 0.1] are the valid scores for the 5 classes. The softmax score is a standard computation (softmax function) used to determine the predicted probability of each class among multiple classes. Exemplary softmax functions include:

[0021]

[0022] Among them, l i,c It is the logit of each class c, for example, i.

[0023] In step 104, a precision-recall tradeoff is used to achieve high speed and high accuracy, classifying cells into classes or pseudo-classes. As described herein, cells are not classified solely based on the highest soft-max score, but rather by comparing two highest soft-max scores (e.g., the highest score minus the second highest score), and if the difference between the two highest soft-max scores is equal to or greater than a threshold, then the cell (image) is classified into the class with the highest soft-max score. If the difference is less than the threshold, then the cell (image) is classified into a pseudo-class. In some embodiments, the order of steps is modified. In some embodiments, fewer or additional steps are implemented.

[0024] Figure 2A diagram illustrating a neural network architecture according to some embodiments. The convolutional neural network receives cell image inputs. The convolutional neural network uses a rectified linear activation function (ReLU), which directly outputs the input if it is positive; otherwise, the output is zero. Then, max pooling is used to downsample 2x2 data to one value, which is the maximum value of the 2x2 representative picture. There is another layer of the convolutional neural network and ReLU. Then, global average pooling is implemented. Finally, soft-max scores are obtained. Soft-max is a normalized score similar to probability. Soft-max scores can be used as confidence level thresholds. Soft-max scores are numbers from 0 to 1, including 0 and 1. Soft-max scores are the standard output of any multi-class classifier.

[0025] Figure 3 A flowchart illustrating trading recall for precision based on soft-max scores according to some embodiments. In step 300, soft-max scores are generated and / or received. In step 302, it is determined whether the classification determined based on the soft-max scores is ambiguous. If S1 - S2 < t, then it is determined to be ambiguous. In the equation, S1 is the largest soft-max score, S2 is the second largest soft-max score, and t is a threshold determined by the test set and the target precision. For example, according to previous tests and the desired precision, the threshold can be determined to achieve the desired precision (e.g., a threshold of 0.5 determined from previous tests / historical data results in a precision of 90%, while a threshold of 0.6 results in a precision of 99%, which is the target in that example). In another example, the threshold can be determined and set by the user. In another example, learning and artificial intelligence can be used to adjust the threshold. If the class to which the cell belongs is ambiguous, then in step 304, the cell is classified as a pseudo-class. If the class to which the cell belongs is not ambiguous, then in step 306, the cell is classified as the class corresponding to the maximum score. In some embodiments, the order of the steps is modified. In some embodiments, fewer or additional steps are implemented.

[0026] Figure 4 A chart illustrating an exemplary classification according to some embodiments. In the first row under the title, the S1 - S2 difference of the exemplary data is 0.9, which is very high (higher than the threshold), so it is clear that the system has confidence in classifying the cell (e.g., in class 3). In the second row, the S1 - S2 difference of the exemplary data is 0.1, which is very low, so it is ambiguous and difficult to classify, and the cell is classified as a pseudo-class. By rejecting cells with low differences between soft-max scores, a precision of around 99% and a recall of 98.5% can be achieved.

[0027] Figure 5A block diagram of an exemplary computing device according to some embodiments is shown, configured to implement a neural network for cell image classification. The computing device 500 can be used to acquire, store, compute, process, transmit, and / or display information such as images and videos. The computing device 500 can implement any neural network for cell image classification. Generally, a suitable hardware architecture for implementing the computing device 500 includes a network interface 502, a memory 504, a processor 506, one or more I / O devices 508, a bus 510, and a storage device 512. The choice of processor is not critical, as long as a suitable processor with sufficient speed is selected. The memory 504 can be any conventional computer memory known in the art. The storage device 512 can include a hard disk drive, CD-ROM, CDRW, DVD, DVDRW, high-definition optical disc / drive, ultra-high-definition drive, flash memory card, or any other storage device. The computing device 500 can include one or more network interfaces 502. Examples of network interfaces include network interface cards (NICs) connected to Ethernet or other types of LANs. One or more I / O devices 508 may include one or more of the following: keyboard, mouse, monitor, screen, printer, modem, touchscreen, button interface, and other devices. One or more neural network applications 530 for implementing one or more neural networks for cell image classification are likely stored in storage device 512 and memory 504 and processed like typical applications. Figure 5 The components shown, more or less, can be included in the computing device 500. In some embodiments, neural network hardware 520 is included. Although Figure 5 The computing device 500 includes an application 530 and hardware 520 for a neural network for cell image classification, but the neural network for cell image classification can be implemented on the computing device in hardware, firmware, software, or any combination thereof. For example, in some embodiments, the neural network application 530 is programmed in memory and executed using a processor. In another example, in some embodiments, the neural network hardware 520 is programmable hardware logic that includes gates specifically designed to implement a neural network for cell image classification.

[0028] In some embodiments, one or more neural network applications 530 include several applications and / or modules. In some embodiments, a module further includes one or more sub-modules. In some embodiments, fewer or additional modules can be included.

[0029] Examples of suitable computing devices include microscopes, personal computers, laptops, computer workstations, servers, mainframes, handheld computers, personal digital assistants, cellular / mobile phones, smart devices, game consoles, digital cameras, digital camcorders, camera phones, smartphones, portable music players, tablets, mobile devices, video players, video disc burners / players (e.g., DVD burners or players, HD disc burners / players, UHD disc burners / players), televisions, home entertainment systems, augmented reality devices, virtual reality devices, smart jewelry (e.g., smartwatches), vehicles (e.g., autonomous vehicles), or any other suitable computing device.

[0030] Figure 6 A schematic diagram illustrates the overall configuration of a biosample analyzer according to some embodiments.

[0031] Figure 6 An example configuration of the biosample analyzer disclosed herein is shown. Figure 6 The biosample analyzer 6100 shown includes: a light irradiation unit 6101 that irradiates a biological sample S flowing in a flow channel C with light; a detection unit 6102 that detects the light generated by irradiating the biological sample S; and an information processing unit 6103 that processes information about the light detected by the detection unit. For example, the biosample analyzer 6100 is a flow cytometer or an imaging cytometer. The biosample analyzer 6100 may include a sorting unit 6104 that sorts specific biological particles P from the biological sample. For example, the biosample analyzer 6100 including the sorting unit is a cell sorter.

[0032] (Biological sample)

[0033] Biological sample S can be a liquid sample containing biological particles. For example, biological particles can be cellular or non-cellular biological particles. Cells can be living cells, and more specific examples include blood cells (such as red blood cells and white blood cells) and germ cells (such as sperm and fertilized eggs). Furthermore, cells can be cells collected directly from a sample (such as whole blood) or can be cultured cells obtained after culturing. Non-cellular biological particles are extracellular vesicles, or particularly, for example, exosomes and microvesicles. Biological particles can be labeled with one or more labeling substances, such as dyes (especially fluorescent dyes) and antibodies labeled with fluorescent dyes. Note that particles other than biological particles can be analyzed by the biosample analyzer of this disclosure, and beads, etc., can be analyzed for calibration, etc.

[0034] (Flow channel)

[0035] The flow channel C is designed to form a flow of biological sample S. Specifically, the flow channel C can be designed to form a flow in which the biological particles contained in the biological sample are substantially arranged in a single row. The flow channel structure including the flow channel C can be designed to form laminar flow. Specifically, the flow channel structure is designed to form laminar flow in which the flow of biological sample (sample flow) is surrounded by a flow of sheath fluid. The design of the flow channel structure can be suitably chosen by those skilled in the art, or a known design can be employed. The flow channel C can be formed in a flow channel structure (such as a microchip (a chip with micron-sized flow channels) or a flow cell). The width of the flow channel C is 1 mm or less, or specifically, it can be not less than 10 μm and not more than 1 mm. The flow channel C and the flow channel structure including the flow channel C can be made of materials such as plastic or glass.

[0036] The biosample analyzer of this disclosure is designed such that a biological sample flowing in a flow channel C, or specifically, biological particles in the biological sample, is irradiated with light from the light irradiation unit 6101. The biosample analyzer of this disclosure can be designed such that the irradiation point on the biological sample is located within the flow channel structure in which the flow channel C is formed, or it can be designed such that the irradiation point is located outside the flow channel structure. An example of the former could be a configuration in which light is emitted onto the flow channel C in a microchip or flow cell. In the latter case, for example, biological particles leaving the flow channel structure (particularly its nozzle portion) can be irradiated with light, and an air-jet flow cytometer can be used.

[0037] (Light Illumination Unit)

[0038] The light illumination unit 6101 includes a light source unit that emits light and a light-guiding optical system that guides the light to the illumination point. The light source unit includes one or more light sources. For example, the type of light source(s) is a laser light source or an LED. The wavelength of the light emitted from each light source can be any wavelength of ultraviolet, visible, and infrared light. For example, the light-guiding optical system includes optical components such as beam splitters, mirrors, or optical fibers. For example, the light-guiding optical system may also include a lens group for converging the light and includes an objective lens. One or more illumination points where the biological sample and the light intersect can be present. The light illumination unit 6101 can be designed to collect light emitted from one or different light sources to one illumination point.

[0039] (Detection unit)

[0040] The detection unit 6102 includes at least one photodetector that detects light generated by emitting light onto the biological particle. For example, the light to be detected can be fluorescence or scattered light (such as one or more of the following: forward-scattered light, back-scattered light, and side-scattered light). For example, each photodetector includes one or more light-receiving elements and has an array of light-receiving elements. Each photodetector may include one or more photomultiplier tubes (PMTs) and / or photodiodes (such as APDs and MPPCs) as light-receiving elements. For example, the photodetector includes a PMT array in which multiple PMTs are arranged in a one-dimensional direction. The detection unit 6102 may also include an image sensor, such as a CCD or CMOS. Using the image sensor, the detection unit 6102 can acquire images of the biological particle (e.g., bright-field images, dark-field images, or fluorescence images).

[0041] The detection unit 6102 includes a detection optical system that directs light of a predetermined detection wavelength to a corresponding photodetector. The detection optical system includes a spectral unit (such as a prism or diffraction grating) or a wavelength separation unit (such as a dichroic mirror or optical filter). The detection optical system is designed to disperse light generated by light irradiation onto, for example, biological particles, and to detect the dispersed light using a photodetector with a greater number of photodetectors than the number of fluorescent dyes used to label the biological particles. A flow cytometer including such a detection optical system is called a spectral flow cytometer. Furthermore, the detection optical system is designed to separate light corresponding to the fluorescence wavelength band of a specific fluorescent dye from the light generated by light irradiation onto, for example, biological particles, and to detect the separated light using a corresponding photodetector.

[0042] The detection unit 6102 may further include a signal processing unit that converts the electrical signal obtained by the photodetector into a digital signal. The signal processing unit may include an A / D converter as a device to perform the conversion. The digital signal obtained by the conversion performed by the signal processing unit can be transmitted to the information processing unit 6103. The digital signal can be processed by the information processing unit 6103 as light-related data (hereinafter also referred to as "optical data"). For example, the optical data may be optical data including fluorescence data. More specifically, the optical data may be light intensity data, and the light intensity may be light intensity data including fluorescence (the light intensity data may include characteristic quantities such as area, height, and width).

[0043] (Information Processing Unit)

[0044] For example, the information processing unit 6103 includes a processing unit that performs processing on various data (e.g., light data) and a storage unit that stores the various data. When the processing unit acquires light data corresponding to a fluorescent dye from the detection unit 6102, the processing unit can perform fluorescence leakage correction (compensation process) on the light intensity data. In the case of a spectroscopic flow cytometer, the processing unit also performs a fluorescence separation process on the light data and acquires light intensity data corresponding to the fluorescent dye. For example, the fluorescence separation process can be performed using the unmixing method disclosed in JP 2011-232259 A. If the detection unit 6102 includes an image sensor, the processing unit can acquire morphological information about the biological particles based on the image acquired by the image sensor. The storage unit can be designed to store the acquired light data. The storage unit can also be designed to further store spectral reference data used during the unmixing process.

[0045] In the case where the biosample analyzer 6100 includes the sorting unit 6104 described later, the information processing unit 6103 can determine whether to sort the biological particles based on optical data and / or morphological information. Then, the information processing unit 6103 controls the sorting unit 6104 based on the determined result, and the biological particles can be sorted by the sorting unit 6104.

[0046] The information processing unit 6103 can be designed to output various types of data (e.g., light data and images). For example, the information processing unit 6103 can output various types of data generated based on light data (e.g., two-dimensional graphs or spectral graphs). The information processing unit 6103 can also be designed to accept various types of data input and to accept user selection processes for graphs. The information processing unit 6103 may include output units (e.g., a display) or input units (e.g., a keyboard) for performing output or input.

[0047] The information processing unit 6103 can be designed as a general-purpose computer and as an information processing device, such as including a CPU, RAM, and ROM. The information processing unit 6103 can be included in a housing that includes a light irradiation unit 6101 and a detection unit 6102, or it can be located outside the housing. Furthermore, various processes or functions to be performed by the information processing unit 6103 can be implemented via a server computer or via a cloud connected to a network.

[0048] (Sorting Unit)

[0049] The sorting unit 6104 performs sorting of biological particles based on the determined result executed by the information processing unit 6103. The sorting method may involve generating droplets containing biological particles through vibration, applying a charge to the droplets to be sorted, and controlling the direction of travel of the droplets by electrodes. Alternatively, the sorting method may involve controlling the direction of travel of biological particles within a flow channel structure. For example, the flow channel structure may have a control mechanism based on pressure (injection or aspiration) or charge. An example of a flow channel structure may be a chip (e.g., the chip disclosed in JP 2020-76736A) having a flow channel structure in which the flow channel C branches downstream to a recovery flow channel and a waste flow channel, and specific biological particles are collected in the recovery flow channel.

[0050] To use a neural network for cell image classification as described in this paper, a device such as a microscope with a camera is used to acquire the content, and the device is capable of processing the acquired content. The neural network for cell image classification can be implemented with user assistance or automatically without user intervention.

[0051] In operation, the neural network used for cell image classification enables high accuracy, high speed, and high recall. High accuracy can reach approximately 99%, high speed includes classifying 2500 images per second, and high recall is approximately 99%. Other implementations do not match the aforementioned accuracy, speed, and / or recall.

[0052] In terms of accuracy in cytology, the accuracy threshold is typically around 99%, meaning the system should be correct 99% of the time to be considered accurate. In an exemplary implementation, the accuracy on a publicly available dataset is 99% and the recall is 98.5%. In terms of speed, the system processes 2500 images per second (with a latency of less than 0.4 ms), for example, on a Titan X GPU.

[0053] Several Implementations of High-Efficiency and Robust High-Speed ​​Neural Networks for Cell Image Classification

[0054] 1. A method comprising:

[0055] Cell images are processed using neural networks, including generating multiple soft-max scores, each soft-max score corresponding to a class.

[0056] Compare the highest soft-max score among the multiple soft-max scores with the second highest soft-max score to determine the difference;

[0057] Cell images are classified as pseudo-classes when the difference is below a threshold; and

[0058] Cell images are classified into classes corresponding to the highest soft-max score when the difference is equal to or higher than a threshold.

[0059] 2. The method as described in Clause 1, wherein the neural network comprises two convolutional layers and one fully connected layer.

[0060] 3. The method as described in Clause 1, wherein the soft-max score is between 0 and 1, inclusive.

[0061] 4. The method as described in Clause 1, wherein the threshold is set based on historical information to generate high precision and high recall.

[0062] 5. The method as described in Clause 1, wherein the threshold is determined by machine learning and artificial intelligence.

[0063] 6. The method as described in Clause 1, wherein the threshold is determined manually by the user.

[0064] 7. The method described in Clause 1 further includes acquiring cell images.

[0065] 8. An apparatus comprising:

[0066] Non-transitory memory for storing applications, which are used for:

[0067] Cell images are processed using neural networks, including generating multiple soft-max scores, each soft-max score corresponding to a class.

[0068] Compare the highest soft-max score among the multiple soft-max scores with the second highest soft-max score.

[0069] Soft-max scores are used to determine differences;

[0070] Cell images are classified as pseudo-classes when the difference is below a threshold; and

[0071] When the difference is equal to or greater than a threshold, the cell image is classified into the class corresponding to the highest soft-max score; and

[0072] A processor, coupled to memory, is configured to process the application.

[0073] 9. The apparatus as described in Clause 8, wherein the neural network comprises two convolutional layers and one fully connected layer.

[0074] 10. The apparatus as described in Clause 8, wherein the soft-max score is between 0 and 1, including 0 and 1.

[0075] 11. The apparatus as described in Clause 8, wherein the threshold is set based on historical information to generate high precision and high recall.

[0076] 12. The apparatus as described in Clause 8, wherein the threshold is determined by machine learning and artificial intelligence.

[0077] 13. The apparatus as described in Clause 8, wherein the threshold is determined manually by the user.

[0078] 14. The apparatus as described in Clause 8, wherein the apparatus is further configured to acquire cell images.

[0079] 15. A system comprising:

[0080] The first device is configured to acquire cell images; and

[0081] The second device is configured to:

[0082] Cell images are processed using neural networks, including generating multiple soft-max scores, each soft-max score corresponding to a class.

[0083] Compare the highest soft-max score among the multiple soft-max scores with the second highest soft-max score.

[0084] Soft-max scores are used to determine differences;

[0085] Cell images are classified as pseudo-classes when the difference is below a threshold; and

[0086] Cell images are classified into classes corresponding to the highest soft-max score when the difference is equal to or higher than a threshold.

[0087] 16. The system as described in Clause 15, wherein the neural network comprises two convolutional layers and one fully connected layer.

[0088] 17. The system as described in Clause 15, wherein the soft-max score is between 0 and 1, inclusive.

[0089] 18. The system as described in Clause 15, wherein the threshold is set based on historical information to generate high precision and high recall.

[0090] 19. The system as described in Clause 15, wherein the threshold is determined by machine learning and artificial intelligence.

[0091] 20. The system as described in Clause 15, wherein the threshold is determined manually by the user.

[0092] The present invention has been described with reference to specific embodiments in conjunction with details to facilitate an understanding of the construction and operating principles of the invention. Reference to specific embodiments and their details herein is not intended to limit the scope of the appended claims. It will be apparent to those skilled in the art that various other modifications may be made to the embodiments chosen for illustration without departing from the spirit and scope of the invention as defined by the claims.

Claims

1. A method comprising: Illuminating biological samples flowing in a flow channel with light; Detecting the light generated when a biological sample is irradiated; Cell images of biological samples are processed using neural networks, including generating multiple soft-max scores, each soft-max score corresponding to a class; Sort the multiple soft-max scores from highest to lowest; Compare the highest soft-max score among the multiple soft-max scores with the second highest soft-max score to determine the difference; Cell images are classified as pseudo-classes when the difference is below a threshold. as well as Cell images are classified into classes corresponding to the highest soft-max score when the difference is equal to or higher than a threshold.

2. The method of claim 1, wherein the neural network comprises two convolutional layers and one fully connected layer.

3. The method of claim 1, wherein the soft-max score is between 0 and 1, including 0 and 1.

4. The method of claim 1, wherein the threshold is set based on historical information to generate high precision and high recall.

5. The method of claim 1, wherein the threshold is determined by machine learning and artificial intelligence.

6. The method of claim 1, wherein the threshold is determined manually by the user.

7. The method of claim 1, further comprising acquiring cell images.

8. An apparatus comprising: The light irradiation unit is configured to irradiate a biological sample flowing in the flow channel with light; The detection unit is configured to detect light generated by irradiating a biological sample; Non-transitory memory for storing applications, which are used for: Cell images of biological samples are processed using neural networks, including generating multiple soft-max scores, each soft-max score corresponding to a class; Sort the multiple soft-max scores from highest to lowest; Compare the highest soft-max score among the multiple soft-max scores with the second highest soft-max score to determine the difference; Cell images are classified as pseudo-classes when the difference is below a threshold. as well as When the difference is equal to or greater than a threshold, the cell image is classified into the class corresponding to the highest soft-max score; as well as A processor, coupled to memory, is configured to process the application.

9. The apparatus of claim 8, wherein the neural network comprises two convolutional layers and one fully connected layer.

10. The apparatus of claim 8, wherein the soft-max fraction is between 0 and 1, including 0 and 1.

11. The apparatus of claim 8, wherein the threshold is set based on historical information to generate high precision and high recall.

12. The apparatus of claim 8, wherein the threshold is determined by machine learning and artificial intelligence.

13. The apparatus of claim 8, wherein the threshold is determined manually by the user.

14. The apparatus of claim 8, wherein the apparatus is further configured to acquire cell images.

15. A system comprising: A first device, configured to acquire cell images, wherein the first device further includes: A light irradiation unit is configured to irradiate a biological sample flowing in a flow channel with light; and The detection unit is configured to detect light generated by irradiating a biological sample; and The second device is configured to: Cell images are processed using neural networks, including generating multiple soft-max scores, each soft-max score corresponding to a class. Sort the multiple soft-max scores from highest to lowest; Compare the highest soft-max score among the multiple soft-max scores with the second highest soft-max score to determine the difference; Cell images are classified as pseudo-classes when the difference is below a threshold; and Cell images are classified into classes corresponding to the highest soft-max score when the difference is equal to or higher than a threshold.

16. The system of claim 15, wherein the neural network comprises two convolutional layers and one fully connected layer.

17. The system of claim 15, wherein the soft-max score is between 0 and 1, including 0 and 1.

18. The system of claim 15, wherein the threshold is set based on historical information to generate high precision and high recall.

19. The system of claim 15, wherein the threshold is determined by machine learning and artificial intelligence.

20. The system of claim 15, wherein the threshold is determined manually by the user.

Images