Single cell identification for cell sorting
Features are extracted from cell images through neural network models to distinguish single cells from noise events, solving the problem that traditional FACS cannot distinguish noise events, achieving high accuracy and high discernment ability single-cell recognition, improving the accuracy of the experiment and Sorting purity.
Patent Information
- Application Number
- CN202380043131.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-11-30
- Filing Date
- 2023-11-12
- Publication Date
- 2025-01-07
AI Technical Summary
Traditional fluorescence activated cell sorting (FACS) cannot distinguish noisy events that affect experimental accuracy and sorting product purity, such as aggregates, fragments, edges, out of focus, separation bubbles and bubbles, resulting in experimental accuracy and sorting purity. Restricted.
Using neural network model to extract features from cell images, distinguish single-cell and noise events, and using the neural network architecture of convolutional layer and fully connected layer, it can classify more than 2,000 images per second, achieving high accuracy and high resolution capabilities. single-cell recognition.
It improves the accuracy and recall of single-cell recognition, can distinguish noise events more accurately, enhances the accuracy of experiments and the purity of sorted products, has high resolution ability and high accuracy, and is suitable for different downstream applications.
Smart Images
Figure CN119278447A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to cell sorting. More particularly, the present invention relates to image-based cell sorting. Background Art
[0002] Traditional fluorescence activated cell sorting (FACS) does not have image information. FACS uses the laser scattering signal of cells to detect noise events. FACS cannot distinguish certain noise events that affect the accuracy of the experiment and the purity of the sorted product. Summary of the invention
[0003] The single cell identification described in this article utilizes cell image information and uses a neural network model to extract cell features to finely distinguish noise events from single cells, allowing the user to choose which different types of noise events to exclude depending on the application requirements. The fast neural network model is able to extract richer and more specific cell features than hand-selected features, which enables the model to have higher accuracy and higher discrimination ability to identify single cells and distinguish noise events in real time. Real-time single cell identification using a neural network model represents a novel technology that has never been applied before. Compared with traditional FACS (fluorescence activated cell sorting), it has high discrimination ability and high accuracy. The usefulness of this technology lies in integration with any bright field (BF) model and fluorescence (FL) model to identify single cells for different downstream applications.
[0004] In one aspect, a method programmed in a non-transitory memory of a device includes receiving an input, processing the input using a neural network to identify noise events or single cells, and classifying the input as a noise event or a single cell. The input includes multiple cell images. The neural network includes two convolutional layers and a fully connected layer. The neural network includes multiple kernels to extract features with different levels of granularity from the input. The noise events are selected from a group including aggregates, fragments, doublets, edges, defocused, detached bubbles, and bubbles. Classifying the input includes annotating each image with a noise event or a single cell. The neural network is configured to classify more than 2000 images per second. The method also includes implementing a bright field model or a fluorescence model in parallel with classifying the input.
[0005] In another aspect, an apparatus includes: a non-transitory memory for storing an application for: receiving an input, processing the input using a neural network to identify a noise event or a single cell, and classifying the input as a noise event or a single cell; and a processor coupled to the memory, the processor being configured to process the application. The input includes a plurality of cell images. The neural network includes two convolutional layers and a fully connected layer. The neural network includes a plurality of kernels to extract features with different levels of granularity from the input. The noise events are selected from a group including aggregates, debris, doublets, edges, out of focus, isolated vesicles, and bubbles. Classifying the input includes annotating each image with a noise event or a single cell. The neural network is configured to classify more than 2,000 images per second. The application is also configured to implement a bright field model or a fluorescence model in parallel with classifying the input.
[0006] On the other hand, a system includes a first device and a second device, the first device being configured to acquire a cell image, and the second device being configured to: receive a cell image, process the cell image using a neural network to identify noise events or single cells, and classify the cell image as a noise event or a single cell. The neural network includes two convolutional layers and one fully connected layer. The neural network includes multiple kernels to extract features with different levels of granularity from the input. The noise events are selected from a group including aggregates, debris, doublets, edges, out of focus, isolated bubbles, and bubbles. Classifying the cell image includes annotating the cell image with noise events or single cells. The neural network is configured to classify more than 2,000 images per second. The second device is also configured to implement a bright field model or a fluorescence model in parallel with classifying the input. BRIEF DESCRIPTION OF THE DRAWINGS
[0007] Figure 1 An image of an exemplary noise event is illustrated in accordance with some embodiments.
[0008] Figure 2 A diagram illustrating a neural network architecture configured to implement single cell recognition in accordance with some embodiments.
[0009] Figure 3 A flow chart of a method for implementing single cell identification for cell sorting according to some embodiments is illustrated.
[0010] Figure 4 Illustrated is a block diagram of an exemplary computing device configured to implement a single cell identification method in accordance with some embodiments. DETAILED DESCRIPTION
[0011] Samples from a flow cytometer will contain the desired single cells in suspension as well as noise events such as aggregates, doublets, and debris. In addition, some cells will be offset from the center, resulting in cells that are out of focus or off the edge of the interrogation area. Finally, some cells will show blebbing (attachment and detachment), which may indicate cell stress and damage. Traditional flow cytometry-based cell sorters have limited ability to exclude noise events from single cells in a sample because it does not have image information. Traditional cell sorting uses the laser scatter signal of the cell to exclude different noise cells (doublets, clumps, debris), but cannot distinguish other noise events (edges, out of focus, blebs, detached blebs). This affects the accuracy of the experiment and the purity of the sorted population. The single cell identification described in this article utilizes cell image information and uses a neural network model to extract cell features to finely distinguish noise events from single cells, allowing the user to choose which different types of noise events to exclude depending on the application requirements.
[0012] The fast neural network model described in this paper is able to extract richer and more specific cellular features than hand-selected features, which enables the model to have higher accuracy and higher discrimination ability to identify single cells and distinguish noise events in real time.
[0013] Real-time single cell identification using neural network models represents a novel technology that has never been applied before. It has high discrimination and accuracy compared to traditional FACS (fluorescence activated cell sorting). The usefulness of this technology lies in integration with any bright field (BF) model and fluorescence (FL) model to identify single cells for different downstream applications.
[0014] Figure 1 An image of an exemplary noise event according to some embodiments is illustrated. The noise event includes aggregates 100, doublets 102, debris 104, edges 106, out-of-focus 108, separation vacuoles 110, and blebs 112 that are distinct from single cells 114. Aggregates 100 include multiple cells, while doublets 102 include two cells. Debris 104 is an artifact of a non-single cell. The edges 106 of a cell may be missing, or the cell may be out-of-focus 108. Separation vacuoles 110 are separated blebs, and blebs 112 are attachment vacuoles or rounded protrusions on the cell surface.
[0015] Traditional FACS has no image information and can only exclude a subset of noise events present in each sample. FACS can distinguish single cells from aggregates, doublets, and debris because it uses the laser scattering signal of the cell to detect noise events. FACS cannot distinguish between other noise events (e.g., bubbles, separation bubbles, out of focus, and edges) and single cells. This affects the accuracy of the experiment and the purity of the sorted product. In other words, FACS can determine whether the image is binary information of a single cell; and in the single cell identification embodiment described in this article, there are different noise categories. Depending on the application, the user may want to allow images with specific noise categories. For example, if the user is performing rare event sorting, and the user wants to sort as many target cells as possible, images with bubble events, edge events, and slightly out of focus events can be included in the sorting because the user does not want to exclude potential useful information. Since images can be more finely classified based on noise events, further processing can also be limited or expanded based on noise events.
[0016] For example, Table 1 illustrates the differences between FACS and potential single cell identification implementations.
[0017]
[0018] Table 1
[0019] Figure 2 A diagram of a neural network architecture configured to implement single cell recognition according to some embodiments is illustrated. Neural network 200 includes a module with two convolutional layers and one fully connected layer. Neural network 200 includes multiple kernels to extract rich features with different levels of granularity. Neural network 200 is capable of processing more than 2000 images per second using a 2080Ti GPU. Neural network 200 is highly accurate: when 5C noise events are excluded on the publicly available WBC2020 dataset, the accuracy of single cell recognition is 98.6% and the recall rate is 97.3%. Neural network 200 can be integrated with a cell type classifier module. Although a specific exemplary neural network architecture is described, neural network 200 is not limited to a specific architecture. For example, neural network 200 can include fewer, additional or different components, such as fewer convolutional layers, multiple GPUs or different GPUs.
[0020] The neural network 200 receives input (e.g., many cell images), processes the input as described herein using machine learning to extract features, and generates outputs, such as labeling each image with one or more noise events or labeling as a single cell. The neural network can be trained using multiple data sets.
[0021] The neural network 200 can be implemented in BF applications by parallelizing with a BF cell type classifier by sharing a feature extractor. The neural network 200 can be implemented in FL applications by parallelizing with a FL cell type classifier on a GPU. The neural network 200 can also be used in conjunction with cell sorting applications such as nuclear translocation, immune flow fish, extracellular vesicles, etc.
[0022] Figure 3 A flow chart of a method for implementing single cell identification for cell sorting according to some embodiments is illustrated.In step 300, input is received at a device.The input can include cell images and / or other content.
[0023] In step 302, the input is processed using a neural network. The neural network includes two convolutional layers and one fully connected layer, where multiple kernels extract features with different levels of granularity from the input. Processing the input includes using machine learning to distinguish noise and identify noise events. For example, the neural network learns to distinguish factors such as detecting more than two circular objects to determine aggregates; detecting points and lines that are not part of a cell to determine debris; detecting two circular objects to determine doublets; detecting mostly circular objects with missing partial circles to determine edges; detecting blurred or distorted lines to determine out-of-focus; detecting separation circles that are smaller than cells to determine separation blebs; and detecting attachment circles that are smaller than cells to determine blebs. The neural network is also able to learn that if no noise events are detected, then single cells are determined.
[0024] In step 304, the input is classified. For example, the input is classified as having noise events (e.g., aggregates, debris, doublets, edges, out of focus, isolated bubbles, blebs) or single cells. For example, each image of the input is annotated as a noise event or a single cell. In some embodiments, if multiple noise events are detected, the image is classified and / or annotated as having two noise events. In some embodiments, if multiple noise events are detected, the image is classified and / or annotated as the more prominent noise event as determined by the machine language.
[0025] The method was able to be implemented such that over 2000 images per second could be classified using a 2080Ti GPU, with single-cell identification accuracy >98.6% and recall >97.3% when filtering out 5C noise events.
[0026] In some embodiments, fewer steps or additional steps are implemented. For example, a method for implementing single cell identification can be integrated with a BF application or a FL application. In another example, a neural network is trained using multiple data sets before processing additional inputs. In some embodiments, the order of the steps is modified.
[0027] Figure 4The block diagram of an exemplary computing device configured to implement a single cell identification method according to some embodiments is illustrated. The computing device 400 can be used to obtain, store, calculate, process, transmit and / or display information such as images and videos including 3D content. The computing device 400 can implement any encoding / decoding aspects. Generally speaking, the hardware structure suitable for implementing the computing device 400 includes a network interface 402, a memory 404, a processor 406, (one or more) I / O devices 408, a bus 410 and a storage device 412. The selection of the processor is not critical, as long as a suitable processor with sufficient speed is selected. The processor 406 can include one or more central processing units (CPUs) and / or one or more graphics processing units (GPUs), such as 2080Ti GPUs. The memory 404 can be any conventional computer memory known in the art. The storage device 412 can include a hard disk drive, CDROM, CDRW, DVD, DVDRW, high-definition disk / drive, ultra-HD drive, flash memory card or any other storage device. The computing device 400 can include one or more network interfaces 402. Examples of network interfaces include network cards connected to Ethernet or other types of LANs. (One or more) HO devices 408 can include one or more of the following: keyboard, mouse, monitor, screen, printer, modem, touch screen, button interface, and other devices. (One or more) single cell identification applications 430 for implementing single cell identification embodiments may be stored in storage device 412 and memory 404 and processed as applications are normally processed. Figure 4 More or fewer of the components shown in can be included in computing device 400. In some embodiments, single cell identification hardware 420 is included. Figure 4 The computing device 400 in the embodiment includes an application 430 and hardware 420 for a single cell identification implementation, but the single cell identification method can be implemented on the computing device using hardware, firmware, software, or any combination thereof. For example, in some embodiments, the single cell identification application 430 is programmed in a memory and executed using a processor. In another example, in some embodiments, the single cell identification hardware 420 is programmed hardware logic including gates specifically designed to implement the single cell identification method.
[0028] In some embodiments, (one or more) single cell identification applications 430 include multiple applications and / or modules. In some embodiments, a module also includes one or more sub-modules. In some embodiments, fewer or additional modules can be included.
[0029] Examples of suitable computing devices include a personal computer, a laptop computer, a computer workstation, a server, a mainframe computer, a handheld computer, a personal digital assistant, a cellular / mobile phone, a smart appliance, a game console, a digital camera, a digital camcorder, a camera phone, a smart phone, a portable music player, a tablet computer, a mobile device, a video player, a video disc writer / player (e.g., a DVD writer / player, a high-definition disc writer / player, an ultra-high-definition disc writer / player), a television, a home entertainment system, an augmented reality device, a virtual reality device, smart jewelry (e.g., a smart watch), a vehicle (e.g., a self-driving vehicle), or any other suitable computing device.
[0030] In some embodiments, a first computing device (eg, a microscope / camera configuration) captures cell images and transmits the images to a second computing device to implement a single cell identification method to process and classify the images.
[0031] To utilize the single cell identification method, the device acquires or receives an image of cells for sorting. The single cell identification method can be implemented with user assistance or automatically without user involvement.
[0032] In operation, the single cell identification described herein can distinguish single cells from any number (e.g., 7) types of noise events, allowing the user to select which different types of noise events to exclude depending on the application requirements. Single cell identification can be implemented as a single cell identification module based on a neural network.
[0033] The neural network based module extracts better features and achieves better performance than hand-selected features for single-cell recognition.
[0034] The high-speed single-cell identification module can be integrated with any BF or FL downstream application.
[0035] Some Examples of Single Cell Identification for Cell Sorting
[0036] 1. A method of programming in a non-transitory memory of a device, comprising:
[0037] Receive input;
[0038] Process inputs using neural networks to identify noise events or single cells; and
[0039] Classify inputs as noise events or single cells.
[0040] 2. A method as described in clause 1, wherein the input includes multiple cell images.
[0041] 3. A method as described in clause 1, wherein the neural network includes two convolutional layers and one fully connected layer.
[0042] 4. A method as described in clause 1, wherein the neural network includes multiple kernels to extract features with different levels of granularity from the input.
[0043] 5. The method of clause 1, wherein the noise events are selected from the group consisting of aggregates, fragments, doublets, edges, out of focus, separate bubbles, and bubbles.
[0044] 6. The method of clause 1, wherein classifying the input comprises annotating each image with a noise event or a single cell.
[0045] 7. The method of clause 1, wherein the neural network is configured to classify more than 2,000 images per second.
[0046] 8. The method of clause 1 further comprising implementing a brightfield model or a fluorescence model in parallel with classifying the input.
[0047] 9. An apparatus comprising:
[0048] Non-transitory memory for storing applications for:
[0049] Receive input;
[0050] Processing the input using a neural network to identify noise events or single cells; and classifying the input as a noise event or a single cell; and
[0051] A processor coupled to the memory, the processor configured to process the application.
[0052] 10. An apparatus as described in clause 9, wherein the input includes a plurality of cell images.
[0053] 11. An apparatus as described in clause 9, wherein the neural network includes two convolutional layers and one fully connected layer.
[0054] 12. An apparatus as described in clause 9, wherein the neural network includes multiple kernels to extract features with different levels of granularity from the input.
[0055] 13. The apparatus of clause 9, wherein the noise events are selected from the group consisting of aggregates, fragments, doublets, edges, out of focus, separate bubbles, and bubbles.
[0056] 14. The apparatus of clause 9, wherein classifying the input comprises annotating each image with a noise event or a single cell.
[0057] 15. The apparatus of clause 9, wherein the neural network is configured to classify more than 2,000 images per second.
[0058] 16. The apparatus of clause 9, wherein the application is further configured to implement a brightfield model or a fluorescence model in parallel with classifying the input.
[0059] 17. A system comprising:
[0060] A first device configured to acquire a cell image; and
[0061] The second device is configured as:
[0062] receiving cell images;
[0063] Processing cell images using neural networks to identify noise events or single cells; and
[0064] Classify cell images into noise events or single cells.
[0065] 18. The system of clause 17, wherein the neural network comprises two convolutional layers and one fully connected layer.
[0066] 19. A system as described in clause 17, wherein the neural network includes multiple kernels to extract features with different levels of granularity from the input.
[0067] 20. The system of clause 17, wherein the noise events are selected from the group consisting of aggregates, fragments, doublets, edges, out of focus, separate bubbles, and bubbles.
[0068] 21. The system of clause 17, wherein classifying the cell images comprises annotating the cell images with noise events or single cells.
[0069] 22. The system of clause 17, wherein the neural network is configured to classify more than 2,000 images per second.
[0070] 23. The system of clause 17, wherein the second device is further configured to implement a brightfield model or a fluorescence model in parallel with classifying the input.
[0071] The present invention has been described according to specific embodiments in combination with details to facilitate understanding of the principles of construction and operation of the present invention. Such reference to specific embodiments and details thereof herein is not intended to limit the scope of the appended claims. It will be readily apparent to those skilled in the art that other various modifications may be made in the embodiments selected for illustration without departing from the spirit and scope of the present invention as defined by the claims.
Claims
1. A method of programming in a non-transitory memory of a device, comprising: Receive input; Process inputs using neural networks to identify noise events or single cells; as well as Classify inputs as noise events or single cells.
2. The method of claim 1, wherein the input comprises a plurality of cell images.
3. The method of claim 1, wherein the neural network comprises two convolutional layers and one fully connected layer.
4. The method of claim 1, wherein the neural network comprises a plurality of kernels to extract features with different levels of granularity from the input.
5. The method of claim 1, wherein the noise events are selected from the group consisting of aggregates, fragments, doublets, edges, out of focus, separate bubbles, and bubbles.
6. The method of claim 1, wherein classifying the input comprises annotating each image with a noise event or a single cell.
7. The method of claim 1, wherein the neural network is configured to classify more than 2,000 images per second.
8. The method of claim 1, further comprising implementing a brightfield model or a fluorescence model in parallel with classifying the input.
9. An apparatus comprising: Non-transitory memory for storing applications for: Receive input; Process inputs using neural networks to identify noise events or single cells; as well as Classify inputs as noise events or single cells; as well as A processor coupled to the memory, the processor configured to process the application.
10. The apparatus of claim 9, wherein the input comprises a plurality of cell images.
11. The apparatus of claim 9, wherein the neural network comprises two convolutional layers and one fully connected layer.
12. The apparatus of claim 9, wherein the neural network comprises a plurality of kernels to extract features with different levels of granularity from the input.
13. The apparatus of claim 9, wherein the noise events are selected from the group consisting of aggregates, debris, doublets, edges, out of focus, separate bubbles, and bubbles.
14. The apparatus of claim 9, wherein classifying the input comprises annotating each image with a noise event or a single cell.
15. The apparatus of claim 9, wherein the neural network is configured to classify more than 2,000 images per second.
16. The apparatus of claim 9, wherein the application is further configured to implement a brightfield model or a fluorescence model in parallel with classifying the input.
17. A system comprising: a first device configured to acquire a cell image; as well as The second device is configured as: receiving cell images; Process cell images using neural networks to identify noise events or single cells; as well as Classify cell images into noise events or single cells.
18. The system of claim 17, wherein the neural network comprises two convolutional layers and one fully connected layer.
19. The system of claim 17, wherein the neural network comprises a plurality of kernels to extract features with different levels of granularity from the input.
20. The system of claim 17, wherein the noise events are selected from the group consisting of aggregates, debris, doublets, edges, out of focus, separate bubbles, and bubbles.
21. The system of claim 17, wherein classifying the cell images comprises annotating the cell images with noise events or single cells.
22. The system of claim 17, wherein the neural network is configured to classify more than 2,000 images per second.
23. The system of claim 17, wherein the second device is further configured to implement a brightfield model or a fluorescence model in parallel with classifying the input.