Cell classification model training method and apparatus, cell classification method and apparatus

Abstract

The present application relates to a kind of cell classification model training method and device, cell classification method and device.The cell classification model training method includes: obtaining the focused state single-cell quantitative phase microscopy image of target cell;The focused state single-cell quantitative phase microscopy image is handled to obtain corresponding cell feature data, and constitutes cell feature data set;Cell feature data set is screened for predictive performance, and the optimal feature data set is obtained;With the optimal feature data set as sample data, with the kind of target cell as label, initial model is trained to obtain cell classification model.The method and device of the present application are high in classification accuracy of multiple cells, and are suitable for being applied to large-scale, multi-class, high-heterogeneity cell group Label-free detection classification.

Description

Technical Field

[0001] This invention belongs to the field of cell characterization and classification technology, specifically relating to a cell classification model training method and apparatus, and a cell classification method and apparatus. Background Technology

[0002] Single-cell research not only advances our understanding of fundamental biological processes such as cell differentiation, individual development, and disease progression, but also offers new possibilities for drug development and precision medicine. Imaging flow cytometry (IFC) captures single-cell images under laser illumination of the detection area, enabling quantitative analysis of cell morphology, structure, and various biomarkers. However, IFC is highly dependent on fluorescent labeling, which can lead to problems such as cytotoxicity, non-specific binding, autofluorescence interference, and complex sample pretreatment. Furthermore, this label-dependent analytical approach is ineffective for cell populations lacking prior knowledge or with frequently changing surface antigens (such as CTC cells).

[0003] In recent years, research teams have integrated different quantitative phase microscopy (QPM) architectures with inductively coupled plasma (IFC) for imaging fluidized cells and for cell characterization and classification. However, compared with platforms based on staining, fluorescence, and other labeling detection, the imaging accuracy and classification precision of current QPM-IFC still lag behind, especially when analyzing large-scale, highly heterogeneous single-cell populations. This is primarily because the rapid flow of microfluidics can cause cell images to become blurred or their shapes to be stretched and distorted. Under non-ideal microfluidic focusing, optical defocusing of suspended cells is also unavoidable. Ultimately, low resolution leads to severe loss of subcellular texture, resulting in the inability to achieve accurate cell characterization and classification. Summary of the Invention

[0004] To overcome the problems existing in the prior art, the present invention provides a cell classification model training method, comprising:

[0005] Acquire focused single-cell quantitative phase microscopy images of target cells;

[0006] The focused single-cell quantitative phase microscopy image is processed to obtain the corresponding cell feature data, which constitutes a cell feature dataset.

[0007] The optimal feature dataset is obtained by performing prediction performance screening on the cell feature dataset.

[0008] Using the optimal feature dataset as sample data and the type of target cell as label, the initial model is trained to obtain the cell classification model.

[0009] Another aspect of the present invention provides a cell classification model training device, comprising:

[0010] The image acquisition module is used to acquire focused single-cell quantitative phase microscopy images of the target cells.

[0011] The image processing module is used to process the focused single-cell quantitative phase microscopy image to obtain corresponding cell feature data, which constitutes a cell feature dataset.

[0012] The feature filtering module is used to filter the cell feature dataset based on prediction performance to obtain the optimal feature dataset.

[0013] The model training module is used to train the initial model to obtain the cell classification model using the optimal feature dataset as sample data and the type of target cell as the label.

[0014] In another aspect, the present invention provides a cell classification method, comprising:

[0015] Acquire focused single-cell quantitative phase microscopy images of cells to be classified;

[0016] The focused single-cell quantitative phase microscopy images of the cells to be classified are processed to obtain the corresponding cell feature data;

[0017] Based on the cell feature data, the cells to be classified are classified using a cell classification model, which is trained according to the cell classification model training method.

[0018] Another aspect of the present invention provides a cell sorting device, comprising:

[0019] The image acquisition subsystem is used to acquire focused single-cell quantitative phase microscopic images of cells to be classified.

[0020] An image processing subsystem is used to process the focused single-cell quantitative phase microscopic image of the cells to be classified to obtain corresponding cell feature data;

[0021] A cell classification subsystem is used to classify the cells to be classified based on the cell feature data using a cell classification model. The cell classification subsystem also includes a cell classification model training device for training and obtaining the cell classification model.

[0022] The beneficial effects of this invention are as follows:

[0023] 1. Using focused single-cell quantitative phase microscopy images as initial data for subsequent cell classification model training not only avoids motion blur and image distortion during cell flow, but also eliminates the need for digital refocusing, greatly shortening the reconstruction time of single-cell quantitative phase microscopy images and saving computational resources.

[0024] 2. By adjusting the flow rate ratio of sheath fluid to sample flow in the microfluidic platform, cells of different properties and sizes can be precisely focused on the detection area in both the vertical and horizontal directions. By aligning the recording focal plane of the digital holographic microscope with the three-dimensional hydrodynamic focusing plane of the sample flow, a three-dimensional hydrodynamically focused single-cell holographic image can be directly obtained, thus yielding a focused quantitative phase microscopic image of a single cell.

[0025] 3. By comprehensively considering the correlation and redundancy of cell features, the predictive performance of the cell feature dataset is screened to obtain the optimal feature dataset. Model training based on this optimal dataset can shorten training time and achieve high accuracy in the resulting classification model. Specifically, a mutual information-based feature selection algorithm and an incremental feature selection algorithm are used to comprehensively evaluate the importance of each feature in the cell feature dataset for cell classification and the redundancy between cell features, thereby achieving predictive performance screening.

[0026] 4. Theoretical simulation and experimental testing confirm that the classification method and device of the present invention achieve 100% accuracy in classifying various cell types, and are suitable for label-free detection and classification of large-scale, multi-category, and highly heterogeneous cell populations.

[0027] 5. The method and apparatus of this invention have a wide range of applications, suitable for any cell or particle of cell size. They can be used for the classification of normal cells, such as red blood cell counting and morphological studies, as well as for the identification and classification of diseased cells (such as tumor cells). The method and apparatus of this invention can be used for the detection of cells in vivo or in vitro in humans or mammals, as well as for the detection of plant, microbial, and algal cells. For example, they can be used for the detection and classification of algal cells in environmental monitoring and ecological research, the detection and classification of plankton communities in aquatic ecosystems, and pollen counting and classification in ecology and environmental science.

[0028] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0029] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same parts in the exemplary embodiments of the invention.

[0030] Figure 1This is a flowchart of a cell classification model training method according to an embodiment of the present invention.

[0031] Figure 2 This is a top view of a microfluidic chip for a cell classification model training method according to an embodiment of the present invention.

[0032] Figure 3 A schematic diagram of a digital holographic microscope structure is shown, illustrating a cell classification model training method according to an embodiment of the present invention.

[0033] Figure 4 (a) to Figure 4 (f) illustrates the theoretical simulation and experimental verification results of three-dimensional fluid dynamics focusing for a cell classification model training method according to an embodiment of the present invention. Wherein, Figure 4 (a) shows the vertically focused computational fluid dynamics (CFD) simulation results (yz plane), where the sample injection rate is 1 μL / min, R v The vertical sheath flow velocity is represented by the ratio of the sample flow velocity to the vertical sheath flow velocity, and the sample flow height (H) is shown as a black streamline. s ), with a scale bar of 30μm. Figure 4 (b) Shows the results of the vertical focusing experiment (yz plane), with conditions consistent with the CFD simulation. The red band represents the sample flow containing the fluorescent dye rhodamine, and the dashed lines represent the upper and lower boundaries of the sample channel. The scale bar is 30 μm. H s The average value is the result of 5 independent repeated trials. Figure 4 (c) Shows the vertical focus height H s With the velocity ratio R of the vertical sheath flow to the sample flow v The curve of H change s (sim) represents CFD simulation data, H s (exp) is the average of 5 independent repeated trials, and the dashed line represents the error. Figure 4 (d) shows the horizontal focusing CFD simulation results (xy plane), where the sample injection rate is 1 μL / min, R v Fixed at 0.5, R h The velocity ratio of the horizontal sheath flow to the sample flow is represented by a black streamline, indicating the sample flow width (W). s ), with a scale bar of 200μm. Figure 4 (e) shows the results of the horizontal focusing experiment (xy plane), with conditions consistent with the CFD simulation. The red band represents the sample flow containing the fluorescent dye rhodamine, the dashed line represents the channel boundary, and the scale bar is 100 μm. s The average value is the result of 5 independent repeated trials. Figure 4 (f) shows the horizontal focus width W s With the velocity ratio R of the horizontal sheath flow to the sample flow h The curve of W's change s(sim) is CFD simulation data, W s (exp) is the average value of 5 independent repeated tests, and the dashed line represents the error.

[0034] Figure 5 (a) to Figure 5 (h) show the digital holographic microscopy imaging results of microspheres in the three-dimensional hydrodynamic focusing state of a cell classification model training method according to an embodiment of the present invention. Figure 5 (a) and Figure 5 (b) respectively show the flow conditions of 10 μm and 20 μm monodisperse polystyrene microspheres recorded in the white light mode in the microfluidic chip under three conditions: no focusing, only vertical focusing (two-dimensional focusing) and three-dimensional hydrodynamic focusing, with a scale bar of 50 μm. Figure 5 (c) and Figure 5 (d) are respectively the reconstructed amplitude (the upper figure in the figure) and phase (the lower figure in the figure) of the digital holographic microscope for 10 μm and 20 μm microspheres in the three-dimensional hydrodynamic focusing state, with a scale bar of 10 μm. Figure 5 (e) and Figure 5 (f) are respectively the reconstructed particle sizes of amplitude and phase of 10 μm microspheres under different focusing states (N=50). Figure 5 (g) and Figure 5 (h) are respectively the reconstructed particle sizes of amplitude and phase of 20 μm microspheres under different focusing states (N=50). Wherein *** indicates p-value<0.001, ** indicates 0.001<p-value<0.01, which is given based on Student's t-Test analysis.

[0035] Figure 6 (a) to Figure 6 (c) show the training results of different types of cancer cell lines of the cell classification model training method according to the embodiment of the present invention. Figure 6 (a) shows the importance scores of 81 features obtained by the mutual information-based MRMR algorithm for classification, sorted from high to low. Figure 6 (b) shows the incremental feature selection curve of each machine learning model, in which the gray shaded part is the feature MRMR score value, the blue curve is the accuracy of the ten-fold cross-validation set, the red curve is the accuracy of the test set, and the accuracy marked in each figure is the accuracy of each model trained by the optimal feature data set on the test set. Figure 6 (c) shows the prediction performance of each machine learning model for different genera of cancer cells, and the indicators include recall rate, precision, F1-score, accuracy and ROC AUC value.

[0036] Figure 7 (a) to Figure 7(d) shows the quantitative phase microscopy image feature analysis of breast cancer multi-subtype cells of a cell classification model training method according to an embodiment of the present invention. Figure 7 (a) shows single-cell quantitative phase microscopy images of 8 breast cancer-related cell lines, including 2 LA-type cell lines (MCF-7 and T47D), 2 LB-type cells (BT474 and MDA-MB-361), 1 Her2+ type cell (SK-BR-3), 2 TN-type cells (HCC1937 and MDA-MB-231) and 1 normal mammary cell (MCF-10A); the color bar represents the scale of the optical path difference (normalized value), and the phase information is represented by the optical path difference. Figure 7 (b) shows violin plots (N=1000) of the numerical distribution of 6 features of 8 breast cancer-related cell lines, including Volume (Bulk), Dry Weight Density (Bulk), Kurtosis (1st Order), Entropy (GLCM), Gray-level Asymmetry (GGCM) and Run-length Difference (GLRLM). *** indicates p-value<0.001, ** indicates 0.001<p-value<0.01, * indicates 0.01<p-value<0.05, and ns indicates p-value>0.05, which are given based on One-way ANOVA analysis. Figure 7 (c) shows the mean heat map and difference significance of 81 quantitative phase microscopy image features of breast cancer-related cell lines. The color bar represents the scale of the mean value (normalized value) of each feature of different cell lines (N=1000). The significance level (α) is 0.05, which is given based on the One-way ANOVA post-hoc multiple comparison test (Tukey test method). Figure 7 (d) shows the heat map of the Pearson correlation coefficient matrix of 81 quantitative phase microscopy image features of breast cancer-related cell lines, where blue indicates positive correlation, red indicates negative correlation, the magnitude of the correlation is represented by the grid size, and yellow boxes are used to divide each feature subset.

[0037] Figure 8 (a) to Figure 8 (c) shows the training classification results of normal mammary cell lines and breast cancer multi-subtype cell lines of a cell classification method according to an embodiment of the present invention. Figure 8 (a) shows the importance ranking of 81 features for classification obtained by the mutual information-based MRMR algorithm. Figure 8 (b) shows the incremental feature selection curve of each model. The gray shaded area is the MRMR score value of the features, the blue curve is the accuracy of the ten-fold cross-validation set, the red curve is the accuracy of the test set, and the accuracy marked in each figure is the accuracy of each model trained with the optimal feature set on the test set. Figure 8(c) shows the predictive performance of each model for different types of cancer cells, with metrics including recall, precision, F1 score, accuracy and ROC AUC value.

[0038] Figure 9 (a) shows a quantitative phase micrograph of cells in blood according to an embodiment of the present invention. Figure 9 (b) and Figure 9 (c) shows the confusion matrix of the LDA model obtained by training the cell classification model according to an embodiment of the present invention. Detailed Implementation

[0039] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0040] This invention provides a method for training a cell classification model, comprising the following steps.

[0041] Step 1: Obtain focused single-cell quantitative phase microscopy images of the target cells.

[0042] In studies combining quantitative phase microscopy (IFC) and imaging flow cytometry (IFC), achieving high-precision IFC imaging of single cells under rapid flow conditions is a major challenge. Because high-speed suspended cells are naturally optically defocused, acquired images are prone to motion blur or elongation distortion. Traditional IFCs can use sheath flow technology to focus cells, but when integrated with IFC, they are bulky, complex to operate, and require additional digital refocusing or image processing algorithms (denoising, etc.) to compensate for the reconstruction accuracy of IFC. However, this is computationally expensive and time-consuming, and does not match the rapid analysis workflow of IFC.

[0043] This invention first acquires a focused single-cell quantitative phase microscopic image of the target cell. This focused single-cell quantitative phase microscopic image refers to a three-dimensional hydrodynamically focused image of a single cell as a single-cell image-plane holographic quantitative phase microscopic image. The characteristic of image-plane holography is that the object or its image is directly located on the holographic recording plane, eliminating the need for optical diffraction propagation operations in a computer during reconstruction.

[0044] In this embodiment, a three-dimensional hydrodynamic focused single-cell image plane holographic image is acquired by a microfluidic chip and a digital holographic microscope. The three-dimensional hydrodynamic focused single-cell image plane holographic image is then processed to obtain a focused single-cell quantitative phase microscopic image. The microfluidic chip is configured to provide a three-dimensional hydrodynamic focused sample flow containing the target cell to the digital holographic microscope, and the digital holographic microscope is configured to record that its focal plane coincides with the focal plane of the sample flow.

[0045] Among them, the three-dimensional fluid dynamics focused sample flow refers to a sample flow that achieves fluid dynamics focusing in both the vertical (longitudinal) and horizontal (lateral) directions; the three-dimensional fluid dynamics focused single-cell image plane holographic image refers to the image plane holographic image of a single cell in the three-dimensional fluid dynamics focused sample flow. By aligning the focal plane of the three-dimensional fluid dynamics focused sample flow with the recording focal plane of the digital holographic microscope, the image plane holographic image can be directly acquired.

[0046] Specifically, such as Figure 2 As shown, the microfluidic chip is provided with a sample channel 1, a first sheath flow channel 2, and a second sheath flow channel 3 arranged along the horizontal plane. The first sheath flow channel 2 and the second sheath flow channel 3 are symmetrically arranged on both sides of the sample channel 1 and are connected to the sample channel through a confluence port 4.

[0047] Two first sheath flows are injected into the sample channel vertically from above and below at symmetrical positions before and after the sample flow injection point 5, respectively, with a second flow velocity. The sample flow is injected vertically into the sample channel with a first flow velocity. Two second sheath flows are injected into the sample channel 1 through the first sheath flow channel 2 and the second sheath flow channel 3 with a third flow velocity. The ratio of the second flow velocity to the first flow velocity is used as the first flow velocity ratio, and the ratio of the third flow velocity to the sum of the first and second flow velocities is used as the second flow velocity ratio. By adjusting the first flow velocity ratio and / or the second flow velocity ratio, the sample flow flows in a straight line within the sample channel along the recording focal plane of the digital holographic microscope. That is, by adjusting the first flow velocity ratio and the second flow velocity ratio, longitudinal and lateral focusing of the sample flow can be achieved, i.e., three-dimensional hydrodynamic focusing. The first and second sheath flows are of the same type and do not mix with the sample flow.

[0048] Both the first and second sheath flow channels are polygonal in shape, and at the confluence, the angles between the first and second sheath flow channels and the sample channel are equal. The first and second sheath flow channels can also be designed in other suitable forms, such as streamlined shapes.

[0049] This embodiment uses a microfluidic chip based on three-dimensional hydrodynamic focusing to achieve strict and precise single-cell manipulation, and builds an image plane pre-magnification off-axis digital holographic microscope based on a Mach-Zehnder interferometer structure to capture the complex wavefront of single-cell samples.

[0050] Specifically, a microinjection pump is used as the power source to continuously and symmetrically compress the sample flow longitudinally through a first sheath flow at the same velocity, thus focusing the sample flow longitudinally. After the three immiscible laminar flows advance along the channel to the confluence, they are symmetrically sandwiched by two second sheath flows that converge at equal angles, thereby achieving lateral focusing of the sample flow within the sample channel. Cells will then flow in an orderly arrangement along the central axis of the sample channel. The sheath flow's encapsulation restricts motion blurring caused by cell floating and stacking, and also reduces the risk of sample contamination and channel blockage due to contact between the sample flow and the channel wall.

[0051] In this embodiment, at the confluence, the angle between the first sheath flow channel and the second sheath flow channel and the sample channel is 30° to 60° (preferably 45°), and the first flow velocity ratio R... v Set the value to 0–1.5 (preferably 0.25), and set the second flow rate ratio R... h The value is set to 0–5 (preferably 2). By changing the flow rate of the sample stream, the changes in the width and height of the focused sample stream are measured. When the flow rate of the sample stream changes abruptly from 1 to 20 μL / min (1.46 m / s), the focused width and height only fluctuate within a small range, proving that the hydrodynamic focusing state remains stable at high speeds.

[0052] Subsequently, the three-dimensional fluid dynamics-focused sample flow was introduced into the digital holographic microscope, the structure of which is as follows: Figure 2 As shown in the figure. The off-axis digital holographic microscope with pre-magnified image plane built in this embodiment, based on a Mach-Zehnder interferometer structure, uses a single-frequency solid-state laser emitting a 632.8nm wavelength laser. The laser is collimated by a set of optical path calibration mirrors and then incident along the center of the optical path. It is then filtered by a spatial light filter composed of an aperture and a collimating lens to produce a uniform parallel spot. After passing through a polarization beam splitter, it is split into object light and reference light. The object light path is equipped with a 40x objective lens with NA 0.75 and a tube lens. After illuminating the sample, the sample is magnified by the microscope objective. The reference light path is equipped with a microscope objective with the same parameters as the object light path to balance surface errors as much as possible and reduce the computational load of subsequent digital compensation processing in the computer. In addition, a filter is added to balance the light intensity attenuation in the object light path caused by the sample. The last two beams of light interfere at a small angle on the imaging plane. The single-cell holographic image of the sample stream is automatically and continuously acquired by a fast, high-resolution camera (sCMOS) at a resolution of 496×496 at 30 fps, with a pixel size of 2.5 μm. Those skilled in the art may also choose other suitable existing digital holographic microscopes, as long as their recording focal plane is adjustable.

[0053] By precisely adjusting the recording focal plane of the digital holographic microscope to coincide with the focal plane of the three-dimensional hydrodynamic focused sample flow provided by the microfluidic chip, a three-dimensional hydrodynamic focused single-cell image plane holographic image can be directly obtained. It can quickly reconstruct an accurate quantitative phase microscopy image without digital refocusing, which usually accounts for about 60% of the time of quantitative phase microscopy image reconstruction and consumes a lot of computing resources.

[0054] The process of processing a three-dimensional hydrodynamic focused single-cell image plane holographic image to obtain a focused state single-cell quantitative phase microscopic image includes:

[0055] (1) Spectrum filtering: The spectrum of the single-cell holographic image and the background holographic image is obtained by Fourier transform. Then, according to the principle of holographic reconstruction, the original image (+1 level term) is filtered out separately and frequency shifted to the center of the spectrum.

[0056] (2) Numerical reconstruction: The amplitude of the object light wave can be obtained by transforming the spectrum to the spatial domain using the inverse Fourier transform, and then the enveloping phase of the object can be extracted by the arctangent operation.

[0057] (3) Phase processing: First, phase filtering is performed using the double exposure phase subtraction method (i.e., cell-wrapped phase - background-wrapped phase). Then, the least squares unwrapping algorithm is used to obtain the continuous phase of a single cell. Finally, the distortion is eliminated by quadratic surface fitting, thus obtaining a quantitative phase microscopic image of a single cell.

[0058] Fabrication and Testing of Microfluidic Chips

[0059] In this embodiment, the microfluidic chip can be fabricated according to the following process: First, the chip channel pattern is drawn using AutoCAD, the pattern is printed as a mask, and then the chip is fabricated following the standard soft lithography process. In this embodiment, each channel (including the sample channel, the first sheath flow channel, and the second sheath flow channel) on the microfluidic chip is 100 μm wide and 35 μm high. Both the first and second sheath flow channels are zigzag-shaped, and the angle between the first and second sheath flow channels and the sample channel at the confluence is 45°. The size of each channel can be adjusted according to the type of target cells or the actual application scenario, which is easily understood by those skilled in the art.

[0060] After obtaining the microfluidic chip, it can be simulated and tested to ensure its functional feasibility. For details, see [link to documentation]. Figure 4 (a) to Figure 4(f) In this embodiment, a simplified channel model is established in the finite element simulation software COMSOL Multiphysics. The three-dimensional fluid dynamics focusing process is simulated using a single-phase flow module to investigate the influence of different flow velocities on the focusing effect of the sample flow. Specifically, a three-dimensional model of the channel in the focusing region is performed, with the channel cross-section (xz) consistent with the actual chip size (100μm × 35μm). A laminar flow physics field is configured, assuming that the laminar flow is incompressible, immiscible, and that all sidewalls are free of slippage. The sample flow boundary is displayed using velocity streamlines. First, the velocity ratio (Ri) between the first sheath flow and the sample flow in the two longitudinal flows is set. v The value changes from 0 to 1.5, increasing by 0.25 in each increment. The height of the focused sample flow (H) in the longitudinal section (yz) of the measurement channel is also considered. s Simulations showed a decrease in focus from 35 μm to 6 μm, conforming to a first-order exponential decay relationship. Subsequently, due to the focusing process being "first longitudinal, then lateral," R was fixed. v The velocity ratio (R) is set to 0.5 between the two second sheath flows and the three laminar flows (front first sheath flow, sample flow, and rear first sheath flow). h The value changes from 0 to 5, increasing by 0.5 each time. The width of the focused sample stream (W) in the cross-section (xy) of the measurement channel is also considered. s Simulations show that the value decreases from 100μm to 6μm, which also conforms to the first-order exponential decay relationship.

[0061] Then, the sample flow was visualized using the fluorescent dye Rhodamine B to detect the three-dimensional hydrodynamic focusing effect within the actual microfluidic chip. A 1mL plastic syringe was used to draw up the sheath flow and sample flow, and the syringe was fixed to the injection pump. A silicone tube (0.4mm inner diameter, 0.9mm outer diameter) was connected to the syringe needle, and a 23G steel needle (0.3mm inner diameter, 0.6mm outer diameter) was connected to the other end of the tube. The steel needle was directly inserted into the pre-reserved inlet (0.6mm aperture) on the PDMS chip, thus connecting the syringe and the microfluidic chip. Three independent microinjection pumps were used to control the injection flow rate at each inlet, injecting the sheath flow and sample flow into the channel at different flow rates. The fluid focusing was observed under an inverted fluorescence microscope, and fluorescence images were captured using software. The results showed that the experimental measurements were basically consistent with the simulated values, proving that by adjusting the flow rate ratio, the three-dimensional focusing effect can be precisely and accurately controlled to adapt to the focusing needs of cells of different sizes.

[0062] <Reliability Testing of Focused Single-Cell Quantitative Phase Microscopy Images>

[0063] This embodiment uses polystyrene microsphere standards to test the accuracy of digital holographic microscopy in reconstructing particle amplitude and phase under different flow focusing states. First, the particle size of the microspheres is measured from the holographically reconstructed amplitude reconstruction image to verify the accuracy of amplitude reconstruction. Then, the particle size of the microspheres is measured from the holographically reconstructed quantitative phase microscopy image to verify the accuracy of phase reconstruction. This is based on the linear relationship between the quantitative phase microscopy image and the actual thickness of the object, as shown in equation (1) below:

[0064]

[0065] Where λ refers to the incident light wavelength, n1 refers to the refractive index of the sample, n0 refers to the refractive index of the surrounding medium, d refers to the actual thickness of the object, φ refers to the quantitative phase, and OPD refers to the optical path difference. Therefore, given the refractive indices of the medium and the object, the true height of the object can be calculated using this formula.

[0066] Specifically, 10μm and 20μm microspheres were diluted to approximately 10 μm using 1X PBS buffer. 6 A concentration of microspheres / mL was prepared, and 0.1% Tween 20 (Aladdin) was added to prevent polymerization. The prepared microsphere solution was used as the sample flow, and 1X PBS buffer containing 0.1% Tween 20 was used as the sheath flow. The sample flow rate was kept constant at 1 mL / mL using the syringe pump.

[0067] The flow rate was adjusted to μL / min, and the flow rate ratio of the sheath flow to the sample flow was adjusted to stabilize at the corresponding focusing state. The CMOS camera was set to continuous acquisition mode to capture and record holographic images of polystyrene microspheres. A quantitative phase microscopic image was reconstructed, and the microsphere particle size was calculated using the above relationship. The deviation was calculated by comparing it with the known particle size parameters to prove the reliability of the phase reconstruction results.

[0068] This embodiment demonstrates label-free digital holographic microscopy detection of cells focused by three-dimensional hydrodynamics. Based on standardized physical properties and a size consistent with various cell types to be analyzed, this embodiment uses 10μm and 20μm polystyrene microspheres to simulate cells, verifying the detection effect of this invention on microparticles focused by three-dimensional hydrodynamics. First, the holographic light path is blocked, and the microsphere focusing process is qualitatively observed under an external white light source. Figure 5 (a) Figure 5 (b)). The concentration of the microsphere suspension is ~10. 6 The flow rate was slowed to 1 μL / min. Without any focusing, the microspheres were stacked longitudinally and floated laterally through the channel, exhibiting varying degrees of defocusing. After applying a first sheath flow constraint in the vertical direction, the microspheres were compressed longitudinally to the same plane, but continued to float and sway laterally. Upon applying a second sheath flow constraint in the horizontal direction, the microspheres were focused onto the central axis of the channel. The focusing conditions for 10 μm microspheres were R... v=0.75(H) s =9.7μm), R h =3(W s =9.6μm), the focusing condition for 20μm microspheres is R v =0.125(H) s =21.3μm), R h =1.25(W) s =19.5μm). By removing the obstructions from the white light and the holographic optical path, a holographic interference image of a single microsphere can be acquired within a preset detection window. After frequency domain filtering and numerical reconstruction, the amplitude reconstruction image of the single microsphere is obtained. Figure 5 (c) Figure 5 (d)); then, through phase extraction, unwrapping, and distortion compensation, a quantitative phase microscopic image of the single microsphere is finally reconstructed. Figure 5 (c) Figure 5 (d)).

[0069] Quantitative comparison of the reconstruction effects of particle amplitude and quantitative phase microscopic images under different focusing states was conducted. First, the length and width of the microspheres in the amplitude reconstruction image were measured, and their average value was taken as the amplitude particle size value of the microspheres. Then, the difference between the highest and lowest phases of the microspheres in the quantitative phase microscopic image was extracted, and the particle size value of the microspheres in the quantitative phase microscopic image was calculated according to equation (1). Fifty microspheres from each of the different focusing states were used for statistical analysis. Figure 5 (e) to Figure 5 (h)) The results show that three-dimensional hydrodynamic focusing is a prerequisite for accurately reconstructing the amplitude and quantitative phase microscopy images of microspheres. Especially for quantitative phase microscopy images, the particle size values ​​of 10μm and 20μm microspheres in the unfocused state are 7.13μm and 16.5μm, respectively, which differ significantly from the true particle size of the microspheres; after three-dimensional hydrodynamic focusing, the particle size values ​​of the quantitative phase microscopy images are 10.4μm and 19.8μm, respectively. Figure 5 f) Figure 5 (h)) is basically consistent with the true value. Furthermore, three-dimensional fluid dynamics focusing further improves measurement consistency. Although the accuracy of measuring microsphere amplitude and particle size in quantitative phase microscopy images appears to have improved when only vertical focusing is performed, the dispersion between measured values ​​is high. Three-dimensional fluid dynamics focusing, through focusing in both the horizontal and vertical directions, strictly limits the trajectory of particles, ensuring the stability of the measurement results. The CV value (i.e., the ratio of standard deviation to mean) comprehensively reflects the accuracy and stability of the fluid focusing and measurement system in IFC. Calculations show that the CV values ​​for amplitude measurement of 10μm and 20μm microspheres in this embodiment are 3.3% and 1.9%, respectively, and the CV values ​​for phase measurement are 4.3% and 4.7%, respectively, which are very close to the 2-3% level of mature commercial flow cytometers.

[0070] Step 2: Process the focused single-cell quantitative phase microscopy image to obtain the corresponding cell feature data, thus forming a cell feature dataset.

[0071] After obtaining a focused single-cell quantitative phase microscopy image of the target cell, it is processed to obtain the corresponding cell feature data, specifically including:

[0072] Step S21: Preprocess the focused single-cell quantitative phase microscopy image, including filling small holes and breaks in the image and smoothing the boundaries.

[0073] Step S22: Convert the preprocessed single-cell quantitative phase microscopy image into a binary image.

[0074] Step S23: Use the largest connected component in the binary image as a mask to segment the cell ROI region from the binary image.

[0075] Step S24: Extract cell feature data based on the cell ROI region. Specifically, the cell feature data includes: 1. Morphological feature data: describing the geometric features of the ROI / cell, including volume, surface area, sphericity, etc.; the dry weight of the cell can also be calculated based on phase information; 2. First-order gray-level feature data (i.e., first-order histogram feature data): describing the statistical characteristics of different gray levels (voxel intensities) and probability distributions of the ROI / cell, but without considering the spatial interaction between them, only based on the global gray-level histogram, including maximum value, minimum value, mean, standard deviation, variance, skewness, kurtosis, entropy, energy, etc. These features reflect the symmetry, uniformity and local intensity distribution changes of the measured voxels; 3. Higher-order texture feature data: describing the spatial distribution relationship of gray values ​​of the ROI / cell, including gray-level co-occurrence matrix (GLCM), gray-level gradient co-occurrence matrix (GGCM), gray-level run length matrix (GLRLM), gray-level size region matrix (GLSZM), neighborhood gray-level difference matrix (NGTDM), gray-level dependence matrix (GLDM), etc. In this embodiment, the relevant features of the first three matrices are selected. Finally, 81 features were extracted for each target cell to construct cellular feature data for each target cell.

[0076] Based on the cellular characteristic data of each cell, a cellular characteristic dataset can be established. The cellular characteristic dataset established in this embodiment covers immune cells, tumor cells, erythrocytes, leukocytes, stem cells, germ cells, algal cells, microbial cells, and planktonic cells. Immune cells include, but are not limited to, T cells, B cells, natural killer cells (NK cells), macrophages, granulocytes, dendritic cells (DC cells), mast cells, monocytes, and platelets. Tumor cells include, but are not limited to, lung cancer cells, colorectal cancer cells, cervical cancer cells, breast cancer cells, human monocytic leukemia cells, prostate cancer cells, skin cancer cells, gastric cancer cells, liver cancer cells, cervical cancer cells, esophageal cancer cells, thyroid cancer cells, bladder cancer cells, lymphoma cells, pancreatic cancer cells, kidney cancer cells, uterine cancer cells, oral cancer cells, melanoma cells, ovarian cancer cells, brain cancer cells, multiple myeloma cells, nasopharyngeal cancer cells, gallbladder cancer cells, glioblastoma cells, and testicular cancer cells. In one embodiment, tumor cells include lung cancer cells A549, colorectal cancer cells HCT116, cervical cancer cells HeLa, normal breast cancer cells MCF-10A, breast cancer cells T47D, MCF-7, BT474, MDA-MB-361, SK-BR-3, HCC1937, MDA-MB-231, and human monocytic leukemia cells THP-1. White blood cells include, but are not limited to, neutrophils, eosinophils, and basophils. Stem cells include, but are not limited to, embryonic stem cells, adult stem cells, induced pluripotent stem cells, mesenchymal stem cells, hematopoietic stem cells, neural stem cells, epithelial stem cells, muscle stem cells, liver stem cells, skin stem cells, pancreatic islet stem cells, and corneal stem cells. Germ cells include, but are not limited to, sperm cells, spore cells, and pollen cells. Microbial cells include, but are not limited to, bacterial cells and fungal cells. Algal cells include, but are not limited to, green algae, diatoms, and cyanobacteria.

[0077] Step 3: Filter the cell feature dataset based on its predictive performance to obtain the optimal feature dataset.

[0078] When there are too many input features, the classification model has a high risk of overfitting and a high overall computational cost. Therefore, it is desirable to further reduce the feature dimensionality while ensuring high accuracy. For some classification models, higher-order features lead to higher accuracy; however, for others, the opposite is true. Therefore, the importance of features to a classification model cannot be generalized solely based on feature category.

[0079] Therefore, this embodiment performs prediction performance screening on the cell feature dataset to obtain the optimal feature dataset, specifically including:

[0080] The importance of all cell features in the cell feature dataset is ranked using a feature selection algorithm based on mutual information.

[0081] Based on the sorting, the cell features used for training are determined by an incremental feature selection algorithm, and a subset of the corresponding cell feature dataset is selected as the optimal feature dataset according to the determined cell features.

[0082] Specifically, this embodiment employs a mutual information-based MRMR (Max-Relevance and Min-Redundancy) algorithm to evaluate the classification performance of each cell feature individually, filtering out low-value features. Compared to some methods that process features discretely, the mutual information-based MRMR algorithm simultaneously weighs redundancy (between feature variables) and correlation (between feature variables and target variable), providing a ranking of feature importance. The ranking results can be found in [link to relevant documentation]. Figure 6 (a).

[0083] Furthermore, combining the incremental feature selection algorithm, based on the aforementioned ranking of feature importance, features are added one by one for testing in descending order of importance, resulting in the incremental feature selection (IFS) curve. Figure 6 (b) As shown in the figure, once the number of features reaches a certain level, adding more features will not improve classification accuracy and may even decrease it, indicating redundancy or overfitting. Furthermore, more features result in longer training time and slower prediction speed. Considering both accuracy and efficiency, the cell features used for training were ultimately determined, and a subset of the corresponding cell feature dataset was selected as the optimal feature dataset for subsequent training.

[0084] Step 4: Using the optimal feature dataset as sample data and the target cell type as label, train the initial model to obtain a cell classification model.

[0085] After determining the optimal feature dataset, the initial model is trained using this dataset as sample data and the target cell type as the label to obtain a cell classification model. The initial model can be a Linear Discriminant Analysis (LDA) model, a Support Vector Machine (SVM) model, a Neural Network (NN) model, a Logistic Regression (LR) model, a Nearest Neighbor (KNN) model, a Naive Bayes model, a tree model, etc. Specifically, the initial model can be a subspace KNN or a Bagging Tree model. The initial model can also be other suitable machine learning models.

[0086] In this embodiment, both the LDA and SVM models achieved an overall classification accuracy of 100% on the optimal feature dataset. Given that the LDA model requires fewer features, has shorter training time, and faster prediction speed, it was determined to be the optimal classification model. Furthermore, recall, precision, F1 score, accuracy, and ROC AUC values ​​can be used to evaluate the classification performance of each model for different cell types (see [link to relevant documentation]). Figure 6 (c)). Overall, after sufficient training, the prediction accuracy of each classification model for various cells is above 90%.

[0087] This invention also relates to a cell classification method, comprising:

[0088] Acquire focused single-cell quantitative phase microscopy images of cells to be classified;

[0089] Focused single-cell quantitative phase microscopy images of cells to be classified are processed to obtain corresponding cell feature data;

[0090] Based on cell feature data, a cell classification model is used to classify the cells to be classified. The cell classification model is trained according to the cell classification model training method described above.

[0091] The steps of acquiring focused single-cell quantitative phase microscopy images of the cells to be classified and processing these images to obtain corresponding cell feature data are similar to those in the aforementioned training method and will not be repeated here.

[0092] This invention also provides a cell classification model training device, comprising:

[0093] The image acquisition module is used to acquire focused single-cell quantitative phase microscopy images of the target cells.

[0094] The image processing module is used to process the focused single-cell quantitative phase microscopy image to obtain the corresponding cell feature data, which constitutes the cell feature dataset.

[0095] The feature filtering module is used to filter the cell feature dataset based on its predictive performance to obtain the optimal feature dataset.

[0096] The model training module is used to train the initial model to obtain a cell classification model using the optimal feature dataset as sample data and the target cell type as label.

[0097] Preferably, a three-dimensional hydrodynamic focused single-cell image plane holographic image is acquired by a microfluidic chip and a digital holographic microscope, and the three-dimensional hydrodynamic focused single-cell image plane holographic image is processed to obtain a focused single-cell quantitative phase microscopic image; wherein, the microfluidic chip is configured to provide a three-dimensional hydrodynamic focused sample flow containing the target cell to the digital holographic microscope, and the digital holographic microscope is configured to have its recording focal plane coincide with the focal plane of the sample flow.

[0098] Preferably, the microfluidic chip is provided with a sample channel, a first sheath flow channel, and a second sheath flow channel arranged along a horizontal plane. The first sheath flow channel and the second sheath flow channel are symmetrically arranged on both sides of the sample channel and are connected to the sample channel through a confluence port.

[0099] Preferably, the sample stream is injected vertically into the sample channel at a first velocity, and two first sheath streams are injected vertically from above and below the sample channel at symmetrical positions before and after the sample stream injection point, respectively, at a second velocity. Two second sheath streams are injected into the sample channel through the first sheath stream channel and the second sheath stream channel at a third velocity. The ratio of the second velocity to the first velocity is used as the first velocity ratio, and the ratio of the third velocity to the sum of the first velocity and the second velocity is used as the second velocity ratio. By adjusting the first velocity ratio and / or the second velocity ratio, the sample stream can achieve three-dimensional hydrodynamic focusing.

[0100] Preferably, both the first sheath flow channel and the second sheath flow channel are zigzag-shaped, and at the confluence, the angles between the first sheath flow channel and the second sheath flow channel and the sample channel are equal.

[0101] Preferably, the first sheath flow and the second sheath flow are of the same type in terms of composition. In microfluidics, the first sheath flow and the second sheath flow exhibit laminar flow mode, so they do not mix.

[0102] Preferably, the cell feature data includes at least one of the cell morphological feature data, first-order grayscale feature data, and high-order texture feature data.

[0103] Preferably, the initial model is one of the following: Linear Discriminant Analysis (LDA) model, Support Vector Machine (SVM) model, Neural Network (NN) model, Logistic Regression (LR) model, Nearest Neighbor (KNN) model, Naive Bayes model, or Tree model.

[0104] Preferably, the image processing module processes the focused single-cell quantitative phase microscopy image according to the following steps to obtain cell feature data:

[0105] Preprocessing of focused single-cell quantitative phase microscopy images;

[0106] The preprocessed single-cell quantitative phase microscopy images are converted into binary images;

[0107] The largest connected component in the binary image is used as a mask to segment the cell ROI region from the binary image;

[0108] Cell feature data are extracted based on the cell's ROI region.

[0109] Preferably, the feature selection module performs prediction performance screening on the cell feature dataset according to the following steps to obtain the optimal feature dataset:

[0110] The importance of all cell features in the cell feature dataset is ranked using a feature selection algorithm based on mutual information.

[0111] Based on the ranking, the cell features used for training are determined by an incremental feature selection algorithm, and a subset of the corresponding cell feature dataset is selected as the optimal feature dataset according to the determined cell features.

[0112] Preferably, the target cells are selected from at least one of immune cells, tumor cells, red blood cells, white blood cells, stem cells, germ cells, algal cells, microbial cells, and planktonic cells, and more preferably at least two.

[0113] The immune cells are preferably selected from at least one of the following: T cells, B cells, natural killer cells, macrophages, granulocytes, dendritic cells, mast cells, monocytes, and platelets.

[0114] The tumor cells are preferably selected from at least one of the following: lung cancer cells, colorectal cancer cells, cervical cancer cells, breast cancer cells, human monocytic leukemia cells, prostate cancer cells, skin cancer cells, gastric cancer cells, liver cancer cells, cervical cancer cells, esophageal cancer cells, thyroid cancer cells, bladder cancer cells, lymphoma cells, pancreatic cancer cells, kidney cancer cells, uterine cancer cells, oral cancer cells, melanoma cells, ovarian cancer cells, brain cancer cells, multiple myeloma cells, nasopharyngeal cancer cells, gallbladder cancer cells, glioblastoma cells, and testicular cancer cells.

[0115] Leukocytes are preferably selected from at least one of neutrophils, eosinophils, and basophils;

[0116] The stem cells are preferably selected from at least one of embryonic stem cells, adult stem cells, induced pluripotent stem cells, mesenchymal stem cells, hematopoietic stem cells, neural stem cells, epithelial stem cells, muscle stem cells, liver stem cells, skin stem cells, pancreatic islet stem cells, and corneal stem cells;

[0117] The germ cells are preferably selected from at least one of sperm cells, spore cells, and pollen cells;

[0118] The microbial cells are preferably selected from at least one of bacterial cells and fungal cells.

[0119] This invention also provides a cell sorting device, comprising:

[0120] The image acquisition subsystem is used to acquire focused single-cell quantitative phase microscopic images of cells to be classified.

[0121] The image processing subsystem is used to process the focused single-cell quantitative phase microscopic images of the cells to be classified to obtain the corresponding cell feature data.

[0122] The cell classification subsystem is used to classify cells based on cell feature data using a cell classification model. The cell classification subsystem also includes a cell classification model training device for training and obtaining a cell classification model.

[0123] Preferably, the image acquisition subsystem includes a microfluidic chip and a digital holographic microscope. The microfluidic chip and the digital holographic microscope are used to acquire a three-dimensional hydrodynamic focused single-cell image plane holographic image to obtain a focused single-cell quantitative phase microscopic image. The microfluidic chip is configured to provide the digital holographic microscope with a three-dimensional hydrodynamic focused sample flow containing cells to be classified, and the digital holographic microscope is configured to record that its focal plane coincides with the focal plane of the sample flow.

[0124] Preferably, the microfluidic chip is provided with a sample channel, a first sheath flow channel and a second sheath flow channel arranged along a horizontal plane. The first sheath flow channel and the second sheath flow channel are located on both sides of the sample channel and are connected to the sample channel through a confluence port. The sample channel is provided with a sample flow injection point.

[0125] Preferably, two first sheath streams are injected into the sample channel from above and below the sample channel at a second velocity in the vertical direction before and after the sample stream injection point. The sample stream is injected into the sample channel through the sample stream injection point at a first velocity in the vertical direction. The two second sheath streams are injected into the sample channel through the first sheath stream channel and the second sheath stream channel at a third velocity. The ratio of the second velocity to the first velocity is used as the first velocity ratio, and the ratio of the third velocity to the sum of the first velocity and the second velocity is used as the second velocity ratio. The first velocity ratio and / or the second velocity ratio are adjusted to achieve three-dimensional hydrodynamic focusing of the sample stream.

[0126] Preferably, the first sheath flow channel and the second sheath flow channel are both zigzag-shaped and symmetrically arranged on both sides of the sample channel. At the confluence, the angle between the first sheath flow channel and the second sheath flow channel and the sample channel is equal. The two first sheath flows are injected into the sample channel from above and below the sample channel at symmetrical positions before and after the sample flow injection point with a second flow velocity in the vertical direction.

[0127] Preferably, the first sheath flow and the second sheath flow are the same type of sheath flow.

[0128] Example 1

[0129] In this embodiment, the cell classification model training method and classification method of the present invention are applied to heterogeneous subtypes of breast cancer cells. This embodiment uses cell lines with the same organ origin and more similar morphological characteristics as target cells, including one normal breast cell line (N=1000), seven breast cancer cell lines of different subtypes (N=1000×7), and a blank background (N=3000). This group of cell lines was chosen because they cover all major breast cancer subtypes (Luminal A / Luminal B / Her2-enriched / TN). In clinical practice, different subtypes guide the selection of clinical treatment plans, affecting patients' sensitivity to treatment strategies and prognostic outcomes. Therefore, precise differentiation among them is of great practical value, enabling patients to benefit more.

[0130] See Figure 7 (a) Based on the quantitative phase microscopic image reconstruction results of the eight cell types, the TN cell line shows certain characteristic similarities to normal breast cells. The average particle size of the T47D, BT474, and SK-BR-3 cell populations is similar and smaller than that of the other cell populations. Some characteristic distributions are visualized using a violin plot. Figure 7 (b) reflects the differences between different subtypes of breast cancer cells and normal breast cells. Figure 7 (c) The characteristic mean distributions of the eight cell types were summarized, and analysis of variance and multiple comparison tests were performed. Compared with the characteristic datasets of different cancer cell lines, the characteristic datasets of breast cancer multi-subtype cell lines are obviously more complex, reflected in the similarity of a single feature among different subtypes (mean differences are not significant). Statistically, only one feature among the 81 cell features showed significant differences in all pairwise comparisons, revealing that the classification task of breast cancer multi-subtypes is more challenging. Figure 7 (d) is the Pearson correlation coefficient matrix between features. High correlation still tends to appear in similar features and higher-order features, and the frequency of high correlation between features is higher than when distinguishing different types of cancer cells, indicating that the degree of feature redundancy has increased and the training difficulty of the classification model is higher.

[0131] We will continue to use a combination of MRMR and IFS for feature selection. Figure 8 (a) Rank the MRMR scores for each feature. The results show that although more features and longer training time are generally required, the LDA model still achieves 100% classification accuracy after sufficient training. Figure 8 (b) The classification performance of neural network models and SVM models also remained at a high level. Compared with the basic KNN and Tree models, the classification accuracy of subspace KNN and Bagging Tree models improved by 3.4% and 13.2%, respectively.

[0132] Evaluate the predictive performance of each classification model for different cell types. Figure 8 (c) The prediction accuracy for different subtypes of breast cancer cells is generally lower than that for different types of cancer cell lines, which is consistent with common sense, as cell groups are more similar to each other. In summary, the strategy of training based on the hierarchical features of quantitative phase microscopy images and MRMR combined with IFS has shown universal applicability to various machine learning models. The above research results further demonstrate the ability of this method to perform fine classification tasks of multiple cell / particle groups.

[0133] Example 2

[0134] This embodiment applies the cell classification model training method and classification method of the present invention to the simulated label-free detection of circulating tumor cells in blood, demonstrating the practical application potential of the present invention. Blood samples were obtained from donations by healthy volunteers; see [link to relevant documentation]. Figure 9 (a) Quantitative phase microscopy images of red blood cells (RBCs) and peripheral blood mononuclear cells (PBMCs) were acquired. MDA-MB-231 cells were used as a model of breast cancer CTCs, and HeLa cells were used as a model of cervical cancer CTCs. The LDA model was trained according to the aforementioned feature extraction and model training methods of this invention, and its confusion matrix ( Figure 9 (b) and (c) show that the cancer cells can be identified and classified from the background blood cells with an accuracy of nearly 100%.

[0135] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for training a cell classification model, characterized in that, include: Acquire focused single-cell quantitative phase microscopy images of target cells; The focused single-cell quantitative phase microscopy image is processed to obtain the corresponding cell feature data, which constitutes a cell feature dataset. The optimal feature dataset is obtained by performing prediction performance screening on the cell feature dataset. Using the optimal feature dataset as sample data and the type of target cell as label, the initial model is trained to obtain the cell classification model; A three-dimensional hydrodynamic focused single-cell image plane holographic image is acquired using a microfluidic chip and a digital holographic microscope. The three-dimensional hydrodynamic focused single-cell image plane holographic image is then processed to obtain the focused single-cell quantitative phase microscopic image. The microfluidic chip is configured to provide the digital holographic microscope with a three-dimensional hydrodynamic focused sample flow containing the target cell, and the digital holographic microscope is configured to have its recording focal plane coincide with the focal plane of the sample flow. The optimal feature dataset was obtained by performing prediction performance screening on the cell feature dataset, including: The importance of all cell features in the cell feature dataset is ranked using a feature selection algorithm based on mutual information. Based on the sorting, the cell features used for training are determined by an incremental feature selection algorithm, and a subset of the corresponding cell feature dataset is selected as the optimal feature dataset according to the determined cell features.

2. The cell classification model training method according to claim 1, characterized in that, The microfluidic chip is provided with a sample channel, a first sheath flow channel and a second sheath flow channel arranged along a horizontal plane. The first sheath flow channel and the second sheath flow channel are located on both sides of the sample channel and are connected to the sample channel through a confluence port. The sample channel is provided with a sample flow injection point.

3. The cell classification model training method according to claim 2, characterized in that, Two first sheath streams are injected into the sample channel from above and below it at a second velocity in the vertical direction before and after the sample injection point. The sample stream is injected into the sample channel at a first velocity in the vertical direction through the sample injection point. Two second sheath streams are injected into the sample channel through the first sheath stream channel and the second sheath stream channel at a third velocity. The ratio of the second velocity to the first velocity is used as the first velocity ratio, and the ratio of the third velocity to the sum of the first velocity and the second velocity is used as the second velocity ratio. The first velocity ratio and / or the second velocity ratio are adjusted to achieve three-dimensional hydrodynamic focusing of the sample stream.

4. The cell classification model training method according to claim 3, characterized in that, The first sheath flow channel and the second sheath flow channel are both zigzag-shaped and symmetrically arranged on both sides of the sample channel. At the confluence, the angles between the first sheath flow channel and the second sheath flow channel and the sample channel are equal. The two first sheath flows are injected into the sample channel from above and below the sample channel at symmetrical positions before and after the sample flow injection point with a second flow velocity in the vertical direction. and / or The first sheath flow and the second sheath flow are the same type of sheath flow.

5. The cell classification model training method according to claim 1, characterized in that, The cell feature data includes at least one of the following: morphological feature data of cells, first-order grayscale feature data, and high-order texture feature data; and / or The initial model is one of the following: Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Neural Network (NN), Logistic Regression (LR), Nearest Neighbor (KNN), Naive Bayes, or Tree Model.

6. The cell classification model training method according to claim 1, characterized in that, The focused single-cell quantitative phase microscopy image is processed according to the following steps to obtain cell feature data: The focused single-cell quantitative phase microscopy image is preprocessed; The preprocessed single-cell quantitative phase microscopy images are converted into binary images; The largest connected component in the binary image is used as a mask to segment the cell ROI region from the binary image; Cell feature data are extracted based on the cell ROI region.

7. The cell classification model training method according to claim 1, characterized in that, The target cells are selected from at least one of immune cells, tumor cells, red blood cells, white blood cells, stem cells, germ cells, algal cells, microbial cells, and planktonic cells.

8. The cell classification model training method according to claim 7, characterized in that, The target cells are selected from at least two of the following: immune cells, tumor cells, red blood cells, white blood cells, stem cells, germ cells, algal cells, microbial cells, and planktonic cells.

9. The cell classification model training method according to claim 7 or 8, characterized in that, The immune cells are selected from at least one of the following: T cells, B cells, natural killer cells, macrophages, granulocytes, dendritic cells, mast cells, monocytes, and platelets. The tumor cells are selected from at least one of the following: lung cancer cells, colorectal cancer cells, cervical cancer cells, breast cancer cells, human monocytic leukemia cells, prostate cancer cells, skin cancer cells, gastric cancer cells, liver cancer cells, cervical cancer cells, esophageal cancer cells, thyroid cancer cells, bladder cancer cells, lymphoma cells, pancreatic cancer cells, kidney cancer cells, uterine cancer cells, oral cancer cells, melanoma cells, ovarian cancer cells, brain cancer cells, multiple myeloma cells, nasopharyngeal cancer cells, gallbladder cancer cells, glioblastoma cells, and testicular cancer cells. The white blood cells are selected from at least one of neutrophils, eosinophils, and basophils; The stem cells are selected from at least one of embryonic stem cells, adult stem cells, induced pluripotent stem cells, mesenchymal stem cells, hematopoietic stem cells, neural stem cells, epithelial stem cells, muscle stem cells, liver stem cells, skin stem cells, pancreatic islet stem cells, and corneal stem cells. The reproductive cells are selected from at least one of sperm cells, spore cells, and pollen cells; The microbial cells are selected from at least one of bacterial cells and fungal cells.

10. A cell classification model training device, characterized in that, include: The image acquisition module is used to acquire focused single-cell quantitative phase microscopy images of the target cells. The image processing module is used to process the focused single-cell quantitative phase microscopy image to obtain corresponding cell feature data, which constitutes a cell feature dataset. The feature filtering module is used to filter the cell feature dataset based on prediction performance to obtain the optimal feature dataset. The model training module is used to train the initial model to obtain the cell classification model by using the optimal feature dataset as sample data and the type of target cell as the label. Specifically, a three-dimensional hydrodynamic focused single-cell image plane holographic image is acquired using a microfluidic chip and a digital holographic microscope. The three-dimensional hydrodynamic focused single-cell image plane holographic image is then processed to obtain the focused single-cell quantitative phase microscopic image. The microfluidic chip is configured to provide the digital holographic microscope with a three-dimensional hydrodynamic focused sample flow containing the target cell, and the digital holographic microscope is configured to have its recording focal plane coincide with the focal plane of the sample flow. The feature selection module performs prediction performance screening on the cell feature dataset according to the following steps to obtain the optimal feature dataset: The importance of all cell features in the cell feature dataset is ranked using a feature selection algorithm based on mutual information. Based on the sorting, the cell features used for training are determined by an incremental feature selection algorithm, and a subset of the corresponding cell feature dataset is selected as the optimal feature dataset according to the determined cell features.

11. The cell classification model training device according to claim 10, characterized in that, The microfluidic chip is provided with a sample channel, a first sheath flow channel and a second sheath flow channel arranged along a horizontal plane. The first sheath flow channel and the second sheath flow channel are located on both sides of the sample channel and are connected to the sample channel through a confluence port. The sample channel is provided with a sample flow injection point.

12. The cell classification model training device according to claim 11, characterized in that, Two first sheath streams are injected into the sample channel from above and below it at a second velocity in the vertical direction before and after the sample injection point. The sample stream is injected into the sample channel at a first velocity in the vertical direction through the sample injection point. Two second sheath streams are injected into the sample channel through the first sheath stream channel and the second sheath stream channel at a third velocity. The ratio of the second velocity to the first velocity is used as the first velocity ratio, and the ratio of the third velocity to the sum of the first velocity and the second velocity is used as the second velocity ratio. The first velocity ratio and / or the second velocity ratio are adjusted to achieve three-dimensional hydrodynamic focusing of the sample stream.

13. The cell classification model training device according to claim 12, characterized in that, The first sheath flow channel and the second sheath flow channel are both zigzag-shaped and symmetrically arranged on both sides of the sample channel. At the confluence, the angles between the first sheath flow channel and the second sheath flow channel and the sample channel are equal. The two first sheath flows are injected into the sample channel from above and below the sample channel at symmetrical positions before and after the sample flow injection point with a second flow velocity in the vertical direction. and / or The first sheath flow and the second sheath flow are the same type of sheath flow.

14. The cell classification model training device according to claim 10, characterized in that, The cell feature data includes at least one of the following: morphological feature data of cells, first-order grayscale feature data, and high-order texture feature data; and / or The initial model is one of the following: Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Neural Network (NN), Logistic Regression (LR), Nearest Neighbor (KNN), Naive Bayes, or Tree Model.

15. The cell classification model training device according to claim 10, characterized in that, The image processing module processes the focused single-cell quantitative phase microscopy image according to the following steps to obtain cell feature data: The focused single-cell quantitative phase microscopy image is preprocessed; The preprocessed single-cell quantitative phase microscopy images are converted into binary images; The largest connected component in the binary image is used as a mask to segment the cell ROI region from the binary image; Cell feature data are extracted based on the cell ROI region.

16. The cell classification model training device according to claim 10, characterized in that, The target cells are selected from at least one of immune cells, tumor cells, red blood cells, white blood cells, stem cells, germ cells, algal cells, microbial cells, and planktonic cells.

17. The cell classification model training device according to claim 16, characterized in that, The target cells are selected from at least two of the following: immune cells, tumor cells, red blood cells, white blood cells, stem cells, germ cells, algal cells, microbial cells, and planktonic cells.

18. The cell classification model training device according to claim 16 or 17, characterized in that, The immune cells are selected from at least one of the following: T cells, B cells, natural killer cells, macrophages, granulocytes, dendritic cells, mast cells, monocytes, and platelets. The tumor cells are selected from at least one of the following: lung cancer cells, colorectal cancer cells, cervical cancer cells, breast cancer cells, human monocytic leukemia cells, prostate cancer cells, skin cancer cells, gastric cancer cells, liver cancer cells, cervical cancer cells, esophageal cancer cells, thyroid cancer cells, bladder cancer cells, lymphoma cells, pancreatic cancer cells, kidney cancer cells, uterine cancer cells, oral cancer cells, melanoma cells, ovarian cancer cells, brain cancer cells, multiple myeloma cells, nasopharyngeal cancer cells, gallbladder cancer cells, glioblastoma cells, and testicular cancer cells. The white blood cells are selected from at least one of neutrophils, eosinophils, and basophils; The stem cells are selected from at least one of embryonic stem cells, adult stem cells, induced pluripotent stem cells, mesenchymal stem cells, hematopoietic stem cells, neural stem cells, epithelial stem cells, muscle stem cells, liver stem cells, skin stem cells, pancreatic islet stem cells, and corneal stem cells. The reproductive cells are selected from at least one of sperm cells, spore cells, and pollen cells; The microbial cells are selected from at least one of bacterial cells and fungal cells.

19. A cell classification method, characterized in that, include: Acquire focused single-cell quantitative phase microscopy images of cells to be classified; The focused single-cell quantitative phase microscopy images of the cells to be classified are processed to obtain the corresponding cell feature data; Based on the cell feature data, the cells to be classified are classified using a cell classification model, which is obtained by training according to the cell classification model training method described in any one of claims 1-9.

20. A cell sorting device, characterized in that, include: The image acquisition subsystem is used to acquire focused single-cell quantitative phase microscopic images of cells to be classified. An image processing subsystem is used to process the focused single-cell quantitative phase microscopic image of the cells to be classified to obtain corresponding cell feature data; A cell classification subsystem is used to classify the cells to be classified based on the cell feature data using a cell classification model. The cell classification subsystem further includes a cell classification model training device according to any one of claims 10-18 for training to obtain the cell classification model. The image acquisition subsystem includes a microfluidic chip and a digital holographic microscope. The microfluidic chip and the digital holographic microscope are used to acquire a three-dimensional hydrodynamic focused single-cell image plane holographic image to obtain the focused state single-cell quantitative phase microscopic image. The microfluidic chip is configured to provide the digital holographic microscope with a three-dimensional hydrodynamic focused sample flow containing cells to be classified, and the digital holographic microscope is configured to have its recording focal plane coincide with the focal plane of the sample flow.

21. The cell sorting device according to claim 20, characterized in that, The microfluidic chip is provided with a sample channel, a first sheath flow channel and a second sheath flow channel arranged along a horizontal plane. The first sheath flow channel and the second sheath flow channel are located on both sides of the sample channel and are connected to the sample channel through a confluence port. The sample channel is provided with a sample flow injection point.

22. The cell sorting device according to claim 21, characterized in that, Two first sheath streams are injected into the sample channel from above and below it at a second velocity in the vertical direction before and after the sample injection point. The sample stream is injected into the sample channel at a first velocity in the vertical direction through the sample injection point. Two second sheath streams are injected into the sample channel through the first sheath stream channel and the second sheath stream channel at a third velocity. The ratio of the second velocity to the first velocity is used as the first velocity ratio, and the ratio of the third velocity to the sum of the first velocity and the second velocity is used as the second velocity ratio. The first velocity ratio and / or the second velocity ratio are adjusted to achieve three-dimensional hydrodynamic focusing of the sample stream.

23. The cell sorting device according to claim 22, characterized in that, The first sheath flow channel and the second sheath flow channel are both zigzag-shaped and symmetrically arranged on both sides of the sample channel. At the confluence, the angles between the first sheath flow channel and the second sheath flow channel and the sample channel are equal. The two first sheath flows are injected into the sample channel from above and below the sample channel at symmetrical positions before and after the sample flow injection point with a second flow velocity in the vertical direction. The first sheath flow and the second sheath flow are the same type of sheath flow.

Images