Cell typing with cell lasing

Abstract

A method is presented for identifying a cell type of a subject using laser emissions. The method includes: directing, by a light source, a light beam towards a biological sample, where the biological sample is disposed in a laser cavity; capturing, by a detector, spatial distribution of laser emissions from the biological sample; and identifying a cell type for the biological sample from the captured spatial distribution of laser emissions using machine learning.

Description

CELL TYPING WITH CELL LASINGCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 624,484, filed on January 24, 2024. The entire disclosure of the above application is incorporated herein by reference.FIELD

[0002] The present disclosure relates to identifying cell types using laser emissions.BACKGROUND

[0003] Circulating tumor cells (CTCs) are tumor cells that shed from the primary tumor and enter the bloodstream, which can serve as biomarkers for cancer diagnosis and prognosis. CTCs have been observed to be extremely rare in peripheral blood of cancer patients, often numbering fewer than one in a million. Accurate detection of CTCs is crucial for their high-purity isolation and for providing precise diagnoses to cancer patients. Numerous strategies have been developed to identify and isolate these rare cells, including immuno-labeling, and label-free microfluidic methods. However, positive immuno-labeling, for example, can only detect the CTCs expressing selected markers and label-free microfluidic methods may yield samples with low purity of CTCs. An accurate and antigen-independent CTC identification method can enhance the purity of existing methods, improve the accuracy of CTC enumeration for clinical diagnosis, and provide a comprehensive spectrum of CTCs without marker selection.

[0004] Biolaser is an emerging tool in biological studies. Using fluorescent labeling in cells as the gain medium, the biolaser exhibits threshold behavior, spectral fingerprints, and lasing patterns that are highly sensitive to intracellular environments. Previous research has demonstrated that the threshold behavior of cell lasers can effectively distinguish between cancerous cells and healthy cells when their chromatins are stained with dyes. This agrees with the findings that chromatin aggregations appear in cancer cells. Studies have shown that chromatin abnormality exists in pancreatic cancer. Furthermore, it is shown that the lasing mode and the pattern may change in cells when the intracellular structure changes. Recently, deep learningtechniques have been employed in cell lasers to distinguish cell types based on cell sizes. However, the accuracy is still far from needed to identify rare cells such as CTCs. Nor do they examine the intracellular environment that cell lasers can probe.

[0005] Here, to address existing challenges and provide an assisting method for current CTC identification strategies, an interpretable-AI assisted biolaser platform is developed that examines the intracellular environment and is capable of detecting CTCs from white blood cells (WBCs) with high accuracy (>95%). Utilizing artificially generated data, a neural network effectively extracts features of the lasing modes from chromatin-stained cells and uses these features to analyze the intracellular environment to identify CTCs from white blood cells. This Al-assisted biolaser platform shows promise for clinical applications, offering both accurate CTC enumeration and insights into the biological characteristics of the cells.

[0006] This section provides background information related to the present disclosure which is not necessarily prior art.SUMMARY

[0007] This section provides a general summary of the disclosure, and is not a comprehensive disclosure of its full scope or all of its features.

[0008] Accurate identification of cell types at the level of single cells holds potential in the field of precision medicine. For instance, the identification of cancer cells can play a crucial role in monitoring cancer progression and assessing the response to drug treatments in patients. This disclosure describes a method to distinguish different types of cells using cell lasing. First, cells are stained with cell-labelling fluorophores and then placed inside a laser cavity, such as Fabry-Perot cavity. An external pump laser is ramped in power and scanned over the cells and excites the lasing emission from these cells. The lasing emission patterns are fed into a neural network (or some other types of artificial intelligence (Al), machine learning (ML) or Deep Learning (DL) models) based image processing algorithms to distinguish between the cell types or for cell sub-typing. Using the AI / ML methods, the inter-cavity variations among different lasing cavities, which may have different cavity characteristics such as Q-factors, can be significantly reduced. Consequently, the cell phenotyping capabilities (such as sensitivity, specificity, and accuracy) can be significantly improved.

[0009] In one aspect, a method is presented for identifying a cell type of a subject using laser emissions. The method includes: directing, by a light source, a light beam towards a biological sample, where the biological sample is disposed in a laser cavity; capturing, by a detector, spatial distribution of laser emissions from the biological sample; and identifying a cell type for the biological sample from the captured spatial distribution of laser emissions using machine learning. The method may further include staining the biological sample with a fluorophore prior to placing the biological sample in the laser cavity.

[0010] In one embodiment, the intensity of the light source is changed to two or more different intensity values and a spatial distribution of laser emissions from the biological sample is captured at the different intensity values of the light source.

[0011] In some embodiments, a feature vector is generated from the spatial distribution of laser emissions from the biological sample by a first encoder, and a cell type for the biological sample is identified from the captured spatial distribution of laser emissions using the feature vector and a neural network.

[0012] In another aspect, the method for identifying a cell type of a subject using laser emissions includes: directing, by a light source, a light beam towards a biological sample, where the biological sample is disposed in a laser cavity; capturing, by a detector, spatial distribution of laser emissions from the biological sample; generating, by a first encoder, a feature vector from the spatial distribution of laser emissions from the biological sample, determining position data of the biological sample in relation to center of the laser cavity; generating, by a second encoder, a position vector for the biological sample from the position data; and identifying, by the neural network, a cell type for the biological sample using the feature vector and the position vector.

[0013] In yet another aspect, a laser imaging system is presented. The laser imaging system comprises: a laser cavity configured to hold a biological sample from a subject; a light source configured to generate and direct a light beam towards the biological sample; a detector configured to capture spatial distribution of laser emissions from the biological sample; and a controller interfaced with the light source and detector. The generates a feature vector from the spatial distribution of the laser emissions and identifies a cell type for the biological sample from the feature vector using a neural network. The controller may further operate to change intensity of the light source to two or more different intensity values and capture spatial distribution of laser emissions at each of the different light intensity values.

[0014] In some embodiments, a motorized stage is configured to move the biological sample relative to the light beam of the light source, where the laser cavity resides on the motorized stage. The controller determines position data of the biological sample in relation to center of the laser cavity; generates a position vector for the biological sample from the position data using a second encoder; and identifies the cell type for the biological sample based in part on the position vector.

[0015] Further areas of applicability will become apparent from the description provided herein. The description and specific examples in this summary are intended for purposes of illustration only and are not intended to limit the scope of the present disclosure.DRAWINGS

[0016] The drawings described herein are for illustrative purposes only of selected embodiments and not all possible implementations, and are not intended to limit the scope of the present disclosure.

[0017] Figure 1 is a diagram of an example laser imaging system.

[0018] Figure 2 is a flowchart depicting the proposed method for identifying cell type using laser emissions.

[0019] Figure 3 is a diagram of a multi-state machine learning approach.

[0020] Figure 4 is a diagram of the workflow of experimental trajectory for human samples.

[0021] Figure 5 is an example deep learning model for classifying cells.

[0022] Figure 6A shows lasing mode patterns from three randomly selected white blood cells (WBCs) from healthy donors. The mode patterns on each row were emitted by the same cell at four selected pumping powers.

[0023] Figure 6B shows lasing mode patterns from three randomly selected Pancreatic 368 cells (labeled P-368) at the same four pumping powers. The lateral size of each square image containing a lasing mode pattern is 35 pm.

[0024] Figure 6C is a graph showing an exemplary lasing threshold of a white blood cell.

[0025] Figure 6D is a graph showing an exemplary lasing threshold of a tumor cell. Dashed lines indicate the cell laser emission intensity (integrated over all pixels within an image) at four pumping powers.

[0026] Figure 6E is a graph showing lasing rate of white blood cells at the four selected pumping powers. It reaches 54.5% at 150 pJ / mm2.

[0027] Figure 6F is a graph showing the lasing rate of Pancreatic 368 cells at the four selected pumping powers. It reaches 95.5% at 150 pJ / mm2.

[0028] Figures 7A-7D are T-SNE visualizations of the feature vectors of 500 randomly selected cells from Pancreatic 368 cell line and 500 WBCs from Donors 1-6 each at four selected pumping powers: 25 pJ / mm2, 45 pJ / mm2, 90 pJ / mm2, and 150 pJ / mm2, respectively. The average feature vector of each cell population is shown in the plot as the large dot of the same color.

[0029] Figures 8A-8D are T-SNE visualizations of lasing mode patterns of the feature vectors of 500 randomly selected cells from each of three CTC-derived cell lines (Pancreatic 368, 322, and 331 ) and 500 WBCs from Donors 1-6 each at four selected pumping powers: 25 pJ / mm2, 45 pJ / mm2, 90 pJ / mm2, and 150 pJ / mm2, respectively. The PScLE was trained on one CTC model (Pancreatic 368) and WBCs. As shown in the figures, populations from the three CTC models form a cluster that is distinct from WBCs at all four pumping power levels, showing the PScLE’s zero-shot generalization ability to distinguish unobserved CTCs from WBCs.

[0030] Figure 9A is a graph showing the receiver operating characteristic (ROC) curve of the DCLC model with an area under the curve (AUC) of 0.9994.

[0031] Figure 9B is a confusion matrix of the classifier model setting threshold to 0.2.

[0032] Figure 10A is a diagram showing process flow of the validation experiments. Pancreatic 368 cells (referred to as CTCs here) were first stained using SYTO 9 and subsequently with DAPI and spiked into WBCs that were stained only with SYTO 9. Fluorescence signals from DAPI-stained Pancreatic 368 cells were used to validate the classification results of the DCLC.

[0033] Figure 10B is a graph showing cell (Pancreatic 368) counts of the DCLC on the biolaser channel against cell counts on the fluorescence channel. The dashed line has a unity slope to guide an eye.

[0034] Figure 10C is a laser emission image showing a detected CTC class cell (marked with a red box) in the spiked-in sample by the deep learning model. Unmarked lasing spots were classified as WBCs by the model.

[0035] Figure 10D is a fluorescence image of the same cavity position, showing that only the CTC was present at the same location as in Fig. 10C.

[0036] Figure 10E is a graph showing sensitivity and specificity of the four spiked-in experiments. Note that for the control set with no Pancreatic 368 spiked in, sensitivity and specificity are plotted as 1 .

[0037] Figure 11 A is a graph showing enumeration of CTCs from pancreatic cancer patients (1-2) and lung cancer patients (3-8), normalized to the number per mL of the original blood sample.

[0038] Figure 11 B is a graph showing a comparison between CTC counts obtained by the biolaser method and the IF method for pancreatic cancer and lung cancer. The solid line has a unity slope in the log-log scale to guide an eye.

[0039] Figure 11 C is images of an observed pancreatic CTC using the biolaser method. The image is cropped from the full slide scans at 90 pJ / mm2. The image size is 300 x 200 pm2. The bounding box is used to indicate the position of the CTC. The bounding boxes for WBCs are omitted for clarity. IN the bottom images, the HLS of that CTC is shown (from left to right: 150 pJ / mm2, 90 pJ / mm2, 45 pJ / mm2, and 25 pJ / mm2). The lateral size of each square image containing a lasing mode pattern is 35 pm.

[0040] Figure 11 D are images of an observed pancreatic CTC from IF staining as identified by DAPI+ (blue), CD45- (green), and PanCK+ (red). The surrounding WBCs from the same sample are DAPI+ / CD45+ / PanCK-.

[0041] Corresponding reference numerals indicate corresponding parts throughout the several views of the drawings.DETAILED DESCRIPTION

[0042] Example embodiments will now be described more fully with reference to the accompanying drawings.

[0043] Figure 1 depicts an example laser imaging system 10. The laser imaging system 10 is comprised generally of a light source 12, a laser cavity 13, a detector 14 and a controller 15. In the example embodiment, the laser imaging system 10 further includes a beam expansion lens set 16, a mirror scanning subsystem 17, a scanning lens set 18, a beam splitter 19 and an objective lens 20. It is to be understood that only the relevant components of the imaging system 10 are discussed in relation to Figure 1 , but that other components may be needed to control and manage the overall operation of the system.

[0044] The laser cavity 13 is configured to hold a biological sample (e.g., cells, tissues, etc.) from a subject, such as a person. In one embodiment, the laser cavity 13includes a first reflection surface and a second reflection surface with the biological sample sandwiched between the first reflection surface and the second reflection surface. The laser cavity 13 resides on a motorized stage 8 or another movable device. The motorized stage is configured to move the biological sample relative to the beam of the light source 12.

[0045] In the example embodiment, the light source 12 is further defined as a laser. The light beam emitted from the laser is directed towards the biological sample housed by the laser cavity 13. More specifically, the light beam is directed by the beam expansion lens set 16, the mirror scanning subsystem 17, the scanning lens set 18 and the beam splitter 19. It is readily understood that other types of light sources as well as other optical components may be used generate and direct the light in the imaging system 10. Example laser cavities 13 may include but are not limited to Fabry-Perot cavity, a ring resonator or a photonic crystal laser cavity.

[0046] Prior to placement in the laser cavity 13, the biological sample may be stained and / or treated with a dye or lasing medium (such as, YO-PRO® or fluorescein isothiocyanate (FITC)) so as to generate fluorescence and / or laser emissions in desired color wavelengths. The fluorophore (i.e., dye molecules) of the treated biological sample absorbs the light from the laser and reflects an emission back through the objective lens 20 and towards the beam splitter 19. The emission in turn passes through the beam splitter 19 and is captured by the detector 14. It is envisioned that the detector 14 may be a camera or another type of imaging device, such as avalanche photodiode, a spectrometer, or an avalanche photodiode and a spectrometer.

[0047] In one embodiment, the controller 15 is implemented as a microcontroller. The skilled artisan will recognize that the logic for the control of the imaging system 100 by the controller 15 may be implemented in hardware logic, software logic, or a combination of hardware and software logic. In this regard, the controller 15 can be or can include any of a digital signal processor (DSP), microprocessor, microcontroller, or other programmable device, which are programmed with software implementing the above-described methods. It should be understood that alternatively the controller 15 is or includes other logic devices, such as a Field Programmable Gate Array (FPGA), a complex programmable logic device (CPLD), or application specific integrated circuit (ASIC). When it is stated that the controller 15 performs a function or is configured to perform a function, it should be understood that the controller 15 is configured to do so with appropriate logic (such as, in software, logic devices, or a combination thereof).The controller 15 may also be in data communication with a personal computer or another computing device.

[0048] Figure 2 further depicts an improved method for identifying a cell type of a subject using laser emissions. A light beam from a light source is directed towards a biological sample as indicated as 22. The biological sample is disposed on a laser cavity. The biological sample may be stained with a fluorophore prior to placing the biological sample in the laser cavity.

[0049] Spatial distribution of laser emissions from the biological sample are captured by the detector as indicated at 24. Preferably, a state of the light source is varied over time and the spatial distributions of the laser emissions are captured at different states of the light source. “State” can refer to but is not limited to intensity, polarization, light pulse duration, light beam incident angle, etc. The captured spatial distributions are passed along to the controller for further analysis.

[0050] From the captured spatial distributions of the laser emissions, a cell type for the biological sample is identified at 26 using machine learning. In one embodiment, the cell type of the biological samples are identified using a neural network. In another embodiment, the cell type of the biological samples are identified using a multi-stage approach as seen in Figure 3. A feature vector is generated by a first encoder 32, where the first encoder 32 extracts features from the captured spatial distribution of laser emissions from the biological sample. A position vector is generated by a second encoder 34, where the second encoder extracts the position vector from the relative position of the biological sample to the center of the laser cavity. The feature vector and the position vector are input to a neural network 36 and the cell type for the biological sample is identified by the neural network 36. Other machine learning techniques are suitable for identifying a cell type from the captured spatial distributions of the laser emissions.

[0051] As a proof of concept, these techniques were applied to circulating tumor cells (CTCs) analysis. With reference to Figure 4, the entire CTC analysis consists of sample preparation (red blood cell (RBC) depletion, CTC enrichment using the microfluidic Labyrinth, and cell staining, single-cell lasing measurements under a homebuilt laser emission microscope, and deep-learning-assisted detection and classification.

[0052] Blood samples from seven healthy donors, two patients with pancreatic cancer, and six patients with lung cancer were collected under the approval of theUniversity of Michigan Institutional Review Board (HUM00173332, HUM00227562, and HUM00042204) with written informed consent provided by all participants. In addition, pancreatic cancer patient CTC-derived cell lines, Pancreatic 368, Pancreatic 322, and Pancreatic 331 cells, were used.

[0053] Healthy donor and patient blood samples were diluted with PBS (1 :1 ratio) and added on top of Ficoll to deplete the red blood cells (4 mL blood solutions mL Ficoll). The resulting solution was transferred to a 15 mL conical tube, which was subsequently centrifuged at 400 g for 40 min. Finally, the monocyte cell layer along with potential tumor cells was collected from the layered solution. The collected cell suspension was further diluted to 1 :5 of initial blood volume. A syringe pump was used to process the sample through the microfluidic Labyrinth device. The Labyrinth device employs an inertial focusing method to sort the blood cells based on different sizes, producing a sample that contains enriched CTCs (if any) and fewer white blood cells. When a healthy donor sample was processed through the Labyrinth device, only the white blood cells existed in the post-Labyrinth enriched sample. The outputs of the Labyrinth were fixed (4% paraformaldehyde, 10 min) and then used for single-cell laser studies. The sample preparation steps are summarized in the top panel of Fig. 4.

[0054] For the patient CTC-derived cell line samples (Pancreatic 368, Pancreatic 322, and Pancreatic 331 cells), no RBC depletion or CTC enrichment was needed.

[0055] A cell membrane permeable nucleic acid stain, SYTO 9 (Thermo Fisher Scientific, USA), was used and which turns into a fluorescent state after binding with DNA / RNA. Cells collected from the Labyrinth device were resuspended in Hank’s Balanced Salt Solution (HBSS) at a until a final concentration of 50 pM was reached. The cells underwent incubation at 4 °C for 25 min, and then were washed twice and resuspended in HBSS.

[0056] The micro Fabry-Perot (F-P) cavity consisted of two 1 ” x 1 ” custommade dielectric mirrors (Evaporated Coatings Inc., USA) with a high reflectivity (R>99.5%) at 510-550 nm, which covered the lasing band for cell lasers. After cells were deposited on the bottom cavity mirror, polystyrene microbeads (Sigma-Aldrich, USA) were placed at the four corners of the bottom mirror to control the cavity length.

[0057] The Fabry-Perot cavity was placed into a 3-D printed cage with spring suspension designed to increase its stability and subsequently transported onto the laser scanning platform. The optical system was a confocal microscopy-like setup. A nanosecond laser (473 nm, 120 ns pulse width) served as the optical pumping sourceand single-cell lasing mode patterns were collected using a camera (DCC3260C, Thorlabs, USA).

[0058] For each cell deposited in an Fabry-Perot cavity, its lasing mode patterns for four different optical pumping powers (25 pJ / mm2, 45 pj / mm2, 90 pJ / mm2, and 150 pj / mm2) were tracked and collected. These pumping powers were chosen as they best show the evolution of the lasing mode patterns for the test tumor cells from the CTC models (Pancreatic 368, 331 , and 322) and achieve a high lasing rate for tumor cells (i.e., the fraction of lasing tumor cells out of the total number of tumor cells) but not for white blood cells. For a cell that did not lase (e.g., at a low pumping power), a dark image of the cell position was taken and still treated as its lasing mode pattern. This dark image indicates that the pumping power is below the lasing threshold of that cell.

[0059] The four lasing mode patterns were stacked for each cell into a tensor for data processing. Since this high dimensional data contains the unique information of that cell, we refer to it as Hyper Lasing Signature (HLS) of the cell.

[0060] The lasing mode pattern of a cell arises from its fluorescently labeled biomolecules, which act as the gain medium. The distribution of the gain medium and its microenvironment, such as the refractive index of organelles within the cell, modulate the cell lasing mode pattern. Unlike the fluorescence images of a cell, when pumping power increases, the lasing mode pattern of a cell at the increased pumping power changes non-linearly. For example, the lasing mode pattern at a higher pumping power may not be obtained by simple linear scaling of the pixels from the lasing mode pattern obtained at a lower pumping power. Therefore, the HLS of a cell can provide insights into the microenvironment of the cell that is beyond one lasing mode pattern at one specific pumping power, as well as the regular fluorescence images.

[0061] To effectively process the HLSs of cells and investigate unique lasing properties of different cell types, a deep learning model called the Deep Cell-Laser Classifier (DCLC) was developed to learn how the internal microenvironment of cells modulates their lasing mode patterns. Figure 5 shows the structure of an example model. The model consists of a parallel single-cell laser encoder (PScLE) and a cell classifier. In this example implementation, the PScLE contains four ResNets or residual network blocks (identical in structure, but with separately optimized parameters), each of which encodes the image-based lasing mode pattern from one pumping power into the feature space. In this way, the tensor form of the HLS of a cell is projected into four feature vectors of the same dimension, each corresponding to one pumping power. Thelasing mode patterns of the cell are represented by feature vectors that can be used to study the microenvironment of the cell.

[0062] While the feature space provides more dimensionalities and robustness for cell phenotyping studies, this example focuses on circulating tumor cells detection. The PScLE is connected to a cell classifier and optimized it for rare cell detection to emphasize high specificity. The cell classifier consists of a fully connected architecture with three layers, each featuring 256, 32, and 8 neurons with ReLU activation. The output layer consists of two neurons using a softmax activation function to provide the probability distribution over the two possible cell types. The DCLC model was implemented using the open-source framework Pytorch and trained on an RTX 3090.

[0063] In this study, white blood cell samples were used from Donors 1-6 (note: the WBCs from Donor 7 were reserved for another use as described later) and the pancreatic cancer patient CTC-derived cell line, Pancreatic 368 cells (referred to as “tumor cells”).

[0064] First, each cell population was deposited into separate Fabry-Perot cavities for single-cell laser measurement. Figure 6A shows the lasing mode patterns from three randomly selected lasing white blood cells. WBC 1 starts to lase at 45 pj / mm2. WBCs 2 and 3 do not lase at a pumping power of 45 pj / mm2or lower. When the pumping power is above the lasing threshold, the lasing signal intensity (summed CCD pixel value) increases. The three randomly chosen tumor cells show a lower lasing threshold than that for WBCs, as shown in Fig. 6B. With increasing pumping power, the lasing mode patterns of the tumor cells change significantly, which indicates the emergence of higher orders of modes. Figures 6C and 6D show an exemplary fitted function of the pump-laser from a selected cell from each cell type, showing that white blood cells have a much higher lasing threshold than tumor cells.

[0065] The fraction of lasing cells (lasing rate) for the two cell types was further examined at the four pumping powers (Fig. 6E and 6F). White blood cells have a lasing rate of 54.5% at the highest pumping power (150 pj / mm2). In contrast, Fig. 6F shows that the lasing rate of tumor cells is higher than that for white blood cell at a given pumping power. At the highest pumping power (150 pJ / mm2), a lasing rate of 95.5% is attained for cancer cells. Therefore, the lasing threshold can be used as one of the filters for white blood cells. For example, it can filter out 45.5% of the non-lasing white blood cells at the expense of only 4.5% of the non-lasing tumor cells at 150 pj / mm2.Here, a cutoff of the threshold filter is set at 150 pJ / mm2because further increasing pumping power no longer improved the lasing rate of tumor cells.

[0066] First, the PScLE was investigated and the encoding process of the lasing mode patterns was examined for a single pumping power level and the corresponding feature vector. Each feature vector has a dimension of 1024. The encoding was performed by a ResNet 101 block trained on the two cell types. The feature vectors represent the activation of neurons that were able to distinguish the lasing modes in WBCs and tumor cells (Pancreatic 368). Figures 7A-7D show the t-distributed stochastic neighbor embedding (t-SNE) of the feature vectors of 500 randomly selected lasing cells of each cell population. The white blood cells from different donors, while demonstrating some degree of variability, overlap strongly among themselves and are distinguishable from the tumor cell cluster, suggesting that, within each pumping power, white blood cells have different lasing mode patterns from tumor cells and this difference can be captured by the PScLE. Note that the sampled lasing mode patterns of each cell population were collected from multiple Fabry-Perot cavities and over multiple staining batches. No large distinct subclusters formed within each cell type, indicating that our single-cell laser measurement system and the PScLE are robust against inter-batch variations, such as fluctuations in cavity parameters (e.g., cavity length).

[0067] At this point, the PScLE had been trained on one CTC model (Pancreatic 368). To demonstrate the ability of the encoder to learn the lasing properties unique to tumor cells and subsequently generalize the knowledge for potentially more heterogeneous tumor cells, the encoder was tested on two more patient CTC-derived cell lines, Pancreatic 322 and Pancreatic 331 , that were not observed by the encoder during the training. The HLSs of these two cell lines were projected into feature vectors and plotted along with white blood cells and Pancreatic 368 cells using t-SNE visualization. Figures 8A-8D show 500 randomly sampled cells from each of the above two previously unobserved CTC models, together with 500 white blood cells from Donors 1-6 and 500 Pancreatic 368 cells. The two new tumor cell populations form their own clusters that overlap strongly with the trained Pancreatic 368 cluster, while still remaining distinguishable from the WBC clusters, which demonstrates that the PScLE can learn the common lasing mode pattern features from one tumor cell model, and then generalize this knowledge to other unobserved tumor cells.

[0068] CTC detection requires a high specificity considering the rarity of CTCs. The DCLC model had a modularized design that allows the tuning of the classifier towards a high specificity mode when performing CTC detection. The trained PScLE was used to encode the original HLS dataset (where each datum contained the four lasing mode patterns of the cell) into the feature vector set. The classifier unit was trained on the feature vector dataset for binary classification. Figure 9A shows the receiver operating characteristic (ROC) curve of the cell classifier with an area under the ROC curve (AUC) of 0.9994. A threshold of 0.2 was set for the classifier to reduce the false positive rate. A threshold lower than 0.5 indicates that the model prefers classifying cells as white blood cells. The confusion matrix for the cell classifier is shown in Fig. 9B, where the sensitivity and specificity here refer to the performance of the deep learning model, without corrections from lasing rates.

[0069] To validate the performance of the proposed method, particularly that the DCLC model can correctly identify CTC class cells, the system was tested with spikedin samples. The process flow is shown in Fig. 10A. CTC class cells (Pancreatic 368) were stained with SYTO 9 and subsequently with DAPI, and mixed with white blood cells collected from Donor 7, whose sample had not been observed by the DCLC model. The white blood cells were stained only with SYTO 9. Therefore, the fluorescence signal of the DAPI channel can be used to identify whether a cell is a circulating tumor cell or a white blood cell. Fluorescence images were taken from the spiked-in sample with a separate optical filter set and measured biolaser signals from the same sample afterward. Finally, fluorescence images of the sample were used, where only CTC class cells have strong fluorescence signals, to validate the classification results of the DCLC model.

[0070] Figure 10B shows the counted CTC class cells in the biolaser channel against the fluorescence channel from three spiked-in samples and one control set. The CTC class cell counts from the two methods agree well with each other. After this, one- to-one validation is performed, where for each cell observed in the fluorescence channel, the lasing pattern of that cell is searched for in the biolaser channel. If that cell was correctly identified as CTC class by the DCLC, then it is referred to as a one-to-one match. Figures 10C and 10D show an example of such one-to-one matching. Define the ratio of matched cell number over the number of the CTCs counted in the fluorescence channel as “sensitivity” and the ratio of matched cell number over the number of the CTCs identified in the biolaser channel CTC count as “specificity”. Asshown in Fig. 10E, most of these rates are close to 1 in all four validation experiments, indicating that the biolaser method indeed performs accurate single-cell level CTC detection. However, note that the sensitivities are generally smaller than the specificities, falling below 0.8 in one experiment set. This is due to a reduced lasing rate of CTC class cells when they were additionally stained with DAPI.

[0071] In Figure 11 A, CTC enumeration on the samples from two patients with pancreatic cancer was performed, and then, as a demonstration of the zero-shot generalization ability of the method, with the samples of six lung cancer patients. Note the anticipated heterogeneity of the lasing modes from patient CTCs. Comparing CTC counts from the biolaser method with the traditional immunofluorescence staining method, a similar trend in CTC counts was found. This shows that the biolaser method can consistently detect the abnormalities of chromatins and differences of the microenvironment in CTCs, and, with the help of the DCLC model, can generalize this knowledge to CTCs from unobserved tumor sources. Figures 11C and 11 D show an example of a CTC detected by biolaser and IF respectively.

[0072] This disclosure focuses on the biolaser platform. In conjunction with the Labyrinth technology, the biolaser platform can achieve generalizable and antigenindependent detection of circulating tumor cells. A Labyrinth device can achieve approximately two orders of magnitude depletion of white blood cells (and up to four order depletion of white blood cells can be achieved with a double-Labyrinth device), and the biolaser method further filters out additional three orders of WBCs. Together, the integrated system accomplishes a cumulative white blood cell depletion of 5-7 orders of magnitude, while maintaining a sensitivity exceeding 90% for circulating tumor cell detection in validation experiments.

[0073] A high pumping power is needed to acquire the desired lasing rate of cells. However, high pumping powers would often cause signal intensities from cells with low lasing thresholds to saturate camera pixels that do not have large dynamic ranges. While saturated patterns of cells lose local information of the mode, they still contribute positively to the HLS of the cell and improve the performance of cell-type classification.

[0074] The foregoing description of the embodiments has been provided for purposes of illustration and description. It is not intended to be exhaustive or to limit the disclosure. Individual elements or features of a particular embodiment are generally not limited to that particular embodiment, but, where applicable, are interchangeable andcan be used in a selected embodiment, even if not specifically shown or described. The same may also be varied in many ways. Such variations are not to be regarded as a departure from the disclosure, and all such modifications are intended to be included within the scope of the disclosure.

Claims

CLAIMSWhat is claimed is:1 . A method for identifying a cell type of a subject using laser emissions, comprising: directing, by a light source, a light beam towards a biological sample, where the biological sample is disposed in a laser cavity; capturing, by a detector, spatial distribution of laser emissions from the biological sample; and identifying a cell type for the biological sample from the captured spatial distribution of laser emissions using machine learning.

2. The method of claim 1 further comprises staining the biological sample with a lasing medium prior to placing the biological sample in the laser cavity.

3. The method of claim 1 further comprise changing intensity of the light source to two or more different intensity values and capturing a spatial distribution of laser emissions from the biological sample at the different intensity values of the light source.

4. The method of claim 1 further comprises generating, by a first encoder, a feature vector from the spatial distribution of laser emissions from the biological sample, and identifying, by a neural network, a cell type for the biological sample from the captured spatial distribution of laser emissions using the feature vector.

5. The method of claim 4 further comprises determining position data of the biological sample in relation to center of the laser cavity; generating, by a second encoder, a position vector for the biological sample from the position data; and identifying, by the neural network, a cell type for the biological sample based in part on the position vector.

6. The method of claim 1 wherein the laser cavity is further defined as one of a Fabry-Perot cavity, a ring resonator, distributed Bragg reflector (DBR) cavity, or a photonic crystal laser cavity.

7. A method for identifying a cell type of a subject using laser emissions, comprising: directing, by a light source, a light beam towards a biological sample, where the biological sample is disposed in a laser cavity; capturing, by a detector, spatial distribution of laser emissions from the biological sample; generating, by a first encoder, a feature vector from the spatial distribution of laser emissions from the biological sample, determining position data of the biological sample in relation to center of the laser cavity; generating, by a second encoder, a position vector for the biological sample from the position data; and identifying, by the neural network, a cell type for the biological sample using the feature vector and the position vector.

8. The method of claim 7 further comprises staining the biological sample with a lasing medium prior to placing the biological sample in the laser cavity.

9. The method of claim 7 further comprise changing intensity of the light source to two or more different intensity values and capturing a spatial distribution of laser emissions from the biological sample at the different intensity values of the light source.

10. The method of claim 1 wherein the laser cavity is further defined as one of a Fabry-Perot cavity, a ring resonator, distributed Bragg reflector (DBR) cavity, or a photonic crystal laser cavity.

11. A laser imaging system, comprising a laser cavity configured to hold a biological sample from a subject;a light source configured to generate and direct a light beam towards the biological sample; a detector configured to capture spatial distribution of laser emissions from the biological sample; and a controller interfaced with the light source and detector, the controller generates a feature vector from the spatial distribution of the laser emissions and identifies a cell type for the biological sample from the feature vector using a neural network.

12. The laser imaging system of claim 11 wherein the controller changes intensity of the light source to two or more different intensity values and captures spatial distribution of laser emissions at each of the different light intensity values.

13. The laser imaging system of claim 11 further comprises a motorized stage configured to move the biological sample relative to the light beam of the light source, wherein the laser cavity resides on the motorized stage.

14. The laser imaging system of claim 13 wherein the controller determines position data of the biological sample in relation to center of the laser cavity; generates a position vector for the biological sample from the position data using a second encoder; and identifies the cell type for the biological sample based in part on the position vector.

15. The laser imaging system of claim 11 wherein the laser cavity is further defined as a Fabry-Perot cavity.

16. The laser imaging system of claim 11 wherein the light source is further defined as a laser.

17. The laser imaging system of claim 11 wherein the controller is further defined as a microcontroller.

Images