Non‐Invasive and Minute‐Frequency 3D Tomographic Imaging Enabling Long‐Term Spatiotemporal Observation of Single Cell Fate

Non‐invasive and rapid imaging technique at subcellular resolution is significantly important for multiple biological applications such as cell fate study. Label‐free refractive‐index (RI)‐based 3D tomographic imaging constitutes an excellent candidate for 3D imaging of cellular structures, but its full potential in long‐term spatiotemporal cell fate observation is locked due to the lack of an efficient integrated system. Here, a long‐term 3D RI imaging system incorporating a cutting‐edge white light diffraction phase microscopy module with spatiotemporal stability, and an acoustofluidic device to roll and culture single cells in a customized live cell culture chamber is reported. Using this system, 3D RI imaging experiments are conducted for 250 cells and demonstrate efficient cell identification with high accuracy. Importantly, long‐term and frequency‐on‐demand 3D RI imaging of K562 and MCF‐7 cancer cells reveal different characteristics during normal cell growth, drug‐induced cell apoptosis, and necrosis of drug‐treated cells. Overall, it is believed that the proposed 3D tomographic imaging technique opens up a new avenue for visualizing intracellular structures and will find many applications such as disease diagnosis and nanomedicine.


Introduction
Long-term spatiotemporal observation of single cells has been a long-standing effort, which helps understand cell fate decisions in biology and medicine. [1] The latest long-term imaging enabled the observation of novel subcellular structures named darkvacuole bodies in the cell division process and found their interplay with organelles before ultimately collapsing into the plasma membrane. [2] 3D optical imaging is the most accessible and convenient means for the observation of cellular structures and components, among many other imaging techniques like ultrasound rotational instability and low operating throughput for the current cell rotating methods by mechanical, [18] hydrodynamic, [19] optical, [20] and electrical [17a] means. In the optics design, phaseshifting, [21] and especially off-axis [22] based QPI methods, are widely used in 3D RI imaging of cells. Since off-axis holograms [23] were developed by Leith in 1962, more and more 2D/3D internal structures have been studied thanks to their solution to the twinimage problem. However, mechanical vibrations and air fluctuations affect the interferometric system sensitivity in retrieving the quantitative phase. To solve this problem, some commonpath QPI methods, such as Fourier phase microscopy [24] and diffraction phase microscopy, [25] which filtered part of the signal field as a reference beam from a diffraction grating, were invented by Michael Feld's lab. Further, the white light diffraction phase microscopy (wDPM) [26] adopted low temporal coherence light as white light illumination combined with common path configuration to reduce speckle noise in the reconstructed phase images. Thus, wDPM excels in imaging stability and phase sensitivity and has demonstrated imaging capability in cancer cells [27] and red blood cells. Generally, the potential of RI imaging has been demonstrated through the limited number of instantaneous frames in cell dynamics, cell growth, blood screening, tissue diagnosis, etc. [28] Some pioneering works based on QPI for RI imaging largely focused on imaging flow cytometry [6b,29] and have made exciting progress. However, the application of 3D RI imaging in long-term observation of cell fate for suspension cells, which are one big cohort, has not been reported, mainly due to the lack of effective integration of both efficient specimen control and high-performance QPI.
To bridge this gap, we report here a long-term 3D RI imaging system incorporating an acoustofluidic device to manipulate and culture single cells in a customized live cell culture chamber, and a cutting-edge wDPM module with spatiotemporal stability. The system intermittently records a stack of interferograms of an individual cell, which is cultured in situ and rolled on demand for more than one revolution in the acoustofluidic chip. Here, based on our previous work, [30] the acoustofluidic device is featured to be able to achieve stable rolling of large-scale cells to satisfy the stringent imaging need. The 2D interferograms are processed to reconstruct the 3D RI image, which is capable of revealing the cell structure by comparing it with confocal fluorescence imaging. We conducted 3D RI imaging experiments for 250 cell samples of different types and extracted RI-relevant properties as a demonstration to specify the cell types. Importantly, we performed longterm and frequency-on-demand 3D RI imaging of intracellular structures for K562 and MCF-7 cancer cells, visualizing their 3D structural changes at unparalleled high temporal resolution during normal cell growth, drug-induced cell apoptosis, and necrosis of drug-treated cells. The results suggest that cell growth, apoptosis, and necrosis exhibit completely different characteristics from the perspective of 3D RI imaging, creating research opportunities in more biological experiments. In particular, the system also features adjustable imaging frequency that can be set on-demand to observe time-critical events, as demonstrated to be 1 min in the 10-min execution phase of apoptosis and the end phase of necrosis. Overall, we believe that this non-invasive and minutefrequency 3D tomographic imaging platform opens up a new perspective for single-cell fate observation by label-free visualiz-ing whole-cell structures for a long time and with spatiotemporal stability.

Results and Discussion
As sketched in Figure 1a, the experimental setup mainly includes an acoustofluidic chip placed in a homemade live cell culture chamber to manipulate live cells, and an inverted microscope with an add-on wDPM optical module to record the interferograms of cells. For 3D imaging purposes, cells should be manipulated to perform out-of-plane rotation when the illumination is projected upright. We adopted an acoustofluidic device (for details, see our previous work) [30] with arrayed horseshoe structure through bubble-induced acoustic streaming fields to capture and rotate massive cells simultaneously (Figure 1b). We customized a wDPM module according to Popescu's work [26] to obtain interferograms of cells, as sketched in Figure 1a. The specifications of the optics were designed to ensure clear and stable fringes and to properly record the interferograms. [31] Once acquiring the interferogram, the phase map can be obtained by the off-axis Fourier-based algorithm. [32] Note that cells are commonly assumed as phase objects in the field, and their phase distribution exhibits great fluctuations. Correspondingly, there may normally exist 2 phase ambiguities between two adjacent points, which need to be resolved through phase unwrapping. Different unwrapping methods (Figures S1 and S2 and Video S1, Supporting Information) are evaluated, and the pathindependent, unweighted least squares phase unwrapping algorithm is adopted for high accuracy and high speed (Tables S1 and S2, Supporting Information). The unwrapped phase maps are then used to reconstruct the 3D RI image ( Figure S3, Supporting Information).

Noise and Accuracy Evaluation of Phase Reconstruction
In order to characterize the noise stability of the wDPM system, we first measured the phase noise over 10 s using the background interferograms with pixel size 500 × 500 (or 21.7 × 21.7 μm 2 ). The overall spatiotemporal noise has a standard deviation of 0.9 nm (timescale 10 s), which is slightly better than but at the same order of magnitude as other wDPM systems. [26] The error of the reconstructed 2D phase was testified with a ϕ3 ± 0.2 μm polymethyl methacrylate (PMMA) bead to be 6.6%. These evaluations (Figure S4, Supporting Information) validate the high stability, sensitivity, and accuracy of wDPM.
We also used atomic force microscope (AFM) imaging to verify the accuracy of the reconstructed phase. To do so, we fabricated a 500-nm tall square micropillar quartz wafer (RI = 1.45) to serve as a standard sample. Note that as the lateral size of the sample increases, the accuracy of the phase measurement decreases due to the insufficient spatial coherence of the optical system. [31] Since most single cells in our experiment are below 20 μm in diameter, the side length of the square micropillar of 20 μm was used in verification. The height distributions calculated by Equation (6) and measured by AFM (MFP-3D-SA, Asylum, USA) are shown in Figure 1c  www.advancedsciencenews.com www.small-methods.com plots the profile curves along the center line of the micropillar for AFM and wDPM imaging. It was found that the wDPM result follows the AFM measurement firmly. Actually, the height value measured by AFM is 500.49 ± 11.88 nm, while the wDPM result is 505.18 ± 51.28 nm. Taking the AFM imaging result as the ground truth, the phase measurement error of wDPM is below 6.7%. Since the phase accuracy of wDPM is inversely related to the lateral size of the sample, [33] for cells or specimens smaller than the square pillar (i.e., 20 μm), the error would be far less than 6.7%. It is worth noting that the halogen lamp commonly used in the wDPM setup would not be able to achieve such high accuracy because it lacks spatial coherence like the supercontinuum laser (SCL) used here. Note here, we did not quantify the spatial coherence since the phase accuracy was high. Readers may refer to the literature [33] for such a quantification method when necessary.
We further evaluate the phase consistency using the same quartz wafer. Phase consistency is defined as the variation of phase values supposed to be uniform at different locations of an even surface, and is mainly related to the poor spatial coherence [33] of the conventional halogen lamp illumination system when the specimen is positioned differently in the field of view. To evaluate the consistency, we moved the micropillar horizontally in the field of view to four different locations and obtained the four corresponding phase maps. The profile curves along the same center line of the micropillar are plotted in Figure 1f, showing subtle differences. The average phase difference caused by the micropillar displacement is 0.09 rad (or 4%). This verifies that our wDPM system with SCL has uniform illumination and high spatial coherence. The benefit is that the phase map is reliably accurate even if the sample (i.e., cells) has translation during rotation, which relaxes the working condition of the rotating platform. Phase consistency may be improved by further reducing the diameter of the 0th-order mask of the spatial filter (SF) to ensure a more uniform reference beam [33] in a compromise with the 1st-order intensity.
The imaging resolution of our system is theoretically calculated to be 0.71 and 3.67 μm in the lateral and axial directions, respectively, according to their equations (Equations (S2) and (S3), Supporting Information). The theoretical values are very close to the experimental values (0.72 μm in the lateral direction and 4.26 μm in the axial direction, see Figure S5, Supporting Information). The spatial resolution in the subsequent 3D RI maps may be further evaluated by using polystyrene beads with known RI values. [14] Generally, we can increase the resolution by using a higher numerical aperture (NA) objective lens or a shorter wavelength of illumination.

Cell Rotation
In the experiment, we injected a large number of cells (5 × 10 6 cells/ml) into the chip, and then kept the fluid still by closing the fluid ports completely, to ensure that at least dozens of cells were stably and purely rolling at the same time in different locations. Cells did not behave to couple displacement with rotation, unlike the cases in flow cytometry where the cells flow forward rapidly. Instead, they rotated stably with the microvortices induced by acoustic fields in the acoustofluidic device. However, to allow single-cell 3D RI image reconstruction, we only focused on one single cell with a region of interest (ROI) of 500 × 500 pixel (or 21.7 × 21.7 μm 2 ) in the field of view and recorded its interferograms for at least one revolution at 200 Hz at an angle interval of ≈0.36°. The cell rotation speed (measurement method, see Figure S6a, Supporting Information) was controlled at about 12 rpm (0.2 rev s −1 ), by setting the voltage (1.7 V, Figure S6b, Supporting Information) applied to the actuator in the acoustofluidic device. Once finishing one cell recording, we switched the ROI to another cell by moving the microscope stage. Note in the experiment, the arrayed acoustofluidic chip ensured that multiple cells roll (12 rpm) at the same time (Video S2, Supporting Information). This allows ample rolling cell samples to be available for video recording and thus improves the efficiency of the experiment.
In this work, we relied on the acoustofluidic device, which was our previous work, [30] for cell rolling. Here, we do not want to give too much detail on the device, since it was reported earlier. Instead, we give some detailed information on the two key issues in rotation, including rotation stability and angle registration, which are critical for 3D RI imaging.
We confirmed that cell rotation has little off-axis and speed fluctuation by analyzing the bright-field image frames of a rotating cell. For off-axis evaluation, we selected one feature point on the equator in the first image frame and tracked its positions by template matching in the two frames after one and two revolutions (Figure 2a). The feature point changed its latitude angle from 0°to 1.15°and 0.87°only. On the other hand, we also examined the cell diameter along a fixed direction (i.e., the initial rotating axis, y-axis) in each image frame during one revolution, and plotted diameter versus time in Figure 2b. These data points have a very small ratio of standard deviation over the mean (1.17%), showing negligible off-axis.
To check how stable the rotation speed was, we deliberately examined a cell with stably visible and thus trackable feature points. These rolling points are hardly identified in normal cells because of complex imaging effects associated with angle-varying diffraction and diffusion. Thus we selected an abnormal cell shown in Figure 2c and tracked it through template matching the feature points selected near the equator in every ten frames. Transferring the position coordinates to the longitude angle, we then obtained the interval angle between every ten frames to be 25.50°, 29.28°, 26.53°, and 25.23°. For these values, the ratio of standard deviation over the mean is 6.94%, indicating a quasi-constant rotation speed. It should be stated that these poor feature points probably cause an adverse effect on rotation speed stability. Actually, if we calculate the rotation speed by matching the whole cell structure after one full revolution, the value is more stable (<3%, Figure  S6a, Supporting Information). Furthermore, we obtained three 3D RI images (Figure 2d) for the same cell during three consecutive revolutions. The correlation coefficient between every two neighboring images is 98.45% and 99.44%, indicating high reproducibility of rotation.
We also performed angle registration for 3D RI reconstruction to compensate for any cell rotation speed variation due to cell heterogeneity. Here, because the rotation was not coupled with displacement, we did not follow the relatively complex method. [17c] Indeed, we simply used all the 2D phase maps during >1 revolution to calculate the structural similarity [34] w.r.t. the first 2D phase map. Once another peak was found, the corresponding The diameter of the cell along the y-axis has a variation of 1.17% during one revolution, showing the stability of the rotating axis too. c) Tracking a feature point on a rolling cell to evaluate its speed stability. The speed variation is 6.94%. d) The three 3D RI images recovered from the 2D phase maps of the 1st, 2nd, and 3rd revolution of a rolling cell. The correlation coefficient between each two neighboring 3D RI images is 98.45% and 99.44%, indicating high reproducibility of rotation. e) Full revolution is determined by finding the structural similarity peak during rotation. The 2D phase maps are used in calculating the structural similarity.
phase map marked a full revolution ( Figure 2e). With the total number of 2D phase maps obtained, all the phase maps in one revolution were linearly assigned a rotation angle.

Cell 3D RI Imaging
We successfully demonstrated label-free 3D RI imaging of suspended cells. Figure 3a-f shows the interim and resultant 3D imaging of one HaLa cell. The interference fringes are clearly seen from the original interferogram (Figure 3a), validating the feasibility of wDPM optics. Actually, the fringes maintained stably due to the use of the common-path geometry, which ensured the long-term 3D observation and measurement of cells. The unwrapped halo-free phase map ( Figure 3b) extracted from the original interferogram shows less noise without the laser speckle, which is attributed to the use of white light illumination. Without laser speckle, phase reconstruction suppresses spurious noise and ensures high accuracy of phase and subsequent 3D RI reconstruction. The 2D RI distribution of the slice is exhibited in Figure 3c, which indicates the decoupling ability of optical diffrac-tion tomography (ODT) to accurately calculate RI from the phase projections. The 3D RI distribution of the cell can be observed in different ways of representations, as shown in Figure 3d-f, which provides insights into the intracellular structure from the aspect of 3D RI distributions (Video S3, Supporting Information, cell rotation speed 12 rpm).
The average RI of the HeLa cell was measured to be 1.374, which is in agreement with the previously published data. [35] By rendering the 3D RI image using three colors to represent three ranges of the RI values, we can get some clues to discern the different cell structures like the nucleus and cytoplasm. As Figure 3e shows, the red region may represent the nucleus, the green may contain endoplasmic protein-rich organelles in the cytoplasm, and the light blue may represent the cytoplasm. Note here the segmentation threshold of 1.37 corresponds to the sample's average RI, while 1.35 is determined as the midway of the low range from [1. 33, 1.37].
We further compared the 3D imaging results of the same HeLa cell by confocal microscopy and our wDPM method. Basically, the HeLa cell was stained in red using RedDot 1 (Biotium, USA) for the nucleus and in green using CMFDA (Yeasen, China) for the  Figure S7, Supporting Information) such that the volume ratio of the nucleus given by RI imaging is the same as confocal imaging. i) Comparison of a particular A-A section (i.e., equator) of (g) and (h) with close values of the area ratio.
cytoplasm. The cell was rotated first to obtain the interferogram stacks for different angles with the wDPM module. The same cell was then placed under a laser scanning confocal microscope (Ti2 A1R, Nikon, Japan) to fully settle down, which took a short period of time (i.e., ≈3 min). Next, the cell was scanned by axial scanning for ≈3 min to obtain 44 dual-channel fluorescence image stacks, which were then reconstructed by a 3D deconvolution algorithm using Imaris 9.2.1 ( Figure S7, Supporting Information) for its 3D confocal image (Figure 3g). To allow a direct comparison, we segmented the 3D RI image into two regions (Figure 3h), to match the two fluorescence labels of the nucleus and cytoplasm in confocal imaging. The threshold for the nucleus was determined through the RI histogram ( Figure S7, Supporting Information) such that the volume ratio of the nucleus given by RI imaging is the same as confocal imaging. Due to the experimental setup limit, the confocal imaging module was not coupled with the 3D RI imaging setup on the same microscopy, so we were not able to align the two 3D images well. But it can still be seen that the RI image roughly resembles the confocal image (Video S4, Supporting Information). This observation is also evident if we compare the slices containing the nucleus extracted from the same equator of the two 3D images (Figure 3i). It is worth noting that the selection of the RI range for the nucleus resulted in a finer boundary compared to the smooth boundary processed by con-focal fluorescent microscopy. This indicates an advantage of RI imaging for revealing the uneven distribution of proteins in the nucleus or other organelles with no need for complex labeling against labeling microscopy.

Cell Geometric and Biochemical Parameters Measurement
Cell geometric parameters have the potential to identify cell types and perform early disease diagnosis. [36] Recent high-quality works [29a,37] on label-free and high-throughput imaging flow cytometry have revealed cell population heterogeneity, showing the great value of QPI in cellular biophysical characterization. [37] Here, we extracted the geometric parameters (e.g., volume, surface area, and sphericity index) from the 3D RI image of the cell. For each rotation angle, the cell contour is extracted from the unwrapped phase image instead of the interferogram, since the cell contour is not clear due to the deformed interference fringes. The cell contours in one revolution are used to form the 3D morphology of the cell by a standard alpha-shape algorithm, which has been reported in our previous research. [38] The volume (V) was obtained by counting the number of voxels inside the cell 3D morphology. The surface area (S) was calculated from the boundary of 3D morphology. The cell sphericity index is defined as SI = (36 V 2 ) 1/3 /S. Cell biochemical parameters are essential for studying certain cell metabolic activities (e.g., cell division, infection). [28b] The RI of a biological specimen exhibits a strong linear dependence on the concentration of organic molecules within the specimen especially the protein concentration, specifically, the dry mass density (DMD) was given [39] by where n(x, y, z) is the cell 3D RI value obtained by ODT reconstruction, n m is the RI of the surrounding medium (here n m = 1.33 for the used medium), and is the average specific RI increment and set to 0.19 mL g −1 according to literature. [40] Note in this paper, to allow easy reading of the RI and DMD values, we present them both although they are linearly related. The to-tal cell dry mass (DM) can then be acquired by integrating dry over the entire volume of the cell. Note that the dry mass quantitatively reflects the content of biological macromolecules, such as proteins, which is an important metabolic indicator, and can be used to quantify cell growth non-invasively. [41] The above cell geometric and biochemical parameters of different cells were obtained to demonstrate the potential of 3D imaging. Enabled by the ability of the system to conveniently obtain data, we analyzed 50 samples for each cell type of MCF-10A, MCF-7, HL60, HeLa, and A549. In particular, for each cell type, all the samples came from the same batch of cells to avoid possible variations in the microfluidic conditions. The geometric and biochemical parameters for the five types of cells are plotted in Figure 4a-f and summarized in Table S3, Supporting Information. For MCF-7, HL60, and HeLa cells, which were reported in the literature, [35,42] the results of mean RI had a very small www.advancedsciencenews.com www.small-methods.com discrepancy (4-6%), verifying the accuracy of our measurement to some extent.
The RI values for the tested cell samples fall in the range of 1.35-1.41 and exhibit both intertype and intratype differences. This may be attributed to the difference in intracellular organelle presence and distribution, specific protein expression, and accumulation. [8b] In addition, the geometric properties of the tested samples are also different, suggesting that any one of the above parameters cannot solely specify the cell type. An interesting finding is that smaller cells (MCF-10A, volume: 1046.2 ± 261.9 fL, RI: 1.3806 ± 0.0065; HL60, volume: 904.9 ± 191.6 fL, RI: 1.3797 ± 0.0080) have larger mean RI compared to larger cells (MCF-7, volume: 2600.9 ± 663.6 fL, RI: 1.3771 ± 0.0043; A549, volume: 2369.7 ± 716.5 fL, RI: 1.3768 ± 0.0043), this may be due to the smaller ratio of the core (cytosolic portion) to shell (cell membrane) for smaller cells. It is worth noting that the geometric parameters of MCF-10A and HL60 cells are not statistically different, but the RI-relevant parameters (RI, DMD, and DM) have noticeable differences (see the p-values in Figure 4a-f). This indicates that multiple biochemical features collectively are more capable of comprehensively characterizing cells.
To demonstrate that multi-parameter characterization of the same cell provides a richer dimension of information to reflect the cell heterogeneity, we presented the original data and their principal component analysis (PCA) results. Figure 4g shows the distribution of the six parameters after normalization for all five types of cells. The results reveal the difference between the five cell types, but the parameters that reflect the difference most are volume, RI, and DM. Actually, the PCA result shown in Figure 4h indicates that the first three principal components of the six parameters are able to distinguish four of the five tested cell types, except for A549 cells from MCF-7 cells. As these two cell types show extremely high similarities in morphology and biochemical parameters, more characteristic parameters, such as electrical (membrane capacitance, conductivity of cytoplasm) and mechanical (shear modulus) parameters, [43] may be required to distinguish them.
3D RI imaging allows us to compare the internal structures of cells. Here, we compared the human breast cancer cell (MCF-7) and its normal counterpart human normal breast epithelial cell (MCF-10A). Their 3D RI images are shown in Figure 4i,j. It was noted that the high-RI distribution is denser in the central region of MCF-10A, while the high-RI distribution is denser in the outer region of MCF-7 (Video S5, Supporting Information). This pattern may or may not be true since the data was based on one single sample, thus needing thorough investigation in the future. But the spatial RI distribution is likely to inspire new trials in the investigation of cancer and non-cancer cells.

Long-Term 3D RI Imaging for Growing Cells
Observing live cells for a long time and monitoring the changes, especially their 3D structures, helps to understand how individual cells regulate their growth. [42,44] We used the leukemia cell line K562 in the experiment as it grows in a suspended state, which facilitates rotation operation. The cell suspension was first injected into the acoustofluidic chip fixed in a petri dish. Then the dish was kept in a homemade chip-compatible cell culture incu-bator ( Figure S8, Supporting Information), which maintained the proper culture conditions of 37°C and 5% CO 2 . During culture, the microfluidic chip was infused at a very slow rate with a fresh culture medium to provide the necessary nutrients for the cells. The petri dish was filled with culture medium and covered with a cover glass to reduce the adverse effect of evaporation. We rolled the K562 cell every 20 min to record the wDPM interferograms to resolve its 3D RI image. The whole experiment lasted for 5 h. The illumination was turned on only during interferogram recording (i.e., 10 s) to minimize the photothermal effect on the cell sample and fluid.
We show the 3D RI images of the K562 cell at an interval of 1 h in Figure 5a (Video S6, Supporting Information). In order to clearly reflect the structural changes during cell growth, we rendered the cell in four different colors, each corresponding to a specific RI range presumably associated with a cell organelle (i.e., nucleus, nucleolus, cytoplasm, high RI organelles in the cytoplasm). In doing so, we determined the RI ranges by imaging one nucleus-labeled K562 cell with our system and confocal fluorescent microscopy. The nucleus and cytoplasm RI range is determined through the RI histogram ( Figure S9, Supporting Information) such that the volume ratio of the nucleus given by RI imaging is the same as confocal imaging. The nucleolus RI range is defined to be the top 15%, 20%, and 25% RI among the nucleus, and the RI range for high RI organelles in the cytoplasm is defined to be the top 3%, 5%, and 7% RI among the cytoplasm. From the resultant nine 3D RI images for hour 0 ( Figure S10, Supporting Information), we observed no big difference in revealing the compartments of the cell with different thresholding values. Thus, we took the 20% RI among the nucleus as the nucleolus, and the top 5% RI among the cytoplasm as high RI organelles as an example, to demonstrate the likely estimation of the dynamic trend of cell fate. We also confirmed the ellipsoidal shape of the nucleolus through staining ( Figure  S11, Supporting Information). As shown in Figure 5a, the nucleolus (in red) changes less, while the high RI organelles in the cytoplasm (in yellow, e.g., endoplasmic reticulum) gradually increase and finally spread in the cytoplasm within 5 h, depicting the chronological protein synthesis in different parts of the cell. Note that these changes are extremely challenging for confocal microscopic imaging because the photobleaching problem would stop long-term observation of the stained cell sample, and are not reported by existing RI imaging techniques. Through our unparalleled long-term, label-free 3D imaging capability, we can dynamically observe the specific mechanism of cells in-depth, and distinguish the subtle life processes of cells such as cell necrosis, apoptosis, infection, division, and differentiation.
We quantified the parameters over time during the growth of the K562 cell. Regarding the morphology (Figure 5b), the volume of K562 increased with time, while the surface area was almost unchanged. This means that the suspension cell became more spherical during the growth process, which is also reflected by the sphericity index. Figure 5c shows that the average RI of the K562 remained almost unchanged (RI fluctuation ≈0.004) in 5 h, while the variance had an increasing trend, which may be related to the local protein synthesis in the cell. The time-varying dry mass curve (Figure 5d) also clearly reveals the cell growth. Interestingly, this leukemia cell K562 appears to be growing at a rate of 18.7 pg h −1 , which is 2.5 times more than HeLa cells. [26] www.advancedsciencenews.com www.small-methods.com

Anti-Cancer Drug Effect Observation at Minute-Frequency for Single Cell
Doxorubicin (DOX) is an anti-tumor antibiotic drug that can inhibit the synthesis of RNA and DNA in different cell lines such as MCF-7. In order to visualize cell-drug interaction, we treated MCF-7 cells with DOX at a concentration of 50 μg ml −1 for 4 h. Then we replaced the DOX with a culture medium for continuous observation of the cell in the homemade cell culture incubator. During the observation of apoptosis, we applied an imaging scheme with a long-to-short time interval for interferogram recording. In particular, initially, we used an interval of 10 min for interferogram recording and checked the change to the unwrapped halo-free phase map. If the change was small, the apoptosis was thought of as not starting and we maintained the interval. Otherwise, we changed the interval to 1 min and kept it for the next 10 min. This is because the MCF-7 cell is known to complete the execution phase of apoptosis within 10 min. [45] We then obtained the 3D image series (Video S7, Supporting Information) for the anti-drug-induced cell apoptosis. It is noted we could not always use a short interval because the laser source may damage the cell or the stability of the air bubble. The on/off switching of the laser source is experimentally verified as effective for cell viability and rotation stability.
In order to accurately analyze the spatiotemporal changes of the cell structure during this 10-min execution phase, we performed voxel matching (Supporting Information) for each of two adjacent 3D RI images of this period. After voxel matching, the 3D RI image was redrawn with the corrected rotation axis. For each time being, its 3D RI image (displayed with two RI ranges for easy analysis) and nine 2D RI slices evenly spaced along the latitude are shown in Figure 6a, where we found and marked two apoptosis-relevant regions by arrows. The two regions exhibit the same mode of time-varying RI transform, that is, from a rather uniform RI to enhanced loci, then to a low RI around the RIenhanced loci. This mode is more obvious for region #2, which was recorded entirely from its onset till the end in the process of 10 min. The first transform (region #2 from 81 to 85 min) can be attributed to the condensed chromatin and DNA fractures during the execution of apoptosis, [46] which is observed via classic fluorescent microscopy. The second transform (region #1 from 81 to 85 min and region #2 from 85 to 87 min) is observable around the granular particles with high RI. The reason remains unclear, but we speculate that it may be related to some mechanism of enzymatic digestion of proteins during apoptosis. It is also nontrivial to point out that apoptosis was found to take place from multiple sites of the cell, and these sites may or may not necessarily synchronize their action, indicated by the asynchronous behavior of the two regions here. Only with our imaging system, we are able to observe the cell structural change at high temporal resolution for the first time, which is not possibly achievable by the state-of-the-art confocal fluorescence technology [47] (≈6 s to scan a cell). Our imaging system may provide new thoughts for biologists to investigate cancer cell development and progression and contribute to drug development and screening.
Apart from the transient events, the long-term experimental results show that after DOX treatment, the MCF-7 cell underwent apoptosis after a period of growth. As shown in Figure 6b, RI, volume, and DM of the cell exhibited an obvious jump along with the apoptosis. Before apoptosis, the cell turned to grow normally, as indicated by the stable or increasing curves of the three variables. Specifically, during apoptosis, RI dropped from 1.3516 to 1.3410, and volume dropped from 6166.8 to 4399.8 fL. Interestingly, DM kept increasing in the first 80 min at a rate of 29.8 pg h −1 , then dropped sharply from 571.47 to 177.55 pg with a total change of 68.93%, and continued reducing.
To compare apoptosis and necrosis, we monitored the DOXtreated MCF-7 cells directly without the live cell culture chamber to induce cell necrosis (Video S8, Supporting Information). Similarly, we conducted the long-to-short time interval imaging and performed voxel matching to correct the rotation axis. We plot the 3D RI images and 2D RI slices in Figure 7a, and the cellular property indicators in Figure 7b. It can be seen that the difference between apoptosis and necrosis is paramount. The most obvious phenomenon is that the RI value keeps dropping while volume keeps rising, both with a sudden acceleration in the end for necrosis. Though more data points are required to draw a conclusion on the biological processes, the capability of our system to differentiate and characterize cell fates would be valuable to biomedical applications.
In the literature, there are contradictory reports on the relationship between nuclear RI and plasmatic RI. Guck group [21a,35] and Wax group [48] reported critical examination on several cell lines including MCF-7, A549, BEAS-2B, HVE, HL60, Jurkat, and HeLa cells. They reached the conclusion that for these cell lines, nuclear RI is greater than plasmatic RI. Their work looked solid. On the other hand, other groups using QPI for 3D RI measurement obtained the opposite conclusion. For example, Ferraro group reported the results for NIH-3T3 cells, [22b] human neuroblastoma cells (SK-N-SH cell line), and MCF-7 cells. [29b] Shaked group reported the results for MCF-7 and three types of white blood cells, [17a] red blood cells, white blood cells, and HT-29 (human colorectal adenocarcinoma) cells, [49] 54 HT29-GFP cells. [50] Izatt group involving Wax reported the results for MCF-7 and HT-29 cells. [51] Park group reported the results for frog erythrocytes. [52] These contradictory data indicate that one cell sample of the same type can have a higher nuclear RI, while the other sample may have a higher plasmatic RI. In particular, the measurement results of 54 HT29 cells [50] by both interferometric phase microscopy and confocal fluorescence microscopy showed that for some cells, the nuclear RI is higher than that of the cytoplasm and for others, it is lower, but in most of the cases, the nuclear RI is higher. In this paper, our focus was not on their exact relationship in values. For demonstration, we conducted confocal microscopic imaging of HeLa ( Figure 3) and K562 ( Figure 5) as a reference to separate the nucleus and cytoplasm. These two cell samples happened to have higher RI values in the nucleus than in the cytoplasm. For other 3D RI images of the cell samples (Figures 4i,j, 6, and 7), we did not have confocal images as a reference, so we were not able to tell exactly where the nuclear region was. Therefore, we did not label the nucleus in those cells and directly presented the RI ranges. For those who want to identify the nucleus, the computational segmentation based on the statistical inference (CSSI) method reported by Ferraro group [29b] is recommended.
The illumination settings in this work did not cause photodamage to the cells. Here, according to our laser source specifications (average power 0.2 W, repetition rate 40 MHz, pulse width 100 ps, beam diameter 2 mm), the irradiance is calculated to be 1592 W cm −2 , seven orders of magnitude lower than the threshold (6 × 10 10 W cm −2 ) for cell blebbing, myelin sheet damage, [53] which has the lowest threshold among other cell damage like morphological cell damage, [54] chromosome dissection. [55] On the other hand, in a systematic photo-damage study with an irradiance of two orders of magnitude higher than ours, cells were reported to sustain without noticeable damage for 15-s exposure. [56] In contrast, the laser-on time for one 3D RI image was 10 s in our study. This was a very safe value. Actually, we did not observe cell abnormality during the 5-h period of culture and imaging (e.g., Figure 5), which was the longest test in this paper. In addition, the fact that the dry mass kept increasing (e.g., Figure 5d, right) during this period provided another evidence that the cell was alive. Based on the above analysis, our current illumination settings can be expected to work very longer than 5 h, with little risk of photodamage. Note the imaging system is independent of the cell state. Like what we demonstrated in the paper, the system worked for the cells normally cultured in the live cell culture chamber, and those deliberately configured toward death via apoptosis and necrosis. Nevertheless, for applications where cell viability is critical, it is recommended to select healthy cell samples by labeling and ruling out dead cells.
The quality of the 3D RI maps can be improved in two main aspects. First, we relied on ODT to calculate the RI stack from the phase map. ODT considers the diffraction effect, but backscattering is not available in the interferogram. Partial information is inevitably lost during this calculation and artifacts would be produced. Moreover, the single-axis rotation would cause missing data in the angular spectrum at the two poles of a spherical object. [17a] This may be the reason for concentric circles. To address this information loss issue, regularized iterative algorithms [57] or deep learning methods [58] may be used. Second, we used Rytov first-order approximation to reconstruct the 3D RI maps. Because the Rytov approximation is a solution under the linear scattering model, the cell samples need to satisfy the assumption of weak scattering. However, the suspended cells were quite thick (10-μm scale), which could have strong scattering. Hence, artifacts were produced. To overcome this drawback, a more accurate nonlinear scattering model, such as the Lippmann-Schwinger model, [59] can be adopted to improve the reconstruction quality.

Conclusion
In this study, we demonstrate a 3D RI imaging configuration that uses wDPM for highly sensitive and stable interferometric tomography along with acoustofluidic rotation of single cells. The acoustofluidic chip simultaneously rolled multiple cells stably, hence facilitating on-chip sample preparation and manipulation. The constructed wDPM system recorded interferograms of the rolling cell for more than 360°with improved spatial and temporal phase sensitivity. The spatiotemporal system noise had a standard deviation of 0.9 nm, the phase measurement error was no more than 6.7%, and the phase consistency error was no more than 4%. The cell 3D RI image reconstructed from the interferograms was verified by confocal imaging to be able to reveal intracellular structures. We then extracted the morphological and biochemical parameters of five types of cells through their 3D RI distribution for comparison in the characterization of cells. Among the results, the normal breast epithelial cells MCF-10A and its corresponding cancer cells MCF-7 showed significant differences in volume and DMD, which demonstrates the potential for cancer diagnosis. We obtained the 3D RI image series of the K562 suspension cell in the course of 5 h of on-chip culture to reveal the structural changes during cell growth. Importantly, longterm and frequency-on-demand 3D RI imaging of MCF-7 cancer cells suggested different characteristics during drug-induced cell apoptosis, and necrosis of drug-treated cells, calling for in-depth investigation in the future.
In a larger context, our 3D RI imaging system highlights the advantage of off-axis interference, partially coherent illumination, and common-path geometry. The off-axis interference allows for a high acquisition rate (200 fps) to capture rapid cellular dynamics. The white-light illumination and common-path geometry allow for speckle-free and nanometer-path-length stability, which ensures long-term imaging with accurate phase reconstruction. Moreover, the non-invasive acoustofluidic manipulation technology allows for long-time stable rolling with little damage to cells. These improvements allow us to observe cell fate at frequency-on-demand (e.g., 1-10 min intervals for more than 5 h), bridging the gap between long-time monitoring (e.g., >5 h), and rapid imaging (e.g., 0.1-5 min). Empowered by these capabilities, the label-free long-term 3D RI imaging approach holds great potential in functional cell imaging that has been recently proposed, [60] by providing rich spatiotemporal information of live cells useful for enhancing the specificity and functions of cell organelles and biological macromolecules. [61] Overall, the configuration offers an effective platform for labelfree 3D imaging and intracellular characterization of single cells and is expected to be a powerful tool in many biological and medical applications such as disease diagnosis and nanomedicine.

Experimental Section
Cell Culture and Preparation: Adherent cells, including HeLa, MCF-7, MCF-10A, A549, and HL60 of 8-18 μm in diameter, were used in the experiment. These cells were cultured using Dulbecco's modified eagles medium (DMEM) (Gibco, Grand Island, NY, USA) with 10% fetal bovine serum (FBS) (Gibco) and 1% penicillin/streptomycin (Gibco) in a 5% CO2 and 37°C incubator (LabServ CO150, Fisher Scientific, USA). Adherent cells were rinsed with PBS (Gibco) twice and then lifted off by treating with trypsin for 3 min. The cell suspension was washed three times by centrifuging at 300 g for 3 min, removing the supernatant with a pipette, and resuspending the cell pellet in PBS solution with different cell concentrations according to experimental requirements. Suspension cell K562 (erythroleukemia cell line) was also used in the experiment. K562 cells were cultured using RPMI-1640 (Gibco, Grand Island, NY, USA) with 10% FBS, and 1% penicillin/streptomycin in a 5% CO2 and 37°C incubator. Unlike adherent cells, K562 cells could be centrifuged directly to obtain a high concentration of about 10 6 cells/mL.
Chip Fabrication and Manipulation: The microfluidic chip in Figure 1b was fabricated using standard soft lithography and replica molding technique, as reported before. [30] First, air bubbles were generated inside the horseshoe voids by flowing through the microfluidic chamber with buffer. Then the cell suspension was injected into the chamber using a syringe by an automated syringe pump (KD scientific, Legato 270, USA), and the piezoelectric buzzer (FT-12 T-18.5E, Yuansheng Electronics Co., Ltd., China) connecting to a function generator (AFG 3052C, Tektronix, USA) was activated with a sinusoidal wave to oscillate the bubbles. Around each vibrating bubble, there would generate microvortices to trap and roll cells. A proper frequency in the range of 20-200 kHz was tuned for the out-ofplane rotational manipulation. Cell rotation speed was controlled to the proper value (12-120 rpm) by changing the voltage applied to the piezoelectric buzzer over 1-10 V pp . For better spatial resolution in imaging, a low rotation speed of 12 rpm (or 0.2 rev s −1 ) was used in the experiment.
wDPM Imaging Optics: The wDPM optical setup is shown in Figure 1a. An SCL source (SC-PRO-15, YSL, China) was employed to obtain spatially coherent white light illumination, with advantages in coherence and power. [33] Through filtering (FGS550, Thorlabs, USA) the invisible light out of the laser source, its center wavelength was tuned to 660 nm. At the image plane of the inverted microscope (Ti-U, Nikon, Japan), a customized diffraction grating (110 lines/mm) was placed to generate multiple diffraction orders containing full spatial information about the object. The 0th and first-order beams were then isolated with a customized pinhole filter as the SF at the Fourier plane of lens L1. The 0th-order beam was spatially low-pass filtered as the reference field, whereas the first-order was completely passed as the object field. The diameter of the pinhole (0th order mask) at the Fourier plane was set as 19 μm to ensure the uniformity of illumination in imaging. Lens L1 (AC508-080-AB-ML, Thorlabs, USA) and L2 (AC508-300-AB-ML, Thorlabs, USA) were achromatic to minimize chromatic dispersion, and they together formed a highly stable Mach-Zehnder interferometer. The two beams were interfered and captured by a CMOS camera (Prime BSI, Teledyne Photometrics, USA) at the image plane of the L1-L2 system. Throughout the experiments, the microscope was equipped with a 40× objective lens with an NA of 0.6. The L1-L2 lens system yielded an additional magnification of 3.75. To prevent stray light, shading lens tubes were used to cover the light path of the wDPM module.
Phase Reconstruction: For a given scene, the interferogram acquired by wDPM can be described as [32] I (x, y) = I 0 (x, y) + I 1 (x, y) + 2 √ I 0 I 1 cos [ x + (x, y)] where I 0 and I 1 are the image intensity of the 0th and first order, and is the phase map. The spatial frequency, = 2 /Λ, is due to the grating, and Λ is the grating period.
To calculate the optical path difference (OPD) of the cell at each angle, first, the background interferogram I bk was obtained without the cell in the scene. Then the interferogram with the cell in the scene at a certain angle was obtained and denoted by I ob . The phase map can be obtained by the off-axis Fourier-based algorithm [32] as www.advancedsciencenews.com www.small-methods.com The unweighted least squares phase unwrapping algorithm in Equations (4) and (5) was developed to solve the phase wrapping problem.
where W[] defines a wrapping operator that wraps all values of its argument into the range (− , ] by adding or subtracting an integral multiple of 2 radians from its argument. To eliminate the halo effect, the Hilbert transform was performed on the unwrapped phase and then directional filtering was performed along three directions (0°, 45°, 90°) one by one, as reported in the literature. [62] Removal of the halo around the specimen enabled more accurate extraction of cell contours.
The OPD map with the cell in the scene is given by