Abstract

The spectral absorption of biomolecules holds immense potential for revealing molecular identity, biological functions, and clinical diagnostics. Here, we present color-resolved third harmonic generation microscopy (cTHGM), a noninvasive method for molecular imaging based on absorption-enhanced THG response. Leveraging one single broadband femtosecond laser beam, cTHGM captures subtle absorption variations, including ~2-nanometer shifts in the Soret band, enabling precise distinction of glycated hemoglobin (HbA1c) from hemoglobin with negligible phototoxicity. We demonstrate that cTHGM measures HbA1c fractions in single red blood cells (RBCs) in vivo and ex vivo, providing noninvasive HbA1c measurement and insights into HbA1c distribution at the cellular level. In addition, cTHGM reconstructs historical glycemic trajectories by decoding single-RBC HbA1c distributions, offering a retrospective view of glycemic variability over months. This innovative method combines label-free imaging, high spatial and spectral resolution, and noninvasive manners, making it a promising tool for diabetes management, glycemic variability monitoring, and broader applications in precision medicine.

INTRODUCTION

In the era of precision medicine, blood glucose monitoring plays a vital role in personalized care, not only in diabetes management (1, 2) but also in broader applications such as tumor risk assessment and areas of health optimization. While notable progress has been made in tracking glycemic levels through various methods, no current available methods are able to measure a patient’s historical glycemic trajectory retrospectively.

Glycemic variability, the fluctuations in patient’s glycemic trajectory, has been increasingly recognized as a key factor contributing to numerous adverse clinical outcomes, including the progression of diabetes-related complications such as cardiovascular disease, retinopathy, and nephropathy (3–7). Furthermore, the importance of understanding a patient’s glycemic trajectory extends beyond diabetes management. Emerging evidence suggests that glycemic variability may also be associated with the risk of various cancers, including gastric cancer (8) and pancreatic ductal adenocarcinoma (9–12). For instance, patients with new-onset diabetes, particularly those with rapidly rising glucose levels before diagnosis, are at substantially higher risk for developing pancreatic ductal adenocarcinoma (13–15). However, there is currently no method capable of reconstructing a patient’s glycemic trajectory—especially in the absence of detailed clinical data—to predict cancer risk or provide a more comprehensive view of their glycemic profile.

Each day, new red blood cells (RBCs) are generated in the body, while older RBCs are gradually removed. Over the course of their lifespan, RBCs accumulate glycated hemoglobin (HbA1c), which reflects the averaged blood glucose levels they have been exposed to. The younger RBCs, having been in circulation for a shorter period, primarily reflect more recent glycemic values, whereas older RBCs, which have been exposed to glucose for a longer period, provide a picture of the longer-term blood glucose averages. By examining the HbA1c fractions in RBCs of different ages, we can reconstruct the glycemic trajectory of an individual. This enables the estimation of blood glucose levels at different times during the RBC lifespan, offering a retrospective view of glycemic fluctuations over the past few months—insights that traditional methods cannot provide.

However, measuring single-RBC-level HbA1c is technically challenging. While transient absorption microscopy (TAM) allows for the measurement of HbA1c at the single-cell level (16), its absorption feature introduces potential photodamage. Furthermore, TAM’s relatively slow imaging speed restricts its throughput, limiting its ability to perform high-throughput HbA1c measurements, which is critical in clinical settings.

In this study, we develop color-resolved third harmonic generation (THG) microscopy (cTHGM) for rapid HbA1c distribution measurement using an ultrabroadband (1200 to 1300 nm) near-infrared femtosecond laser. A single laser beam excites simultaneous THG over a broad wavelength range, enabling a sensitive discrimination of molecular color. With different molecular absorption features matched by different parts of the excitation spectrum, this resonantly enhanced spectral THG metrology achieves unprecedented molecular recognition through the absorption spectrum without the need of physical absorption (Fig. 1A). Spectral responses from different molecules can be differentially visualized by partitioning the epi-collected THG into multiple-color channels (Fig. 1B). Pseudocolor subfemtoliter three-dimensional (3D) imaging could distinctly capture the spectral virtual-absorption features, which uniquely enabled the detection of subtle hemoglobin (Hb) color variations (only ~2-nm peak shift in the Soret band) in single RBCs, both in vivo and ex vivo. This capability provides in vivo access to HbA1c distributions in circulating human RBCs, showcasing the exceptional color sensitivity of our technique.

Fig. 1. Summary of the contrast mechanisms and schematic illustrating the spectral resolving process.

(A) Summary of the contrast mechanism of cTHGM. The solid black line indicates the real energy level, and the solid gray line indicates the virtual energy level. The magnitude is resonance enhanced when a real electron transition exists corresponding to three times the incident photon energy. (B) Schematic illustrating the spectral resolving process. The conceptual diagram shows that RBCs with low HbA1c (<5.7%, corresponding to a nondiabetic state) and high HbA1c (≥6.5%, corresponding to a diabetic state) can be differentiated by their distinct THG spectral profiles and resulting pseudocolors. The curves in the gray dashed square indicate the transmission spectra of the filter sets, where blue, green, and red solid lines represent the transmission curves of channels A, B, and C, respectively. The curves in the blue, green, and red dashed squares indicate the thus-filtered THG spectra of channels A, B, and C, respectively, where rectangles with the corresponding color mark the transmission spectral range of the respective channel. The curves in the black dashed square indicates the spectral reconstruction of the THG signals from channels A, B, and C.

The THG contrast arises from the nonlinear polarization excited by intense electric fields, which is resonantly enhanced when a real electron transition corresponds to three times the incident photon energy, thus carrying the intrinsic absorption contrast information of biomolecules. With the required pulse energy on the subnanojoule level or lower, three-photon absorption will be strongly reduced, thus minimizing the nonlinear photodamage and the required laser excitation energy. Moreover, light scattering is a major issue in deep-tissue imaging; however, the long-excitation wavelength and high-order nonlinearity of THG provide an improved signal-to-noise ratio (SNR) (17, 18). Resonance-enhanced and color-resolved THG is therefore ideal for label-free in vivo molecular imaging compared to other modalities involving only one- or two-photon processes. By tripling the near-infrared excitation frequency (biological window) to the near-ultraviolet region where most biomolecules exhibit absorption features, resonance-enhanced cTHGM imaging enables molecular discrimination with high SNR in deep tissues.

Our method advances the field by offering unprecedented absorption spectrum sensitivity in vivo without the need of physical absorption, enabling the label-free section-free clinical high-3D-resolution distinction of biomolecules with subtle differentiation. In addition, it provides broader molecular identification, capable of detecting at least six different biomolecules simultaneously (fig. S1). The use of a single-beam modality further simplifies the implementation, making it more accessible and practical for both clinical and research applications. Another key advantage of cTHGM is that it is dominated by the virtual transition process with minimized linear and nonlinear absorption. Because of the strong resonance enhancement, the required excitation pulse energy for cTHGM is on the subnanojoule level. Because of the large detuning between the excited photon energy and the transition energy, resonant THG enhancement effectively prevents one-photon absorption, thereby minimizing linear phototoxicity (19). Given that the ground-state population at the focus is limited, resonance-enhanced THG competes with other ground-state absorption processes, including two-photon (2p) and three-photon (3p) absorption (20). With subnanojoule pulse energy, this competition partially provides much stronger signals over those from 2p and 3p absorption, reducing the potential need for nonlinear heat generation at the focus. Previous studies on three-photon microscopy (3 pm), which also involves nonlinear processes but relies on real absorption, have shown that even at a higher pulse energy level of 1 to 2 nJ (21–23), a 1.3-μm femtosecond pulse does not induce nonlinear photodamage or fluorophore saturation. A resonance-enhanced cTHGM system, however, operates with subnanojoule pulse energy, which is much lower than required by 3 pm. Therefore, it is reasonable to assert that cTHGM does not generate measurable nonlinear photodamage. Furthermore, previous studies (24, 25) reported no evidence of photodamage under the laser conditions of the proposed system.

RESULTS

RBC HbA1c distributions can reflect historical glycemic trajectories

Following an RBC HbA1c distribution model (see Materials and Methods), we first numerically examined the relationship between the single-RBC-level HbA1c distribution (called “RBC HbA1c distribution” or “the distribution”) and the temporal glucose trajectory (called “glucose trajectory”) of a subject for the past 150 days, as summarized in Fig. 2.

Fig. 2. Simulated historical glycemic trajectories and their corresponding RBC HbA1c distributions.

Simulated historical glycemic trajectories (A to C and G to I) and their corresponding RBC HbA1c distributions (D to F and J to L). [(A) to (C)] Glucose trajectories representing (A) stable glucose levels, (B) increasing glucose levels, and (C) decreasing glucose levels over 150 days. [(G) to (I)] Glucose trajectories representing a (G) single-event fluctuation, (H) oscillatory patterns with different amplitudes, and (I) oscillatory patterns with different periods. [(D) to (F)] and [(J) to (L)] Corresponding RBC HbA1c distributions derived from the glucose trajectories in [(A) to (C)] and [(G) to (I)], demonstrating how different glucose trajectories shape the HbA1c distributions.

As shown in Fig. 2A, a higher glucose baseline accelerates the glycation of RBCs, resulting in a flatter RBC HbA1c distribution (Fig. 2D). This occurs because more RBCs are glycated to higher HbA1c fractions when exposed to persistently elevated glucose levels over their lifespan. In contrast, a lower glucose baseline produces a narrower distribution dominated by lower HbA1c fractions. When glucose levels increased gradually over time (Fig. 2B), the RBC HbA1c distribution shifted toward higher HbA1c fractions, with the left side of the distribution notably reduced (Fig. 2E). This reduction in lower HbA1c fraction RBCs indicates that newly generated RBCs entering the bloodstream were rapidly glycated because of the higher recent glucose levels. Conversely, when glucose levels decreased gradually over time (Fig. 2C), the RBC HbA1c distribution shifted in the opposite direction, with the left side of the distribution becoming more pronounced (Fig. 2F). For glucose trajectories with single-event fluctuations (Fig. 2G), the RBC HbA1c distribution captures these events as distinct features: A glycemic spike introduces a dip (concavity) in the distribution, while a glycemic drop creates a bump (convexity) (Fig. 2J). Periodic oscillatory glucose trajectories further demonstrate how glycemic variability is encoded in RBC HbA1c distributions. As shown in Fig. 2H, oscillatory patterns in glucose with higher amplitudes are reflected in corresponding periodic features in the distribution, with the amplitude of oscillation encoded as higher peaks (Fig. 2K). Similarly, changing the oscillatory period of glucose fluctuations (Fig. 2I) alters the periodicity of the features in the RBC HbA1c distribution (Fig. 2L). These numerical results indicate that RBC HbA1c distributions can encode and reflect historical glucose trajectories, including baseline trends, single-event fluctuations, and periodic oscillations, making them a powerful tool for reconstructing glycemic variability.

A resonance-enhanced THG response differentiates HbA1c from Hb

As shown in Fig. 3A, the absorption spectra of oxyHbA0 and oxyHbA1c reveal only a slight red shift in the Soret band, with HbA1c’s absorption peak being ~2 nm redder compared to that of HbA0. This minor shift shows the difficulty in distinguishing HbA1c from Hb using conventional optical methods. Adding to this challenge, even in older RBCs that have been exposed to glucose for extended periods, the HbA1c fraction typically remains below 20%, further complicating single-RBC HbA1c fraction measurement.

Fig. 3. A red-shifted THG response of HbA1c enables the differentiation of high- and low-HbA1c RBCs.

(A) Absorption spectra of air-exposed (oxygenated) HbA0 (blue solid line) and HbA1c (orange solid line) solutions and sodium dithionite–treated (deoxygenated) HbA0 (red dashed line) solutions. a.u., arbitrary units. (B) THG emission spectra of HbA0 and HbA1c solutions, showing the normalized THG responses of HbA0 and HbA1c solutions (2.0 mg/ml) at different excitation wavelengths (scatterplot). The peak THG emission for HbA0 occurs at 1230-nm excitation, while HbA1c shows maximum THG emission at 1240-nm excitation. The shift indicates a red-shifted THG response when compared to HbA0. The solid lines represent the normalized absorbance of HbA0 and HbA1c. These measurements, with an overlap of the linear absorption and nonlinear THG emission spectra, taken with a tunable laser under consistent conditions, highlight the distinct resonance-enhanced THG response nature and its capability to differentiate HbA1c from HbA0. (C) Statistics of the 1240/1230 ratios of the older RBCs and younger RBCs. The scatterplot illustrates the 1240/1230 ratios of individual RBCs, with older RBCs denoted in red and younger RBCs in blue. Each dot represents the ratio of an individual cell. After the 1240/1230 ratio was logarithmically transformed to improve its distribution characteristics, the mean and SD of each group and the P value between the two groups were calculated. The red horizontal line within each box indicates the exponential of the mean [exp(mean)] in the log domain for each group. The boxes represent the 95% confidence interval [exp(mean − 1.96*SD) to exp(mean + 1.96*SD)]. The black lines extending above and below each box mark the range from exp(mean − 3*SD) to exp(mean + 3*SD). Significant differences can be observed between these two groups (P value <1 × 10−15). The P value was calculated by Origin 2023B. ***P < 0.001.

However, we observed that HbA1c’s THG emission spectrum exhibits a noticeable red shift compared to that of HbA0. As shown in Fig. 3B, the peak THG emission for HbA0 occurs under 1230-nm excitation, which is three times the wavelength of the emission, while HbA1c’s emission peak is shifted to be under 1240-nm excitation, consistent with the absorption-based resonant enhancement principle. This red-shifted response indicates that the subtle spectral differences between HbA1c and HbA0 can be captured through the THG emission spectrum, providing a sensitive approach to molecular differentiation because of the three times excitation wavelength expansion of the THG process.

To validate whether this sensitive THG response can reflect the subtle HbA1c fraction difference in single RBCs, we performed an animal study using continuous Percoll gradient centrifugation to separate older and younger RBCs from the same mouse. Older RBCs, having experienced prolonged exposure to elevated blood glucose, exhibit higher HbA1c fractions compared to younger RBCs. THG imaging was conducted at excitation wavelengths of 1230 and 1240 nm to calculate the 1240/1230 intensity ratios for individual RBCs. As shown in Fig. 3C, older RBCs displayed notably higher 1240/1230 ratios than younger RBCs (P < 1 × 10−15), reflecting their higher HbA1c fractions. These findings establish that through resonance enhancement, plural excitation wavelength THG imaging analysis is capable of addressing the challenges posed by minimal spectral differences and low HbA1c fractions within RBCs, enabling sensitive molecular identification and further precise glycemic history reconstruction.

Quantitative measurement of HbA1c fractions in single RBCs using cTHGM imaging

To achieve rapid single-RBC HbA1c measurement, we implemented a broadband Cr:forsterite laser as the excitation source (fig. S2) of cTHGM, allowing plural excitation wavelength THG spectroscopy and imaging in one shot using a single-beam modality. This cTHGM design substantially simplifies the system complexity, facilitating functional imaging of RBCs both ex vivo and in vivo with high sensitivity and higher throughput than TAM. With high sensitivity to minute shifts in the absorption peak, we explored the capability of cTHGM to distinguish oxyHbA1c from oxygenated Hb (oxyHb) within single human RBCs on the basis of their THG color. To avoid potential false-positive HbA1c analysis because of low oxygen saturation (sO2) levels, we designed a filter set that spectrally isolates interference from deoxygenated Hb (deoxyHb). Specifically, the possible resonant THG signal from deoxyHb (peaking at ~430 nm) (26, 27) is optically directed to channel C, ensuring that our analysis channels, channel A for monitoring oxyHb and channel B for monitoring oxyHbA1c, are free from oxygenation-state artifacts (see Materials and Methods). cTHGM could identify the color shift of RBCs in the whole blood of patients with diabetes mellitus (DM) relative to those of non-DM (NDM) volunteers (Fig. 4, A and B). RBCs with a low HbA1c fraction presented a stronger channel A intensity (blue). Our results indicated the capability of cTHGM to distinguish oxyHbA1c from oxyHb with a subfemtoliter resolution. We translated cTHGM into an in vivo clinical application. In our previous research, a femtosecond Cr:forsterite laser caused no photodamage in clinical trials (24, 25). To solve tissue reabsorption, which will change the observed THG color, we calibrated reabsorption by measuring the THG color of collagen fibers ex vivo and the in vivo THG color of collagen fibers near the moving RBCs within the same images. We examined oxygenated RBCs flowing in the papillary capillary loop between the dermis and basal layers of the volunteers. The DM and NDM RBCs had different THG colors (Fig. 4, C to F), whose shift was consistent with that of ex vivo RBCs.

Fig. 4. Assessment of HbA1c fractions in RBCs with cTHGM.

(A and B) Pseudocolored ex vivo cTHGM images of the whole-blood samples from (A) a patient with diabetes who has a high HbA1c = 12.7% and (B) a patient without diabetes who has a normal HbA1c = 5.3%. Blue: THG of channel A; green: THG of channel B; red: THG of channel C. Scale bar, 10 μm. The channel A intensity (blue channel) is enhanced 10 times to provide better contrast for the naked eye. (C to F) In vivo SHG, THG, and cTHGM images of RBCs in human cutaneous microvasculature from [(C) and (D)] two patients with diabetes who have HbA1c = 9.3 and 11.3% and [(E) and (F)] two patients without diabetes who have HbA1c = 4.8 and 5.4%. SHG (channel D) is presented in a green pseudocolor, and THG (channels A, B, and C combined) is presented in a magenta pseudocolor. The reabsorption-calibrated pseudocolored cTHGM image is presented in blue for the THG of channel A, green for the THG of channel B, and red for the THG of channel C. Scale bars, 10 μm. The channel A intensity (blue channel) is enhanced 10 times to provide better contrast for the naked eye. (G) Correlation plot of individual-level B/A ratio and HbA1c fraction with Pearson correlation coefficient = 0.9399, P < 0.0001 (two-tailed). (H) Statistics of the individual-level B/A ratios from five normal-HbA1c (HbA1c < 5.7%) volunteers and five abnormal-HbA1c (HbA1c > 6.5%) volunteers. A significant difference can be observed between the two groups (P = 0.004, based on the one-tailed Mann-Whitney U test). **P < 0.01. The P value was calculated using Origin 2023B software.

We further extended cTHGM application to provide the quantitative information of a patient’s RBC distribution on HbA1c fraction. The B/A ratio (channel B intensity/channel A intensity), which represents the contrast of the resonance enhancement provided by oxyHbA1c and oxyHb, was used as an indicator of the HbA1c fraction. The HbA1c fraction in individual RBCs is influenced not only by glucose levels but also by the age of the RBCs (16, 28, 29). Here, we first used the average B/A ratio across all RBCs measured by cTHGM from one individual (called individual-level B/A ratio) for comparative analysis with his/her whole-blood results obtained via mass spectrometry. A notable difference (P < 0.01) was observed between the averaged B/A ratios from NDM and DM RBCs in vivo and ex vivo (Fig. 4, G and H), while the scatterplot shows a clear linear trend of the individual-level B/A ratio with the HbA1c fraction (R = 0.9399). This linear relationship indicated that cTHGM B/A ratios could reflect the HbA1c fraction. The key parameters for the 10 volunteers recruited for this study, including their conventionally measured whole-blood HbA1c fractions and our optically measured individual-level B/A ratios, are summarized in table S1.

Single-RBC-level HbA1c distributions and potential historical glycemic trajectory reconstruction

We further defined the HbA1c index (see Materials and Methods) to represent the HbA1c fraction at a single-RBC level and individual levels. Besides the individual-level values, the normal-HbA1c volunteers’ single-RBC-level HbA1c distributions were quite different from the those of abnormal-HbA1c volunteers’ (Fig. 5, A to C). The HbA1c fraction inside a single-moving RBC in humans was noninvasively quantified by an in vivo label-free imaging method, with a high potential for noninvasive DM diagnosis. In contrast to the smooth theoretical curves in Fig. 2, the experimental distributions in Fig. 5 appear more coarse grained. This is a direct consequence of finite sampling. The limited number of measured RBCs necessitates wider histogram bins to ensure statistical significance. This practical constraint means that while long-term glycemic trends are robustly captured, finer details corresponding to high-frequency fluctuations may be averaged out because of the limited number of RBCs sampled. The achievable temporal resolution is therefore dependent on the total RBC count, as detailed in the Supplementary Materials (Supplementary Text S3 and figs. S6 to S8).

Fig. 5. Single-RBC-level HbA1c distributions of 10 volunteers.

(A to C) Distribution probability density versus RBC HbA1c index for (A) three ex vivo measured normal-HbA1c (HbA1c < 5.7%) volunteers, (B) three ex vivo measured abnormal-HbA1c (HbA1c > 6.5%) volunteers, and (C) four in vivo measured volunteers (two normal-HbA1c volunteers and two abnormal-HbA1c volunteers), for whom representative single-RBC images are shown in Fig. 4 (C to F). The grouping width of in vivo measured distributions is 5%. The grouping width of ex vivo measured distributions is 1%. (D) Heatmap of KS distances between the CDFs of RBC HbA1c indices for the normal-HbA1c group (NDM1, NDM2, and DM1) and abnormal-HbA1c group (DM2, DM3, and DM4). The KS distance quantifies the maximum difference between CDFs, with higher values indicating greater dissimilarity. (E) Best fitting (dark gray dot-dashed line) with fluctuation of DM1 in comparison with the measured distribution. (F) Best-fitted glucose trajectories (dark gray line) with limited fasting glucose data (red solid squares) from clinical blood tests. The fitting glucose trajectory matched the fasting glucose data.

To further quantify the distinctions in HbA1c distributions between normal-HbA1c and abnormal-HbA1c groups, we calculated the pairwise Kolmogorov-Smirnov (KS) distances (30, 31) between the cumulative distribution functions (CDFs) of RBC HbA1c indices for three normal-HbA1c (NDM1, NDM2, and DM1) and three abnormal-HbA1c (DM2, DM3, and DM4) volunteers. The resulting heatmap (Fig. 5D) revealed consistently larger KS distances between the normal-HbA1c and abnormal-HbA1c groups compared to distances within the normal-HbA1c and abnormal-HbA1c groups, highlighting notable differences in RBC HbA1c distributions between normal-HbA1c and abnormal-HbA1c individuals. This finding reinforces the hypothesis that diabetes, especially bad glucose control, induces measurable alterations in the RBC HbA1c distribution.

An exciting aspect of this study is the ability of cTHGM to provide a noninvasive alternative to traditional HbA1c measurements. While conventional HbA1c testing reflects average glucose levels over 3 months, and glucometers offer momentary snapshots, both approaches are invasive. Our results demonstrated that cTHGM enables noninvasive quantification of patients’ HbA1c fractions, substantially reducing discomfort and potential risks associated with blood sampling.

Moreover, cTHGM introduces an unprecedented capability to retrospectively infer glycemic trajectories from single-RBC HbA1c distributions. This technique leverages the variation in HbA1c fractions across RBCs, combined with statistical modeling to estimate RBC ages (see Materials and Methods). Specifically, on the basis of the knowledge gained from our simulations (Fig. 2), a measured RBC HbA1c distribution exhibiting higher HbA1c fractions among younger RBCs indicates a recent substantial increase in blood sugar levels. Notably, we observed this exact pattern in volunteer DM1, whose mean HbA1c fraction was similar to nondiabetic volunteers but showed an elevated HbA1c fraction among younger RBC subgroups (Fig. 5A, middle panel). Guided by this principle, we reconstructed the glycemic trajectory of volunteer DM1 by fitting their measured HbA1c distribution to glucose fluctuations using our statistical model (Fig. 5E). The reconstructed glucose trajectory closely matched the limited available fasting glucose records (Fig. 5F), providing robust validation of the method’s capability to accurately capture detailed glycemic fluctuations. To experimentally validate this crucial link between a rapid glycemic rise and this specific distribution pattern, we conducted a controlled study in diabetic mouse models. The results confirmed that an acute hyperglycemic event in T1DM mice reproduces a distribution pattern notably similar to that of volunteer DM1, providing independent corroboration for the validity of our approach (see Supplementary Text S5 and figs. S10 and S11). This ability to discern subtle historical variations offers substantial clinical potential, enhancing personalized diabetes management by retrospectively evaluating glycemic control over extended periods.

While this methodology paper does not include interventional clinical trials, our preclinical results underscore the potential of cTHGM as a transformative tool for personalized diabetes management and the study of glycemic variability. Future research should focus on expanding these findings through larger clinical studies to fully explore the diagnostic and prognostic applications of cTHGM in diabetes care.

DISCUSSION

In this study, we first demonstrated through THG spectroscopy that resonance-enhanced THG enables the distinction of Hb and HbA1c. On the basis of this absorption-contrast, yet absorption-free resonant enhancement effect, here, we proposed and developed cTHGM. Compared to other molecular imaging methods such as TAM and sum-frequency generation microscopy, cTHGM offers not only superior 3D spatial resolution, imaging speed, and spectral sensitivity but also a notably simpler and more streamlined implementation.

In contrast to TAM (16), cTHGM offers improvements in both biosafety and imaging speed, which are critical factors in clinical and research settings. TAM, which relies on the actual absorption of light, can induce photodamage and phototoxicity owing to the energy absorbed by the target sample. This not only limits its repeated use but also poses a risk to live specimens. In contrast, cTHGM uses a noninvasive absorption-contrast but absorption-free approach, substantially reducing phototoxicity and making it a safer option for long-term and repeated studies. With a subnanojoule required pulse energy at the 1200- to 1300-nm spectral region, cTHGM also minimizes the resonance-induced three-photon absorption effect. In addition, cTHGM substantially enhances imaging speed, offering a voxel dwell time of ~0.07 μs, over 100 times faster than TAM’s pixel dwell time of 10 μs. This remarkable improvement in speed is crucial for capturing dynamic biological events in real time, thereby greatly enhancing its application in live-cell and in vivo studies. For extended discussion on the cTHGM biosafety issue, please see Supplementary Text S6.

In comparison with sum-frequency generation microscopy (26), cTHGM stands out not only due to its simplified system design but especially due to its ability to image multiple biomolecules simultaneously within the same field of view. This multiplexing capability is critical when complex biological interactions and pathways must be studied in situ. Furthermore, the superior spectral sensitivity of cTHGM enables finer discrimination between closely related molecular states, such as the subtle differences between oxyHb and oxyHbA1c. Such imaging with high spectral sensitivity is indispensable for quantitative molecular diagnostics and monitoring, particularly in settings where subtle biochemical changes can dictate clinical outcomes. Recent studies and meta-analyses have identified glycemic variability as an independent risk factor for various macrovascular and microvascular complications in patients with diabetes, even after adjustment for mean HbA1c levels (32–34). Assessment of HbA1c variability using this demonstrated cTHGM method may provide complementary information to state-of-the-art continuous glucose monitoring (35), offering additional insights for optimizing glycemic control strategies.

In conclusion, cTHGM is a label-free, noninvasive, slide-free, ultrahigh-3D-resolution molecular imaging technology that maps endogenous optical absorption contrasts while avoiding physical absorption of the excitation laser light. The color-sensitive feature of cTHGM makes it well suited for quick translation toward in vivo color-related molecular imaging applications. The presented technology could also make in vivo and ex vivo 3D submicrometer resolution imaging possible with absorption labeling dyes of different colors.

MATERIALS AND METHODS

cTHGM system

The cTHGM system for human clinical study (fig. S5) was adapted from a commercial scanner (MPM-SCAN4; Thorlabs) combined with a homemade upright microscope to suit human clinical trials. The laser beam was focused by a high–numerical aperture (NA) water-immersion objective (UAPON 40XW340; 40×/water/1.15 NA; Olympus) with a max 90-mW average power after the objective (<0.86 nJ), comparable to our previous studies (24, 25). cTHGM uses lab-written software to achieve laser synchronous triggering (36, 37). The 3D reconstruction was realized by laterally scanning the focal point with a fast resonant mirror (CRS 8 kHz; Cambridge) and a galvanometer mirror (6210H 6-mm Y Mirror; Cambridge) with 930 by 1024 pixels (0.07-μs voxel dwell time) or 472 by 512 pixels (0.14-μs voxel dwell time) and axially scanning the objective with a z-axial stepping motor (TSDM40-15x; Sigma Koki) under the bidirectional scan mode. Furthermore, an XY-motorized stage (MLS203; Thorlabs) was used as a complement to the scanner for wide-range 2D imaging. The THG and second harmonic generation (SHG) emissions at the focal point were epi-collected through the same objective, guided through a dichroic beam splitter (FF705-Di01-25x36; Semrock), filtered by a color glass filter (CG-KG-5-50; CVI), split into four channels (annotated channels A, B, and C from shorter wavelengths to longer ones for THG signals and channel D for SHG signals), and detected by four side-on photomultipliers (R4220P, Hamamatsu, for THG channels A, B, and C; R928P, Hamamatsu, for SHG channel D). The filter set was designed to optimize the sensitivity of oxyHb, HbA1c, and deoxyHb. Channel A was separated from channel B by a 412-nm-edge long-pass dichroic filter (T412lpxt; Chroma), channel B was separated from channel C by a 416-nm-edge long-pass dichroic filter (FF416-Di03; Semrock), and channel C was separated from channel D by a 593-nm-edge long-pass dichroic filter (FF593-Di03; Semrock). Four band-pass filters (405-10 OD4, Alluxa, for channel A; FF01-417/60-25, Semrock, for channel B; FF01-425/26-25, Semrock, for channel C; and FF02-617/73–25, Semrock, for channel D) were used to further partition signals and filter out unwanted wavelengths. The obtained 14-bit signals of channels A, B, C, and D were independently amplified by setting the supply voltages of photomultiplier tubes to 900, 800, 800, and 800 V and pseudocolored with blue, green, red, and sea green for presentation, respectively. The scanner, motors, and acquisition system were synchronized by lab-written software (36, 37). The measured lateral and axial resolutions were 0.46 and 3.4 μm, respectively, by taking a 1266-nm-excitation-wavelength, 1.15-NA objective in skin tissues (average refractive index, 1.4).

Sample preparation for absorption spectral measurement

An ultraviolet-visible spectrometer (JASCO V-770) was used to measure the absorption spectrum. The sample used in this work is as follows: resuspended oxidized aqueous Hb solution [0.1 mg/ml; lyophilized Hb (H0267; Sigma-Aldrich) in 1× phosphate-buffered saline (PBS)], resuspended oxidized aqueous HbA1c solution [0.1 mg/ml; lyophilized Hb (H0267; Sigma-Aldrich) in 1× PBS], and resuspended deoxidized aqueous Hb solution [0.1 mg/ml; lyophilized Hb (H0267; Sigma-Aldrich) in sodium dithionite (0.14 mg/ml; 157953; Sigma-Aldrich) in 1× PBS]. The oxidized Hb sample was exposed to air for 30 min before measurement to ensure complete oxidation. For the deoxidized Hb sample, a dithionite concentration of 2.65 mg per 100 mg of Hb was used to maintain oxidization less than 1% for 30 min (38), and each experimental procedure (including mounting and data acquisition) was done within 20 min.

Protocol for animal study

Animal research ethics approval was obtained from the Animal Facility, University of Macau (protocol ID: UMARE-009-2023). Six- to eight-week-old male BALB/c and ICR mice were obtained from the breeding source of Animal Facility of University of Macau, and all procedures were performed in strict accordance with the guidelines of the Animal Research Ethics. Four to six mice were housed per acrylic cage and kept in a room maintained at constant temperature and humidity under 12-hour light and dark cycles and fed a regular ad libitum diet with clean drinking water.

The fresh blood sample is obtained with cardiac puncture, and mice are anaesthetized with 0.2 ml of 2.5% Avertin per 10 g body weight by intraperitoneal injection. Blood is extracted with a syringe [1 cc/ml Terumo, with AGANITM needle 27Gx 1/2″ (0.4 by 13 mm)], which is filled and dispensed with sodium heparin (0.2 ml of 10 IU/ml each mouse). Mice are euthanized by cervical dislocation.

The collected blood is ejected to a 15-ml centrifugation tube with the syringe needle removed. Five hundred microliters of whole-blood sample is loaded into a 15-ml tube and resuspended with 1× PBS. The ratio of blood sample and PBS is 1:9. For 500 μl of whole-blood sample, 4500 μl of PBS should be added. Then, the sample is put into a low-speed centrifuge (Eppendorf 5702) at 3000 rpm at room temperature for 10 min. This separates the plasma from the RBCs, forming a pellet of RBCs at the bottom of the tube. The supernatant is removed with a pipette (80% of the upper solution). This series of centrifugation should be repeated for three times to obtain a pure RBC sample. The whole-blood sample was then sandwiched within two cover glasses. To prevent the detriment of RBCs, 0.05-mm-thick double-sided sticker (IS201, SUNJinLab) was affixed between two cover glasses.

Protocol for continuous Percoll gradient centrifugation of RBCs

Percoll is a low-viscosity-density gradient medium that is used for isolating cells with different densities in centrifugation. Percoll is a nontoxic medium that is almost chemically inert and would not damage the RBC membrane during centrifugation. The normal RBC density lies around 1.10 g/ml, and the Percoll gradient can form within the density range of 1.0 to 1.3 g/ml. RBCs in different ages vary in density, and in normal situations, older RBCs are relatively smaller than the younger RBCs because of the loss of cytoplasmic volume, usually a smaller and denser cell; thus, Percoll centrifugation can create a separation between different aged RBCs.

Percoll (Cytiva 17-0891-01, Sigma-Aldrich) is diluted with saline before use. Saline (8.5%) is added to Percoll at a ratio of 1:9, and then Percoll is diluted with 0.85% saline into the corresponding concentrations of 70, 68, 66, 64, and 62%. In a new 15-ml centrifugation tube, 2 ml of Percoll solution is loaded from the highest concentration and slowly pipetted on the edge of the tube to allow a distinct Percoll layer to form. The resuspended blood sample is layered on top of Percoll without mixing. The tube is then centrifuged at 1200 rpm for 15 min at room temperature. A continuous gradient of RBCs is formed, with the upper layer having a lower density of younger RBCs and the lower layer having a higher density of older RBCs (39). The young and old groups are pipetted, avoiding the supernatant; washed with PBS; and centrifuged at 3000 rpm for 10 min. The supernatant is discarded, and the steps are repeated twice to remove any Percoll in the sample.

Protocol for ex vivo human RBC imaging

The air-exposed 1:40 diluted human whole-blood samples (National Taiwan University Hospital, approval number 201709032RINB) in 1× PBS were used in imaging of RBCs considered as having a high oxygen level. The whole-blood sample was added on a slide (Microscope Slide, Frosted One End, MARIENFELD) and sandwiched with cover glass (High Precision Cover Glass, Square, NO.1.5H, MARIENFELD) on top. To prevent the detriment of RBCs, 0.05-mm-thick double-sided sticker (IS201, SUNJinLab) was affixed between the slide and cover glass. All imaging processes were performed under the protocol reviewed and approved by the Research Ethics Committee of National Taiwan University Hospital with obtained informed consent from each subject. Six volunteers with coronary artery disease were involved in ex vivo measurement. All volunteers were older than 18 years without regard to sex. Four volunteers were diagnosed as diabetic, and two volunteers were not. The processes of sample transportation, sample preparation, and imaging were finished within 30 min after blood sample collection from patients. The average laser power after the objective was less than 60 mW (<0.57 nJ) for ex vivo RBC imaging.

Protocol for human in vivo imaging

This trial was conducted in accordance with the principles of the Declaration of Helsinki. The study protocol was approved by the Research Ethics Committee of National Taiwan University Hospital (no. 200903064D). Written informed consent was provided by all volunteers before the experiment. A total of four volunteers, including three females and one male aged 20 to 55 years, was investigated. None of the volunteers had skin disease in the imaged areas. Volunteers who were pregnant or breastfeeding were excluded.

The in vivo cTHGM images of the right ventral forearm were obtained at about 15 cm proximal to the volunteer’s wrist. In this system, subjects had to place their forearm under the objectives with the ventral side upward, and the imaging depth was controlled by the z-axial motor with a 1-μm resolution. The imaging frame rate was 15 frames per second (fps). The average laser power after the objective was less than 90 mW. The total accumulated exposure time to laser for one volunteer was ≤30 min. There were no inflammatory symptoms, no skin color change, no pigmentation, no wound, no blister formation, and no ulceration reported. According to the subjects’ experiences, they felt no discomfort during the experiment. For the collagen samples used in the reabsorption calibration process, a bovine collagen solution (BM-5133; FibriCol) was cultured to form films ~40 μm in thickness at a collagen concentration of 8 mg/ml. These collagen samples provided stable references to correct the reabsorption effects encountered during human in vivo RBC imaging.

Quantitative single-RBC-level HbA1c fraction analysis

The cTHGM images underwent color contrast enhancement solely to improve visual clarity without affecting the subsequent quantitative analysis. Specifically, for ex vivo cTHGM images, the intensity of channel A (blue channel) was enhanced by a factor of 10. For in vivo cTHGM images, color calibration was first conducted using the THG colors of collagen fibers adjacent to the microvasculature as well as ex vivo–cultured bovine collagen samples. After calibration, the intensity of channel A (blue channel) was also enhanced 10-fold. It should be emphasized that these color enhancements were applied exclusively for better visual contrast and did not influence the image analysis.

The cTHGM images were analyzed by lab-written MATLAB codes. The images of ex vivo human whole-blood samples were 10-frame averaged. With whole-blood samples exposed to air, all RBCs were oxygenated. For segmentation of ex vivo RBCs, the region of interest (ROI) was provided by thresholding (Otsu) the THG signals in channel C using the lab-written code, while each ROI will represent one RBC. In principle, the THG channel C intensity does not include the resonance-enhanced contrast provided by oxyHb and oxyHbA1c. Single-RBC detection was based on the RBC morphology (40). To improve the SNR, we excluded the 20% proportion of ex vivo RBCs with the weaker combined THG signals of channels A, B, and C before the quantitative analysis.

The in vivo human skin images were acquired under 15 fps without frame averaging. We examined the RBCs with high oxygen levels flowing toward the skin surface in the papillary capillary loop (41). For segmentation of in vivo RBCs, the ROI was provided by thresholding the combined THG signals of channels A, B, and C using Otsu’s method. By considering that RBCs mainly comprise Hb, the primarily clustered intensities in the intensity histogram were chosen as Hb-relevant pixels.

To analyze the HbA1c fraction of every single RBC, the ratio of signal strength in channel B to channel A, denoted as the B/A ratio, was applied. The B/A ratio of each RBC (single RBC level) was obtained by averaging the B/A ratio of each considered pixel (sub-RBC level) within the ROI. To compensate for the reabsorption effect, the single-RBC-level B/A ratio of in vivo RBCs was calibrated by comparing the B/A ratios of the ex vivo collagen fibers and the in vivo collagen fibers next to the moving RBCs within the same image. These collagen signals were considered reliable internal references, experiencing similar reabsorption conditions to neighboring RBCs. In addition, an ex vivo–cultured bovine collagen sample served as an external calibration reference to further correct the THG colors obtained from in vivo RBC imaging, ensuring robust and precise compensation of tissue reabsorption artifacts.

To link the B/A ratios to clinically relevant parameters, we modeled the HbA1c index as a linear transformation of the B/A ratio based on the linear relationship (Fig. 4G) between the individual-level B/A ratios and the HbA1c fractions

HbA1c index=p1∗B/Aratio+p2

(1)

For estimation of parameters ( p1,p2 ), we formulated an optimization problem. The optimization problem solution is the estimated parameter values. We defined the cost function to describe the estimation error of the individual-level HbA1c index

Loss(p1,p2)=∣HbA1c fraction−HbA1cindex∣=∣HbA1c fraction−p1∗B/Aratioindividual−level−p2∣

(2)

To avoid the fitting distortion, the HbA1c fractions of reticulocytes (RBC precursors) should serve as the boundary condition. Reticulocytes have the lowest HbA1c fractions (single RBC level) of all circulating RBCs. It has been reported (28) that about 1.2% of circulating RBCs are reticulocytes, with the HbA1c fractions being about 0.737% in nondiabetes. The optimization problem can be written as

Minimizep1,p2loss(p1,p2)

(3)

Subject to11.2%·N∑1.2%·NHbA1cindexreticulocyte=0.737

(4)

where N is the number of imaged RBCs.

With the boundary condition, the loss(p1,p2) could be optimally minimized, whose p1,p2 were then estimated. This nonlinear optimization allows the single-RBC-level HbA1c index to reflect the critical information regarding single-RBC-level HbA1c fractions. As a result, p1=0.9201,p2=−4.839.

Quantitative single-RBC-level 1240/1230 ratio analysis

The THG images at different excitation wavelengths were acquired using a tunable laser (Insight X3, Spectra Physics) without frame averaging, and the microscopy system used here has been previously reported (42). The image-processing pipeline for single-RBC detection followed the protocol described in the “Quantitative single-RBC-level HbA1c fraction analysis” section.

Briefly, THG images were binarized using Otsu’s thresholding method to maximize interclass variance, effectively distinguishing RBC signals from background noise. The resultant binary images were segmented into discrete contours corresponding to potential RBCs. To ensure robust identification, these segmented regions underwent morphological filtering based on RBC-specific metrics.

To accurately compute the 1240/1230 ratio, single-RBC detection was independently performed on images obtained at excitation wavelengths of 1230 and 1240 nm. Mean THG intensities within the segmented single-RBC masks at each excitation wavelength were calculated to derive the 1240/1230 ratio for each RBC. Only RBCs exhibiting negligible spatial displacement during wavelength tuning were included in the analysis. To eliminate RBC movement during wavelength switching, the RBC samples were sandwiched between two cover glasses.

RBC HbA1c distribution model

RBCs undergo glycations with glucose over time, forming HbA1c without the need for enzyme catalysis. This chemical reaction is irreversible, making each RBC a cumu