The advancement of laser technology has significantly revolutionized the field of biomedical imaging, particularly with the emergence of photoacoustic microscopy (PAM). PAM integrates optical excitation, achieved through a laser source with a specific wavelength, and acoustic detection, carried out by an ultrasonic transducer. This combination enables label-free, high-resolution, and depth-resolved imaging, leveraging the contrast provided by endogenous tissue chromophores [1–5]. Lipids, a major constituent of biostructures, are of particular interest because the dysregulation of lipid metabolism is linked to numerous pathological conditions [6–10]. In particular, the accumulation of lipids in the liver, such as in non-alcoholic fatty liver disease (NAFLD), can be the hallmark risk factor for severe complications with significant morbidity and mortality burdens, including cirrhosis, end-stage liver disease, and hepatocellular carcinoma (HCC) [11,12]. Existing non-invasive imaging techniques for intrahepatic lipid detection, such as ultrasound (US), computed tomography (CT), and magnetic resonance imaging (MRI), offer qualitative and quantitative assessments of total lipid accumulation, or steatosis, in the liver [11–14]. However, these modalities suffer from poor spatial resolution and low imaging signal-to-noise ratio (SNR) [11,13]. While biopsy remains the gold standard for detecting steatosis levels, it is an invasive technique that is also prone to sampling errors [13,15]. Thus, quantitative lipid-based PAM can bridge the gap between macroscopic non-invasive imaging techniques and high-resolution histopathological methods for detailed compositional analysis of hepatic tissues.

With the advent of suitable laser sources, PAM of lipids in the shortwave infrared (SWIR) window (from 1000 to 2000 nm) has garnered much attention [5,16–20]. The SWIR region offers a new opportunity for deep tissue imaging. It benefits from reduced scattering compared to the visible region and exhibits lower water absorption than the mid-infrared region, making it particularly advantageous for biomedical applications. In this region, lipid absorption is characterized by the two primary absorption peaks at 1.7 and 1.2 μm, corresponding to the first and the second overtones of the carbon-hydrogen bond (C–H bond), respectively [21–23]. The first overtone at 1.7 μm is far more advantageous for PAM imaging of lipids compared to the 1.2-μm band for two key reasons. First and foremost, the 1.7-μm window presents a higher absorption coefficient of lipid compared to 1.2 μm, producing six to seven times stronger PA signals [21,22,24]. Second, given the water-rich environments of biological tissues, a major roadblock for bioimaging applications at 1.7 μm is the higher water absorption at longer wavelengths. However, the enhancement factor of the PA signal overrides the increased water absorption at the 1.7-μm absorption window [25]. Therefore, PAM for lipids can achieve higher contrast and SNR by exploiting the absorption peak located around 1725 nm under the condition that a suitable laser source is employed. The longer wavelength also offers the added benefits of longer penetration depths and lowered phototoxicity, making it more well-suited for biomedical applications [26,27].

To realize quantitative lipid PAM at the 1.7-μm absorption band, an ideal laser source must meet several criteria: (i) a narrow emission bandwidth centered at the lipid absorption peak for specificity, (ii) high spectral purity and OSNR for enhanced sensitivity and efficient photoacoustic generation, and (iii) highly stable emission for imaging reliability and accuracy. These attributes would enhance the overall quality and contrast of lipid-based PAM such that it can highlight minute lipid-rich structures in biosamples for quantitative analysis. This also brings more energy margin for the increased repetition rate to achieve fast imaging. High imaging contrast can facilitate quantitative analysis of PAM images by improving the accuracy and reliability of image segmentation techniques, as lack of sufficient imaging contrast is known to be a barrier to accurate segmentation results [28]. It would also eliminate the need for extensive pre-processing prior to more complex analysis, such as via deep learning [29–31]. This can significantly streamline image analysis and/or virtual histology workflows, especially in a clinical setting entailing time pressure. Hence, a tailored laser source at 1.7 μm for high-contrast PAM imaging would have important implications for quantification of lipids for ex vivo and potential in vivo analysis.

Previous studies of lipid-based PAM techniques at the 1.7-μm absorption band have mainly utilized optical parametric oscillators (OPOs), with supercontinuum sources also being used to a lesser extent [16,19,25,32–41]. However, the key limitation to the application of commercial OPOs to lipid-based PAM lies in the spectral characteristics of the sources. While OPOs can provide broad wavelength tunability and mJ pulse energies, their spectral energy density, OSNR, and spectral purity are limited. These limitations result in restricted imaging contrast, particularly within the 1.7-μm window where water absorption is a concern. Moreover, OPOs typically operate at kHz repetition rates, significantly constraining the achievable imaging speeds for PAM applications. Supercontinuum sources have also been developed for lipid PA applications at 1.7 μm. However, the major drawback of pulsed supercontinuum lasers lies in their relatively low spectral energy density. So, in order to achieve high enough pulse energies to excite PA signals with sufficient SNR, the filtered bandwidth of the spectrum remains on the order of several tens of nanometers, hampering the specificity of lipid detection [41,42].

In this study, we present a HOPE operating at 1725 nm for lipid-based PAM at the first overtone of the C−H bond. The 1725-nm HOPE leverages the dual gain scheme combining parametric gain and thulium-doped fiber (TDF) gain, building on the HOPE theory established in our previous works [4,43]. While earlier studies focused on the 1930-nm wavelength regime where TDF gain is more easily achieved [44], our current work successfully extends this theory to the shorter wavelength operation of TDF at 1.7 μm. Despite the quasi-three-level behavior of the TDF, we demonstrate that the 1725-nm HOPE can attain high spectral purity and suppress the TDF noise floor at longer wavelengths. The 1725-nm HOPE emission achieves a 3-dB spectral bandwidth of 1.4 nm with a high OSNR of $\sim 34\mathrm{dB}$ sans any additional filter in the cavity loop. The source can achieve a spectral energy density up to 480 nJ/nm post-amplification as well as an exceptionally high pulse repetition rate (PRR) of 194 kHz. The performance of the HOPE allows for highly sensitive lipid detection with a high imaging contrast ratio of 23.6:1.

The 1725-nm HOPE is integrated with a traditional transmission-mode PAM system to image human liver samples with dichotomized steatosis grades. The high sensitivity and contrast provided by the 1725-nm-HOPE-based PAM system allowed precise mapping of lipid distribution within the liver samples. This facilitated a clear distinction between the healthy and the fatty liver samples through both qualitative and quantitative assessments. With regard to the latter, a segmentation-based image analysis protocol using Otsu’s method was employed to quantify the percentage of steatosis in the PAM images. To the best of our knowledge, this is the first report of a segmentation-enabled PAM method for intrahepatic steatosis quantification at 1.7 μm. Thanks to the high contrast ratio of the PAM images, the segmentation algorithm could be applied to the images without any contrast enhancement pre-processing. The quantification results closely reflect the clinically assigned steatosis grades for the liver samples, demonstrating the capability of the 1725-nm HOPE in enabling high-sensitivity and high-contrast quantitative PAM of intrahepatic lipids to quantify liver steatosis levels.

The schematic of the 1725-nm HOPE is shown in Fig. 1. The 1725-nm HOPE is based on the HOPE principle developed in our previous works [4,43]. The 1725-nm HOPE cavity operates on a dual gain scheme constituting parametric gain by the highly nonlinear fiber (HNLF) and gain provided by the TDF. The dual gain in the cavity forms a gain filter effect, which has been touched on in the previous HOPE works [4,43]. Briefly, the gain filter in the cavity is formed by a narrowband parametric gain, which is boosted up by additional TDF gain in order to surpass the threshold of the oscillation loop. As a result, the HOPE cavity can inherently achieve a narrowband emission with high spectral purity and spectral energy density without the need of an additional bandpass filter within the cavity. In the cavity loop, the HNLF also has the role of producing the idler in the desired SWIR range through degenerate four-wave mixing, given by $\mathrm{\Delta }\omega ={\omega }_p-{\omega }_i={\omega }_s-{\omega }_p$, where $\mathrm{\Delta }\omega$ is the frequency shift, and ${\omega }_p$, ${\omega }_s$, and ${\omega }_i$ are pump, signal, and idler angular frequencies, respectively. The residual pump remaining after the parametric process acts as the in-band pump for the TDF to boost the idler wavelength. Hence, at first glance, it may seem feasible to shift the wavelength operation of the HOPE scheme from 1.9 to 1.7 μm by shifting the pump wavelength of the cavity.

However, the major challenge associated with the 1725-nm HOPE is the accessibility of short wavelength gain in the TDF. The emission of TDF has been shown to span a broad spectral range from 1.6 to 2.1 μm. As previously mentioned, our existing HOPE works have exploited TDF gain at 1930 nm. However, it is much more challenging to achieve efficient laser emission at 1.7 μm compared to 1.9 μm. This is due to the quasi-three-level behavior of TDF, whereby the three-level nature of the TDF becomes pronounced in the shorter wavelength regime of its emission spectrum [45–52]. In quasi-three-level transitions, as in the ${}^3{\mathrm{F}}_4-{}^3{\mathrm{H}}_6$ transition in TDF, the lower laser levels are very close to the ground state, leading to the lower laser levels being partially occupied at thermal equilibrium (room temperature). This culminates in reabsorption losses at the laser wavelength, as laser photons can be absorbed by ions in the lower laser level to move to the upper laser level [48–50]. Hence, a much higher population inversion is required for substantial net gain at 1.7 μm. Unfortunately, such high inversions will also lead to enhanced gain at the 1.8–2 μm range. This leads to increased amplified spontaneous emission (ASE) noise and associated parasitic lasing, which can further hinder the 1725-nm operation of the HOPE cavity. Notably, the length of the TDF also needs to be selected carefully because there is a trade-off relationship between the gain and the absorption loss in the TDF, thanks to the small emission cross section and high reabsorption characteristic at 1.7 μm [44,45,53].

To overcome the barrier of achieving high spectral purity and stable operation in the 1725-nm HOPE scheme, we utilized the complementary relationship between parametric gain and TDF gain with respect to the pump power. The gain from parametric amplification is related to the peak power of the pump, whereas TDF gain is related to its average power [4]. In the 1725-nm HOPE scheme, a distributed feedback laser diode (DFB-LD, BOX Optronics Tech) at 1562 nm produced a continuous-wave output that was modulated to pulses by a high-speed optical amplitude modulator (AM, JDS Uniphase). The operating point of the AM was carefully selected by adjusting the applied direct current bias, which in turn changed the modulation depth of the signal. Within a specific range of the chosen operating point of the AM near its quadrature point, increasing the modulation depth was associated with an increase in the peak power and a decrease in the average power, and vice versa. To maintain the AM’s operating point and mitigate the effects of bias drift, a proportional-integral-derivative (PID) controller system was employed. This ensured stable modulation and consistent performance of the 1725-nm HOPE scheme, enabling high spectral purity and reliable operation.

To achieve a narrow bandwidth in the 1725-nm HOPE scheme using the theory of the gain filter, the modulation depth was decreased to lower the parametric gain in the cavity loop. However, reducing the parametric gain necessitated an increase in the TDF gain to surpass the threshold for laser oscillation. Typically, this can be achieved by (i) increasing the length of the gain fiber and (ii) increasing the average pump power. Due to the trade-off between gain and reabsorption in the TDF, the TDF length cannot be increased too much lest the net gain was lowered. Additionally, significantly increasing the average pump power risks the onset of parasitic lasing and ASE at longer wavelengths. To mitigate these issues, we empirically adjusted the length of the TDF using the cut-back method while simultaneously fine-tuning the modulation depth of the pump signal. This allowed us to experimentally arrive at a point where the parametric gain was optimally reduced for narrow bandwidth emission and the TDF gain was sufficiently increased for threshold breakthrough with suppressed ASE. Hence, the gain filter effect could be successfully implemented in the 1725-nm regime by accessing the shorter wavelength gain of the TDF emission.

Given the abovementioned implementation of the dual gain scheme in the laser cavity, the overall setup of the 1725-nm HOPE can be summarized as follows. The seed of the 1725-nm HOPE was provided by a DFB laser diode at 1562 nm. This continuous-wave seed was modulated to pulses by an AM. A PID-controlled feedback scheme was employed at the AM to control the modulation depth of the pulse signal and eliminate the effects of bias drift through negative feedback. The pulsed signal was subsequently amplified to an average power of $\sim 1\mathrm{W}$ by an optical low-noise amplifier (OLNA). The OLNA comprised an erbium-doped fiber pre-amplifier (EDFA), followed by an optical bandpass filter to eliminate ASE noise, and a second-stage amplification by a high-power EDFA. A polarization controller was used to change the polarization state of the pump signal for phase matching in the HNLF before the OLNA. The amplified pulsed signal was then fed into the laser cavity via a 1560/1730-nm wavelength-division multiplexer (WDM). The parametric amplification and wavelength conversion through four-wave mixing took place in a spool of 50-m-long HNLF (HNLF-SPINE ZDW-1566, OFS) with a zero-dispersion wavelength of 1566 nm. A 30-cm-long TDF (TmDF200, OFS) in the cavity utilized the residual pump after the parametric amplification process to boost the overall gain. A spool of 1-km-long single-mode fiber (SMF) was used to introduce delay into the cavity, while a variable delay line (VDL) was used to fine-tune the cavity length to precisely match the repetition rate. A 50:50 coupler led 50% of the signal out of the loop, and the rest was looped back into the cavity for oscillation.

The output from the cavity was further amplified for imaging by a two-stage optical amplifier with in-band pumping at 1550 nm. A 50-cm length of TDF (TmDF200, OFS) was utilized in the first stage, followed by an optical isolator (ISO). A bandpass filter (BPF) formed by a circulator and a chirped fiber Bragg grating (CFBG) with a bandwidth of 4 nm was used to suppress the ASE noise. In the second stage, the signal was further amplified by a 1-m-long TDF, which was followed by another ISO and two cascaded 1560/1730-nm WDMs to filter out residual pump signals. The optimal lengths of the TDFs for the amplification process were also determined empirically through the cut-back method to minimize the reabsorption loss in the TDF, which is prominent in the shorter wavelength regime of its emission band [45].

The output and the performance of the 1725-nm HOPE characterized in the spectral and time domains are shown in Fig. 2. Figures 2(a) and 2(b) present the emission spectrum of the HOPE output before amplification by the two-stage optical amplifier. The HOPE spectrum is centered at 1725 nm with a 3-dB bandwidth of 1.4 nm, demonstrating the effectiveness of the dual gain scheme in maintaining a narrow gain bandwidth to ensure high selectivity of the lipid absorption band. Additionally, the dual gain setup significantly suppresses ASE noise, as evidenced by the low noise floor across the spectrum beyond 1.7 μm in Fig. 2(a). Notably, there is also negligible ASE noise near 1.9 μm, where the highest gain of the TDF is typically observed. By minimizing ASE noise, the system achieves an OSNR of 34 dB with high spectral purity. The combination of high spectral energy density and OSNR allows the source to generate PA signals from lipids with improved efficiency, resulting in enhanced sensitivity and specificity. This leads to the system’s ability to map the lipid distribution in tissues with high dynamic ranges or contrast ratios at lower pulse powers, minimizing the risk of photodamage and enhancing the safety aspect in bioimaging applications.

In the time domain, the HOPE achieves pulsed emission at a high PRR of 194 kHz, as illustrated in the inset of Fig. 2(c). The pulses have a full width at half maximum (FWHM) duration of 7.35 ns, which lies within the optimal pulse duration window for efficient PA generation. The stability of the pulsed emission is quantified by calculating the pulse width jitter and the standard deviation (std) to mean ratio of the peak pulse power. The pulse width jitter is a measure of the variation in pulse width from one pulse to another, while the std/mean ratio describes the stability of the peak pulse power. Using 3000 consecutive pulses, the pulse width jitter is found to be 2.2 ps, which accounts for 0.03% of the pulse width. Meanwhile, the std/mean ratio of the peak power is 1.87%, implying a high pulse-to-pulse power stability. High stability in pulse width and peak power is crucial for PAM because it minimizes variations in PA signal generation efficiency due to power fluctuations. This results in higher contrast and improved accuracy of imaging, enabling more reliable and detailed mapping of lipid distributions in tissues.

Figure 2(d) shows the output power from the HOPE system after amplification by the two-stage optical amplifier. The pump power for the first stage of the amplifier was fixed at 600 mW, while the pump power for the second stage was varied from 200 to 1500 mW to generate a plot of output power versus pump power at the second stage. The selection of 600 mW for the first stage was based on experimental findings that provided an optimal trade-off between gain and ASE within the system. Increasing the pump power beyond 600 mW in the first stage resulted in only a marginal increase in the amplified signal power. The TDF-based amplifier system can achieve a slope efficiency of 9.3%, with a maximum peak pulse energy of 670 nJ being achieved at a pump power of 1.5 W at the second stage amplifier. The maximum spectral energy density is found to be approximately 480 nJ/nm after amplification by the two-stage optical amplifier, with the 3-dB bandwidth of the spectrum remaining virtually unchanged after the amplification process. The performance parameters at 1725 nm are up to par with the latest report of the 1930-nm new HOPE in Ref. [4] in spite of the challenges of achieving TDF gain in the shorter wavelength regime. The performance of the HOPE in achieving a high OSNR and spectral purity allows it to induce efficient PA signal generation, and the ultra-stable emission in terms of the power and the pulse width variation mitigates artifacts and enhances accuracy. These factors allow for a high contrast ratio within the PAM images, allowing minute features to be distinguished clearly and reliably, thus facilitating quantitative analysis.

However, despite its distinguished performance, there are some limitations to the current 1725-nm HOPE architecture. Firstly, as the parametric gain in the cavity relies on precise phase-matching conditions for efficient four-wave mixing, the laser cavity remains vulnerable to environmental interference that may affect the polarization state of the signal. While efforts have been made to mitigate this by protecting the cavity with an isolation hood, complete elimination of the effects of unforeseen environmental stimuli cannot be guaranteed. In the future, we plan to incorporate a motorized fiber PC in combination with a PID controller system in order to minimize polarization changes, making the HOPE immune to environmental interference and further enhancing its long-term stability. Another limitation of the current system lies in its wavelength tunability. While, by the theory of the dual gain scheme, it is possible to achieve broadband tunability of the HOPE output, the tunability is restricted under the current architecture by the bandwidths of the passive components. In the future, the tunable operation of the HOPE will be expanded to include more wavelengths in the 1.7-μm window by upgrading the system with passive components that support more broadband operation.

The amplified output from the HOPE was employed in a traditional transmission-mode PAM system (detailed in Ref. [4]). The depth imaging capabilities of the PAM system were verified by the standard test depicted in Fig. 3(a). A prism-shaped block of agar gel (length, $\sim 6.5\mathrm{mm}$; height, $\sim 1.8\mathrm{mm}$) was submerged in a photoacoustic chamber filled with olive oil. The prism of agar gel emulated the water-rich tissue environments, acting as an attenuating medium. The sloped surface of the prism created a changing depth profile. This would allow the verification of the system’s ability to precisely track the depth-associated time delays in the lipid-induced PA signals, demonstrating its depth-resolving ability.

150 A-lines were collected along the sloping edge of the agar gel at 0.05-mm intervals with a fixed focal length. The delay in the lipid-induced PA signal allowed the visualization of the depth of the agar-oil interface, as shown in Fig. 3(b), where the depth relative to the lowest delay is plotted against the scan distance along the cross-section of the agar prism. The reconstructed slope of the interface closely follows the physical agar-oil interface, confirming that the system accurately measures the depth information based on differences in signal timing.

The normalized PA signals at three different points along the scanned line are shown in Fig. 3(c), with the waveforms numbered as per the correspondingly labeled locations along the interface in Fig. 3(b). The PA signal at the lowest point of the agar prism [labeled i in Fig. 3(b)] has the highest SNR due to the shortest optical path length through the attenuating medium. It should be noted that the waveforms display a weaker peak following the initial PA transient. This arises from the reflection of the PA signal at the optical window due to its high acoustic impedance. As the excitation beam moves along the length of the agar prism, the path length through the agar prism increases, resulting in a decrease in the SNR. However, even above a millimeter of agar depth, the SNR of the signal remains sufficiently high to distinguish the PA transient from the noise floor. This demonstrates the capability of the 1725-nm-HOPE-based PAM system for millimeter-level depth imaging notwithstanding the water absorption factor.

The PAM system was employed in the imaging of human liver samples with dichotomized steatosis grades assigned through clinical analysis. The fatty liver sample was designated as ‘severe’ with $>30\%$ steatosis, while the healthy liver sample was classified as ‘normal’ with 0% excess steatosis. The imaging was carried out with a pulse energy of 60 nJ incident on the samples. This corresponds to a fluence of $9\mathrm{mJ}/{\mathrm{cm}}^2$ per pulse, which is far below the $1.0\mathrm{J}/{\mathrm{cm}}^2$ ANSI safety standard for skin exposure to 1.7-μm laser illumination [54]. While the HOPE is capable of providing pulse energies of up to 670 nJ, the low pulse energy of 60 nJ was employed mainly for two reasons. Firstly, due to the high repetition rate of the source and mechanical scanning system utilized in the PAM setup, multiple laser pulses were incident on a single pixel point before translation to the next location. Thus, a low pulse power was used to minimize the risks of localized heating and associated microstructural changes in the lipid distribution in the sample. Secondly, reducing the pulse energy is effectively advantageous in demonstrating the capability of the 1725-nm HOPE in achieving sufficient SNR and excellent image contrast even under low-power conditions. This highlights the high sensitivity of the system for lipid-based PAM imaging while maintaining the integrity and safety of the biosample.

The results of the imaging are presented in Fig. 4. The true pictures of the normal and fatty liver samples after fixation on the photoacoustic chamber are shown in Figs. 4(a)i and 4(b)i, respectively. The PA signals acquired from raster scanning the healthy and the fatty liver samples under the 1725-nm-HOPE-based PAM system are reconstructed as both two-dimensional and three-dimensional images. To this end, the waveform at each pixel point was Hilbert transformed, and the 2D images were reconstructed from the maximum amplitudes of the transformed signals.

From an exemplary A-line within the fatty liver image, the SNR is found to be 12 dB. The maximal contrast ratio within the images is derived to be 23.6:1. The high contrast ratio allows a precise mapping of the lipid content within the samples, showcasing lipid localizations within distinct globular structures in the liver tissue. Given the high sensitivity and contrast ratio provided by the 1725-nm-HOPE-based PAM system, the fatty liver can be distinguished from the normal liver through plain visual comparison. Expressed as percentages, the RMS contrasts of the healthy liver and the fatty liver PAM images were found to be 9% and 19%, respectively. This is in line with the healthy liver sample having much lower variations in its pixel intensities, considering that it exhibits far sparser lipid content compared to the fatty liver sample. As can be seen in Fig. 4(b)ii, the fatty liver shows a much more widespread distribution of lipid signals compared to the normal liver [Fig. 4(a)ii]. There are seven large globular lipid-rich lesions distributed over the whole sample as well as dispersed lipid signals near the edges. On the other hand, only two lipid globules are apparent within the normal liver sample. There are also far less dispersed lipid signals over the whole scanned area. The lipid globules in the fatty liver are larger than the normal liver, with more intense PA signal amplitudes. This shows that the fatty liver sample indeed displays much more severe lipid accumulation compared to the healthy liver.

The PAM images are also reconstructed in three dimensions [Figs. 4(a)iii, 4(a)iv, 4(b)iii, and 4(b)iv]. The global alpha offset (transparency) of the 3D images has been adjusted to the same level for clearer visualization of the three-dimensional lipid deposits within. The lateral cross-sections of the normal and fatty liver show thicknesses of 0.96 and 1.09 mm, respectively, which correspond closely to the real thicknesses of the human liver samples. This demonstrates the capability of the 1725-nm-HOPE-based PAM system in achieving millimeter-level depth imaging. Along the same line, the three-dimensional structures of the fat deposits within the liver samples are also depicted within the PAM images. In both the fatty and the normal liver samples, the fat lesions take bulbous shapes, as is apparent in Figs. 4(a)iii and 4(b)iii. This shows the ability of the system to not only map the concentration and distribution of the lipid content within biological samples in one single plane but also allow the visualization of the three-dimensional structure of the lipid deposits. This capability can potentially be extended to study the structural formation of lipid-rich structures within biological tissues. Potentially, this can aid current medical and diagnostic insights into the onset and progression of NAFLD and other lipid-related disorders.

The intrahepatic lipid contents of the normal and the fatty liver samples were quantified using Otsu’s thresholding method. The ROI pairs and the corresponding segmentation results are shown in Fig. 5.

In each image, three pairs of ROIs at non-overlapping locations are identified. Each pair constitutes an ROI suffixed ‘-liver’ that determines the liver-to-water threshold $T_{\mathrm{liver}}$, and another ROI suffixed ‘-lipid’ that produces threshold $T_{\mathrm{lipid}}$ to segment the lipid-rich steatotic areas from the liver tissue. The ‘-liver’ ROIs are chosen to encompass only water/liver regions with low lipid signals to produce a threshold that can segment the liver sample holistically from the water background. On the other hand, the ‘-lipid’ ROIs are selected to cover the delineations between the regions with the lipid-rich lesions and the surrounding low-lipid tissue region. Particular care was taken to include as much of the highest lipid signal areas as possible to ensure that $T_{\mathrm{lipid}}$ does not overestimate the lipid content in the tissue.

The ROI pairs in the healthy liver were labeled H1, H2, and H3 and assigned the colors yellow, cyan, and green, respectively, as shown in Fig. 5(a). The segmented images produced by each of the ROI pairs are shown in the corresponding colors in Fig. 5(c). Similarly, the ROI pairs in the fatty liver samples were labeled F1, F2, and F3 (yellow, cyan, green) as seen in Fig. 5(b), with the segmented images presented in Fig. 5(d). In the segmented images, black represents the water background, gray represents the liver region, and the white pixels represent the steatotic areas. To illustrate an example of the steatosis percentage analysis from the segmented images, we can consider the ROI pair F1 in Fig. 5(b). In this pair, ‘F1-liver’ produces a threshold of 0.115 via Otsu’s method between the water background and the liver region. The second ROI of the pair, ‘F1-lipid’, produces a threshold value of 0.459 to differentiate the lipid-rich regions from the surrounding liver region. From this, we find the number of pixels above $T_{\mathrm{liver}}$ to be 9277 (sum of gray region, $N_{\mathrm{liver}}=9277$), and the number of pixels above $T_{\mathrm{lipid}}$ to be 2194 (sum of white region, $N_{\mathrm{lipid}}=2194$) out of a total of 20,800 pixels in the image. Using Eq. (3) defined in Section 5, the percentage of steatosis is then found to be 23.6%. In addition, the images are also segmented using the mean of the threshold pairs, as shown in Figs. 5(e) and 5(f).

Table 1 and Table 2 present results from the ROI analyses of the healthy and fatty liver samples, respectively.Table 1.

Quantification of Steatosis Level from the PAM Image of Healthy Liver Sample

From the results of the segmentation and subsequent analysis, the mean steatosis percentages are found to be 5.17% for the healthy liver sample and 23.6% for the fatty liver sample. The difference of more than 18% between these two values demonstrates a clear and significant relative variation in fat infiltration between the samples labeled ‘normal’ and ‘severe’, highlighting the ability of the 1725-nm-HOPE-based PAM system to effectively differentiate intrahepatic steatosis levels. It is important to note that the method used in the scope of this work quantifies steatosis based on the percentage of lipid-containing area rather than the proportion of hepatocytes with lipid accumulation as in histopathological techniques.

The 5.17% steatosis determined for the healthy liver sample aligns with the expected value, as lipids constitute about 5% of the total weight of normal livers [55]. Therefore, a steatosis level near 5% is generally considered within normal ranges by imaging and histological criteria according to the US NAFLD management guidelines [56,57]. The capability of the 1725-nm-HOPE-based PAM in detecting close to baseline levels of fat infiltration holds important implications for potential diagnostic applications. This is because ultrasound and CT, which are primary among the conventional modalities for the assessment of liver fat, suffer from low specificity and sensitivity for mild steatosis, which may lead to compromised diagnostic accuracy in the early stages of NAFLD [12,13,15]. The highly sensitive detection of baseline fat infiltration offered by the 1725-nm PAM system holds potential for complementing existing modalities in addressing this gap.

On the other hand, quantification of the fatty liver sample yielded a mean value of 23.6% steatosis. While this demonstrates a significant increase in lipid content compared to the healthy liver sample, the steatosis level remains lower than the sample’s clinically assigned grade of $>30\%$. This discrepancy may stem from differences in modality-specific calibration, such as those observed among MRI, ultrasound, and histological methods, which operate on individual grading scales. Despite this numerical difference, the substantial gap between the healthy and fatty liver samples underscores the ability of the 1725-nm-HOPE-based PAM system to discern varying degrees of steatosis in a completely non-destructive and fully quantitative manner. This suggests that PAM-derived steatosis quantification can provide a distinct yet complementary means of assessing liver fat, supplementing semi-quantitative histopathological assessments, which are susceptible to overestimation errors [58–60]. However, to ensure clinical applicability, the 1725-nm PAM-based quantification method must be benchmarked against current gold standard techniques. Therefore, in future studies, we will conduct imaging on a larger cohort of samples, in conjunction with histological analysis, to establish the correlation between the 1725-nm-HOPE-based PAM quantification method and the widely accepted reference standards.

The percentage difference in the steatosis levels obtained from the healthy liver sample from the three ROIs is 6.2%, while it is 3.39% for the fatty liver. Hence, there is less than a 10% deviation margin in the steatosis values obtained from different ROI pairs, indicating the reliability of the segmentation and analytical protocol used. The means of the steatosis levels obtained from the three threshold pairs also match the steatosis level deduced with the mean threshold pairs (0.19% deviation for healthy liver sample from 5.17% to 5.18%). Additionally, for both the healthy and fatty liver samples, the segmented liver regions (gray areas) closely match the actual tissue areas. However, some regions within the liver region were misclassified as water background due to pixel values falling below the $T_{\mathrm{liver}}$ threshold. This misclassification likely resulted from regions within the liver sample being out of focus of the 1725-nm illumination because of surface unevenness. This effect was more prominent in the healthy liver sample than in the fatty liver, possibly explaining the slightly elevated steatosis value obtained from the healthy liver sample.

In this study, we demonstrated the development and application of a 1725-nm HOPE for lipid-specific PAM at the 1.7-μm absorption window. By accessing the short wavelength gain of TDF, the 1725-nm HOPE achieves nanosecond emission with a narrow spectral bandwidth of 1.4 nm and a high OSNR of $\sim 34\mathrm{dB}$, enabling high sensitivity and imaging contrast. The source can also attain a PRR of 194 kHz, which is conducive to potential ultrafast PAM applications. The proposed source facilitated precise PAM imaging of lipid distribution in human liver samples with dichotomized steatosis grades, achieving a maximal contrast of 23.6:1. By employing Otsu’s method for segmentation-based quantification, the steatosis levels determined from the PAM images closely aligned with variations of steatosis levels from clinical assessments, underscoring the capability of this system for quantitative analysis. In summary, this study establishes a proof-of-principle demonstration of the 1725-nm-HOPE-based PAM as a high-sensitivity, high-contrast imaging tool for intrahepatic lipid quantification, effectively bridging the gap between macroscopic imaging and histopathological methods. Future work will incorporate comparisons with histological and gold standard methods to further validate its capabilities, which can potentially lead to the development of more robust quantification algorithms. Moreover, a limitation of the current system lies in the imaging speed bottleneck, whereby each image takes around 40 min to acquire. In the future, we aim to equip the PAM setup with a faster data acquisition system in order to shorten the imaging time and realize the full potential of the ultrafast imaging capabilities lent by the high PRR of the 1725-nm HOPE. We will explore the application of the 1725-nm HOPE for quantitative PAM imaging of lipids in the brain so as to gain insights into disordered lipid changes in neurodegenerative disorders [61–64]. Moreover, we will explore automated segmentation methods to facilitate high-throughput lipid quantification using the 1725-nm-HOPE-based PAM technique.

Sample preparation for PAM. The use of human liver samples is approved by the Ethics Committee Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong Western Cluster (UW 05-359 T/1022). The samples were initially in a frozen state at $-80°\mathrm{C}$. Shortly before the imaging was carried out, the samples were thawed and fixed in 4% paraformaldehyde solution (PFA) in phosphate-buffered saline (PBS) (pH 7.4). Immediately before the imaging, the samples were placed on top of the optical window in the photoacoustic chamber, bathed in PBS, and firmly covered with a thin PVC film to prevent them from coming into contact with the water coupling medium. The PAM images were acquired by raster scanning with a mechanical $X\text{-}Y$ linear stage (LTA-HS, Newport).

PAM image contrast analysis. The acquired data from the healthy and fatty liver samples were reconstructed into two- and three-dimensional PAM images, with most of the analysis being focused on the former. In order to analyze the contrast presented by the PAM images, we consider two metrics: (i) the contrast ratio of the PAM image and (ii) the root-mean-square (RMS) contrast. These two parameters provide complementary information about the 1725-nm-HOPE-based PAM imaging performance. The contrast ratio is a measure of the dynamic range of the image and, subsequently, the image quality. This reflects the capability of the 1725-nm-HOPE-based PAM system in highlighting the lipid-rich regions as opposed to the surrounding tissue and enhancing the visibility of tissue details: $$\text{Contrast ratio}=\frac{I_{\max }}{I_{\min }},$$(1)where $I_{\max }$ and $I_{\min }$ are the maximum and minimum values of the normalized pixel intensities in the PAM images.

The RMS contrast takes all the pixel values within the image into consideration, hence providing a holistic measure of the overall variation in brightness of the image. This is more conducive to comparative contrast analysis between the heathy and fatty liver images. The RMS contrast is defined as the standard deviation of the image pixel intensities, given by [65] $$\text{RMS contrast}={\left(\frac{1}{N}\sum _{i=0}^{N-1}{(I_i-I_{\mathrm{mean}})}^2\right)}^{\frac{1}{2}},$$(2)where $N$ is the total number of pixels, $I$ is the normalized pixel intensity $(0\le I_i\le 1)$, and $I_{\mathrm{mean}}$ is the mean normalized pixel intensity over the whole image.

Steatosis quantification protocol. A segmentation-based image analysis protocol was formulated to quantify the steatosis grades presented by the PAM images. This quantification protocol serves to validate the reliability of the 1725-nm-HOPE-based PAM method while enabling a more objective distinction between the two liver samples by quantitatively measuring the lipid content. This not only supplements the visual comparison of the PAM images but also enhances the accuracy in correlating lipid distribution with clinically established steatosis grades. Our intrahepatic lipid quantification protocol mimics the assessment of liver steatosis through histological analysis. Histological analysis deduces the level of steatosis by counting the number of hepatocytes containing microscopically discernible lipid vacuoles in the specimen [11,13,66]. In parallel to this approach, we look to establish the percentage of the steatotic ‘lipid-rich’ regions within the PAM images, which would correspond to lipid-containing hepatocytes under histological analysis. To this end, we employ a classical threshold-based segmentation approach based on Otsu’s method [67]. Otsu’s method is one of the most prevalent thresholding methods in use in image processing [68]. In its simplest form, this method selects an intensity threshold that minimizes intra-class intensity variance, while maximizing inter-class intensity variance [67]. In our PAM images, there are three constituent classes that need to be differentiated: (i) the water background, (ii) the liver tissue, and (iii) the intrahepatic steatotic lesions. This requires two thresholds: $T_{\mathrm{liver}}$ to differentiate the water background from the liver region, and $T_{\mathrm{lipid}}$ to differentiate the steatotic lesions from the liver tissue.

We manually select three pairs of regions of interest (ROIs) in the PAM images of the healthy and fatty liver samples. All ROIs are chosen to be $20\times 20\mathrm{pixels}$ in dimension. In each pair of ROIs, one ROI spans the boundary between the water background and the liver region. This ROI, denoted by the suffix ‘-liver’, produces the threshold value $T_{\mathrm{liver}}$. The second ROI in the pair, denoted by the suffix ‘-lipid’, encompasses the delineation between the lipid-rich areas and the non-lipidic regions. This produces the second threshold value $T_{\mathrm{lipid}}$. Three pairs of ROIs are chosen from non-overlapping locations within the images so as to ensure reliability and provide effective averaging. For the purpose of this work, the percentage steatosis as presented by the PAM images is defined by $$\text{Percentage steatosis by area}=\frac{N_{\text{lipid}}}{N_{\text{liver}}}\times 100\%,$$(3)where $N_{\mathrm{lipid}}$ is the number of pixels with intensity values above $T_{\mathrm{lipid}}$ and $N_{\mathrm{liver}}$ is the number of pixels with intensity values above $T_{\mathrm{liver}}$. In effect, this is the ratio of the lipid-containing area to the total area of the liver sample.

Acknowledgment. The authors would like to thank Mr. Mingsheng Li for valuable discussions and insights.