2.1.

Biology of Human Blood

Human blood consists of plasma (about 55 vol.%) and cells ($\sim 45$ vol.%) in which 99% of the cells are red blood cells (RBCs) and the remaining 1% are leukocytes and thrombocytes.^9,10 Plasma is a complex composition of dissolved ions (electrolytes), lipids, sugars, and proteins.¹⁰ RBCs, also known as erythrocytes, have a flat biconcave shape and a mean volume of $90\mu \mathrm{m}$,^{3 9,11} and they contain about 30 pg of hemoglobin, a globular metalloprotein responsible for oxygen transport throughout the body.^9,11 Hemoglobin concentrations range from 134 to $173\mathrm{g}/\mathrm{L}$ in whole blood and 299 to $357\mathrm{g}/\mathrm{L}$ in RBCs, varying according to age, gender, and health status. For example, anemia, cancer, or hereditary hemochromatosis decrease hemoglobin levels in the blood.^9,12

Each hemoglobin molecule contains four heme groups that can be bound to oxygen; the unbound state is referred to as deoxygenated hemoglobin, Hb, and the saturated bound state is considered oxygenated and denoted by ${\mathrm{HbO}}_2$.¹⁰ Sometimes the oxygen saturation of arterial blood, ${\mathrm{SaO}}_2$, is differentiated from that of peripheral blood, ${\mathrm{StO}}_2$, because arterial blood oxygenation should be the same throughout the body. In contrast, peripheral blood has varying oxygenation levels as oxygen is absorbed from the blood by peripheral systems.^11,13,14 Hemoglobin is also able to bind to other molecules forming carboxyhemoglobin, which arises during carbon monoxide inhalation;^15,16 sulfhemoglobin, which arises due to the irreversible binding of sulphur in the presence of sulfonamides;¹⁷ and carbaminohemoglobin, which results from the binding of hemoglobin in venous blood to carbon dioxide.¹⁸ A further variant state of hemoglobin is methemoglobin,^19–21 in which hemoglobin binds iron in the ${\mathrm{Fe}}^{3+}$ state (unlike normal hemoglobin that binds ${\mathrm{Fe}}^{2+}$), which prevents the binding of oxygen. Methemoglobin occurs naturally in blood at $\sim 1\%$ to 2% concentration,¹⁵ but it can be elevated due to side effects of medication or environmental factors.²² Finally, genetic variants of hemoglobin, such as hemoglobin S, which causes sickle cell anemia, can influence hemoglobin structure and binding properties.²³ Myoglobin, a chromophore commonly found in muscles, binds oxygen with a higher affinity than hemoglobin.²⁴ Myoglobin has a similar spectral response to hemoglobin and may be found in the bloodstream following muscle injury.^24,25

2.2.

Optical Properties of Biological Tissue

Tissue is a complex turbid medium composed of different cell types and protein-rich extracellular matrix, which strongly impact the propagation of light.^11,26,27 Absorption is the transformation of light energy to some other form of energy, such as heat, sound or fluorescence, as light traverses tissue and is quantified by the absorption coefficient, ${\mu }_a$ (${\mathrm{cm}}^{-1}$). Absorption is the primary optical interaction that is exploited to measure hemoglobin biomarkers. In turbid media, scattering is a major contributor to light attenuation and can confound attempts to measure hemoglobin absorption because blood is a highly scattering liquid with strong anisotropy. Scattering refers to a change in the direction of light propagation and is quantified by the scattering coefficient, ${\mu }_s$ (${\mathrm{cm}}^{-1}$), together with directional factors such as the scattering phase function and anisotropy factor, which further relate to the tissue refractive index.

2.2.1.

Optical properties of RBCs, hemoglobin, and its derivatives

The absorption coefficient of hemoglobin is a function of wavelength and the binding state (Fig. 1).³² Hemoglobin is usually oxygen-bound, whereas other variants mentioned above can modify the absorption spectrum and should be considered if relevant to the pathology being assessed because they can confound the measurement.^{10,13,23,33,34} It should be noted that the spectra of hemoglobin and its variety of physiologically relevant bound states are often measured after extracting the hemoglobin protein from RBCs, so they do not consider variation that arises due to scattering from different RBC geometries and orientations.

Fig. 1

The optical absorption of hemoglobin and associated variants. Representative spectra are shown for oxygenated,²⁸ deoxygenated,²⁸ carbamino,¹⁸ carboxy (visible²⁹ and NIR^19,20), methemoglobin (visible³⁰ and NIR^19,20), and sulfhemoglobin.³¹

In the absence of shear stress, human RBCs are biconcave discs with a diameter of 7 to $8\mu \mathrm{m}$, maximal thickness of 2 to $3\mu \mathrm{m}$, and minimal thickness of 0.8 to $1.5\mu \mathrm{m}$.^35,36 Concentrations of Hb within an RBC are high, on the order of 300 to $360\mathrm{mg}/\mathrm{mL}$, and the total refractive index of the cell is well approximated by that of pure hemoglobin through the application of the Kramers–Kronig relations on the total absorption spectrum.^37,38 When light is incident upon a single RBC, scatter from the near and far membrane interfaces leads to a characteristic oscillatory scattering spectrum and phase function.^36,39 The oscillatory spectral shape makes oximetry of a single RBC impossible without a priori knowledge of the precise shape and orientation of the RBC; however, averaging over many cells with random orientation can smooth this oscillation, which permits measurement of oxygen saturation of RBCs in capillaries.^39,40 In addition, hematocrit, the volume fraction of RBCs in whole blood, the presence of blood plasma, and other factors affect the absorption spectra.^18,41–44

The most common hemoglobin-derived optical imaging biomarkers are total hemoglobin (referred to as THb hereafter) and oxygen saturation (referred to as ${\mathrm{sO}}_2$ hereafter). THb is often evaluated using a single wavelength absorption measurement taken at an isosbestic point of ${\mathrm{HbO}}_2$ and Hb (i.e., when their absorption coefficients are equal). ${\mathrm{sO}}_2$ requires an absorption measurement to be made at multiple wavelengths (at least 2), usually spanning regions where either ${\mathrm{HbO}}_2$ and Hb dominate the absorption properties. Data are then often analyzed by applying multivariate statistical approaches for spectral unmixing⁴⁵ to extract the ${\mathrm{sO}}_2$ value. The absorption coefficients of ${\mathrm{HbO}}_2$ and Hb are related through ${\mathrm{sO}}_2$ to the overall optical absorption ${\mu }_a$ as^11,45

Eq. (1)

$${\mathrm{sO}}_2=\frac{[{\mathrm{HbO}}_2]}{[{\mathrm{HbO}}_2]+[\mathrm{Hb}]},$$

Eq. (2)

$${\mu }_a(\lambda )=c_{\mathrm{Hb}}\{\frac{{\mathrm{sO}}_2}{100}{\mu }_{{\mathrm{aHbO}}_2}(\lambda )+(1-\frac{{\mathrm{sO}}_2}{100}){\mu }_{\mathrm{aHb}}(\lambda )\},$$

where $c_{\mathrm{Hb}}$ is the concentration of hemoglobin in the tissue.

In addition to hemoglobin, many other molecules interact with light depending on their concentration and distribution throughout the tissue,⁴⁶ which can disrupt light propagation and also introduce spectral coloring at depth in tissue, confounding attempts to measure THb and ${\mathrm{sO}}_2$. Furthermore, the scattering and refractive index properties of blood are affected by hemoglobin concentration, erythrocyte volume, shape, and aggregation, each of which can be modified in disease. Unlike the absorption coefficient, the anisotropy and the scattering coefficients of blood are not dependent on the changes in oxygenation.^9,47

Hemoglobin-derived biomarkers often rely on variations in the absorption or scattering by hemoglobin molecules or RBCs, and the hardware used to make these measurements varies, including the use of LEDs or lasers for illumination with optical sensors (both arrays and point sensors) or ultrasound sensors for detection. The methods described in this review are summarized in Table 1.

Table 1

Overview of noninvasive hemoglobin monitoring and imaging.

Technology	Clinical application(s)	Spatial resolution	Benefits	Limitations	Clinical status and future potential	References
Pulse oximetry	Used to identify hypoxemia in a range of healthcare settings.	Single point measurement with no spatial resolution.	•Real-time monitoring of arterial saturation. •Extremely simple and quick to use in a clinical setting. •It can be used at the patient’s bedside.	•Single point location measurement. •It can be inaccurate due to calibration assumptions •Errors are associated with variations in hemoglobin and poor perfusion on the tissue measured.	Commonly used in primary through to tertiary care. Potential for increased deployment at-home through wearables and low-cost devices.	13,4849.50.51.52.53.54.55. 56.57.58.–59
Nailfold capillaroscopy	Diagnosing and monitoring systemic scleroderma and other arthritic conditions.	Spatial resolution varies from 0.1 to $10\mu \mathrm{m}$ laterally dependent upon the imaging NA, with corresponding imaging focal depth (Rayleigh range).	•Real-time imaging of capillaries and blood flow is relatively easy. •Typically, the design is portable, so it can be used at the bedside if needed.	•Lack of standardization around the quantification of capillary parameters. •Measures only structural, not functional, information about the capillaries. •Restricted to only imaging the nailfold capillaries.	Used in tertiary care. Emerging methods include quantification of blood flow, blood cell counts, and oxygenation imaging.	5,6061.62.63.64.65.66.67. 68.69.–70
Spectral reflectance imaging (MSI and HSI)	Narrowband endoscopic imaging.	Spatial resolution varies from 5 to $400\mu \mathrm{m}$ depending on the application because most systems use lenses for magnification. Reflectance signal restricted to superficial to $200\mu \mathrm{m}$ of tissue due to scattering.	•A versatile method that can image multiple hemoglobin biomarkers as well as other proteins of interest. •Relatively high spatial resolution is possible.	•Processing of images can be complicated and is not always possible in real time •Snapshot methods are fast but sacrifice spectral and spatial resolution •Scanning methods are slow but have a higher spectral resolution.	Commonly used in tertiary care for endoscopic surveillance. Otherwise, used in research and small-scale clinical trials with promising applications in cancer detection, skin lesions, e.g., burns, and surgical tissue health.	2,6,8,30,7172.73.74.75. 76.77.78.79.80.81.82.83. 84.–85
PAI	Breast lesion evaluation.	PAM can have a spatial resolution of $\sim 30\mu \mathrm{m}$ with an imaging depth of around 2 to 6 mm. PAT can have a resolution of $\sim 200\mu \mathrm{m}$ with a penetration depth 2 to 3 cm.	•Can assess the relative hemoglobin saturation together with other biomarkers at depth while maintaining a reasonable spatial resolution. •It is a versatile technique scaling spatial resolution with depth of imaging required.	•Trade-off between resolution and depth. •Requires acoustic contact between the tissue and detectors. •Systems can be costly, though LEDs can be used at the expense of imaging quality. •The use of high-power pulsed light can present additional safety considerations to ensure that stray light does not damage patients or clinicians’ eyes	Large scale clinical trials ($n>2000$) for the diagnosis of breast cancer. Small scale clinical trials or proof of concept for other applications, such as severity assessment in Crohn’s disease and dermatological conditions.	32,8687.88.89.90.91.92.93. 94.95.96.97.98.99.100.101.102. 103.104.105.106.107.108.–109
Spectroscopic OCT	Structural OCT is a clinical standard of care for retinal imaging. Spectroscopic OCT not yet clinically approved.	Lateral resolution determined by illumination optics, typically tens of μm but as low as 1 to $2\mu \mathrm{m}$ is achievable. Axial resolution determined by source/detector bandwidth, on the order of tens of μm down to $1\mu \mathrm{m}$. Penetration depth is limited to $\sim 1\mathrm{mm}$ in medium-scattering tissue.	Real-time tomographic imaging with high resolution ideal for resolving tissue layers, structural characteristics in 3D.	•Optical scattering in tissue limits imaging depth to 1 to 2 mm for most tissues. •Laser scanning of sample introduces potential for motion artifacts in image, ophthalmic visible OCT implementation limited by safety threshold and patient aversion.	Structural OCT widely used in primary and tertiary centers for ophthalmology, dermatology, and dentistry, also in tertiary centers for cardiology. Spectroscopic OCT for oximetry in preclinical development.	39, 110111.112.113.114.115.116. 117.118.119.120.121.–122
DOI or NIRS	Assessment of brain activity (fNIRS or DOT). Monitoring of tissue oxygenation (cerebral oximetry/NIRS).	1 to 30 mm spatial resolution; high resolutions are only possible at shallow imaging depths ($<2\mathrm{cm}$). Low-resolution imaging is possible up to 10 cm into the tissue.	Capable of determining blood oxygenation in tissue and other chromophores such as melanin, lipids, cytochrome-c-oxidase, and water.	•Relatively low resolution; obtaining high resolution requires prior knowledge of tissue composition and a high density of optodes. •DOI techniques may be combined with other imaging modalities such as MRI or ultrasound to assess tissue composition, resulting in increased system cost. •Computationally-intensive and time-consuming, resulting in limited real-time imaging (spectroscopy does not have this problem).	Clinical approval and small-scale clinical use of tissue oximetry/NIRS for assessment of brain oxygenation. fNIRS/DOT used commonly as a research tool to monitor brain activity. Small scale clinical trials or proof of concept for other applications, such as breast cancer diagnosis, and joint inflammation.	123124.125.126.127.128.129. 130.131.132.133.–134

3.1.

Clinical Applications and Research Studies

Pulse oximetry has been extensively reviewed elsewhere.^{13,48,49,135,136} Pulse oximetry is deployed in many medical applications, from at-home first aid to clinical intensive care units and surgical theaters¹³, and it has found particular utility in assessing hypoxemia in COVID-19 patients¹³⁷ as nonspecialists with minimal training can efficiently operate pulse oximeters.^49,138 Pulse oximetry research in the clinic focuses on its use in treating and diagnosing diseases, such as optimizing oxygenation of ventilated patients, screening neonates for congenital heart diseases, and monitoring patients with sleep apneas.^{50,51,139,140} Despite widespread use, it is also well established that pulse oximetry can suffer racial bias, which results in less accurate oxygenation readings for patients with more skin melanin content, a trait associated with darker skin. The impact of such bias is severe as it has been shown to result in less adequate medical treatment of such patients, meaning that it is important for clinicians to be aware of this limitation and it is also an important area for future research and development.^141,142 Although this bias has been well known for some time, it has been increasingly studied as a result of COVID-19 and the increased clinical use of pulse oximeters to treat respiratory conditions.

3.2.

Technology

Light absorption in pulse oximetry is typically measured using alternating illumination by LEDs at two different wavelengths.^135,143,144 Because the wavelength of the illuminating light is altered with time, oxygenation measurements are susceptible to motion artifacts, which change the area of tissue being illuminated and coupling to the tissue, resulting in inaccuracies. Commonly used wavelength pairs are 660 and 940 nm or 665 and 894 nm,^13,52–55 which are applied in two different modes.

Fig. 2

Pulse oximetry. Schematic illustration of pulse oximetry in the two different operation modes: (a) transmission and (b) reflection. The detected light is cyclic due to the pulsatile nature of blood in the peripheral vascular system. Both transmission and reflection modes have alternating components (AC) and direct components (DC). In tissue, the transmission and reflection of light vary based on the changes in absorption due to blood volume and oxygenation. That is $\mathbf{R}+\mathbf{A}+\mathbf{T}\approx 1$, when $\mathbf{R}$, $\mathbf{A}$, and $\mathbf{T}$ are the normalized reflection, absorption, and transmission intensities, respectively. For this reason, in reflection pulse oximetry, the peak intensity of light will be off by half a cycle from that of the transmission cycle. Examples of pulse oximeter devices include (c) transmission-based devices widely used in a clinical setting. Reproduced with permission from Ref. 145. (d) Low-power devices in development that adhere to the skin and use flexible OLED illumination. Reproduced with permission from Ref. 51. (e) Battery-free pulse oximeters in development that use near field communication for power. Reproduced with permission from Ref. 146.

Evolving from the traditional fingertip pulse oximeters [Fig. 2(c)], the current development of the technology mainly targets wearable devices and focuses on low power usage [Fig. 2(d)], optimization of signal detection, reduction of motion artifacts, flexible illumination and detection [Fig. 2(e)], low-cost devices, miniaturization, and calibration techniques.^51,56–59

3.3.

Analysis

The theory of oximetry analysis has been extensively reviewed by Mackenzie and Harvey,¹⁴⁸ so it will be only briefly introduced; readers are referred to the prior review for a more detailed description. The total extinction coefficient for blood is denoted as $\epsilon$ and related to ${\mathrm{SaO}}_2$ as¹³

Further analysis of the extinction coefficient is needed to isolate the signal from arterial blood because venous blood also absorbs the light, along with other chromophores that appear in the light path, such as melanin (in the skin). It is possible to exploit the cyclic nature of the extinction coefficient due to the pumping of blood by finding the ratio of the variable component (AC) and constant component (DC) at two different wavelengths [Figs. 2(a) and 2(b)], where the difference in light absorption is rather large^13,48,135

The ratio $R$, also known as the modulation ratio, is then related to ${\mathrm{SaO}}_2$ through a calibration procedure using best-fit analysis according to the equation

3.4.

Limitations

Pulse oximetry measurements typically have an error of 3% to 4% depending on the device and calibration used, which is actually sufficient to impact patient care in some cases.^150,156 In addition, standard pulse oximeters cannot detect the presence of methemoglobin, carboxyhemoglobin, or hemoglobin mutations, although their presence and concentrations outside of the expected range will affect the oxygenation readings.^15,23 Some targeted pulse oximeters are able to detect methemoglobin and carboxyhemoglobin, but they are usually used in specific scenarios in which high levels of these derivatives are expected due to exposure.¹⁵⁷ Hemoglobin $F$, present in fetuses and infants under 6 months, has a greater oxygen affinity, which allows the fetus to absorb oxygen from the mother’s bloodstream;¹⁰ in infants, its presence can increase the error in pulse oximetry by a further 3% in addition to the typical errors.¹³ Additionally, when there is poor perfusion to tissues, pulse oximetry can be limited, and if there is not a significant pulse detected, the technique will have increased inaccuracies.⁵⁵ Finally, pulse oximetry has been found to suffer racial bias in two large cohorts, in which black patients had nearly three times the frequency of occult hypoxemia not detected by pulse oximetry as white patients.^141,142 Skin pigmentation leads to an overestimation of arterial oxygen saturation in dark-skinned individuals, which could seriously impact medical decision making and long-term outcomes.¹⁵⁸ These limitations merit increased attention in research and development given the potential for long-term and widespread use of pulse oximetry in COVID-19 patient management and the interest in deployment of the technology in the wearable setting.

4.1.

Clinical Applications and Research Studies

A widespread application of reflectance hemoglobin imaging is in gastrointestinal endoscopy. Virtual chromoendoscopy methods adapt the light source of the endoscopy to focus on two wavelength bands (415 and 540 nm), where hemoglobin absorbs strongly (Fig. 1), thus providing a high contrast morphological image of the tissue vasculature to the clinician.^2,168 Narrowband imaging is the most widely established of these methods and meets the ASGE thresholds for targeting biopsies when imaging patients with Barrett’s esophagus for early signs of cancer.¹⁶⁹ More recently, clinical research studies have demonstrated that, by expanding the number of wavelengths captured in endoscopy,^72,73,170 it is possible to derive hemoglobin biomarkers of THb and ${\mathrm{sO}}_2$ from spectral information to classify disease status;¹⁷¹ however, further clinical study is needed to demonstrate efficacy.

Capillaroscopy is another reflectance-based imaging technique; it is used to image the blood vessels in the finger nailfold to diagnose disease, particularly to identify rheumatic diseases such as systemic scleroderma. Capillaries are microblood vessels from which oxygen and other nutrients are exchanged with the surrounding cells. In the finger nailfold, the capillaries are oriented in loops parallel to the skin, allowing for full visualization at high resolution in reflectance imaging mode. Capillary walls, formed from a single layer of endothelial cells lining the vessels, can rarely be detected during capillaroscopy, whereas the RBC column is visible, and morphological features associated with the capillary can be measured using monochrome, narrowband, and RGB imaging.⁶⁰ Capillary blood flow in the finger ranges in velocity from 0.67 to $4\mathrm{mm}/\mathrm{s}$ depending on physiological factors and the cyclic nature of perfusion.^172–174 Morphological dysfunction of the capillaries is easily identified using the current techniques, but current methods do not make oximetry measurements related to this dysfunction.

Reflectance-based oximetry imaging has been widely explored in retinal imaging because the retina is one of the most metabolically active tissues in the human body.¹⁷⁵ Commercial retinal oximeters are applied to fundus cameras and use dual-wavelength illumination, akin to pulse oximetry, to acquire images simultaneously at one isosbestic wavelength and one sensitive to ${\mathrm{HbO}}_2$. Abnormal retinal oxygenation has been shown to detect diabetic retinopathy, age-related macular degeneration, and glaucoma.^74,75,176 In addition, the retina has similar vascular properties to the brain, making it a perfect window for understanding and diagnosing neurological disease in addition to ocular disease.⁷⁵ Nonetheless, retinal oximetry has yet to find routine clinical application, limited primarily by the impact of fundus pigmentation and vessel size on quantification.⁷⁵ Although some hyperspectral imaging technologies have been explored in an attempt to overcome these limitations, to the best of our knowledge, they have not found clinical application.

In smaller scale studies, clinical trials of reflectance-based hemoglobin imaging have been applied in areas from the skin to the brain. For example, the hemodynamic response of the human cortex has been visualized during open-cranium surgery using hyperspectral imaging combined with multivariate analysis.⁷⁶ Evaluation of hemoglobin oxygenation and melanin content in the face has been of interest for dermatological treatments such as the detection of skin cancer, assessment of scar healing, and evaluation of skin thickness.^177–180 Spectral imaging of hemoglobin in burns,⁸ wounds,⁶ and bruises^30,77 has been explored to assess the progress of healing in a quantitative manner. Moreover, spectral imaging of tissue ${\mathrm{sO}}_2$ has found application in intraoperative imaging.^161,181 Postoperative imaging can also provide clinicians with information on how tissue is healing such as following breast reconstruction in which water, hemoglobin concentration, and oxygen saturation are key indicators for tissue perfusion, an important factor in recovery.¹⁸²

4.2.

Technology

The simplest imaging oximetry methods include two or three wavelengths akin to pulse oximetry, which can be applied using sequential illumination by the target wavelength bands and imaging with a single camera or simultaneous illumination of all wavelengths (e.g., with broadband illumination) and capture using a spectrally resolved method, such as image splitting through band pass filters applied in front of two cameras. Expanding the spectral range of wavelengths to capture more spectral features of Hb or ${\mathrm{HbO}}_2$ (in the visible but also near-infrared range¹⁸³) requires more complex hardware.

There are three main types of multi or hyperspectral systems used in medical imaging that can be applied for unmixing oxygenation of tissue.^78,79,159 In spatial scanning, a spectrograph records the spectral dimension ($\lambda$) while being scanned across a sample. A one-dimensional spectrograph may be point-scanned or a two-dimensional (2D) spectrograph may be line scanned. The approach sacrifices temporal resolution and hence requires minimal movement of the sample to avoid motion artifacts. In spectral scanning, a 2D camera records the spatial dimensions $(x,y)$, while the spectral dimension is scanned, e.g., by changing the illumination wavelength or by filtering the imaging light path (e.g., with a filter wheel or tunable filter). This also has a scanning time associated with it, so it requires minimal movement to prevent spectral artifacts.

Finally, in snapshot methods, a system outputs spectral and spatial dimensions $(x,y,\lambda )$ simultaneously without scanning.^78,79 Snapshot systems may use beam splitting with dichroic filters, volume holographic optical element splitters, image replicating spectrometers, or multispectral filter arrays in the imaging path. Snapshot methods avoid motion artifacts, which makes them an exciting prospect for clinical application; however, they often exhibit poor optical efficiency. They also require a trade-off between spatial and spectral resolution, although this is less problematic for hemoglobin measurements, in which the target spectra are well characterized, and extensive evidence exists for the application of optical measurements. Nonetheless, optimization of the target wavelengths for snapshot imagers can substantially improve their performance.^{166,184–187}

In addition to these intensity-based imaging methods, spatial frequency domain imaging (SFDI) can be used to interrogate hemoglobin-related biomarkers by exploiting modulated illumination of the tissue. SFDI typically uses sinusoidal patterns and extracts the demodulation of these patterns reflected from tissue to calculate absorption and scattering properties; ^123,188 low frequencies are more sensitive to absorption whereas high frequencies are more sensitive to scattering.¹²³ Multiwavelength illumination¹⁸⁹ can then be used to calculate hemoglobin content and oxygenation,^190,191 e.g., in monitoring of peripheral circulation and vascular diseases,¹⁹² for which there are FDA-cleared devices, as well as ulcers,¹⁹³ burns,¹⁹⁴ or tumor margin detection.^195,196 SFDI provides relatively high-resolution images, but conventional methods can be sensitive to motion artifacts¹⁸⁸ and computational processing can limit the rate of image display. More recent reports have shown that it is possible to overcome these limitations using single snapshot of optical properties methods, which can achieve video-rate imaging.¹⁹⁷

4.3.

Analysis

Two- or three-wavelength imaging methods may be viewed qualitatively and interpreted by the operator, as in narrowband imaging, or processed to output quantitative THb or ${\mathrm{sO}}_2$ biomarkers in a manner similar to pulse oximetry, through calculation of image ratios and calibration of the results. For spectral imaging methods, analysis can be time-intensive due to the large amount of data collected, which can be problematic for clinical translation in which the real-time display of biomarker data is often desired. Analysis methods vary depending on the biomarkers targeted and the type of tissue imaged from the simplest techniques such as linear spectral unmixing^198,199 to more complex methods such as multivariate analysis and machine learning.^{26,72,164,165} Linear spectral unmixing determines the type and concentration of chromophores present based on input reference spectra for oxy and deoxy hemoglobin, from which oxygen saturation can be calculated.^{76,170,200,201} Spectral signatures can also be used directly in classification of disease status, e.g., cancer.^{171,181,202,203} Sometimes, a combination of classification and unmixing techniques can produce the optimal results, allowing for data corrections to be applied in certain tissue types.^{26,76,166,170} Similar methods are also used in depth-resolved imaging, but data may need to be corrected based on the imaging depth and the associated level of optical absorption and scattering. Machine learning methods have shown promise in this regard, enabling more accurate determination of hemoglobin oxygenation, particularly at depth, than classic linear spectral unmixing.²⁰⁴

4.4.

Limitations

Imaging tools for the assessment of hemoglobin can be subject to the same limitations as pulse oximetry. In addition, constraints that presently prevent clinical adoption of reflectance-based spectral imaging include cost, reliability of THb and ${\mathrm{sO}}_2$ measurements, clinical evidence for sensitivity and specificity of THb and ${\mathrm{sO}}_2$ in the diseases shown to be of interest in small-scale studies, and the need to process data in real time.^2,71,167 These challenges are common in the clinical translation of optical imaging biomarkers,² though fortunately, in the case of hemoglobin biomarkers, many devices have already navigated the pathway to the clinic, enabling initial studies to be undertaken.

5.1.

Photoacoustic Imaging (PAI)

5.1.1.

Clinical research studies

By far, the most explored clinical PAI application is human breast cancer detection, extensively reviewed elsewhere,^208,209 due to the enhanced angiogenesis of breast cancers compared with background breast parenchyma and the scalability of PAI geometry allowing for a broad view of the area [Fig. 5(a)].^{86,89,208,209,213,214} Multicenter clinical trials have recently been concluded covering >2000 women; these established the ability of PAI to increase the specificity of ultrasound imaging using a real-time map of relative Hb and ${\mathrm{HbO}}_2$.^213,215 PAI has also been explored in other cancer types, considering that neoangiogenesis is a hallmark of cancer, including thyroid,²¹⁶ prostate,²¹⁵ and melanoma among, others.²¹⁷ Beyond applications in cancer, PAI has found a wide range of potential applications in which depth-resolved information is required, e.g., in endoscopic procedures;^90,91 for evaluation of inflammation, such as arthritic joints,^88,92 foot ulcers,⁹³ and Crohn’s disease;²¹⁸ vascular imaging;⁹⁴ and for the guidance of interventional procedures, such as in fetal placentas.⁹⁵

Fig. 5

Tomographic imaging of the human breast for cancer detection (a) PAI of ${\mathrm{sO}}_2$ in breast with infiltrating ductal carcinoma (IDC); S-factor was defined to account for system accuracy and fluence compensation. Reproduced with permission from Ref. 210. (b) DOI of breast IDC (indicated by the red box) resolves ${\mathrm{sO}}_2$, THb, ${\mathrm{H}}_2\mathrm{O}$, lipid, concentrations of which serve to highlight the tumor. Reproduced with permission from Ref. 211. (c) DCS of blood flow relative to an ultrasound image of low-grade carcinoma; the tumor is circled in yellow. These images are referenced to positions s1 and s2 to compare the ultrasound, 3D reconstruction, and cross-section. Reproduced with permission from Ref. 212.

5.2.

Diffuse Optical Spectroscopy and Imaging

5.2.1.

Clinical research studies

Diffuse optical spectroscopy (DOS) is commonly referred to as near-infrared spectroscopy (NIRS) because it uses light in the near-infrared range; the term functional NIRS (fNIRS) is also commonly used but is usually restricted to applications monitoring functional responses to stimuli in the brain via neurovascular coupling. Quantifying and monitoring changes in oxygenation of blood in the brain has found many applications that range from understanding seizures²²³ to detecting brain damage.⁷ Unlike reflectance hemoglobin imaging [Figs. 6(a)–6(d)], in which an open cranium is required (see Sec. 4), fNIRS typically achieves imaging depths of up to 15 mm through the skull [Figs. 6(e) and 6(f)], which covers the outer cerebral cortex in healthy adults.^{127,223–225} fNIRS imaging of the brain to identify intracranial hematomas due to brain trauma has been clinically approved by the FDA recently; however, it has yet to be widely deployed in clinical settings.⁷

Fig. 6

Hemoglobin imaging of the human brain. (a)–(d) Reflectance spectral images of the brain in an adult undergoing epileptogenic tissue resection (a) reference RGB rendering, (b) change in oxygenated hemoglobin over a single timeframe, (c) change in deoxygenated hemoglobin over a single timeframe, and (d) change in total hemoglobin over a single timeframe. Reproduced with permission from Ref. 76. (e), (f) DOI of a neonate during a seizure: (e) changes in HbT concentration mapped throughout the onset of a seizure and (f) average changes in Hb, ${\mathrm{HbO}}_2$, and tHb postonset of the seizure. Reproduced with permission from Ref. 223.

Similar to PAI, DOI has also been deployed in clinical trials to detect cancer, particularly in the breast^226,227 [Figs. 5(b) and 5(c)] and thyroid,¹²⁷ where it has also been used to monitor response to therapy. DOI tends to be lower in spatial resolution than PAI (Fig. 5), but it can often resolve other biomarkers in addition to hemoglobin. PAI typically has a lower temporal resolution compared with DOI.^228,229 In addition, DOI has been used to image muscle tissue such as the forearm, peripheral tissue, and joints.^127,225 Several reviews have been published illustrating the importance of DOI.^230–233

5.3.

Spectroscopic OCT

6.1.

Trade Offs

Choosing the optimal hemoglobin imaging technique for a particular application involves consideration of several factors that often require trade-offs, including signal-to-noise ratio, spatial and temporal resolutions, target depth, and route to integration with existing clinical practice. Optical imaging techniques are ultimately restricted by the maximum permissible exposure at the illumination site, which places a fundamental limit on the signal-to-noise ratio available in the clinical setting. Some techniques are further restricted in the type of illumination used such as OCT, which requires coherent light, and PAI, which requires short, pulsed light; this can add to the complexity of safety considerations in the clinic.

Considering the factors of resolution and depth, reflectance-based imaging can achieve high spatiotemporal resolution in applications in which depth resolution is not vital, for example, with retinal, endoscopic, or intraoperative imaging. One could argue that depth-resolution will become increasingly prevalent with the emergence of more advanced solutions from spectroscopic OCT, PAI, and DOI, particularly as costs decrease. Nonetheless, depth-resolution typically implies a sacrifice of spatial or temporal resolution, which must be determined early in the discovery and development phase of the associated device. Furthermore, different approaches to achieving depth resolution have different strengths and weaknesses. PAI may suffer from shielding effects due to the absorption of overlying tissue and tends to resolve only larger vessels, whereas the use of multiple scattering by DOI enables better detection of capillary oxygenation, despite the overall poorer spatial resolution. This trade-off may explain why PAI has developed more rapidly in clinical translation for breast cancer detection, whereas DOI is more developed for imaging the brain. Given the common contrast source across many of the techniques described, it may also be desirable to combine multiple approaches in a single device for validation purposes or to provide views of the same tissue at different resolution scales, overcoming some of the limitations of the existing technologies.²⁶³ Combining two or more optical imaging techniques can be beneficial because they reduce the imaging limitations that arise as a result of a single technique. A good example of this is the enhanced perfusion oximetry system that combines diffuse reflectance spectroscopy with LDF to quantify blood oxygenation and flow rate for imaging of microvasculature and burns.^264–266

6.2.

Clinical Implementation

There are many factors to consider regarding the route to integration of new techniques into clinical practice. For example, if imaging is required to be non-contact, for example, in delicate targets like the eye, this may restrict the implementation of methods such as PAI or DOI, in which contact with the tissue is typically required. The pathway to clinical adoption may be smoother in the case of an existing optical imaging solution already being deployed. An obvious example is the large-scale deployment of OCT for ophthalmic applications, which provides a direct route for adoption of spectroscopic OCT in the community. Another factor in clinical translation is the assessment of the precision, accuracy, and bias of new biomarker measurements.² These factors can be affected by both device operation and data interpretation, requiring standardization of data acquisition and careful consideration of any signal processing or analysis methods applied before presenting the data to the interpreter, whether this is a human or a machine. Because reflectance-based imaging techniques are usually less computationally intensive compared with the depth-resolved methods, the development of systems and software for real-time imaging is more easily attainable; this is seen most effectively in narrowband imaging of blood vessels in endoscopy.

Further clinical considerations arise later in the translational pathway when the question of biomarker efficacy in decision-making finally arises. Instrument prototypes are often used first in pilot clinical trials at a single clinical center to gather initial data for validation and may be subject to multiple design iterations at this stage. Having successfully passed the first translational gap, which may include CE marking or FDA approval for the device and multicenter clinical trials, the technique is then subject to advanced qualification and ongoing technical validation to determine clinical utility in the healthcare setting, whereby the measurement can be used in clinical decision making. These larger-scale clinical trials help to determine sources of variation that will influence the classification and diagnosis of disease, providing clinical evidence of the ability to change patient management. They also provide extensive reference data sets that can be used to improve interpretation, particularly when machine learning-based methods are involved.

Reflectance-based spectral imaging techniques are still largely in the earlier stages of development with first in-person trials, whereas fNIRS of the brain and PAI of the breast are being used in multiple centers as part of larger-scale clinical trials and DOI has found some level of adoption into the clinic.^{71,87,208,228,231,267} Data arising from these trials is extremely valuable and making annotated datasets open source for the community in the future will not only help accelerate the development of new algorithms but could also enhance our understanding of the biology of hemoglobin oxygenation in disease.

6.3.

Disease Monitoring

Hemoglobin imaging methods could find further applications in monitoring disease, to detect treatment efficacy or disease relapse. The noninvasive nature of these techniques can allow for continuous or periodic monitoring, for which there are several excellent examples that have been highlighted. Pulse oximetry can be applied to a patient for long periods so that clinicians can observe if there are any changes to overall arterial oxygenation. DOI and NIRS also can be used longitudinally to monitor changes in brain function to assess if there is improved brain activity.²⁶⁷ Furthermore, periodic monitoring of diseases such as scleroderma using nailfold capillaroscopy can indicate progression in the stage and severity of the disease, which may influence the treatment provided or indicate if further medical intervention is required.

Periodic monitoring can be applied in the short term, for monitoring wound or burn healing, or in the longer term in the context of endoscopic cancer surveillance in at-risk patient groups such as those with Barrett’s esophagus that are at increased risk of developing cancer. The frequency of monitoring applied is determined by the disease being observed and the rate at which change is expected, as well as by the training required for instrument use and data interpretation. For example, applying pulse oximetry is relatively quick with nonspecialists able to do it, and some more straightforward versions of capillaroscopy can be done using handheld devices. Conversely, endoscopies, OCT, DOI, and NIRS typically require training of specialist operators; hence it can be more expensive to conduct the procedures, meaning they are typically used for less frequent monitoring.

6.4.

Outlook

The acceptability and relevance of new hemoglobin sensing and imaging technologies to clinicians will be driven by various factors, including cost, complexity, and physical size of the systems, as well as the ease of use and data interpretation. The commonplace use of pulse oximetry means that the clinical community is already well aware of the use of systemic ${\mathrm{sO}}_2$ as a disease biomarker. Ongoing technological developments that lead to miniaturization of light sources, optical components, and cameras, as well as decreasing their cost, mitigate some of the technical limitations highlighted in this review. Advances in image processing, including convolutional neural networks, promise to aid in distilling rich datasets to actionable clinical information, enabling imaging systems to be more easily integrated into clinical care. Research questions remain regarding the sensitivity of hemoglobin sensing techniques in diverse populations and the diagnostic power of hemoglobin-derived biomarkers in the wide array of disease presentations included in this review, which will only be answered through comprehensive clinical trials. Hemoglobin imaging techniques add a new dimension of knowledge in a range of clinical settings, from capillaroscopy and endoscopy to intraoperative imaging; emerging technologies are well placed to further enhance these areas of existing clinical practice, but are also likely to contribute to the decentralization of healthcare to tertiary care centers and through the deployment of wearable technologies for self-monitoring in the home.