3.1.

Excitation Light Source

Choice of excitation light source is based on the spectral bandwidth, solid angle of output beam, output efficiency, and regulatory considerations. Commonly used excitation sources are (i) filtered broadband lamps, (ii) laser diodes, and (iii) light-emitting diodes (LEDs). Ease of filtering at detection, illumination of large field of view (FOV), output fluence rate, mounting requirement, and cost are the main factors that influence the choice of excitation source type. Among available options, filtered lamps have the lowest efficiency, largest spectral bandwidth, and largest solid angle, and hence least spatial and spectral confinement. Furthermore, a large fraction of the output photons is rejected at the excitation filter resulting in high heat dissipation, thus making their use cumbersome and suboptimal. Such excitation source setups are seen currently in surgical microscopes.

Laser diodes have the highest spatial and spectral confinement of all sources. The low spectral bandwidth allows maximum excitation light filtering at the fluorescence detection camera. High-power options are available to deliver the best light fluence rate for low fluorophore concentrations, but safety concerns related to maximum permissible exposure limits for skin and eyes complicate the regulatory approval of systems that use these. Moreover, beam expanders would be necessary to ensure illumination of large FOVs. Laser diodes also require precise temperature and current control to ensure fidelity of the output spectrum and power, thus necessitating additional hardware, and remote mounting away from the patient. Such systems would use fiber coupling from the light source to the illumination head. These are seen in the Fluobeam and Curadel Lab-Flare systems.

LEDs provide a trade-off among output power, efficiency, cost, and spectral bandwidth. With the growth of the LED market, it is becoming increasingly economical to produce high-power LEDs. To ensure homogeneity of the excitation field, LEDs would need to be combined into an array. However, one of the major drawbacks of using LEDs is that when fluorophores with small Stokes shifts are used, there will be leakage of excitation light past the emission filter leading to reduced SBR. An excitation filter would thus need to be used to confine the output of the light source. Some of the newer imagers such as PerkinElmer Solaris and SurgVision Explorer Air system utilize LED-based excitation, and we should expect to see further increase in their use.

3.2.

Collection Optical Components and Emission Filters

Multiple trade-offs exist when discussing collection optics. These are FOV size, depth of field, lens $F$-number, and operating distance. Collection optics may be designed for fixed magnification or variable magnification depending on the need to have variable field size and operating distance. Most common imagers have a working distance of 10 to 30 cm with a maximum FOV size of about $15\times 15{\mathrm{cm}}^2$, but the tolerance for focus errors varies among manufacturers and end users.

Emission filter design and choice are critical in maximizing detection sensitivity by limiting background light. Filter choice is influenced by spectral overlap among the reflected excitation light, ambient lighting, and the Stokes shift of fluorophore of interest. Fluorophores with large stokes shift such as PpIX, which can be excited in the blue range ($\sim 405\mathrm{nm}$) and detected in the red range ($\sim 635\mathrm{nm}$), pose few problems, whereas those that have significant overlap between absorption and emission spectrum such as ICG, require more careful selection (See Fig. 3). Longpass, bandpass, and notch filters are broadly the categories for selection. Interference filters generally provide superior out-of-band rejection and transmission in the passband as compared to absorption-based filters. Spectral characteristics of filters also vary with incidence angle of light and there will be transmission of undesirable excitation light at high incidence angles. This along with the excitation leakage through the rejection band determines the noise floor of a device and hence affects its sensitivity.

3.3.

Imaging Sensor

The factors that influence detection performance of the camera are dynamic range, read-out rates, pixel resolution, and on-chip gain. For quantitative imaging, bit depths of 10-bit or more are desirable to provide signal detection of 2 to 3 orders of magnitude while maintaining a low noise floor. Charge-coupled device (CCD)-based cameras are used in fluorescence imagers almost ubiquitously, but suffer from low quantum efficiency in the far-red and NIR-wavelength range and slow read-out time ($<30\mathrm{Hz}$). They can achieve low read-out noise when cooled, generally have high resolutions, and can be used with $16\text{-}\mathrm{bit} \mathrm{A}/\mathrm{D}$ converters. Further improvements in sensitivity are achievable with electron multiplied-CCDs (EMCCD) and intensified-CCDs (ICCD); these can provide analog gains of over $1000\times$ and are of benefit when signal intensity is very low, but signal quality can degrade at high noise levels; these trade-offs would need to be considered carefully. The available quantum yield of sensors in the far-red- and NIR-wavelength regime can drop off significantly from the visible region, and this too will affect appropriate camera/sensor selection on a system. Almost all commercial fluorescence imagers use CCD cameras and the SurgVision system uses an EMCCD camera. The scientific-CMOS (sCMOS) cameras are now a contender against CCD cameras as they provide high read-out rates, high bit depth, and the specific advantage of many more pixels per frame. In addition to the high read-out rate and high bit depth, the compact size, light weight, and low read-out noise make sCMOS cameras the sensor of next-generation fluorescence imagers. One of the drawbacks, however, is the higher cost as compared to standard CMOS technology. Among the commercial systems available, the PerkinElmer Solaris is the only system that uses an sCMOS camera. More sophisticated camera configurations such as multispectral cameras, though rarely used in most systems, will be discussed briefly in Sec. 4.1.

3.4.

Software Control, Computing, Data Storage, and Display Hardware

Software designs vary in the degree of user customizability; those that target clinical use generally have the least flexibility while investigational systems and research-oriented systems allow a good deal of user customization. With growing bit depths, high-speed data transfer and data storage become important considerations. Imaging and storing large video sequences, potentially from multiple cameras, can result in several gigabytes of imaging data per hour—storing one hour of single channel $1024\times 1024{\text{pixel}}^2$ 8-bit fluorescence data at 30 fps for an hour would need $>100\mathrm{GB}$ of space—posing a significant data management challenge. One approach to tackle this is to store compressed video files only and save data into an 8-bit format, even if the camera provides $>10\text{-}\text{bits}$ per pixel. Alternative strategies include saving user-prompted snapshots from a continuous video stream. The ability to customize storage and export on to external drives or servers may be a solution as well, but local protected storage is likely the best candidate for clinical systems to maintain Health Insurance Portability and Accountability Act compliance. However, data storage may be critical in research settings, to allow postprocessing, and image analysis postacquisition. Systems such as the PDE Neo lack on-board storage options and provide only screen captures, which is nonideal and nonquantitative.

As devices grow in sophistication, software control of instruments becomes more complex especially when multicamera systems are used. On-board GPUs are often necessary for simultaneous overlay and streaming. It is important that systems be customizable yet easy to use for clinical staff and surgeons. The software functionality also directly ties-in with data visualization and display optimization. Need for ROI intensity measurement tools, and on-screen window-level and compression options⁵⁰ will increase as systems are used for quantitative or semiquantitative imaging, and as image bit depths exceed the display bit depths. Software design is the most understated aspect of fluorescence imagers, but since their use during surgery would need seamless integration with the surgical protocol, robust, intuitive design is key.

4.1.

Real-Time Overlay of White-Light Reflectance and Fluorescence Images

Most FDA-approved imagers, such as Novadaq SPY, Hamamatsu PDE Neo, and Fluoptics Fluobeam 800, are single channel fluorescence video/image display systems. However, imagers that have the ability to provide the fluorescence image overlaid on a white-light illuminated RGB image in a real-time video stream would provide richer and more complete information to a surgeon/user. To produce such images in real time is significantly more complex than a single channel fluorescence video stream. While there are several approaches for achieving overlaid data, wavelength-based separation of fluorescence ($>650\mathrm{nm}$) and visible white light ($<650\mathrm{nm}$) is the main principle upon which imager designs are based.

The most commonly employed technological approaches for achieving simultaneous white-light and fluorescence imaging are use of beam splitters and multiple cameras or multispectral cameras that separate visible and far-red and/or NIR-wavelengths within the camera itself using prisms and multiple CCD or CMOS sensors. The Flare intraoperative prototypes from the Frangioni laboratory use three cameras,⁴¹ and Curadel’s Lab-Flare imager, based on the Flare prototype, uses three CCD sensors within a single camera body to simultaneously image two fluorescence channels and a white-light RGB channel, by removing the $\sim 800\text{-}\mathrm{nm}$ (NIR) fluorescence component, the $\sim 700\text{-}\mathrm{nm}$ far-red fluorescence component, and using the remainder to produce the RGB image, as shown in Figs. 2(b) and 2(d). The Quest Spectrum system (previously called Artemis) achieves similar wavelength-based separation using prisms within a single camera, and can simultaneously image and overlay two NIR channels (700 to 830 nm and 830 to 1100 nm) on the white-light RGB image stream. However, expansion to more channels would require additional cameras/sensors for simultaneous acquisition. VisionSense Iridium also uses two CCD sensors to produce simultaneous white-light RGB and NIR fluorescence images from the $\sim 800\text{-}\mathrm{nm}$ emission channel and merges the two in real time. These approaches work well in practice, allow independent gain adjustments for each channel, and do not require sequential pulsing of excitation lights. The SurgVision and PerkinElmer Solaris systems, on the other hand, feature two cameras, one for white-light image acquisition and one for fluorescence acquisition, with overlay capability. A screenshot of the Solaris display in Fig. 2(a) shows “white-light RGB” and “white-light RGB + fluorescence overlay” images showing fluorophore uptake in murine lymph nodes. Additional considerations for multicamera setups include coregistration of the various video streams and magnification corrections. For example, Novadaq SPY uses two cameras that are not coregistered and have different pixel dimensions and zoom, as such no overlay functionality has been implemented, and white-light images are available only in snapshot mode.

Fig. 2

Shows various white-light and fluorescence overlay schemes. (a) Shows a screenshot from the PerkinElmer Solaris imager during lymphatic imaging (image courtesy of PerkinElmer). The imaging windows display white light and the fluorescence overlaid on the white-light images simultaneously. User processing controls, such as ROI and display gain adjustments, are also available. (b) Shows the commonly used wavelength-based separation of collected light using dichroic mirrors and filters as seen in the Flare prototype system (reprinted from Troyan et al.⁴¹ with permission of Springer), Curadel Lab-Flare uses a similar setup with slightly different wavelength specifications on beam splitters and emission filters. (c) The modified Bayer filter is an alternative approach to perform simultaneous NIR detection, though this approach limits the active area for the fluorescence channel, reducing sensitivity.⁵¹ (d) Shows an example of simultaneous imaging and display of 700-nm (red) and 800-nm (green) fluorescence channels from the Flare prototype with the mesenteric lymph nodes highlighted by methylene blue (brackets) and a sentinel node (arrow) highlighted by ICG, reprinted from Troyan et al.⁴¹ with permission of Springer.

Other related approaches employ Foveon X3 sensors (HyperEye Medical System, Mizuho Medical, Japan)⁵² to detect unabsorbed NIR light or modify the Bayer filter pattern and filter the NIRF signal at the entrance to the sensor,⁵¹ thus separation happens within the camera itself [see Fig. 2(c)]. These approaches, while overcoming the problem of merging multiple streams, can be limited in their sensitivity to weak fluorescence emissions, which is most significant for targeted tracers at low concentrations in tissues.

The optimization of fluorescence visualization and displays is often underreported as systems are only now beginning to exploit overlay-based displays. With the growth in use of 10-to 16-bit acquisition cameras, proper scaling and mapping of displays on traditional display monitors are an additional concern. Application of appropriate transparency functions to the fluorescence overlay, choice of appropriate colormaps, and need for compression⁵⁰ techniques to display high bit depth images are important areas with only limited discussion in the literature. The field of high dynamic range (HDR) imaging has allowed optimal display of HDR data, yet this approach has not penetrated into medicine much as of today. As fluorescence imagers incorporate simultaneous multiple channel imaging, this aspect will gain greater importance. Elliott et al.⁵³ provide a set of guidelines for effective visualization of fluorescence during surgery using surgical microscopes, which are applicable to open surgery as well.

4.2.

Fluorescence-Mode Operation with Ambient Room Lighting Present

For an imaging system to be easily translatable into a surgical suite or clinical environment, it is desirable that it operates under room lights and provides reasonable SBR. This issue is critical, especially when working with low fluorophore concentrations and when quantitation is necessary. Figure 3(e) shows a plot of the most common room light sources, such as tungsten bulbs, halogen bulbs, compact fluorescence lights, and the newer LED lights. It can be seen that tungsten and halogen lamps have significant output in the 600- to 850-nm range and may thus contribute a major portion of the detected signal during any red to NIR fluorescence imaging. Use of these lamps in rooms is seeing a declining trend, to the ultimate benefit of fluorescence imagers. As a general guideline for researchers and other users of such systems, use of tungsten and halogen lamps, or sunlight, should be avoided completely. Both LEDs and compact fluorescent lights (CFL) have minimal signal contribution over 780 nm and thus imaging of ICG and similar NIR fluorophores in rooms lit with these sources should be attainable with simple filtering techniques. However, it should be noted that CFL lights can often emit in the 700- to 800-nm wavelength range during the warm-up phase, which can last 5 to 10 min (data not shown), and thus contamination of detected fluorescence can occur at these times. For imaging fluorophores in the visible to far-red window, i.e., 500 to 750 nm, normal room lighting would contribute to the detected signal and sophisticated background removal methods are necessary. Pulsing the LED or laser diode excitation light sources in a manner that is synchronized to a gated- or shuttered-detector system, such as CMOS or ICCD camera,⁵⁴ is one technique that may be employed to address background contamination. Similar background mitigation can be achieved using frequency modulation and lock-in detection.⁵⁵ Finally, additional considerations, such as operating in a sunlit room, would require further mitigation to reduce background contamination of the signal for best performance.

Fig. 3

(a)–(d) The chemical formulas are shown along with absorption and emission spectra of the major FDA-approved fluorescent dyes such as fluorescein, PpIX, ICG, and methylene blue. (e) The normalized emission spectra are shown for common light sources used in surgery.

Given the above information about spectral contribution, the ideal surgical room lighting would be white-light LEDs. The PerkinElmer Solaris system performs background correction by pulsing the excitation sources to sequentially image fluorescence emission and background light leakage to make on-line corrections. It has been shown that systems that perform some kind of background correction tend to perform better than those without;⁵⁴ this improvement in performance can also be seen from the sensitivity and linearity tests data in Sec. 4.3.

4.3.

High Sensitivity to Fluorophore of Interest and Ability to Quantify In Situ

The concentrations of fluorescent agents in tissue vary by orders of magnitude depending on their distribution and targeting to specific disease biomarkers. Nonspecific agents, such as ICG, are usually administered intravenously and generally remain in the tissue at concentrations in the low micromolar range. Target-specific agents, on the other hand, are generally given time to clear normal tissue, thus will usually be present in midlow nanomolar concentrations in tissue. This poses a challenge when devices designed for ICG imaging are used to image targeted probes, as the sensitivity limits are not always optimized for low-concentration probes. All available systems were evaluated for their ability to detect and quantify signals from IRDye 800CW (LI-COR Biosciences, Lincoln, NE) in phosphate-buffered saline, through the logarithmic concentration range from 3 pM to $25\mu \mathrm{M}$. The samples were imaged individually to allow the user to modify any available gain and exposure settings and maximize the ability to detect fluorescence emission. We grouped the systems into FDA-approved imagers [Fig. 4(a)] and preclinical imagers [Fig. 4(b)]. Plots showing “${\log }_{10}(\text{fluorophore concentration})$” versus “${\log }_{10}(\text{normalized fluorescence signal})$” are shown in Fig. 4. A handful of imagers also performed imaging in the 700-nm channel, so IRDye 680RD samples were used to evaluate them. As a reference, the performance was compared to the LI-COR Pearl Impulse preclinical imager, which provides over 20-bits of dynamic range, and performs imaging in an ambient-light-free chamber. Slopes of linear fits to the log–log data ($\text{slope}=1$ for linearity on these log–log plots) and the lowest detectable concentrations are shown in the table of Fig. 4(d).

Fig. 4

Plots of ${\log }_{10}(\text{F}\text{luorophore Concentration})$ versus ${\log }_{10}(\text{N}\text{ormalized Fluorescence})$ are shown for measurements of IRDye 800CW using FDA-approved imagers (a). Panel (b) shows IRDye 800CW measurements on imagers that are not approved for clinical use. Measurements from the LI-COR pearl impulse preclinical imager are shown for comparison. Note that large variability exists in dynamic range and detection sensitivity among FDA-approved imagers. Panel (c) shows similar plots for all systems with far-red emission imaging capability when IRDye 680RD samples were tested. A handful of imagers also performed imaging in the 700-nm channel, so IRDye 680RD was tested on these. In (d), the fitted slopes and the lower limit of detection are shown. *Fluoptics has two distinct imagers, Fluobeam700 and Fluobeam800, for imaging in the 700-nm and 800-nm emission bands, respectively. **The Li-COR Pearl imager was included simply as a standard of linearity and sensitivity achievable using an enclosed light-tight imager.

All imagers, under ideal conditions and dimmed lighting were able to detect down to a surface concentration of $\sim 10\mathrm{nM}$ for both IRDye 800CW and IRDye 680RD (for dual-channel systems that have this channel). Thus, per this criterion, all systems seem suitable for imaging high concentrations of ICG, as intended. For imaging lower concentrations of fluorophores, it is observed that systems with higher bit depths, variable electronic gain settings, and/or background-light correction during acquisition had the best sensitivity. The VisionSense Iridium system outperforms all other instruments in terms of sensitivity to low concentrations owing to its high camera bit depth and gain adjustment (1 to $200\times$) capability. The Solaris was a close second in terms of sensitivity, again due to the high bit depth and background correction functionality based on pulsing excitation light; but the lack of gain adjustment likely limits its sensitivity below $\sim 1\mathrm{nM}$. The Fluobeam800 system with the ability to manually vary exposure time could achieve sensitivity down to $\sim 5\mathrm{nM}$, but this comes at the expense of long-exposure times on the order of seconds, which may not be feasible within a clinical setting. Similar sensitivity is obtained with the Novadaq SPY system in real-time video mode. The Quest Spectrum though equipped with a 14-bit camera compresses the image data to 8-bits at the camera output resulting in reduced sensitivity and dynamic range, which severely affected the overall system performance. In terms of quantitative ability, the closer a log–log fitted-slope (Fit equation: ${\log }_{10}y=m·{\log }_{10}x+C$ which is equivalent to $y={10}^C·x^m$, where $m$ is the slope) was to 1, the more reliable a system could be for linear reporting of concentration, which was seen well with the Solaris system, owing to background correction. The Quest spectrum and VisionSense Iridium devices utilize 14-bit and 12-bit cameras, respectively, but ultimately map their data to 8-bit thus resulting in nonlinear compression and an observed slope of $<1$. While VisionSense uses smart image processing algorithms to produce a wide dynamic range, the Quest Spectrum lacked such features, thus the range of detection suffers, and sensitivity is about $\sim 10\mathrm{nM}$. It should be noted that for data from VisionSense and Fluobeam fluorescence signal measurements were scaled by the gain settings and exposure times, respectively. Aside from fluorescence intensity, some systems, e.g., Novadaq SPY, may also provide perfusion and flow rate as quantitative endpoints; these measurements are usually made in software based on the fluorescence intensity data, and would hence be affected by the linearity and quantitative performance of the system as described herein. Curadel’s Lab-Flare R1 was excluded from Fig. 4 as a final commercial system was unavailable for testing at the time of publishing this paper, although based upon the design the sensitivity, it was expected to be comparable to the other advanced systems and reported to be in the single nM range or better. The SurgVision system was also excluded from Fig. 4 as the sensitivity tests on this system were not available in the same methodology as used for all the other systems. However, the sensitivity to IRDye 800CW reported by the company is $\sim 60\mathrm{pM}$, which would make it among the most sensitive of the systems evaluated. This is most likely attributable to their use of an EMCCD capable of single-photon detection.

4.4.

Ability to Image Multiple Fluorophores Simultaneously

Barring excitation sources and emission filters, a large part of the optics and instrumentation of an imager is more or less independent of the fluorophore being imaged within the tolerance limits on the optics. Some systems have been designed to house excitation sources for multiple excitation wavelength bands along with emission filter sets, to allow for multifluorophore imaging, either simultaneously or by switching between channels using a filter wheel. This multichannel functionality certainly adds to the cost of the device, but such a system can be a worthy investment for a research group working with multiple imaging agents, and combinations of targeted and untargeted tracers for quantification of disease biomarkers.⁴⁸ A total of three out of the eight imagers we compared are capable of multiple fluorophore imaging; the Curadel Lab-Flare R1 and the Quest Spectrum can image in the $\sim 700$- and $\sim 800\text{-}\mathrm{nm}$ channels simultaneously, while the Solaris is capable of imaging $\sim 470$-, $\sim 660$-, $\sim 750$-, and $\sim 800\text{-}\mathrm{nm}$ channels independently (nonsimultaneous). Figure 2(d) shows an example of simultaneous white-light, 700-, and 800-nm fluorescence imaging with the merged display available on the Flare imagers.

4.5.

Maximized Ergonomic Use

As a general principle, compact, portable units are easier to deploy in a surgical suite than large roll-in systems. However, the computer, display monitors, and illumination unit contribute significantly to the size and usability of a system. Studies have shown that choice of display, their location, and setup can significantly affect surgical tasks and their outcomes.⁵⁶ Currently, fluorescence-imaging systems are either compact, handheld systems such as VisionSense Iridium, Fluobeam, and PDE Neo, or larger, overhead, wheel-based systems with significant footprints such as the Solaris and Curadel systems. The former do provide mounting arms and carts to users who need them while the latter can be large enough to need a 36- to $100\text{-}{\mathrm{ft}}^2$ room for storage. Stability of images and impact of vibrations vary among devices, with larger heavier devices built for greater stability as compared to handheld systems with optional mounting hardware. Figure 5 shows photographs of the various commercial systems. Though not to exact scale, the systems can be compared for size and physical footprint.

Fig. 5

Images of the leading fluorescence guidance systems evaluated here, targeted for open surgery use, shown with relative approximate size comparison. The PerkinElmer Solaris, Curadel ResVet Lab-Flare, and SurgVision Explorer Air are not 510(k) cleared for human use, while the others are for ICG procedures. All have capability to image ICG in surgical trials, with differing levels of sensitivity and features. Images from left to right are from Solaris™ Open-Air Fluorescence Imaging System, Printed with permission, (c)2015-2016 PerkinElmer, Inc., all rights reserved; NOVADAQ Spy-Elite™, copyright 2016 Novadaq Technologies Inc.; Quest Spectrum™, copyright Quest Medical Imaging; Fluobeam(R), copyright 2016 Fluoptics; Hamamatsu PDE-Neo™; Lab-FLARE(R) Model R1 copyright CURADEL; Visionsense Iridium™, copyright Visionsense; SurgVision Explorer Air prototype, image courtesy of SurgVision. All images have been printed with permission from copyright holders.

Handheld systems can provide better access to complex tissue geometries, such as around the head and neck or inside limbs, and are also highly mobile, whereas the larger systems generally have a wide range of functionalities such as multiple fluorophore capability, large FOV size, and large working distance. Both the PerkinElmer Solaris and Curadel Lab-Flare system have a large range of FOV sizes over which focus errors are minimal; the Lab-Flare in particular has been optimized to maintain parfocality from $0.9\times 0.9{\mathrm{cm}}^2$ to $25\times 25{\mathrm{cm}}^2$ over working distances from 12 to 18 in., enabling its usage in a wide range of surgical applications. The appropriate working distance depends directly on the intended usage, but it is interesting to note that the Solaris is the only system with a fixed working distance of 75 cm, which keeps the imaging head well out of the way of a surgeon’s workspace. Along with the use of multiple excitation angles, this system attempts to provide a highly ergonomic solution to imaging fluorophores during surgery, replacing the surgical light as well. Nevertheless, as surgical applications are highly varied, ranging from inside the abdominal cavity to under the armpit, there is no single optimal design, and selecting a system will require careful consideration of its intended use. With large FOV options available on some systems, the image uniformity and quality across the field can be an important factor to consider. Qualitatively, these aspects were not significantly different across the large roll-in systems, but the handheld devices designed with lightweight optics were of acceptable but slightly poorer performance. This may or may not be significant depending on the intended clinical application, but could be important especially when fluorescence guidance is being used to save or resect fine structures, such as nerves, demanding high resolution and quality across the entire FOV.