5.1.

Alternative Pathways to Translation

There are two distinct pathways that can be followed in developing clinical technologies: technology push or clinical pull. The first often includes technologies that were developed initially for nonclinical purposes and then possible clinical applications are sought to expand the market. In the second approach, the technology is conceived and developed in response to a clearly-defined existing unmet clinical need. Table 3 highlights a few of the advantages and risks/challenges of each approach.

Table 3

Pros and cons of alternative developmental pathways to clinical translation.

In a recent paper by Beswick et al.,² it is stated that “…only once the clinical need be understood in detail can the invention process begin,” but this ignores the fact that OSI in particular has often been driven by new and disruptive technologies from the broad photonics sector as well as by advances in molecular sciences/chemistry (e.g., biomarker-targeted, optically-active “reporters”) and materials science (e.g., photonic nanotechnologies for use in imaging or treatment): i.e., the technology came first and opened up possibilities for addressing the clinical needs. In any case, the unmet need is most likely identified first by a physician who then seeks out research scientists and technologists and/or companies to develop OSI technologies to address the need. Knowing exactly (or even roughly) the intended final clinical use and having strong guidance from the “end-users” are important, especially since “inventors” are not necessarily skilled and experienced in clinical-needs assessment. Productive interactions between scientists/engineers and clinician require significant effort to break down the interdisciplinary knowledge and language barriers, especially if the former are based in a purely academic environment rather than, say in a medical research institute that is part of an academic hospital where there are many more opportunities for ongoing interactions: Popp et al.³ have also discussed the benefits of multidisciplinary research centers established specifically for developing and translating (photonics-based) technologies and give examples of such initiatives and how they might be funded. Regardless of the translational model, in most cases it is the physicians rather than the patients who determine whether a technology will be adopted, but considering the patient perspective is also valuable, preferably at an early point in the translation process.

A good example of clinical pull for OSI technologies is the use of in vivo Raman spectroscopy for surgical guidance of (brain) cancers.^6,40 The unmet clinical need for rapid and accurate tumor tissue detection and localization during surgery was identified by clinical users and the OSI technology, based on point Raman spectroscopy, was then developed and tailored to address this specific need [Fig. 1(a)] and a company was formed to commercialize the technology (ODS Medical Inc., Canada) [Fig. 1(a)]. The use of clinical pull can also facilitate the running of clinical trials through having preexisting collaborative effort and motivation on the clinical side. Examples of successful technology pushes also exist and include technology, which was initially developed without a specific clinical objective in mind but was nevertheless able—through rapid technological advances and productive relationships developed between the technical and medical sectors—to find specific clinical application. OCT is a good example of this, where important clinical applications were identified such as in ophthalmology and gastroenterology. Several companies were created that are now manufacturing OCT clinical systems.

Fig. 1

(a) Raman spectroscopy system for surgical guidance of brain cancers, illustrating the use of a Raman spectroscopy probe in situ to distinguish normal brain from cancer (courtesy of ODS Medical Inc.). The bottom right shows a volumetric rendering of the MRI-visible tumor, relative to invasive cancer that extends into the normal brain. (b) Diagnosis of cervical cancer using optical imaging with conventional colposcope (left) and tampon-like device (right). (c) Comparison of image quality between the two technologies in B (B and C courtesy of Dr. Nirmala Ramanujam, Duke University). (d) Device utilizing fluorescence imaging and PDT for skin cancer (D courtesy of MM Optics, Brazil).

5.2.

Barriers to Clinical Adoption

With consumer technologies, whether they be cell phones or automobiles, the primary determinants of adoption (apart from luck and marketing) are “Do I need or want this?,” “Can I afford it?,” and “Is this the best model?” In considering the barriers to a new clinical technology being widely adopted, it is worth listing the multiple factors involved, some of which are common to consumer products while others are distinctly different and involve a much more diverse set of stakeholders, namely inventors, academic institutions and medical centers, investors, physicians, patients, health-care providers, and governments that need to be engaged at different stages.² The barriers include: scientific and technical, clinical, socio-economic, geographic, cultural and educational, as can be illustrated by a recent and striking example. Cervical cancer is a significant health-care problem worldwide but in particular, in the words of the World Health Organization, it is “an avoidable cause of death in sub-Saharan Africa,” where the incidence is five times higher than in the USA and the death rate is nine times higher. An existing optical technology, colposcopy [Figs. 1(b) and 1(c)], is used in developed countries to diagnose cervical cancer at a treatable stage, in which the surface of the cervix is viewed through an optical probe placed close to the cervix to identify areas of abnormality. Hence, the scientific and technical barriers have already been overcome. What then are the barriers to adoption to address the clear unmet clinical need in sub-Saharan Africa and other parts of the developing world? The first is a combination of clinical and educational factors; there are not enough trained gynecological oncologists to meet the workload. The second is socioeconomic, since a colposcope costs around USD$20,000, which is beyond the means of many local health-care systems. Geographic barriers then come into play because of the difficulty in providing high-tech medical services at the “village”-level across large areas. Another major barrier is cultural, in that there is fear and reluctance on the part of patients and their families to have this examination, especially if performed by a male doctor. Recently, a radical alternative has been developed [Figs. 1(b) and 1(c)], in which the optical system has been markedly reduced in size and cost and is in the form of a tampon that it inserted by the woman herself.⁴¹ The digital images may then be sent via cell phone to a central expert reader for diagnosis. Histological images of tissue samples taken under this image guidance could also then be generated locally using a cell phone-based microscope and transmitted to a central expert pathologist. This is a compelling example of an optical technology being developed/adapted for a defined unmet clinical need rather than vice versa. Clearly, the approach is relevant to a wide range of possible clinical needs and OSI solutions.

Another example of designing optical technologies to meet clinical needs at locally affordable cost, and that are practical for dissemination across large countries with geographically variable medical infrastructure, is in PDT, i.e., the use of compounds that are activated by light to kill abnormal cells or organisms.^16,17 A good example is a group at the University of Saõ Paulo in Brazil that has re-engineered a LED-based light source for PDT treatment,⁴² used in combination with a simple fluorescence imager to delineate the tumor and guide treatment, into a low-cost, reliable, user-friendly package about the size and format of a desk telephone [Fig. 1(d)]. This has been disseminated, with appropriate training, to about 80 local clinics in Brazil and the program has spread to other countries in South and Central America. Both examples demonstrate that overcoming these barriers also provide new opportunities for research, development and commercialization.

6.1.

Preclinical Phases

The technology development and intellectual property (IP) protection leading up to the demonstration of proof-of-principle is usually carried out in preclinical models with relevant biological specimens, and these data are often used to support patenting. Issues relating to preclinical animal models are discussed in more detail in Sec. 9.2 as follows. Although the preclinical steps in translation can be completed relatively quickly and without large capital expenditure for some simpler OSI technologies,² this is not usually the case for “high-end” technologies that require building complex prototypes in the case of medical devices (e.g., surgical robots or multimodal imagers) and/or optically active imaging of therapeutic agents, where the synthesis and testing costs may be very high.

6.2.

Clinical Trials

First clinical trials may include demonstrating that performing the OSI procedure on ex vivo human tissue samples (e.g., biopsies) does not compromise subsequent clinical histopathological diagnosis. Clinical trials require explicit institutional approvals with informed patient/subject consent and, depending on the level of potential risk (and, for optical systems, the laser classification), may also require government regulatory approval. Prototypes will likely be redesigned based on the results of the clinical trials. Optimal synchronization of the clinical trials with the transition from prototype to product is a critical challenge that has significant cost implications. Often, multicenter trials are required for final approval, both to recruit enough patients in a reasonable time and to mitigate against user bias. As well as demonstrating safety, the efficacy of the technology compared with other “predicate” technologies, if they exist, must be demonstrated with statistical validity. The number of patients and/or samples required in these trials can be large: for OSI, late-phase trials typically involve many tens or even hundreds of patients, and there is increasing recognition that this number is often seriously underestimated.^43,44 The number of patients or samples/measurements required depends on the particular technology, what biological characteristics are being measured, the magnitude of expected effects, and the clinical endpoints of the trial. A major factor that drives up the patient numbers is the inter- and intrapatient variability. Given that one is usually investigating a disease, it might be expected that the biochemical or structural changes could vary markedly within the clinical trial population and depend also on the stage of disease. The nondisease background signals, both between patients and within an individual, can also be very high and variable and the impact of this must be minimized within the clinical trial design. Additionally, data processing (calibration, normalization, choice of classification model) can help correct for these variations. For example, the variability between spectra of the “normal” host tissue taken in different patients or even at different locations in the same patient can mask the differences due to disease, so an effective strategy is to normalize the spectra from suspected tumor to that taken in the adjacent normal tissue in the same patient. The requirement to have large numbers of samples is also generally more severe in “label-free” OSI techniques than when using an exogenous contrast agent. This is especially the case when the biological information is distributed across the optical spectrum, as for example, in Raman spectroscopy. It is less of a problem if there is a distinct spectral feature to differentiate disease from normal, as for example, the red-to-green ratio that has been used in some autofluorescence systems.^45–47 For regulatory approval of a commercial product, the so-called registration trials upon which approval is granted must use a device that is substantially equivalent to that commercial product. Note also that the requirements for approval of “devices” are different from those for “drugs.” Many OSI technologies are device-only based but if, for example, an exogenous fluorescence contrast agent or photoactive agent is also used, then drug-device combination will require approval, even if the compound is already approved for another purpose.

6.3.

Failure in Clinical Adoption

Unfortunately, there are examples of OSI technologies that have been taken all the way through approvals only to find that there is no market. The possible reasons for this failure of adoption are several:

8.1.

Technical Factors

With regard to the scientific and technical trade-offs, the first concerns the quality of the information obtained in the clinical procedure. In the case of optical spectroscopy, this includes the signal-to-noise and signal-to-background ratios and the spectral and spatial resolutions. These need to be balanced against the time required to collect spectra or spectral images. In turn, these technical performance factors are likely to be reflected in the cost of the instrument, for example, the quality (sensitivity, linearity, dynamical range of detection, speed, etc.) of the spectrometers and photodetectors used, as well as their size and complexity. In the second trade-off, the main issue around quantity of information is the sampling density of the measurements, i.e., the number of discrete points on the sample/tissue at which spectra are collected, or the field-of-view and pixel density in the case of spectral images. This often has critical clinical implications, for example, to minimize the risk of missing diseased regions due to under-sampling. If the spatial resolution and/or imaging field of view for sampling are incompatible with the typical feature size of the disease, there is increased risk of missing disease due to partial volume sampling. The aforementioned performance factors are primarily related to the hardware in OSI systems.

8.2.

Analysis Factors

The third factor is the choice of algorithms used in the spectral analysis software. Commonly, these are applied to the spectra or images acquired from an individual patient or biospecimen to classify cells or tissues as normal or abnormal (e.g., tumor). The algorithms are often chemometric in nature, i.e., the measured spectrum from the unknown specimen is used as input to a “black box” algorithm that has been trained on spectra taken from a range of prior samples in which the disease classification is known by an independent “gold standard” technique, such as histopathology or biochemistry. The optimum choice of algorithm (e.g., possibly based on principal components analysis, neural network, support vector machines, etc.) is far from trivial and can depend strongly on the nature of the data. There are also many rules and best practices in the training, validation, and implementation of the algorithms that must be followed to ensure that the resulting classification is accurate and robust.^55,56 Common pitfalls include: inadequate sample numbers for training, use of the same data for training and validation, bias in the training and/or validation datasets, inadequacy of the gold standard, and using the algorithm on samples whose spectra or classification lie outside the range used in training the algorithm.³⁰ A common trade-off is the accuracy of the algorithm versus its robustness and speed, the latter being of most concern in the case of spectral imaging where the datasets may be very large: for example, in the case of intravascular OCT imaging, the initial implementation of algorithms used to classify arterial tissues based on the textural features initially took several hours compared with a few seconds for the actual data acquisition in the patient.⁵⁷ Specialized software and hardware (e.g., graphic processing units) may be required to achieve clinically useful data analysis speed, with resulting impact on the system cost. Many other areas of modern medicine involve the acquisition, analysis, management, and archiving of increasing volumes of digital data (“big data”). This is becoming an expensive challenge and the same is increasingly true for some OSI techniques, especially spectral imaging. However, the advent of more advanced machine-learning/artificial intelligence, computer-vision algorithms, and computational capability presents a significant advantage to OSI. Single wavelengths or metrics often do not capture all the clinically relevant information from multiple biomolecular or structural signatures. Faster and better machine-learning algorithms are able to rapidly identify patterns in spectral data and, thereby, provide improved spectral analysis techniques to optimize diagnostic performance.^57–59 Computational power has reached the point where powerful feature extraction and classification techniques can be used, particularly for imaging data where important features in the images can be automatically identified using computer vision techniques.^57,60

8.3.

Workflow and Clinical Factors

The scientific trade-off of information content versus cost/complexity speaks to the need for making the OSI measurements fast enough to not unduly disrupt the clinical workflow. This can be particularly challenging in endoscopic and surgical procedures and it is worth determining early in the development phase what is the maximum acceptable measurement and analysis time once the technology is adopted into routine clinical use: these are usually much shorter than the time that can be tolerated by the clinical investigative team during the clinical trial phase. The target population for optimizing diagnostic performance should also be determined early, since it can significantly affect the clinical trial design and execution. For example, fluorescence-guided resection of brain tumors using the fluorophore protoporphyrin IX induced by administration of the prodrug aminolevulinic acid (ALA-PpIX) is substantially less accurate for low grade than for high-grade gliomas, because of the much lower PpIX concentration in the former, so that the fluorescence signal falls below the limit of detection of current commercial OSI clinical instruments. On the other hand, some low-grade glioma may be curable if all tumor tissue is resected, so this represents a high impact subset of patients. One solution to address this trade-off is to combine fluorescence and diffuse reflectance spectroscopies: the latter is used to determine the tissue absorption and scattering spectra, which are then applied to the measured fluorescence spectrum to correct for the variable attenuation of the excitation and emission light, thereby lowering the detection threshold of the fluorophore concentration in the tumor.⁶¹ This is an example where improved OSI performance extends the clinical utility and expands the potential market. It illustrates that extending the design increases the impact. This is not necessarily true in all cases and is worth determining before getting too far into clinical trials.

Balancing the size of trials against their time and cost is very critical and is usually impossible to determine a priori, since the statistical accuracy (e.g., sensitivity and specificity for disease detection) is only known once the trials are complete. Some initial guidance may be obtained by preclinical studies, especially in animal models that are close to the clinical situation either in the disease and/or in the organ size and geometry. Usually, data from early-phase trials are used to perform a statistical power calculation: that is, data from phase I are used to inform the likely number of patients needed in phase II, from which the results are then used to perform an evidence-based estimate of the numbers needed in phase III. There are established statistical methods to do these calculations, but guidance can also be obtained from published clinical trials of competing technologies for similar clinical applications. Significant effort can be saved by consulting with professional biostatisticians experienced in clinical trial design and analysis, especially if they are also experienced in medical device trials. It is also advisable to consult with more than one clinical end-user, to understand and address potential biases: this input should include consultation with clinicians (physicians, nurses, and other clinical support staff) across the full spectrum of intended clinical environments for the product and spanning the relevant range of expertise. One caution is that community hospitals, local clinics, and individual practitioners who may determine the ultimate market for the technology can have very different perspectives from the “thought leaders” at large academic centers who are often involved in the R&D phases.

8.4.

Commercialization and Technology

There are, of course, many obstacles that comprise the so-called valley(s) of death³ that must be crossed to get new technologies into the hands of clinicians. Here, we will only give a few examples where the commercial and R&D/clinical domains intersect directly. The first is what the technology claims to do, from a regulatory perspective. This could be as simple as “…this optical imaging device characterizes tissue…” without claiming any specific application, disease, or clinical impact. At the other extreme would be, e.g., “…this near-infrared fiberoptic fluorescence spectroscopy device and algorithm reduces the rate of second surgeries in breast cancer lumpectomy procedures by at least 15%….” Clearly, in the first case, the barriers to obtaining regulatory approval, assuming that all safety requirements are met, are much lower than in the second case, which would likely require a multicenter trial with a substantial number of patients. On the other hand, how large is the market for the first technology and who would reimburse its use in the clinic? Nevertheless, it may be advantageous to use the first approach as a stepping stone to the second. Moreover, the choice of an appropriate predicate can help facilitate initial regulatory approval.

In terms of consumables, many OSI technologies are comprised of fixed hardware (e.g., a spectroscopy system or multiwavelength imager) and operate in “label-free” mode, i.e., are based on the intrinsic optical properties of the biosample.⁶² This includes both linear optical techniques and emerging nonlinear-based OSI approaches. However, potential consumables, i.e., components that are used only once or a small number of times, may be required to achieve adequate sensitivity and/or specificity. These include fiberoptic probes/catheters for in vivo spectroscopy, cuvettes for lab-on-a-chip devices used in a fixed read-out instrument, as well as exogenous “reporters,” such as fluorophores, nanoparticles, and photosensitizers. The relative advantages and limitations of label-free versus exogenous agents are many and complex and have been the subject of ongoing debates in the biomedical optics community.^62–64 From a commercial perspective, consumables are usually highly desirable since they represent a continuing revenue stream, but their use often makes the regulatory process much more challenging. This is especially the case if the exogenous agent is actually administered to the patient rather than being used as an ex vivo “stain” as for histopathology, since safety (usually requiring arms-length toxicology studies in multiple animal species including large animals such as dogs, and performed by certified companies in multiple animal species), biodistribution, pharmacokinetics, etc. must all be determined and the materials must be produced under good-manufacturing-practice conditions.

9.1.

Safety

The most important aspect of any new technology intended for clinical use is that it should do no harm, either to the patient or personnel using it. This applies also in the clinical trials. The potential harm may be direct, as in damage caused by a laser beam in the OSI system or toxic effects of optically active exogenous compounds, or indirect as in causing changes to tissue samples such that histopathology readout and hence clinical diagnosis is compromised. Typically, in clinical trials the OSI information should not be used to alter patient care until after it has been demonstrated that the information is reliable: e.g., a new optical biopsy device should not be used to guide the surgeon to remove more tumor until there is an acceptable level of confidence that the technique does indeed reliably detect malignant tissue, based on prior “passive” (i.e., noninterventional) trials. It is particularly important to confirm which of false-positive or false-negative findings in individual patients, in clinical trials and beyond, could place the patient at higher risk: leading to over treatment or under treatment, respectively.

Although most diagnostic OSI applications use low optical power and energy (densities) that are within the usual exposure standards, these standards are based primarily on data from skin and eye exposure. The applicability of these to exposure of internal organs is not well determined, and UV or multiphoton exposure may be of particular concern because of potential mutagenesis.⁶⁴ Published clinical studies of endoscopic techniques can provide useful insight and baseline information on this issue, at least for some wavelength ranges and optical pulse widths.

9.2.

Preclinical Models

Preclinical R&D to optimize and test OSI systems for ultimate use in patients often involves the use of cellular and/or animal models of disease. Especially for in vivo clinical applications, there may be a requirement for safety, biodistribution, and toxicity and effectiveness testing in animals. The basic requirement is to replicate as closely as possible the intended human use with respect to the specific OSI technology and the intended clinical utility claims. This last point is often overlooked. For example, there are many publications on OSI for early cancer detection using transplanted tumors in rodents but there are several typical shortcomings of these models: unlike human tumors, they are not spontaneous; the tumors are often placed accessibly under the skin or in muscle, whereas the detection in patients is on a background of the organ of origin; the tumors are often not of human origin and so may not have the true biochemical or structural characteristics; and the models do not represent early or precancerous lesions that are often the clinical target, at least for diagnostics. Thus, for example, a sensitivity and specificity above 90% for detecting transplanted mouse tumors relative to mouse muscle may have little relevance to detecting premalignant areas in the esophagus of patients. However, such experiments may be informative with respect to the technical performance of the OSI system, and poor performance in these less challenging situations can provide an early indicator of whether a technology is worth pursuing. If for no other reasons than expense and time, it is always worth asking whether it is better to use an animal model in vivo or to utilize ex vivo human tissues (e.g., fresh biopsy or from a tissue bank or archive) or to begin directly with patients, assuming that safety has been demonstrated and there is reason to believe that the technique may be effective.

9.3.

Gold Standards

In validating in vivo clinical OSI techniques the standard of comparison is often the current clinical technique, either biochemical assay of a body fluid, cytology of cell samples, or histopathology of biopsied or resected tissue taken concurrently with the optical spectra or images. Indeed, this may be an absolute requirement for institutional ethical approval of clinical trials and ultimate regulatory approval. Such established methods are considered the gold standard. However, it should be realized that these may not be as reliable as the name implies, especially for conditions, such as early cancer detection, for which the sensitivity and specificity may be rather poor and the agreement between pathologists can be low. A further problem is that, for example, in “optical biopsy” the measurement may not be made in exactly the same location as the biopsy is taken due to unavoidable tissue motion or operator error. Hence, the optical measurement may be misinterpreted as false-positive or false-negative relative to the biopsy. Further, even if the tissue location is identical, the highest grade of tumor (which is what the pathologist usually reports) may represent only a small fraction of the sampling volume of the optical measurement, thereby underestimating the sensitivity of the optical method. This may be partially circumvented by requiring a more detailed histopathological analysis and report (e.g., estimating the fraction of the sample that shows disease) but at added cost and time. Accounting for tissue heterogeneity relative to biopsy sampling volume is a challenge, due to the effect of tissue optical properties on sampling volume and the difficulty in providing weighted volumetric histopathology results.

A further issue in using the current gold standards as the reference is that they are intrinsically limited in their ability to stratify patients, i.e., they can provide a diagnostic of disease $X$ with grade $Y$, but there are subsets of those patients that do not, for example, respond equally well to the therapy. Developing OSI to its full potential will then rely on validation against new approaches, either in the form of genetic testing or more specific molecular-pathology techniques and/or new bio data processing approaches, such as the application of artificial intelligence (machine learning and deep learning). The overall objective is to establish statistically rigorous models to improve patient stratification and predict outcome.

9.4.

Technology Standardization

Unlike more established clinical radiological imaging technologies (e.g., MRI, CT, and nuclear imaging), there is currently a lack of well-established and widely accepted methods of standardizing and calibrating OSI techniques, i.e., there are few if any standard operating practices (SOP). For example, Raman spectroscopy for in vivo detection of cervical cancers has been carried out by various groups using different instruments and algorithms in diverse populations to show improved classification efficiency, but it would be more fruitful in further studies to establish SOPs for clinical implementation.⁶⁵ In part, this is because the techniques and technologies (devices, software, and photoactive compounds) are very diverse, as is the industry base, which has many early-stage companies and only a few large multinational companies and institutions. The professional societies for institution- and industry-based researchers have not typically had strong participation in clinical translation, while conversely those of the clinical users have had little involvement in OSI technology development and promotion. This is beginning to change, as evidenced by recent initiatives by government agencies and professional bodies, both in Europe and the USA, to develop guidelines for clinical biomedical optics: for example, in the clinical uses of fluorescence-guided surgery and the minimum performance levels of the corresponding technologies.⁸

9.5.

Multimodality Optical Spectroscopy and Imaging

Finally, it is most often the case that a single OSI technology does not adequately address the unmet clinical need, frequently because its sensitivity and/or specificity are too low for clinical adoption. There is then the option to combine the technique with one or more other techniques, which may be optical or nonoptical and either established or also under development, to provide the needed complementary information or therapeutic capability. For example, for the unmet clinical need of identifying dysplastic (premalignant) tissues in the esophagus of patients with chronic inflammatory disease (Barrett’s esophagus), many OSI methods have been investigated:^66–69 some have good sensitivity but poor specificity (e.g., tissue autofluorescence), while others have both high sensitivity and specificity but are slower and provide only point sampling (e.g., Raman spectroscopy). Subsurface imaging to stage the disease is also of high added value, since it determines the treatment options in cases where dysplasia is found: potentially, OCT or photoacoustic imaging may have utility for this purpose. Combining some of these techniques has been shown to be more effective than any one technique alone, as has been the case for various OSI tools and applications.^70,71 Of course, a major challenge then is how to combine the techniques into a single device. “Bottom-up” integration is preferred over simply running the techniques in parallel, but can pose considerable technical difficulties, for example, if different optical fiber characteristics are needed for the different methods. A second challenge in multimodality OSI, from a commercialization perspective, is that the IP ownership is likely to be distributed between multiple individuals, institutions, or companies, so that combining these into a common enterprise may not be possible. Moreover, clinical adoption of technologies is often incremental and most readily achieved by creating techniques that complement the current standard of care. Therefore, any opportunity to create a multimodal system combining an optical technique with an already-adopted modality (e.g., MRI, CT, etc.) may be easier to translate to the clinic than combining two unproven technologies. Ideally, new technologies should provide valuable information to clinical users while conforming to standard practice with minimal disruption to the clinical workflow.