Optical imaging technologies for in vivo cancer detection in low-resource settings

Abstract

Cancer continues to affect underserved populations disproportionately. Novel optical imaging technologies, which can provide rapid, non-invasive, and accurate cancer detection at the point of care, have great potential to improve global cancer care. This article reviews the recent technical innovations and clinical translation of low-cost optical imaging technologies, highlighting the advances in both hardware and software, especially the integration of artificial intelligence, to improve in vivo cancer detection in low-resource settings. Additionally, this article provides an overview of existing challenges and future perspectives of adapting optical imaging technologies into clinical practice, which can potentially contribute to novel insights and programs that effectively improve cancer detection in low-resource settings.

Introduction

Cancer presents a high global burden, disparately impacting those living in low- and middle-income countries (LMICs). The incidence rate of cancers in LMICs is anticipated to grow, increasing the need to expand cancer screening and early detection programs that enable early treatment with good patient prognoses [1,2]. Yet, inequities exist in access to cancer screening and early detection programs due to limited infrastructure and shortages of trained personnel, particularly in low-resource settings (LRS). As a result, patient care is compromised, leading to several challenges, including late-stage diagnosis, lack of targeted interventions, and lack of continuity of care (Figure 1). Optical imaging has emerged as a promising technology for in vivo cancer detection and diagnosis, which offers real-time imaging capabilities, noninvasiveness, automated diagnostic performance, and potential for cost-effectiveness [3]. Implementing optical imaging in LRS, however, usually presents several challenges, including high upfront cost, lack of infrastructure, trained personnel, maintenance and technical support, connectivity and data management, and difficulties in integration within existing healthcare systems. Addressing these challenges in LRS requires a multi-faceted approach, including financial investments, capacity building, infrastructure development, and collaboration with relevant stakeholders to ensure successful implementation and sustainable use of optical imaging technologies. In recent years, low-cost, high-quality, and easy-to-use optical imaging systems tailored for LRS have been developed. These tools aim to provide reliable and accurate cancer diagnosis in a simple format, while overcoming financial, infrastructure, and resource limitations. Furthermore, integrating artificial intelligence (AI) with optical imaging modalities opens new opportunities to improve global cancer detection by providing accurate computer-aided diagnosis in real time and reducing reliance on clinical expertise. For example, a colposcope is an expensive piece of equipment used for cervical examination that requires a trained colposcopist to interpret clinical images. This has partly been addressed in low-resource settings with the development of low-cost, battery powered alternatives that reduce cost and the need for infrastructure. In the future, the use of automated algorithms to interpret colposcopic images could reduce the need for expertise. This review explores recent advancements in low-cost optical imaging technologies for in vivo cancer detection in LRS, highlighting benefits, prospects, and remaining challenges to improve cancer care in underserved populations.

Figure 1.

Enabling early cancer detection in low-resource settings (LRS) through in vivo optical imaging. The top panel illustrates the global burden of selected cancers that can potentially be addressed by optical imaging, and current challenges in early detection and its impact on patient care. The bottom panel shows important considerations when designing optical imaging technologies and the potential of these imaging systems, when combined with computer-aided diagnosis, for early cancer detection in LRS. Human anatomy figure created with BioRender.com.

Low-cost optical imaging technologies for low-resource settings

Macroscopic imaging

Macroscopic optical imaging utilizes light at various wavelengths to illuminate a target area that usually spans more than several centimeters. One advantage of macroscopic optical imaging is its ability to provide real-time, non-invasive visualization of tissue structures over a wide area and therefore provide a comprehensive view of the suspicious area [4]. Key innovations in macroscopic imaging have significantly advanced the field. Two notable areas of recent progress are 1) affordable and accessible imaging instrumentation coupled with deep learning (DL) methods and 2) integration of multispectral and multimodal widefield imaging techniques. Macroscopic imaging modalities have become more accessible with enhanced imaging capabilities, potentially enabling more cost-effective imaging strategies. Additionally, DL methods have been applied to interpret macroscopic images, which holds promise to improve diagnostic performance and reduce the dependence on expert interpretation [5]. The integration of multimodal macroscopic imaging has expanded imaging capabilities and enhanced diagnostic accuracy. By combining different imaging modalities, such as autofluorescence imaging (AFI), multispectral imaging (MSI), or narrow-band imaging (NBI), more comprehensive assessments of tissue can be achieved [4]. Furthermore, macroscopic imaging clinical validation studies have demonstrated its effectiveness in various resource-limited settings. The following section focuses on various macroscopic imaging modalities including white light and multispectral imaging, and highlights key technological advancements, real-time image analysis, AI integration, and multimodal imaging approaches.

White light imaging: integrating low-cost instrumentation with deep learning

White light imaging (WLI) is increasingly affordable and accessible with advances in optics, sensors, and smartphone cameras that provide improved image quality from cost-effective imaging equipment. This is exemplified by digital colposcopy. Historically, standard-of-care examinations were conducted with visual inspection by trained professionals using colposcopes and, at most, captured film cervicograms. More recently, advances in cost-effective and miniaturized optics and camera sensors gave rise to the development of various low-cost, digital, portable colposcopes. For example, in a study of 129 patients in Peru, Mueller et al. showed that the Pocket Colposcope achieves comparable performance for identifying precancerous and cancerous cervical lesions when compared to the standard of care [6].

The capability to acquire high-quality digital images opens new opportunities for image interpretation via telemedicine or machine learning, a critical feature for technology dissemination in LRS. For example, the validation of a portable colposcope, Gynocular, with scoring in a clinical study in India demonstrated reliable detection of precancerous lesions. A total of 495 images from 94 patients were examined and assessed by clinicians in real-time and also separately by remote specialists/clinicians [7]. Notably, this still relies on time and availability of expert clinicians that might not always be feasible in LRS.

Digital image acquisition has enabled image capture for real-time analysis, specifically using point-of-care-friendly platforms that also utilize DL. Although macroscopic WLI has proven to be useful for cancer detection, these images still require trained experts for accurate interpretation. DL can play a role to reduce the need for clinical expertise. For example, an automated algorithm developed on digitized cervicograms and digital colposcopy images collected with a smartphone colposcope, MobileODT, was trained, validated, and tested on a combined dataset of 17,013 images from 9462 women across five different studies [8]. The best-performing algorithm showed promise in classifying cervical images in a repeatable and reliable manner for the intended test population. Similarly, a smartphone endoscope and a DL algorithm were developed to distinguish abnormal cervical lesions from normal tissue [9]; results from 20 patients showed good agreement with histology, suggesting the potential to use a simple smartphone endoscope for cervical cancer screening in low-resource settings.

Also of key importance is the integration of various forms of clinical data, such as multimodal and temporal data to develop DL algorithms that approximate clinical decision-making and provide a more comprehensive approach to diagnose diseases with available data. This is illustrated by development of diagnostic algorithms using relevant datasets collected with digital colposcopes that combine multimodal and multi-contrast images [10], utilize time series acetic acid images, green filter images, and Lugol’s iodine images [11], and combine cross-modal information [12].

With increased use of DL, there need to be effective methods to evaluate the accuracy of these diagnostic algorithms. DL is a rapid driver of innovation in macroscopic imaging that lacks substantial oversight and requires comprehensive methods of evaluation at all points in data collection and algorithm development. Development oversight becomes increasingly important in the context of deployment of algorithms to LRS.

There is an important need to develop standardized guidelines for evaluating datasets and models and to standardize the metadata collected for datasets, especially publicly available datasets. Datasets should be collected with validated endpoints and include the methods used to collect them - such as instrumentation used. They should reflect the diversity of the population for its intended use. Although transformative progress has been made in DL for cancer detection, this progress often comes at the cost of biased and less accurate results for underrepresented populations. Dataset and algorithm bias can be present throughout the development process and should be evaluated and identified at each stage of development [13]. Researchers must examine the effort devoted to the evaluation of algorithms and standardization of datasets to ensure development of reliable algorithms that are generalizable in their intended use-case setting. For example, overfitting on training data is a significant problem that can lead to poor performance in test settings and, as such, should be identified before deployment using evaluation metrics such as repeatability [8]. Several tools and models currently exist in the Medical Imaging and Data Resource Center, Microsoft’s Datasheets for Datasets, and a skin cancer specific evaluation checklist [14,15], but field-specific evaluation metrics for other applications should also be established. The World Health Organization (WHO) guidelines for AI also highlight the need for a systematic and standardized approach to evidence generation for AI-based medical devices. It provides a framework outlining key considerations and steps to ensure development, validation, and evaluation of safe and effective AI technologies in healthcare [16].

Multispectral and multimodal macroscopic imaging

MSI, AFI, and NBI have shown significantly enhanced imaging capabilities compared to WLI by providing additional information about tissue characteristics and highlighting specific features of interest, such as increased optical scattering, loss of autofluorescence, and increased vascularization. By combining data from these different imaging modalities, a more comprehensive assessment of an area of interest can be achieved, leading to improved clinical diagnosis and decision-making. This multimodal approach holds great potential for enhancing cancer detection, biopsy guidance, and informing treatment decisions in various clinical settings.

MSI often involves bulky instrumentation and requires specialized optics that allow for spectral scanning processes, such as filters and diffraction grating elements. Advancements in optics and electronics, such as the rapid development of smartphone technologies, have enabled a move toward integration into point-of-care-friendly platforms. MSI has been implemented in various smartphone platforms, which demonstrate the potential for low-resource applications [17–20]. A low-cost commercial instrument called OralScan implements MSI and was tested in India for screening and detection of oral cancers. The evaluation of this MSI device demonstrates the potential for such a device to be used as an adjunct in LRS for lesion assessment and biopsy guidance [21].

While recent efforts have translated MSI to point-of-care friendly platforms, low-cost NBI remains to be implemented with great potential for being adapted for LRS. Advances in DL for NBI have enabled its use for detection, classification, and real-time analysis of precancers and cancers in the gastrointestinal tract, which could be translated to the point of care [22]. For example, an algorithm developed in a large-scale study with 1777 narrowband images from 295 cases collected across three centers showed comparable performance to senior endoscopists and improved performance to junior endoscopists in detection of early gastric cancer, demonstrating the potential for this tool to aid in cancer diagnosis [23]. While the diagnostic performance of the algorithm was assessed on still images, it has the potential to be applied to video frames collected during the endoscopic imaging session.

Integration of multiple imaging modalities such as WLI, AFI, and MSI has proven to be beneficial in improving the sensitivity and specificity of cancer detection. For example, a smartphone-based device with two imaging modalities (WLI and AFI) and a DL algorithm was tested in a large cohort in India [24,25]. Here, the potential for the algorithm to be used by non-expert clinicians was demonstrated. In another case, a bimodal system integrating AFI and MSI with a cloud-based machine learning algorithm demonstrated potential for real-time screening and detection of potentially malignant oral lesions and for providing biopsy guidance [26]. A dermascope that combines WLI and laser speckle contrast imaging for blood flow imaging showed distinct profiles for two different skin lesion types [27]. Overall, the combination of various imaging modalities shows improved diagnostic performance, and the integration of computer-aided diagnosis shows the potential to reduce reliance on expertise.

Microscopic imaging

Besides in vivo macroscopic imaging, promising technical advances have been achieved in developing in vivo microscopic imaging techniques for early cancer detection. The integration of in vivo microscopy (IVM) and computer-aided algorithms has shown potential to enable real-time, non-invasive, accurate point-of-care diagnosis. This approach can potentially address several challenges associated with standard-of-care histopathology, which are particularly significant in LRS, including increased cost, patient risk resulting from invasive biopsies, inaccurate diagnosis due to examining limited areas of suspicious lesions, and extensive needs for infrastructure and expertise [28]. Recent efforts to promote the implementation of IVM in LRS are highlighted in three aspects: 1) advancements in technical design to enhance system portability and reduce cost, as well as field implementation for clinical performance evaluation in LRS; 2) leveraging computer-aided algorithms to enhance imaging capabilities and facilitate clinical data interpretation; 3) integration of multiscale, multimodal imaging for improved diagnostic capability. In this section, we review and discuss key innovations in these aspects of in vivo microscopic imaging that aim to improve cancer screening, diagnosis, and treatment in LRS. While this review primarily focuses on in vivo imaging techniques, it is important to acknowledge recent advances in ex vivo imaging techniques for histopathology, such as digital pathology and AI-based analysis [29]. The clinical adaptation of both novel ex vivo and in vivo imaging techniques in LRS can improve the screening, diagnosis, and treatment process, thus addressing the global health challenges in cancer prevention.

Low-cost system development and clinical validation

Recent progress has been made in the development of portable, affordable, and easy-to-use in vivo microscopes. Several systems have shown promising preliminary results in pilot studies. Currently, IVM reveals microscopic features based on endogenous contrast, such as autofluorescence and reflectance, or exogenous contrast agents, such as fluorescein. Microscopic autofluorescence imaging can be performed using several modalities, like multiphoton microscopy [30] and light-sheet microscopy [31]. However, the effectiveness of using these systems for in vivo cancer detection still requires further clinical evaluation. Although there are some efforts to improve usability, such as development of a probe-based two-photon microscope [32], most existing systems are built in a large form factor with a high cost. Future work can focus on clinical validation and adapting the technology for application in LRS.

Reflectance, another endogenous source of contrast, is more commonly used in clinical settings. Reflectance confocal microscopy (RCM) is an established, commercially available imaging technique. Its clinical performance has been validated in clinical studies over decades, especially for skin cancer detection [33]. However, the large size and high cost of RCM hinder its application in LRS. Recent progress in eliminating the need for expensive mechanical scanning components in RCM, such as the use of chromatic dispersion-based techniques to enable lateral scanning [34] or depth scanning [35], opens new opportunities for designing low-cost, portable RCM systems.

Similar to RCM, optical coherence tomography (OCT) is a costly, commercially available technique and has been extensively studied for in vivo cancer detection. Portable, lower-cost OCT systems have been developed recently [36,37]. Despite advances in system development, only a limited number of clinical studies have been performed to evaluate the diagnostic performance. In a study of 232 patients in India, a low-cost OCT system delineated oral cancer with a sensitivity of 93% and a specificity of 74% when incorporating an automated image processing algorithm [37]. To facilitate the clinical translation and improve cancer diagnosis in LRS, future studies are necessary to further validate clinical performance of these systems.

In vivo microscopic imaging with exogenous contrast agents addresses several challenges associated with imaging with endogenous contrast, including low signal-to-noise ratio and limited resolving capability for cellular structures. Probe-based IVM using exogenous contrast agents have been developed to enable real-time lesion assessment at the point of care. These systems include commercially available options such as confocal laser endomicroscopy (CLE) [38], and other research prototypes with improved imaging capabilities, such as higher imaging speed [39,40]. Despite the capability to resolve subcellular features in real-time, the high cost of CLE (>$100,000) hinders its broad applicability in LRS. As an alternative, a portable and low-cost fluorescence fiber-optic high-resolution microendoscope (HRME) has been developed for in vivo cancer detection and validated in LRS. In a recent study conducted in Brazil, HRME was shown to detect high-grade cervical precancerous lesions with a sensitivity of 92% and a specificity of 60% in 1486 patients when integrated with computer-aided diagnosis, which is comparable to colposcopy interpreted by expert clinicians [41]. Besides system development, an additional challenge of implementing IVM with exogenous agents is that only a limited number of contrast agents have been approved for human use. Future advances in development of novel contrast agents can promote clinical application of IVM in different clinical settings.

Algorithm development to facilitate clinical applications

The rapid advances in development of novel algorithms, especially DL-based algorithms, facilitate the application of IVM in LRS in various aspects. First, recent algorithms enable cost-efficient ways to address the limitations of IVM and improve imaging capabilities. With enhanced computational power and high imaging speed, sophisticated mosaicking algorithms have been developed to significantly enlarge the field of view of IVM while preserving cellular resolution [40]. DL algorithms are also implemented to overcome the intrinsic tradeoffs among different imaging specifications. For example, a DL-based denoising algorithm has been employed in a low-cost RCM system to increase the signal-to-noise ratio at a high imaging speed [42]. Although not yet demonstrated in IVM or cancer detection, a novel DL algorithm can enhance the resolution even beyond the diffraction limitation in multiple fluorescence microscopy modalities [43]. In another non-IVM application, the DL algorithm enables virtual three-dimensional refocusing in a single two-dimensional widefield image without physical scanning, thus digitally increasing the depth of field in widefield fluorescence microscopy [44]. Similar concepts can be applied to enhance the imaging performance of IVM and diagnostic capability for cancer detection without adding additional cost, which is particularly advantageous for potential implementation in LRS.

Second, novel optical system designs can be provided by leveraging AI algorithms at the system development stage. This hardware-software co-design concept has derived design insights to improve imaging capabilities in cost-efficient ways. For example, a microscope with an extended depth of field has been developed by adding an inexpensive phase mask into the optical path and incorporating a DL reconstruction algorithm [45]. In another example, the point-spread-function modulation was designed and optimized with a neural network to enable three-dimensional reconstruction in the microscopy system [46]. These strategies represent promising directions for the development of the next generation of IVM systems for cancer detection that enhance imaging performance while maintaining cost-effectiveness and adaptability for LRS.

Finally, AI-based diagnostic algorithms effectively facilitate the field implementation of IVM in LRS. AI-based algorithms have demonstrated great potential in providing rapid, accurate, and automatic point-of-care diagnosis by analyzing high-resolution imaging data. This alleviates challenges associated with translating novel imaging techniques into clinical practice and ultimately reduces the need for clinical expertise in LRS. AI-based diagnostic algorithms have been incorporated with several low-cost IVM systems and shown promising results in cancer detection [37,41]. AI-based diagnosis will continue to play an important role in increasing accessibility of high-quality health care and improving cancer diagnosis globally. While promising results have been reported, the sample sizes of most studies are limited, and whether the performance will generalize in the absence of expert practitioners remains to be answered. As a result, the performance of AI algorithms needs to be further evaluated prospectively, especially in varied patient populations and clinical settings.

Multiscale and multimodal imaging

Integrating macroscopic and microscopic imaging techniques for multiscale, multimodal imaging (MMI) creates new possibilities to further improve diagnostic capabilities. First, MMI provides an intuitive and accurate method for real-time clinical guidance. A significant barrier to implement IVM in clinical practice is the correlation of microscopic assessment with corresponding lesion location, especially when a substantial amount of microscopic imaging data is acquired over a large tissue area. Introducing physical landmarks onto tissue surfaces can facilitate the placement and position tracking of the microscopic probe for improved correlation [47], but this approach is limited by the accessibility of anatomic sites. Enabled by the multiscale imaging approach, microscopic images are correlated to anatomic locations by tracking the microscopic imaging probe via macroscopic imaging. This provides the spatial context necessary for assessment of suspicious lesions over a large field of view with co-registered microscopic information. Preliminary ex vivo studies have demonstrated promising results in utilizing multiscale imaging for real-time tumor margin delineation in skin [48]. In a recent pilot study, a low-cost multiscale and multimodal mobile colposcope was developed, which integrates a commercially available low-cost colposcope and a high-resolution fiber-optic microendoscope [49]. By implementing algorithms for real-time diagnosis, probe tracking, and image co-registration, the mobile colposcope correlates diagnostic scores to anatomic location and shows potential for improving cervical cancer detection in LRS.

Second, MMI can potentially enable the entire screening, diagnosis, and treatment procedure in a single visit in LRS. Macroscopic imaging provides a rapid examination of suspicious lesions at low resolution, which then guides microscopic investigations of specific regions of concern at high resolution for diagnosis. Multiscale and multimodal algorithms can be developed to provide automatic lesion detection and treatment guidance at the point of care, thereby significantly mitigating reliance on clinical expertise in LRS. Finally, additional diagnostic information provided by different imaging modalities in different scales can be potentially integrated to further improve diagnostic performance. In a pilot study, an MMI system that integrates macroscopic autofluorescence imaging and high-resolution imaging shows improved sensitivity of high-grade oral dysplasia compared to clinical evaluation by an expert [50]. Despite the potential benefits and promising preliminary results, future clinical studies are necessary to further validate the diagnostic capability of MMI systems in LRS.

Discussion and outlook

Recent advances in the development of low-cost optical imaging systems, along with computational algorithms, show significant promise in addressing the global health challenge of cancer prevention in LRS. The rapid progress in optical technology has enabled development of several portable, easy-to-use imaging systems that offer high-quality imaging at a much lower cost. Some of these innovative imaging tools have been validated in clinical studies, demonstrating their adaptability and ability to overcome infrastructure limitations in LRS. Development of algorithms for computer-aided diagnosis, especially the rapid advent of DL-based approaches, holds great potential for providing accurate, real-time, diagnostic guidance, potentially reducing reliance on clinical expertise. Furthermore, integration of macroscopic and microscopic imaging techniques opens new opportunities to improve cancer detection capabilities and enable screening-diagnosis-treatment programs in a single visit in LRS. Table 1 summarizes the low-cost optical imaging technologies discussed in this review, highlighting hallmarks of cancer [51] that can potentially be detected, stage of clinical validation, and clinical applicability in LRS for each imaging modality.

Table 1.

Low-cost optical imaging technologies for in vivo cancer detection in low-resource settings.

Imaging modality	Macroscopic imaging			Microscopic imaging			MMI
	WLI	AFI	MSI	RCM	OCT	HRME
Hallmarks of cancer [51] that can potentially be detected with the modality	• Inducing angiogenesis • Tumor-promoting inflammation	• Activating invasion and metastasis • Deregulating cellular energetics	• Inducing angiogenesis • Tumor-promoting inflammation	• Sustaining proliferative signaling • Evading growth suppressors • Resisting cell death • Enabling replicative immortality	• Sustaining proliferative signaling • Evading growth suppressors • Resisting cell death • Enabling replicative immortality • Activating invasion and metastasis • Tumor-promoting inflammation	• Sustaining proliferative signaling • Evading growth suppressors • Resisting cell death • Enabling replicative immortality	• Sustaining proliferative signaling • Evading growth suppressors • Resisting cell death • Enabling replicative immortality • Activating invasion and metastasis • Deregulating cellular energetics • Inducing angiogenesis • Tumor-promoting inflammation
Stage of clinical validation	Clinically evaluated in large studies	Clinically evaluated in large studies	Tested in pilot studies	Clinically evaluated in large studies	Clinically evaluated in large studies	Clinically evaluated in large studies	Tested in pilot studies
Comments on clinical applicability in LRS	• Low-cost FDA cleared devices available • DL algorithms in development for diagnosis • Easy to obtain high quality images with simple devices	• FDA cleared devices available • DL algorithms in development for diagnosis • Improved sensitivity, but reduced specificity • Must be performed in a dark environment	• Provides additional spectral characteristics for cancer detection • Requires further clinical evaluation	• High-cost FDA cleared devices available • Low-cost systems in development • No contrast agent needed	• High-cost FDA cleared devices available • Low-cost systems in development • No contrast agent needed	• Low-cost systems in development • DL algorithms in development for diagnosis • Contrast agent needed	• Facilitate real-time clinical guidance • Potentially enable the entire screening, diagnosis, and treatment procedure in a single visit • Relatively higher system complexity

WLI: white light imaging; AFI: autofluorescence imaging; MSI: multispectral imaging; RCM: reflectance confocal microscopy; OCT: optical coherence tomography; HRME: high-resolution microendoscope; MMI: multiscale, multimodal imaging; LRS: low-resource settings; DL: deep learning.

Although these efforts hold promise, several major challenges remain to be addressed for improved cancer diagnosis worldwide. Currently, only a limited number of low-cost imaging technologies have been evaluated in clinical studies conducted in LRS. Continued efforts should be made to facilitate technological advancements for LRS. Clinical barriers to technology adoption should be considered during development stages. While AI-based diagnostic algorithms are widely implemented in clinical practice, new challenges have also been introduced, including standardization of training datasets, interpretation of training outcomes and potential artifacts, and ensuring algorithm robustness. Nevertheless, with ongoing efforts to address these challenges, AI-based algorithms will continue to play an increasingly important role in cancer detection. Rapid development of powerful, low-cost optical imaging tools and computer-aided diagnostic algorithms presents opportunities to enable accurate point-of-care cancer diagnosis, thus significantly improving detection of precancer and cancer at early stages globally.

Data availability

No data was used for the research described in the article.