Development of a multispectral imaging apparatus for cost-effective fundus disease detection

Abstract

Fundus spectral imaging (FSI) integrates fundus photography with spectral techniques, providing both spatial and spectral information for retinal imaging. Whereas existing FSI systems have demonstrated advantages in structural and functional imaging, their widespread adoption is hindered by high costs and complex optical designs. To address these challenges, we propose a low-cost multispectral fundus camera with a simplified optical design, built from off-the-shelf optics, 3D-printed parts, and equipped with fiber-bundle-coupled multi-wavelength LED illumination source (470–740 nm). Additionally, the proposed multispectral imaging apparatus incorporates a coaxial non-separated polarization-based reflection suppression technique, using orthogonal polarizers to suppress corneal reflections without pupil-plane separation. To the best of our knowledge, this is the first application of such an architecture in the context of FSI. Experimental results demonstrate that the developed system achieves high-quality FSI under low-cost conditions, validating its feasibility as a practical solution. Clinical validation validates its diagnostic capability for diabetic retinopathy, choroidal pigmented nevus, and, notably, the first reported spectral imaging of peripapillary atrophy. The system achieves performance comparable to conventional color fundus photography while enabling superior diagnosis of deep fundus conditions such as choroidal lesions, offering a cost-effective and practical FSI solution for broader deployment in resource-limited settings.

1. Introduction

Fundus spectral imaging (FSI) integrates fundus photography with spectral techniques, enabling the acquisition of both spatial and spectral information of the fundus [1]. In recent years, FSI has gained widespread attention in retinal disease detection and has been demonstrated to offer significant advantages [2–4]. The spatial information facilitates structural imaging of the fundus, while the spectral information provides insights into its functional characteristics [5–7].

In terms of structural imaging, FSI exhibits high sensitivity and specificity in detecting choroidal tumors and other retinal diseases. It enables the visualization of tumor boundaries [2] that cannot be observed with conventional color fundus photography (CFP). Moreover, compared to fluorescein angiography (FFA) and indocyanine green angiography (ICGA) [8–11], FSI offers superior contrast while remaining completely non-invasive, as FFA and ICGA require contrast agents, making them invasive procedures that may cause discomfort or allergic reactions.

For functional imaging, FSI has been applied in retinal blood oxygen saturation analysis [12–19] and the detection of retinal biomarkers associated with Alzheimer's disease [20], as well as molecular spectral analysis of retinal pigments [21]. Additionally, it can be used for retinal reflectance evaluation [22] and the construction of retinal topography [23], providing comprehensive insights into both retinal structure and functionality. Therefore, FSI has emerged as a promising next-generation retinal examination and diagnostic technology, with the potential to replace CFP, FFA, and ICGA in the future.

Various spectral imaging techniques have been employed in FSI, including scanning-based, staring-based, transform-based, and snapshot imaging methods. Scanning FSI captures a spectral data cube of the fundus by combining slit scanning with grating dispersion, offering high spectral resolution [24,25]. However, its data acquisition time (typically 2-8 seconds) far exceeds the physiological response time of the pupil to light stimulation, often necessitating pharmacological dilation to ensure imaging quality. Fourier transform-based FSI also relies on slit scanning and utilizes a Sagnac lateral shearing interferometer, Fourier lens, and cylindrical lens to acquire interferograms of the fundus, reconstructing the spectral data cube via computational algorithms [26]. Due to its dependence on slit scanning, its temporal resolution is similar to the scanning FSI, facing comparable acquisition delays. Snapshot FSI obtains a complete spectral data cube in a single exposure, significantly improving imaging speed. However, it is often constrained by spatial resolution or field of view (FOV). Techniques such as the single-disperser coded aperture snapshot spectral imager (SD-CASSI) [27] and filter array-based [28,29] methods enable high-speed imaging but typically compromise spatial sampling or FOV. Staring FSI captures different spectral slices of the fundus by discretely illuminating the scene, which can be achieved using switchable filters [30], tunable filters [31,32], tunable light sources [33,34], or multi-wavelength LEDs [35,36]. Among these, multi-wavelength LEDs [5,6,11], despite their limited spectral bands, allow for rapid acquisition within the pupil’s light response time through optimized hardware design and control, making them the most widely used FSI technology today.

However, most existing FSI systems are based on modifications of traditional fundus cameras, typically by integrating spectral imaging modules into the optical path or utilizing relay optics for secondary spectral imaging. Additionally, some research efforts have focused on developing entirely new optical designs for FSI systems, but these devices are often expensive and difficult to deploy, especially in rural and underserved areas where accessibility to retinal disease screening remains a challenge.

To address these limitations, this study proposes a low-cost multispectral fundus imaging system that incorporates commercial off-the-shelf optical components, a 3D-printed mechanical structure, and a fiber-bundle-coupled multi-wavelength LED illumination source spanning 470-740 nm. This fiber-based illumination design not only ensures uniformity across the fundus but also facilitates spectral scalability, making it adaptable to future diagnostic needs. Furthermore, the system integrates a coaxial non-separated polarization-based reflection suppression technique, which, to the best of our knowledge, is applied for the first time in the context of FSI. This technique employs a pair of orthogonally aligned polarizers to effectively suppress corneal and internal surface reflections without requiring separation of illumination and imaging paths at the pupil plane, thereby simplifying the optical design while maintaining high imaging quality.

Beyond the hardware design, this system has been clinically validated in multispectral imaging of various retinal conditions, including diabetic retinopathy (DR), peripapillary atrophy (PPA), and choroidal pigmented nevus (CPN). Notably, our work presents the first spectral imaging report on PPA, offering new insights into its spectral characteristics. The system has received clinical approval and is currently undergoing further evaluation at Ningbo Eye Hospital to assess its feasibility and diagnostic utility. The remainder of this article is organized as follows: Section 2 presents the system architecture and imaging principles. Section 3 discusses the experimental results, including system imaging performance and its application in clinical disease detection. Finally, Section 4 summarizes the findings and explores potential future improvements.

2. Principles and methods

2.1. System prototype

The structural diagram of the system is shown in Fig. 1. All optical components are aligned on the same horizontal plane as the pupil center. The system consists of two optical paths: the illumination path and the imaging path. First, a customized fiber bundle (Beijing Shouliang Technology Co., Ltd., China) is coupled with LEDs of different wavelengths (Chundaxin Inc., China), and is directed into the illumination optical system. A Köhler illumination manner is employed, focusing the light at the pupil and forming a uniform illumination region on the fundus. The emitted light first passes through a polarizer (P₁), then a collimating lens (L₁), a polarization beam splitter (PBS), and a retinal objective lens (L₂). L₁ is a lens with a diameter of 25.4 mm and a focal length of 50 mm, while L₂ is a C-mount industrial lens with a focal length of 25 mm. L₂ focuses the collimated light onto the pupil plane, forming a 4 mm illumination spot (as shown in Fig. 2(C) in Section 3.1), thereby providing uniform illumination across the fundus. To ensure image quality, the subject’s pupil diameter must exceed 4 mm. After passing through the pupil, it illuminates the fundus. The light reflected from the fundus travels back through the retinal objective (L₂), the PBS, a lens (L₃), and a second polarizer (P₂), before passing through the imaging objective (L₅) and reaching the surface of a complementary metal-oxide-semiconductor (CMOS) detector (D₂, MER2-503-36U3 M, Daheng Imaging, Inc., China). This detector captures the total light intensity at each pixel, generating a monochromatic image. Additionally, a separate lens (L₄) and detector (D₁, MER2-503-36U3 M, Daheng Imaging, Inc., China) are used for pupil alignment assistance. L₃ is a lens with a diameter of 25.4 mm and a focal length of 50 mm, while both L₄ and L₅ are C-mount industrial lenses with a focal length of 25 mm.

Fig. 1.

Optical diagram of the fundus multispectral imaging system.

Fig. 2.

Spatial resolution and FOV testing of the prototype system. (A) Results of spatial resolution and FOV testing. (B) The prototype system. (C) A circular illumination spot comprising all selected wavelengths focused at the pupil plane.

We introduce a coaxial non-separated polarization-based reflection suppression technique to effectively attenuate corneal and other forward surface reflections in fundus imaging. Unlike conventional systems that rely on illumination-imaging path separation at the pupil plane, this method employs a coaxial configuration enhanced by a pair of orthogonally aligned polarizers (P₁ and P₂), achieving robust reflection suppression without the need for pupil-based separation. Without these crossed polarizers, reflections from the cornea and ghost images from internal lenses would significantly degrade the imaging quality. The use of a polarization beam splitter prism, instead of a conventional beam splitter, enhances light source efficiency by achieving total reflection.

2.2. Radiation transfer model

We define ${\varphi }_{LED}$ as the luminous power of each LED. The radiation transmission efficiency at the coupling interface between the LED and the optical fiber input is ${\phi }_1$ , while the transmission efficiency of the optical fiber itself is ${\phi }_2$ . So, the light power entering the lighting system is ${\varphi }_1={\phi }_1{\phi }_2{\varphi }_{LED}$ . Let us denote the radiative transfer efficiency of the lighting system, the ocular media (including the cornea, aqueous humor, lens, and vitreous humor), and the imaging system as ${\phi }_3$ , ${\phi }_4$ and ${\phi }_5$ , so the LED illumination power reaching the fundus is represented as ${\varphi }_2={\phi }_1{\phi }_2{\phi }_3{\phi }_4{\varphi }_{LED}$ . The Monte Carlo model of the fundus assumes that there are five different layered structures, each layer containing different pigments (i.e., macular pigments in the retina, hemoglobin in the retina, melanin in the retinal pigment epithelium (RPE), melanin in the choroid, and hemoglobin in the choroid) [36]. The concentration of these five pigments is indicated by $C_{MP}$ , $C_{RH}$ , $C_{RM}$ , $C_{CM}$ , and $C_{CH}$ . The Monte Carlo model computes the reflectance $R(\mathit{p},\lambda )$ at the wavelength $\lambda$ of tissue described by a parameter vector $\mathit{p}(x,y)=g(C_{MP},C_{RH},C_{RM},C_{CM},C_{CH})$ . In the presence of noise ω, the measured value f of the CMOS detector can be expressed as follows.

$$f=\alpha \int \int \int {\phi }_4{\phi }_5{\varphi }_{2}R(\mathit{p},\lambda )Q(\lambda )dxdyd\lambda +\omega =\alpha \int \int \int \mathrm{\Phi }R(\mathit{p},\lambda )Q(\lambda )dxdyd\lambda +\omega$$ (1)

In Eq. (1), the constant α serves to scale the signal to the detector's range and is determined by the factory calibration and $\mathrm{\Phi }={\phi }_4{\phi }_5{\varphi }_2$ is the transmission efficiency of the whole system apart from the fundus reflectivity $R(\mathit{p},\lambda )$ and the CMOS detector quantum efficiency $Q(\lambda )$ . In theory, we could utilize the image data captured by the CMOS detector to infer the model parameters that characterize the reflectivity of the tissue. However, in reality, not all of these factors are known, thus requiring a more thorough examination of the issue. We use P LEDs with different wavelengths, denoted as $1,\dots ,P^{\mathrm{\prime }},\dots ,P$ . The wavelength range of each LED is $({\lambda }_{P^{\mathrm{\prime }}}-{\tau }_{P^{\mathrm{\prime }}}/2,{\lambda }_{P^{\mathrm{\prime }}}+{\tau }_{P^{\mathrm{\prime }}}/2)$ , where ${\lambda }_{P^{\mathrm{\prime }}}$ is the central wavelength of the P, ${\tau }_{P^{\mathrm{\prime }}}$ is the full width at half maxima (FWHM) of $P^{\mathrm{\prime }}$ . A series of fundus images captured at different wavelengths are recorded as $F=f^{(1)},\dots ,f^{(P^{\mathrm{\prime }})},\dots ,f^{(P)}$ .

2.3. Human subjects and ethical considerations

This study was approved by the Ethics Committee of Ningbo Eye Hospital, and written informed consent was obtained from all participants prior to imaging. All procedures adhered to the ethical principles outlined in the Declaration of Helsinki.

The study was conducted in two phases. In the first phase, nine healthy volunteers with no history of ocular disease were recruited in a laboratory setting. A multispectral imaging system was used to acquire fundus images of both eyes at nine different wavelengths (474 nm, 500 nm, 518 nm, 553 nm, 593 nm, 610 nm, 630 nm, 660 nm, and 740 nm), aiming to assess the system’s fundamental imaging capabilities and spectral characteristics. In the second phase, 300 additional healthy participants meeting the same criteria were recruited at Ningbo Eye Hospital. Fundus images were acquired using the prototype system at five key wavelengths (474 nm, 553 nm, 610 nm, 660 nm, and 740 nm). Additionally, to evaluate the differences between multispectral imaging and conventional clinical methods, color fundus images were captured using a commercial fundus camera (Reticam 3100, Syseye, Inc., Shenzhen, China) for comparative analysis. Additionally, the system operates in a non-mydriatic mode and does not require pupil dilation during image acquisition. Multispectral images are captured using a sequential scanning approach. Considering that the human pupil typically reacts to light within 0.2-0.5 seconds, we designed the control system and software to complete the entire image acquisition process within 0.35 seconds. This timing effectively minimizes pupillary response to wavelength changes, ensuring consistency across the multispectral images. Besides, participants were instructed to stabilize their head using a forehead rest and remain still while following the physician’s guidance to complete the imaging process.

3. Results and discussion

3.1. Testing for model eyes

As shown in Fig. 2, the prototype system is mounted on a three-axis translation stage and equipped with a headrest to ensure imaging stability. We developed a C++ based MFC software interface to precisely coordinate the synchronization between the light source and detector. To evaluate the imaging performance of the system, we first assessed its FOV and spatial resolution using a model eye (Rowe Technical Design, Inc.). By adjusting two polarizers, we effectively suppressed corneal reflections, while manual focusing was used to optimize the imaging clarity. Each annular region of the model eye corresponds to a 15° viewing angle, resulting in a total FOV of approximately 30°, which meets the general requirements for fundus imaging. Additionally, we measured the system's resolution using a 1951 USAF spatial resolution test chart. The results showed that within the central FOV, the system could resolve up to Group 5 Element 6, corresponding to a spatial resolution of 8.77 μm. This high resolution demonstrates the system’s capability to capture fine retinal structures, providing a solid foundation for high-precision imaging. Additionally, by manually adjusting the lens focus, the system can compensate for the patient's spherical refraction up to ±15 D.

To further assess the system’s stability and repeatability, we conducted multiple measurements under varying lighting conditions. The results indicated that the spatial resolution remained stable, confirming the system's imaging reliability. Compared with mainstream commercial fundus cameras, such as the Reticam 3100 used in Section 3.5, which has a resolution of 6 μm and a FOV of 45°, our system has a slightly smaller FOV but achieves comparable spatial resolution, making it suitable for retinal pathology detection. Notably, with a resolution of 8.77 μm, the system has significant potential for identifying early biomarkers of DR, such as microaneurysms (30-100 μm) and hard exudates (25-50 μm). This capability underscores the feasibility of a low-cost, high-precision solution for advanced fundus screening. Next, we will integrate clinical samples to further validate the system's applicability in pathological detection.

3.2. Illumination safety

Illumination safety is a crucial consideration in fundus imaging, as it is essential to prevent potential photochemical and thermal damage to the eyes. International standards, such as ISO 10940-2009, ISO 15004-1-2020, and ISO 15004-2-2007, provide detailed guidelines on the protection against optical hazards in fundus cameras. These standards serve as key references for evaluating the safety of illumination systems.

To assess the illumination safety of our system, we used a spectroradiometer (HPC350FR, Shenzhen Tianfangdiyuan Instrument Co., Ltd) to measure the spectral power output of different LEDs (see Fig. 3). Based on the acquired data, we further calculated the maximum output power density (P_MAX) of nine different LEDs (see Table 1). The ocular transmittance (OT) values in Table 1 were derived from the transmission rates of the cornea, aqueous humor, lens, and vitreous, as reported in the literature. The exposure time (ET) was determined by software settings, while the weighting factors were obtained from the ISO 15004-2-2007 standard.

Fig. 3.

Normalized spectral curves of nine used LEDs.

Table 1. Light safety estimation ^a .

Wavelength (nm)	474	500	518	553	593	610	630	660	740
PMAX (W/cm2)	0.808E-3	0.998E-3	0.948E-3	0.434E-3	0.526E-3	0.738E-3	1.009E-3	1.157E-3	1.576E-3
OT (%)	73.71	76.03	77.04	79.22	80.15	82.72	82.90	83.65	84.58
ET (s)	0.1	0.1	0.1	0.1	0.05	0.05	0.05	0.05	0.05
APHWF	0.55	0.1	0.05	0.01	0.0016	0.001	0.001	0.001	<0.001
THWF	1	1	1	1	1	1	1	1	0.83

^a OT is ocular transmittance; ET is exposure time; APHWF is Aphakic photochemical hazard weighting factor; THWF is Thermal hazard weighting factor.

In practical applications, a 740 nm LED was used for continuous illumination during the observation, whereas pulsed light at different wavelengths was employed for multispectral fundus imaging. According to the ISO 15004-2-2007 standard for Group 2 instruments, the 740 nm LED requires evaluation under the continuous illumination mode for its weighted retinal visible and infrared thermal irradiance (E_VIR-R). In contrast, the nine pulsed wavelengths need to be assessed for their weighted retinal visible and infrared radiation exposure (H_VIR-R). The illumination safety evaluation of our system yielded an E_VIR-R value of 1.11E-3 W/cm² under continuous illumination, which is significantly below the standard limit of 0.7 W/cm². Under pulsed illumination, the H_VIR-R values for the nine wavelengths were 0.044E-3 J/cm², 0.759E-4 J/cm², 0.759E-4 J/cm², 0.344E-5 J/cm², 0.337E-6 J/cm², 0.305E-6 J/cm², 0.418E-6 J/cm², 0.484E-6 J/cm², and 0.666E-6 J/cm², all of which are well below the standard limit of 2 J/cm². Therefore, the illumination system of our fundus imaging apparatus totally complies with ISO safety requirements.

3.3. Multispectral imaging of fundus

In this study, we used 740 nm wavelength light for alignment and preliminary observation to ensure the accuracy and stability of fundus image acquisition. The primary reason for selecting 740 nm as the alignment light source is that it falls within the red-light spectrum and is close to the near-infrared region, minimizing visual discomfort. This allows for prolonged illumination without causing significant irritation, thereby reducing blinking and eye movement interference and improving the imaging stability.

Once alignment was completed, the system began capturing multispectral fundus images. Figure 4 presents nine wavelength fundus images and a registered image of a volunteer's right eye. Among them, the registered image consists of 3 × 3 = 9 regions, each of which was cropped and reconstructed from regions of images acquired at different wavelengths, allowing for an intuitive visual assessment of the registration accuracy. Different wavelengths emphasize distinct structural features of the fundus at various depths. In the short-wavelength range (blue and green light), retinal arteries and veins appear more prominently due to the limited penetration of short-wavelength light. In contrast, in the long-wavelength range (red light), the choroidal structures become more distinct, as light penetration depth increases with the wavelength, allowing deeper illumination into the choroidal layer. Thus, longer wavelengths exhibit greater penetration in fundus tissues, making red and near-infrared light more suitable for visualizing deep choroidal structures.

Fig. 4.

Fundus images at nine different wavelengths and the registered image of a volunteer’s right eye. The registered image was cropped and reconstructed from regions of images acquired at different wavelengths.

Notably, different wavelengths also affect vessel contrast. For example, at 553 nm, the extinction coefficients of oxygenated hemoglobin (HbO₂) and deoxygenated hemoglobin (Hb) are relatively similar, resulting in minimal brightness differences between arteries and veins. However, at 630 nm, the extinction coefficient difference between HbO₂ and Hb becomes more pronounced, significantly enhancing the brightness contrast between arteries and veins, which facilitates clearer vascular differentiation.

Multispectral imaging also demonstrates potential in detecting fundus abnormalities. Many lesions that are difficult to identify in conventional color fundus images may become more apparent at specific wavelengths. This is because different pathological tissues exhibit distinct optical absorption and scattering properties across different wavelengths, providing richer optical information for improved diagnosis.

3.4. Relative spectral curves of fundus

Since the system employs a time-division multiplexing approach for image acquisition, misalignment may occur between retinal images captured at different wavelengths. Therefore, before conducting spectral analysis of the retina, image registration is necessary. During the registration process, the retinal image at 554 nm is used as the reference, and images at other wavelengths undergo rigid transformations including translation, rotation, and scaling based on feature points to achieve precise alignment. The registered image is shown in Fig. 4. After registration, the aligned retinal images from different wavelengths are stacked to form a BSQ (Band Sequential) format spectral data cube. BSQ is a commonly used format for storing multispectral or hyperspectral data, where the image data is stored sequentially by spectral bands, with each band’s data continuously arranged.

As different exposure times were used during the image acquisition to maintain uniform grayscale levels across different wavelengths, the initial images are normalized to a common exposure time. This transformation is expressed as: $F^{\mathrm{\prime }}=F/t$ , where F represents the original image intensity, and t is the exposure time in milliseconds. From the spectral data cube, grayscale values at specific spatial locations can be extracted and normalized to their maximum value, yielding a grayscale curve g. In Section 3.2, we obtained the P_MAX for different wavelengths. By dividing g by the maximum-normalized P_MAX, we derive the relative reflectance spectra of retinal tissues. As illustrated in Fig. 5(B), we extracted relative reflectance spectra from different retinal structures such as arteries, veins, the optic disc, the nerve fiber layer, and choroidal vessels for volunteers. To visualize inter-individual variability, we plotted box plots and fitted spectral curves, where the fitted spectra were obtained by connecting the mean values of the data points. The relative spectral curves can be used to evaluate the visibility of different retinal tissues at various wavelengths. At shorter wavelengths (474 nm), veins and arteries are more prominent, while at longer wavelengths (660 nm), the choroid becomes more noticeable. Referring to Fig. 4 and Fig. 5, the optic disc gradually shifts from dark to light.

Fig. 5.

(A) Fundus tissues at 610 nm and (B) their corresponding spectral curves.

3.5. Comparison with CFP

Traditional CFP have been widely used in clinical practice for the diagnosis and treatment of retinal diseases. However, with advancements in technology, spectral imaging has emerged as a novel diagnostic tool. By capturing images across multiple wavelengths, spectral imaging offers unique advantages in detecting a variety of retinal pathologies. To assess the performance of FSI in clinical applications, we conducted extensive testing at Ningbo Eye Hospital, comparing them with conventional CFP. By selecting representative cases of DR, PPA, and CPN, we demonstrated that spectral imaging provides richer and more detailed pathological information at different wavelengths. This section will explore the imaging characteristics of these three pathologies and discuss the potential advantages of spectral cameras in disease detection.

DR is a common complication in diabetic patients. Chronic hyperglycemia causes damage to the retinal blood vessels and nerve tissue, leading to a series of retinal lesions. DR is generally categorized into proliferative and non-proliferative stages, with Stage VI of proliferative diabetic retinopathy (PDR) being the most severe form. PDR often leads to rapid vision loss and, in some cases, blindness. In the early stages of DR, the condition is characterized by microaneurysms and small vessel leakage in the retina. Common imaging features include microaneurysms, hard exudates (see in Fig. 6), and cotton wool spots. As the disease progresses, there will be neovascular or vitreous hemorrhage, and even proliferative membrane formation, complicated by traction retinal detachment, further exacerbating vision loss. As illustrated in Fig. 6, traditional color fundus imaging can detect DR, whereas FSI at specific wavelengths (610 nm and 660 nm) enhances the contrast of both the retinal surface and deeper structures, making the identification of hard exudates more distinct.

Fig. 6.

Fundus images of hard exudates in a case of DR. (A) Conventional color fundus image showing the location of hard exudates. (B-E) Corresponding multispectral images acquired at different wavelengths. The white arrows indicate the locations of the hard exudates.

PPA is a common fundus finding in older, glaucoma or highly myopic patients, thus it is also referred to as myopic conus. They are closely associated with the elongation of the ocular axial length and changes in the shape of the eyeball, resulting from the asymmetric separation of the retinal pigment epithelium (RPE), the Bruch membrane, and the choroid from the optic disc. Although PPA does not significantly affect vision at early stages, without timely monitoring and treatment, they may lead to serious complications such as macular disease, macular hole and cystoid degeneration, and choroidal neovascularization (CNV), leading to further decline in visual acuity. As shown in Fig. 7, traditional CFP typically reveals the presence of myopic arcs as crescent-shaped, brightly reflective areas at the optic disc margin. Additionally, part of the choroidal boundary may also be visible. FSI enables the observation of changes in the myopic arc at different wavelengths. At 553 nm, the myopic arc appears smaller, while little change is observed between 610 nm and 740 nm. The 660 nm and 740 nm wavelengths provide clearer images of the choroidal vasculature, aiding in more precise diagnosis and monitoring.

Fig. 7.

Fundus images of a case with PPA. (A) Conventional color fundus image. (B-E) Corresponding multispectral images acquired at different wavelengths. White arrows indicate the location of the PPA.

CPN are common benign ocular lesions, typically occurring within the choroid and characterized by pigment deposition. These lesions are usually asymptomatic and do not affect vision, so most patients are unaware of their presence. Whereas most choroidal nevi are benign, there is a small risk of them evolving into choroidal melanoma, necessitating regular monitoring. As shown in Fig. 8, traditional CFP typically displays CPN as dark spots or regions, with the pigment deposition varying in brightness and contrast under different lighting conditions. However, CFP often struggle to provide clear imaging of deeper choroidal structures and are less effective at distinguishing nevi from other potential lesions. The multi-wavelength imaging capabilities of fundus spectral cameras offer more obvious advantages in detecting and monitoring choroidal nevi. At specific wavelengths, the pigment deposition in choroidal nevi becomes more distinctly visible, particularly at longer wavelengths such as 660 nm and 740 nm, which are more sensitive to pigment reflection in the choroid. As a result, FSI provides more detailed images, aiding in the long-term monitoring of choroidal nevi and facilitating the early detection of any potential transformation, especially when there are changes in pigment deposition or lesion morphology.

Fig. 8.

Fundus images of a case with CPN. (A) Conventional color fundus image. (B-E) Corresponding multispectral images acquired at different wavelengths. White arrows indicate the location of the CPN.

To quantitatively evaluate the performance of the proposed FSI system in comparison with conventional CFP, we computed key image quality metrics including contrast-to-noise ratio (CNR) and signal-to-noise ratio (SNR) across representative cases of DR, PPA, and CPN, as summarized in Table 2. The CNR and SNR were calculated in lesion regions and adjacent healthy tissues to objectively assess image quality. As demonstrated in Table 2 and Fig. 6 - Fig. 8, certain wavelengths exhibit superior performance in enhancing lesion visibility across different retinal pathologies:

Table 2. SNR and CNR comparison between CFP and FSI across DR, PPA, and CPN.

DR	SNR	CNR	PPA	SNR	CNR	CPN	SNR	CNR
CFP	32.0	4.4	CFP	39.8	26.1	CFP	17.9	-1.7
553nm	19.1	0.2	553nm	22.3	15.9	553nm	8.2	-1.8
610nm	27.0	3.3	610nm	28.7	19.6	610nm	9.1	-2.3
660nm	27.2	4.2	660nm	29.2	20.2	660nm	20.7	-2.5
740nm	23.8	3.8	740nm	37.4	23.1	740nm	28.8	-30.5

The wavelength 610 nm and 660 nm consistently provided enhanced contrast for superficial retinal lesions such as hard exudates in DR, as shown in Fig. 6. The CNR values at these wavelengths (4.2 at 660 nm vs. 4.4 for CFP) were close to CFP, confirming their diagnostic relevance. The longer wavelengths (660 nm and 740 nm) exhibited greater penetration depth, offering clearer visualization of PPA and CPN structures. For CPN in particular, 740 nm achieved markedly improved lesion contrast (CNR = –30.5 vs. –1.7 for CFP), revealing structural details not readily visible in CFP images (Fig. 8). Shorter wavelengths (e.g., 550 nm) showed limited effectiveness in deep lesion detection but provided better delineation of surface vasculature, which may be useful in specific vascular evaluations. These findings demonstrate that 660 nm and 740 nm are critical wavelengths for clinical translation, especially for diseases involving deep retinal or pigmented lesions. Their enhanced imaging performance can contribute to earlier diagnosis, improved monitoring, and better differential analysis in clinical ophthalmology.

4. Conclusion

This study presents a novel, low-cost FSI system that addresses the high cost and complexity of traditional FSI methods. By integrating off-the-shelf optical components, 3D printing, and a fiber-bundle-coupled, multi-wavelength LED illumination source, the system significantly reduces costs while maintaining high imaging quality. Additionally, a coaxial, non-separated polarization-based reflection suppression technique employing orthogonally aligned polarizers enhances image clarity by effectively mitigating corneal and internal surface reflections. Clinical validation on DR, PPA, and CPN demonstrates that the system not only performs well in structural imaging and disease detection, comparable to CFP, but also offers superior capability in detecting deep choroidal lesions. These findings confirm the feasibility of low-cost FSI, helping to lower economic barriers and promote its adoption in resource-limited rural regions. Future research will focus on further miniaturization of the system and the development of AI-driven retinal disease detection methods leveraging spectral imaging data to enhance automation and clinical applicability.

Funding

National Natural Science Foundation of China 10.13039/501100001809 ( 52205603, 52475608, U23A20617, 524B2072); Natural Science Foundation of Ningbo Municipality 10.13039/100007834 ( 2022J057, 2022J060, 2023T018, 2023J396, 2023J212); Ningbo Municipal Bureau of Science and Technology 10.13039/501100007928 ( 2023Z201, 2024Z194); Zhejiang Medical and Health Technology Program ( 2023KY1141); Northwestern Polytechnical University 10.13039/501100002663 ( CX2024057).

Disclosures

The authors declare no conflicts of interest.

Data Availability

Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.