1.1.

Dermoscopy

The practice of dermoscopy has evolved significantly over the past century.¹ Dermatologists use dermoscopes to visualize skin findings and improve diagnostic accuracy; this practice has steadily grown to be widely accepted. Dermoscopy courses, residency training, algorithms, atlases, working groups, and other academic resources are readily available. There is a broad selection of dermoscopes on the market, each with distinct technical specifications. However, it is uncommon to find the technical specifications, such as spectral power distribution (SPD), color rendering index (CRI), or even Kelvin temperature, made available to users of dermoscopic devices, although this information is accessible for lighting in other sectors.

The clinical utility of dermoscopy has been demonstrated in terms of improved diagnostic accuracy.^2,3 Dermoscopy has also been the subject of several high-profile studies on artificial intelligence (AI) and machine learning (ML) approaches to diagnostics.^4,5 An important diagnostic criterion is color, both for pigmented lesions and other skin findings.^1,6 In fact, color itself can be used as a feature for deep learning algorithms used for AI in dermatology.^7–10 Polarized light dermoscopy (technically cross-polarized) has been shown to improve the visibility of colors,^11,12 amongst other features. Color consistency, fidelity, reproducibility, and standardization are needed for both clinical diagnostics as well as computer vision and image processing applications.^13,14

In the past, halogen bulbs were common,¹⁵ whereas today most devices make use of light-emitting diodes (LED). The SPD of halogen (tungsten) generates a continuous distribution of light across the visible spectrum, with most of the energy concentrated in the red and infrared portion of the spectrum, giving older dermoscopes their characteristic “warm” light. The Kelvin temperature of dermoscopes may eventually be incorporated into metadata fields available for DICOM attribute descriptors.¹⁶ Some dermoscopes now integrate multispectral and colored LED, xenon, and UV sources into devices;^17–20 however, most dermoscopes use white LEDs. A lack of uniformity in illuminating chromophores contained in skin could potentially reduce diagnostic accuracy.²¹

1.2.

LED Advances

LED technology itself has grown tremendously in the past two decades, both in terms of science and applications related to medicine²² and imaging. In 2014, the invention of the blue LED won a physics Nobel prize. There is considerable and ongoing effort to improve LEDs efficiency, color rendering, and fidelity.²³ The Nobel committee recognized the challenge of producing blue LEDs, noting its importance to producing white LEDs. Phosphor doping of the blue LED is a common method for producing white light; however, this approach has been shown to produce light that is insufficient for medical lighting in a number of scenarios. In 2000, the surgeon performing the first surgery using LED loupes noted that it was difficult to distinguish arteries from veins.²⁴ The two main issues with phosphor-conversion of LEDs are that energy is lost when photons are converted from shorter to longer (i.e., blue to red), and some of the energy is wasted outside the visible spectrum.²⁵ To make “good” white light, it should match (or generate metamers of) daylight. The chromaticity of the light source is best judged side-by-side with other light sources, and depending on SPD, it may appear as “white” or having a “tint.” Practically, the chromaticity of an LED may not be “white,” although this was the manufacturer’s intention. There is an ongoing effort to improve spectral matching for LEDs;²⁶ however, the issue of colors rendering inconsistently was evident in our experiment.

1.3.

Color Science

Great advances have been achieved since early color theories such as those described by Newton as far back as the 1700s. Both SPD and color reflection can be measured quite precisely; however, despite these ever-advancing technical solutions for color vision, it is not possible to ensure that two individuals perceive the same colors identically, regardless of the accuracy of measurement devices.²⁷

The International Commission on Illumination (CIE) defines color spaces and formulas for measurement and develops reference standards and indexes, among many other functions.^28,29 The CIE defines color rendering as “the effect of an illuminant on the color appearance of objects by conscious or subconscious comparison with the color appearance under a reference illuminant.” The CIE categorizes illuminants into series, having designations for halogen (A), daylight (D) series, fluorescent series (F), and a newer addition of LED series. There have been some efforts to match the SPD of LED more precisely to other color spaces and standard illuminants;^26,30 this research reinforces that it may not be possible to manufacture a LED that perfectly matches a desired SPD, although much progress has been made to reduce spectral mismatching of LED. The goal of reviewing LEDs in dermoscopes is not to determine a perfect white LED for skin illumination, but rather to highlight differences in rendering ability compared with one another and when compared with standard illuminants.

The common CRI is a helpful, but somewhat limited, system for the measurement of color. This is because CRI is known to be prone to manipulation or “gaming” in which metamerism is leveraged to improve rendering values. Further, the values do not always correlate well to visual evaluation, prompting the development of additional fidelity indexes aimed at improving the range in scientific use.³¹

Developed by the Illuminating Engineering Society, TM-30-18 measures fidelity and chroma of light; it was developed as an alternative reference to CRI for increased accuracy.³² This index includes an expanded reference that recently was shown to outperform other evaluation criteria when “naturalness” is the most important aspect.³³

Some research has suggested that color accuracy or the perceptual rendering of colors as they appear to the human eye is not necessarily the crucial aspect for visualization of specific features under dermoscopy;³⁴ however, this research did not compare the degree of differences, nor did it address how two different colors might appear identical when color fidelity is suboptimal. It is feasible that the observation by Avanaki et al. of “less colorful images enhancing visualization of certain details” actually describes ancillary (but important) aspects of feature perception as a function of simultaneous contrast and conspicuity. Simultaneous contrast in both color and greyscale images may reduce feature conspicuity, a well-described phenomenon in the field of radiology.^35,36 A number of publications appear to dispute the claim that perceptual rendering of color is inconsequential in dermoscopy images,^13,37,38 with one notable publication³⁹ actually comparing the features that are perceptible from different dermoscopes and illustrating how some light can completely obfuscate specific findings. Additionally, the ability of ML algorithms to leverage color data^9,10 may rely upon consistency and rendering capability.

Measuring differences in color and limiting the tolerance for these differences can be described with the use of DeltaE2000 ($\mathrm{\Delta }E$): the Euclidean distance between two colors in a given color space. The most recent color difference formula is the CIEDE2000⁴⁰ (Fig. S2 in Supplemental Material). Generally, the perceptible $\mathrm{\Delta }E$ difference between two colors that is noticeable is 1, with opposite colors at 100 (e.g., green and red). An acceptable difference in color, or tolerance, for industrial, commercial, or entertainment industries applications may indeed be wider or more permissive than those in science and medicine. The tolerance of a given application may be in the range of one to seven, with greater perceptible differences having a higher number. No standard tolerance for acceptable color difference has yet been defined for skin imaging; however, it is thought to lie between two and four in the field of dentistry^41,42 as color pertains to matching dental ceramics. The difference in colors perceived in skin is arguably more critical than the color difference in teeth, considering the purpose of matching colors in dentistry is primarily cosmetic in nature, whereas the colors in skin imaging may have diagnostic significance. In the field of skin imaging and skin color matching, various approaches to quantify color have been used,⁴³ with some use of novel facial color contrast measurements.⁴⁴ Spectrophotometers are used for some skin color variation research, which employs pulsed xenon-arc light. In manufacturing and printing, greater degrees of LED differences can reduce the aesthetic appeal of an image or product; in science and medicine, a greater difference may indicate device calibration errors, defective equipment, or in our case, a lack of standards.

The color space chosen to perform the measurements can improve the ability to quantify perceptual differences, reducing the effects of metamerism. The CIELAB color space (or LAB) is a three-dimensional, device-independent model that is intended to be perceptually uniform and describes relative luminance independently from chroma (Fig. 1).

Fig. 1

CIELAB or L*a*b* color space with a graphical representation of $\mathrm{\Delta }E$. A device-independent, three-dimensional color space that is intended to be perceptually uniform relative to human color vision, such that a given numerical change is correlated to a perception of change in color. L*a*b* values are correlated to lightness, chroma, and hue, where L* ranges between 0 (black) and 100 (white), a* describes the red–green axis, and b* describes the blue–yellow axis, akin to the opponent pair of the human visual system. The distance between a pair of colors is reported at the color difference $\mathrm{\Delta }E$. CIELAB and $\mathrm{\Delta }E$ were developed by the Commission Internationale de l’Eclairage (CIE) with the intent of measuring the Euclidean distance within a given color space. Intended to be a difference calculator, $\mathrm{\Delta }E$ of 1 represents the smallest noticeable change in color. $\mathrm{\Delta }E$ is illuminant dependent; a change under differing illuminants illustrates a metamerism index, and two colors that appear the same under differing illuminant sources are considered metamers.

CIELAB measurements can be correlated to skin color and cutaneous findings, decreasing subjectivity in tasks such as classification of Fitzpatrick skin type.⁴⁵ A reference light source of D50 (also developed by the CIE) is commonly used for color matching applications. D50 is a theoretical light source that closely approximates white light and is considered to be an acceptable industry reference standard.

2.1.

Spectral Power Distribution

Spectrometer measurements were taken from (1) Heine (Germany) DELTA10 (halogen), (2) DELTA20T, (3) DELTA30 as well as 3Gen DermLite (California, United States), (4) DL4, and (5) DL4 with Orange Boost. 3Gen’s DL3 and Foto II were also measured (not shown), and the LED SPD matched the other DermLite products.

We measured SPD of three additional sources that may represent cutaneous imaging in real-world settings: Apple iPhone 11 LED flash, xenon-arc (emitted from a Nikon flash), and daylight (southern hemisphere at noon).

Kelvin temperature and CRI were measured for the four dermoscopes and the additional sources. Reference illumination was set to D50 (or CCT 5000 K) for the closest standard approximation to white light.

TM-30 graphical vectors were generated and compared to demonstrate color saturation and chroma fidelity.

2.2.

Reflection Measurements

ColorChecker images were acquired in DNG format, uncorrected (no white balance settings were applied) using an Apple iPhone 11 set to manual with the same exposure settings for four dermoscopes and two references; see Fig. S1 in Supplemental Material. The ambient light was dimmed prior to the serial acquisition of a micro ColorChecker with each dermoscope, at an exposure setting that was sufficient for the varying intensity of the different dermoscopes. The iPhone camera lens was placed directly against the dermoscopes, or a magnetic adapter was used to secure the device to the phone (when available), and the ColorChecker was placed upon a piece of foam to flatten the image field.

ColorChecker images were set to 5000 K, and L*a*b* values were recorded for specific colors. ColorChecker images of three colors (orange–yellow, blue sky, and light skin) were cropped and placed next to each other for qualitative visual analysis.

$\mathrm{\Delta }E$ was measured using Bruce Lindblooom’s calculator⁴⁶ for dermoscopes 3 and 4, representing two of the more commonly used current devices. The formula for DeltaE2000 is provided in Fig. S2 in Supplemental Material.

Dermoscopes 3 and 4 were used to acquire two images of a dermal nevus. The exposure settings were matched, and the images were placed side-by-side for qualitative analysis. The images’ histograms, as well as a single L*a*b* measurement value from the same area, are visible in both images (Fig. 7).