Fast Mid‐Infrared Spectral Probe Decisions Match H&E Stain Results for Keratinocytic Carcinoma

ABSTRACT

We designed a handheld and fast mid‐infrared fiber‐optic spectral probe using a quantum cascade laser (QCL) and a reduced range of wavelengths, to see if keratinocytic carcinoma (KC) could be distinguished from adjacent nonmalignant tissue using discarded skin tissues from Mohs surgery. This study employed two adjacent frozen sections of discarded tissue: one was stained with H&E (the gold standard for skin cancer diagnosis) to identify the location of cancer by a pathologist, while the other was left unstained for mid‐infrared spectral probing on and off cancer as guided by the adjacent H&E stain. A total of 346 spectra from 18 consenting patients were collected during Mohs surgery. After adding a dehumidifier, an accuracy of 95% was obtained on a case sample basis. It will be worthwhile to assess the probe's utility at the surface of live human skin (study approved by the Advarra Institutional Review Board [PRO00044823]).

A handheld and fast mid‐infrared fiber‐optic spectral probe is touched to tissue recording a mid‐infrared spectrum that can distinguish keratinocytic cancer from normal skin.

Abbreviations

ATR

attenuated total reflectance

BCC

basal cell carcinoma

Dx

diagnosis

FN

false negative

FP

false positive

FS

Frozen Section

FTIR

Fourier transform infrared (spectroscopy)

H&E

hematoxylin and eosin

KC

keratinocytic carcinoma

MeCdTe

mercury‐cadmium‐thallium infrared detector

NNB

number of biopsies needed ratio

NT

nontumor

QCL

quantum cascade laser

SCC

squamous cell carcinoma

TN

true negative

TP

true positive

1. Introduction

Skin cancer is diagnosed with a combination of physical examination by visual inspection and biopsy of suspicious lesions—the latter being the current gold standard for diagnosis. Biopsies typically involve sectioning of the tissue, staining with Hematoxylin and Eosin [1] (H&E), and examination under the optical microscope by a pathologist. Routine histology practice typically has biopsies sent to a lab for paraffin‐embedded sections with H&E staining, and results can take days to weeks. On the other hand, frozen sections are stained in‐house for more rapid histological results, such as with Mohs surgery examination of margins. The primary aim of this study is to show that IR spectral decisions agree with H&E stains regarding medical diagnostic practice.

There is demand for more objective and noninvasive examination methods that may support clinicians in their decision whether to biopsy [2, 3, 4, 5] or whether there is cancer at the Mohs surgeon's margin [6, 7, 8]. Several spectroscopic approaches [3, 9, 10] have been investigated such as, optical coherence tomography [11], reflectance confocal microscopy [12], multiphoton microscopy [13], electrical impedance spectroscopy [14, 15], electron scattering spectroscopy [16, 17], Raman spectrsocopy [18], and Fourier transform infrared fiber‐optical evanescent wave spectroscopy [19, 20, 21, 22, 23] for skin cancer. However, high costs, steep learning curves, high irradiance, and/or low specificities impair clinical application [24, 25]. Note that Fourier transform mid‐infrared (FTIR) spectroscopic studies are distinguished by revealing molecular vibrational “fingerprints” with changes that differentiate normal from neoplastic tissue [21, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35], including skin cancer [36]. Noting the success of mid‐infrared (mid‐IR) studies, a fast, noninvasive, mid‐IR probe was designed, but it has a reduced spectral range and will probably use an increased spectral stepsize (relative to FTIR) in order to be fast enough for use in practice.

The incidence and prevalence of keratinocytic carcinoma (KC), including basal cell carcinoma (BCC) and squamous cell carcinoma (SCC), exceeds all other cancers worldwide [37]. Additionally, KC incidence is increasing, with BCC accruing twice as frequently as SCC [8, 38, 39, 40]. However, more than 4400 patients die each year from SCC. The standard approach to skin lesions suspicious for KC is visual inspection, followed by a biopsy. One study [41] found that 75.8% of biopsies were ordered by dermatology physicians leaving a nonnegligible fraction to others (primary care physicians, physician assistants, nurse practitioners, and residents). A variety of studies note variability in the accuracy of this process [41, 42, 43, 44]. Diagnostic accuracy is often given as the number needed for biopsy ratio (NNB) [45] which divides the total number of biopsies by the number of biopsied cancers. It is correlated with specificity (not sensitivity) [45], so a proposed device needs to display both high sensitivity and specificity regarding skin cancer. NNB has been reported [42] as 3.4 and 21.4 for any skin cancer and melanoma, respectively. “Underuse of biopsies may promote misdiagnosis, and overuse will increase cost and morbidity” [42]. Statistics on “clinical suspicion” improving (reducing) NNB [46] suggest to us that a probe might help. For the patient, there is the potential for psychosocial distress due to the unwanted biopsy‐related scars and aesthetics, the lack of timely communication of biopsy results, and the possibility of additional surgery. As of 2014, in settings of the primary care and dermatologist offices, the estimated annual cost of KCs was $4.8 billion, with an additional $3.3 billion for melanoma [47]. A device to speed feedback and reduce the overuse of biopsies would help to increase the quality and contain the rising cost of healthcare. This first pilot study of the mid‐IR probe attempts to differentiate KCs from adjacent nonmalignant tissue using tissue from patients with biopsy‐proven KC undergoing treatment at the time of Mohs surgery. Distinguishing KC from nontumor is not a pressing problem in dermatology, so agreement with H&E diagnosis is a necessary but insufficient criterion for further use by dermatologists and primary care professionals in distinguishing various types of skin cancers and other skin processes. Success in this study warrants future studies with live human skin and differentiation between different types of cancer and other skin conditions.

2. Methods and Experimental

The study called for ex vivo mid‐IR spectroscopic probing of fresh frozen tissue from patients with residual, biopsy‐proven KC. We conducted the pilot study at the Center for Surgical Dermatology (Westerville, OH). All study documents received approval from the Advarra Institutional Review Board (Columbia, MD). Eighteen patients were enrolled and consented—12 male and six female—ranging from 61 to 90 years of age (mean 75 years) on the day of Mohs surgery for either BCC and/or SCC as detailed in Table 1. All patients who were enrolled in the pilot clinical trial fulfilled the following inclusion criteria: men or women of any ethnic group, age 18 and above, biopsy‐proven BCC or SCC of the skin with clinical evidence of residual tumor post‐biopsy.

TABLE 1.

Case summary of pilot study.

Case #	Clinical Dx	Frozen section Dx	NT spectra per case	NT wrong per case (FP)	Tumor spectra per case	Tumor wrong per case (FN)	${\overline{d}}_{\mathit{NT}}$	${\overline{d}}_{\mathit{\text{Tumor}}}$
1	BCC	BCC	6	3	6	3	−0.05 ± 0.29	+0.32 ± 0.76
2	BCC	BCC	7	3	6	2	+0.32 ± 2.29*	+0.15 ± 0.95
3	BCC	BCC	17	9	7	1	+0.46 ± 1.08*	+1.24 ± 0.98
4	SCC	SCC	9	3	7	0	−1.99 ± 3.15	+1.81 ± 0.96
5	BCC	NTS	9	0			−1.10 ± 1.40
6	SCC	SCC	14	4	10	0	−0.74 ± 1.75	+1.55 ± 1.09
7	BCC	NTS	2	0			−2.98 ± 0.46
Subtotal			64	22	36	6
Dehumidifier added
8	BCC	BCC	7	0	6	0	−0.34 ± 2.61	+1.84 ± 1.54
9	BCC	BCC	10	0	5	0	−1.09 ± 0.29	+1.18 ± 0.26
10	SCC	SCC	10	7	10	1	+0.57 ± 1.19*	+0.92 ± 0.45
11	SCC	NTS	48	1			−1.65 ± 0.80
12	SCC	SCC	10	0	5	0	−2.30 ± 1.22	+1.65 ± 0.75
13	BCC	NTS	10	1	20	1	−0.74 ± 0.77	+1.82 ± 1.03
14	BCC	BCC			10	0		+2.24 ± 0.71
15	BCC	BCC	6	0	24	0	−1.84 ± 1.91	+2.06 ± 0.80
16	BCC	BCC	6	2	6	1	−0.02 ± 1.08	+0.70 ± 0.50
17	SCC	SCC	19	4	13	2	−0.62 ± 0.86	+1.40 ± 1.06
18	BCC, SCC	BCC, SCC			21	1		+1.52 ± 0.91
Subtotal			126	15	120	6
Total			190	37	156	12

Note: Clinical and frozen section diagnosis (Dx) is given in the second and third columns. The numbers of nontumor (NT), keratinacytic cancer [Tumor, basal cell carcinoma (BCC) and/or squamous cell carcinoma (SCC)] spectra by case along with IR Medtek probe misclassifications of spectra using a Training with all spectra are given in columns 4–7. Cases 1–7 had evidence of water vapor problems, while Cases 8–18 had the addition of a dehumidifier. Wrong decisions are indicated with an asterisk.

The Mohs surgery followed standard procedures, and the Mohs surgeon considered all discarded tissue submitted for study to be of no clinical value to the patient. The experimental procedure is illustrated in Figure 1. The tissue included remnants of the deepest portion of the frozen section (FS) block, bulk tissue from the surface of the tumor, and normal skin from the Burrow's triangles taken at the time of wound closure. The fresh tissue specimens were collected and quickly frozen in a cryostat (Figure 1A,B). An FS of ~5 μm thickness was stained with hematoxylin and eosin (H&E) for imaging elsewhere at high microscope magnification in case this was needed (Figure 1C bottom). Then two adjacent FSs were obtained: (1) a ~5 μm FS was stained with H&E and examined by a pathologist on site who offered guidance and marked appropriate areas of tumor and nontumor on the slide (Figure 1C), and (2) a ~20–40 μm thick probe slice was unstained and mounted (Figure 1D) on a flexible, semi‐transparent silicone slide for mid‐IR probing of tumor and nontumor regions (unblinded). The softness of this slide allowed the probe to be pushed into the tissue for better IR spectra. The marked H&E stain slide and the thick probe slice were overlaid and aligned so that the pathologist's markings accurately guided the placement of the spectral probe on the thick unstained sample.

FIGURE 1.

Pilot clinical study design: (A) We obtained discarded tissue from a patient with a biopsy proven KC who consented to the use of discarded tissue from Mohs surgery. (B) The tissue was collected and frozen, and a microtome was used to create adjacent ~5 μm and ~20 μm thick slices. (C) The pathologist examined a 5 μm H&E stained FS and inked the locations of cancer and noncancer on the slide. (D) A 20 μm thick, unstained, FS was mounted on a silicone slide. The H&E‐stained FS provided a guide for probing, with the collection of multiple mid‐IR spectra recorded on cancer and noncancer regions.

The foundation of the probe is attenuated total reflectance (ATR) IR spectroscopy. The fiber optic‐based device consists of a mid‐IR source [tunable quantum cascade laser (QCL) from Block Engineering] that was programmed to step from 1340 to 1870 cm⁻¹ in steps of 10 cm⁻¹, i.e., a range including protein amide I and II bands and lipid spectral features. This range and step were obtained by analyzing full range FTIR imaging spectra of a frozen section study on SKH1 murine skin cancer [48]. The full range results were compared to that obtained with a reduced range (matching the QCL) and increased step size (to speed measurement). The chosen range and step allow for rapid accumulation of spectra and analysis with minimal sacrifice of sensitivity and specificity. The QCL laser pulses travel through a fiber optic cable to a fiber‐loop probe sensor [49, 50] (Art Photonics) as shown in Figure 2 (see also photo at bottom of Figure 1D), where some of the IR travels a few wavelengths outside of the fiber‐loop probe as an evanescent wave. The IR and evanescent components travel out of the fiber‐loop through a fiber optic cable to a fast IR detector (Vigo thermal MeCdTe) and digital oscilloscope (Picoscope 4224) which is controlled and operated by a proprietary MATLAB program. After a background spectrum is recorded with the fiber‐loop in air, the fiber‐loop is touched to skin which absorbs IR from the evanescent wave yielding a sample spectrum. The negative log of the ratio of the spectrum signal (touching skin) to the background (air) gives an absorption spectrum of skin tissue. Not knowing the required signal‐to‐noise for cancer discrimination, the mid‐IR probe was programmed to record a single fast spectral scan in ~6 s, but these were too noisy. We rewrote the program to average 10 such fast scans, i.e., each spectrum collected in this work (Table 1) is an average of 10 fast scans requiring a minute to be recorded. Such spectra were recorded at different positions within the regions designated by the pathologist. A total of 16 nontumor and 15 KC specimen sets were collected from 18 consenting patients during Mohs surgery producing 190 nontumor and 156 KC spectra (Table 1). Since the first seven cases showed evidence of strong water vapor lines, a dehumidifier was added to the probe/laser system for Cases 8–18. This reduced the water vapor signal and kept it much more constant, allowing for better subtraction and improved results. The 346 spectra constituted a library of predictors, and the corresponding pathologist assessments provided the response variables for Training machine learning decision equations using the MATLAB programming environment. Many popular models produced overtrained results, so a linear SVM model (MATLAB's “fitcsvm” function) was chosen to represent these results and minimize overfitting, i.e., high accuracy on Training, but low accuracy on new predictions and a failure to generalize. It was critical to not have spectra from the same case in both Training and Testing, so partitioning was done on a case basis [51, 52, 53]. The linear and scaled SVM decision equation is

$$d_k=b+⟨\frac{\mathit{Tes}{t}_{k,j}-{\overline{\mathit{\text{Train}}}}_j}{{{\sigma }_{\mathit{\text{Train}}}}_j}|{\beta }_j⟩$$ (1)

where $j$ is an index for spectral steps, $k$ identifies the test spectrum such as $\mathit{Tes}{t}_{k,j}$. ${\overline{\mathit{\text{Train}}}}_j$ and ${{\sigma }_{\mathit{\text{Train}}}}_j$ are the mean and standard deviation of the Training set which importantly allow scaling (standardization option in SVM). The SVM program determines optimized values for the bias offset $b$ and the beta spectrum ${\beta }_j$ which has the form of an IR spectrum. Values of $d_k$ greater than zero predict KC, and less than or equal to zero predict nontumor. Once obtained, the decision equation itself can be extracted and used independently of MATLAB to classify new spectra. Note also that the linear SVM model can use ${\beta }_j$ and average class spectra to determine the contribution of each spectral wavenumber to the decision equation value on average. Preprocessing was useful and included both subtraction of a baseline offset using the average absorbance between 1770 and 1870 cm⁻¹ and normalization of the resulting spectra to control for different pressures when applying the probe to skin tissue. Our programs tracked the linear SVM predictions relative to the pathologist's assessment (supported by previous biopsy‐proven KC). True positives (TP) counted KCs that were correctly predicted, and true negatives (TN) were a count of correct nontumor predictions. False negatives (FN) were a count of true KC spectra that were incorrectly predicted as nontumor, while false positives (FP) counted true nontumor spectra that were incorrectly counted as KC. All these quantities can be extracted from the information in Table 1. In this context, accuracy is $\left(\mathit{TP}+\mathit{TN}\right)/\left(\mathit{TP}+\mathit{TN}+\mathit{FP}+\mathit{FN}\right)$, sensitivity is $\mathit{TP}/\left(\mathit{TP}+\mathit{FN}\right)$, specificity is $\mathit{TN}/\left(\mathit{TN}+\mathit{FP}\right)$.

FIGURE 2.

Schematic of the fast, mid‐IR spectral probe.

3. Results and Discussion

Each patient presented for Mohs surgery with a history of biopsy‐proven BCC or SCC, and one case had both. All but one of the tumors arose from the head and neck region. Since 31 KC or nontumor samples from 18 consenting patients are quite limited, multiple spectra were recorded for each tumor or nontumor case sample from both different specimens of the case and nearby locations on a particular specimen. The averages of 156 KC (red, labeled Tumor) and 190 nontumor (green) spectra after preconditioning are given in Figure 3 with error bars representing ± one standard deviation about the mean. The difference, i.e., nontumor minus KC, is shown with a black trace. Notice that the average differences are considerably bigger than the errors in the mean differences, so spectral scanning conditions are working. Before conducting machine learning analysis, one can observe that there are distinguishing features. Also, the cyan arrows indicate the strongest water vapor lines. Once the dehumidifier was added for cases 8–18, the spectra were cleaner and more specific.

FIGURE 3.

Differentiating basal cell carcinoma and squamous cell carcinoma based on IR Medtek probe spectra. The nontumor trace is green, while the keratinacytic trace is red (Tumor is BCC and SCC). The difference of nontumor and tumor is black. The error bars are ± one standard deviation of the mean. The cyan arrows indicate strong water vapor transitions.

The full spectral library of 346 spectra was used to train a spectral decision equation with (i) a cost ratio of 2.1 for misclassification of KC relative to nontumor and (ii) not allowing spectral data from a case to be in both Training and Testing. A spectrum‐based tally of correct and incorrect predictions on a case basis is shown in Table 1. The spectral data were divided into two sets (cohorts) in Table 1: Cases 1–7 had a considerable water vapor problem (see the cyan arrows in Figure 3), showing a spectral accuracy of 72%, while Cases 8–18 with the dehumidifier had a spectral accuracy of 92%. Clearly, the dehumidifier is important.

Using linear SVM on the whole spectral library (Cases 1–18) gave an accuracy of 86% (overall error of 14%) with a spectral sensitivity of 92%, a spectral specificity of 81%, and a 5‐fold Cross‐validation error of 23% ± 3%. Results can be extracted from Table 1 (TP = 144, TN = 153, FN = 12, FP = 37). To get a better handle on errors, the spectral library was divided randomly by case using 2, 3, 4, and 5‐fold partitions with 20 random iterations. The Training and Cross‐validation errors smoothly extrapolate [vs 1/(number of folds)] to the full library values; however the Testing errors are high (one run found 47% ± 7%, 45% ± 9%, 47% ± 9%, and 39% ± 14% for the 2, 3, 4, and 5‐fold partitions, respectively). When the model was changed to the popular nonlinear radial basis function (RBF) SVM, the Training error dropped to 0.0%, but the 5‐fold Cross‐validation error increased to 45% ± 2% (linear SVM was 23% ± 3%). The Testing errors were 57% ± 5%, 61% ± 11%, 62% ± 14%, and 63% ± 13% using 2, 3, 4, and 5‐fold partitions, respectively (linear SVM was 45% on average). The linear SVM was the least overtrained model, i.e., the gap between Training and Cross‐validation errors is smallest (9% vs. 45%) and RBF SVM is clearly overtrained. The linear SVM Testing errors are smallest for this model, but still fairly large, suggesting that more cases are needed to close the gap between Training and Testing.

Medical practice prefers examination on a case‐by‐case basis, but the number of cases is small. We have only 18 cases, and not all have both tumor and nontumor tissue samples. If we distinguish “case samples” from cases by counting nontumor and tumor samples separately, then there are 16 nontumor case samples and 15 tumor case samples for a total of 31 case samples. The average of linear SVM decision equation values (${\overline{\mathit{d}}}_{\mathit{NT}}$ or ${\overline{\mathit{d}}}_{\mathit{\text{Tumor}}}$) for all spectrat recorded on each case sample are shown in Table 1 (cost ratio penalty of 2.1 for misclassifying KC relative to nontumor). The uncertainties are the first standard deviation of all spectral values for a case sample. Recall that KC (tumor) has positive decision equation values while nontumor has negative values. Cases 1–7 with excess water vapor have two wrong average decision values out of 12 case samples for a case sample accuracy of 83%. If all of the data is used, then there are three wrong out of 31 case samples for a case sample accuracy of 90%. Using only cases 8–18 with the dehumidifier, there is only one wrong out of 19 for a case sample accuracy of 95% with zero errors on KC samples. Note that the cost ratio for misclassification can be varied to increase the sensitivity of KC detection at the cost of a lower specificity. There are variations with the type of lesion being studied [41, 54, 55, 56] and the imbalance of the number of spectra in each class. The mid‐IR probe shows both high sensitivity and specificity although more case numbers are needed to quantify the errors.

The study had two inherent limitations. The first is that the design of this proof‐of‐concept study does not reflect a real clinical situation. This is revealed by the fact that a pathologist defined the specific tumor and non‐tumor areas on the tissue sections for ex vivo probing, which precludes blinding the prober. Secondly, the sample size is small. However, this number is based on a pre‐study consultation with a statistician that 18 patients would provide sufficient data to demonstrate the probe's accuracy.

4. Conclusions

The mid‐IR region has fundamental vibrations of biomolecules (like protein and lipid) that give bigger and narrower changes in absorption signal than other regions of the electromagnetic spectrum when a cell becomes cancerous. While there has been considerable research using FTIR, there has been limited application to clinical practice due to instrumental complexity, i.e., interferometers with moving optical mirrors and liquid‐nitrogen‐cooled detectors. Using a quantum cascade laser (QCL) as the source, we designed a handheld and fast mid‐IR fiber‐optic spectral probe, which employed a reduced range of wavelengths and an increased step size as compared to FTIR, which is the recognized standard in mid‐IR work. It must also be safe for human use, so it uses only an average of ~4 mW of QCL output which satisfies FDA recommendations [ANSI Z136.1–2014 American National Standard for Safe Use of Lasers, a maximum of 0.1 W/cm² for scans longer than 10 s] assuming a QCL spot size of 4 mm × 2 mm which corresponds to a limit of 8 mW. Water vapor was shown to be an important consideration and the addition of a dehumidifier to the device considerably improved results. This is a long‐standing problem in FTIR with the most common solution being to make the unavoidable water vapor the same in both sample and background. The cohort of Cases 1–7 showed a case sample accuracy of 72%, while the cohort of Cases 8–18 with the dehumidifier showed a case sample accuracy of 95%. Clearly, the dehumidifier is important, and water vapor needs to be controlled.

A statistical approach was possible using linear SVM on the whole library of 346 spectra without allowing spectra from a case to be used in both Training and Testing. A spectral accuracy of 86% was obtained with a 5‐fold Cross‐validation error of 23% ± 3% using a cost ratio of 2.1 for misclassification of KC. The spectral sensitivity was 92% and the spectral specificity was 81%, but improvements can be made by averaging. An analysis on a case sample basis was also accomplished by averaging the decision equation values of all spectra within the set for the case sample. There were three wrong out of 31 case samples for an accuracy of 90% using the whole library, and there was only one wrong out of 19 case samples for an accuracy of 95% using only Cases 8–18 with the dehumidifier. In an attempt to characterize errors with this small number of cases, the spectral library was divided randomly by case using 2, 3, 4, and 5‐fold partitions with 20 random iterations. The Training and Cross‐validation errors show smooth trends toward the above‐mentioned, full‐library values; however the Testing errors ran high (~45%). But the Testing and Cross‐validation errors ran even higher for other models, so linear SVM was the least overtrained model, i.e., the gap between Training and Cross‐validation errors was only 9% for linear SVM, while it was 45% for RBF SVM.

This study was able to show that it is possible for IR probe spectral decisions to agree with H&E stains regarding KC from discarded skin tissues from Mohs surgery with biopsy‐proven KC, i.e., there was 1 wrong out of 19 case samples for the cohort with a dehumidifier. This is an accuracy of ~95%, but more work is needed with a larger number of cases in order to put an error bar on the accuracy. High sensitivity (92% by spectra, 100% by case sample with dehumidifier) is always important, but high specificity (81% by spectra, 89% by case sample with dehumidifier) is correlated with low NNB and is essential for a new probe. High specificity is also noteworthy since sensitivity can be increased at the expense of specificity by adjusting the costs of misclassification or balancing the numbers within the classes. In other words, sensitivities of 95% are readily possible with high specificity near ~80%. We showed that averaging spectral decision equation values within a case sample yields better decisions. At the current spectral probe signal‐to‐noise levels, it is recommended that at least four spectra (each requiring ~1 min) be recorded per suspected lesion or nontumor region, with examination of the average of the decision equation values.

The mid‐infrared fiber‐optic spectral probe is noninvasive, portable, and fast. The cancer was touched directly by a probe in this study which may have applications in cancer resections or excisions; however our ultimate application is to probe live skin in the clinician's office. This study's success in demonstrating agreement with H&E stain results, warrants a new study on live skin (on and off suspected KC lesions) which is a different analytical problem with applications in the clinician's office rather than surgery. Future success will support additional study adding melanoma, nevi, and other nontumor skin conditions to the keratinocytic cancers (basal and squamous cell carcinoma) on live skin. Given success on live skin, the mid‐IR probe has the potential to provide real‐time, quantitative information assisting clinicians in whether to biopsy a suspicious skin lesion or not.

Conflicts of Interest

James V. Coe and Charles Hitchcock were coinventors, with fractional royalty rights to the Ohio State University patents licensed to IR Medtek LLC. Dr. Angela Casey and Dr. Ronald Siegle were the principal investigators for this study and are stock owners. Rebeca Bradley and Maria Vazquez de Vasquez were research scientists employed by IR Medtek LLC.

Bradley R. C., de Vasquez M. G. V., Hitchcock C. L., Casey A. S., Coe J. V., and Siegle R., “Fast Mid‐Infrared Spectral Probe Decisions Match H&E Stain Results for Keratinocytic Carcinoma,” Journal of Biophotonics 19, no. 3 (2026): e202500311, 10.1002/jbio.202500311.

Data Availability Statement

Research spectral library data are not being shared. Readers can contact the corresponding author to see if the data sharing status changes. All patients gave consent for their medical information to be published in print and online, with the understanding that this information may be publicly available. This study and research protocol were approved by the Advarra Institutional Review Board (PRO00044823).