# Identification of Skin Lesions by Snapshot Hyperspectral Imaging

^{1}

^{2}

^{3}

^{4}

^{5}

^{6}

^{*}

## Abstract

**:**

## Simple Summary

## Abstract

## 1. Introduction

^{2}) to extensive involvement of the back, chest, flanks, or lower body. Moreover, these lesions present a wide array of morphologies, from superficial erythematous streaks to deeper epidermal disruptions, some of which may be complicated by infection and bleeding. Given these factors, the automatic localization of lesions by using spatial information alone is not deemed essential. The static nature of the photographic data further diminishes the necessity for employing object-detection algorithms such as YOLOv5 or CNN-based approaches. Clinicians are typically able to identify the location of lesions with relative ease on the basis of surface examination. Most deep-learning CNN algorithms leverage spatial information for feature recognition. However, this approach may not yield definitive results due to the diverse anatomical distribution and morphological heterogeneity of the lesions. Despite these challenges, distinguishing between these three disease types based solely on skin color during clinical examination can be arduous. Consequently, a spectral discriminant method that capitalizes on color space conversion techniques and utilizes the spectral domain is urgently needed. Color space conversion involves translating a color’s representation from one basis to another to ensure that an image, when converted from one color space to another, retains its original appearance as closely as possible. Such an approach could differentiate the extent of invasive lesions on the basis of skin color variations, thus offering a more discernible advantage. The adoption of advanced methods, such as converting from the color space to the spectral domain, represents a noninvasive diagnostic solution that has the potential to achieve high classification accuracy.

## 2. Materials and Methods

#### 2.1. Sample Preparation

#### 2.2. Segmentation Task

#### 2.3. Model HSI-Feature Extraction Task

#### 2.3.1. Spectrometer to XYZ Color Space Conversion

_{spectrometer}) and regulated in the XYZ color gamut space by using Equations (1)–(3) as follows:

#### 2.3.2. Non-Linear XYZ Correction

^{2}) value. The R

^{2}value for the third-order polynomial regression was 0.99997, which was higher than that of linear and quadratic polynomial regressions. Consequently, the camera’s nonlinear response can be effectively corrected through a third-order polynomial model. The correction factor for the nonlinear response is denoted as V

_{non-linear}, as shown in Equation (5).

_{dark}, as defined in Equation (6) as follows:

_{color}, as delineated in Equation (7).

_{color}served as the foundational element for the application of non-linear response correction, through multiplication with V

_{non-linear}. The resultant product was then normalized to the third order to preclude the possibility of over-correction. Subsequently, the constant V

_{dark}was incorporated, leading to the formation of the variable matrix V. This process is systematically delineated in Equation (8). The variable matrix V was then reintroduced to derive the correction matrix C, which constituted the final corrective framework for the system.

#### 2.3.3. Correction Matrix C and the Calibration Camera and Spectrometer

_{camera}, and [XYZ

_{spectrometer}] is the matrix created by X, Y, and Z components obtained from the spectrometer.

_{camera}matrix to the V matrix, resulting in the corrected values of X, Y, and Z (denoted as XYZ

_{correction}), as delineated in Equation (10). Given that the wavelength band utilized in this study falls within the visible light spectrum (380–780 nm), the outcome of this correction can be characterized in terms of chromatic aberration.

_{camera}, and [C] is the correction matrix obtained from Equation (9).

#### 2.3.4. Principal Component Analysis of Reflectance Spectrum

_{correction}), which are acquired subsequent to camera calibration, into spectral data, principal component analysis (PCA) was employed on the reflectance spectrum data (R

_{spectrometer}) of the standard 24-color chart. This analysis yielded the principal components and their associated scores (including eigenvalues) of the reflectance spectrum for the 24-color checker chart.

_{spectrometer}) demonstrated that the initial 12 principal components (EVs) accounted for 99.99% of the total variance within the data (Figure 6) and their density details in the spectrum dataset (Figure S2). The cumulative explained variance plot displays the total variance captured by successive principal components. When the plot reaches its 12th point at 99.99%, it signifies that the first 12 components collectively explain 99.99% of the dataset’s variance. Dimensionality reduction was achieved by leveraging these 12 principal components, thereby extracting the principal component scores. The subsequent multivariate regression analysis utilized these scores, with V

_{color}serving as a predictor variable. V

_{color}was chosen due to its comprehensive delineation of the correlations among X, Y, and Z values. This analytical approach facilitated the derivation of the transformation matrix M, which bridges the measurement gap between the camera and the spectrometer, as explicated in Equation (11).

_{color}is defined as Equation (7) with X, Y, and Z components derived from XYZ

_{spectrometer}; [score] is the principal component scores obtained from 12 sets of principal components for dimensionality reduction of R

_{spectrometer}; and R

_{spectrometer}is the spectral signal obtained from the spectrometer of 24-color patches ranging from 380 nm to 780 nm with a 1 nm resolution.

#### 2.3.5. The Hyperspectrum

_{spectrum}, as indicated in Equation (12):

_{color}is defined as Equation (7) with X, Y, and Z components derived from XYZ

_{correction}; [EV] is the 12 principal components obtained from the dimensionality reduction of R

_{spectrometer}; R

_{spectrometer}is the spectral signal obtained from the spectrometer of 24-color patches ranging from 380 nm to 780 nm with a resolution of 1 nm; and [M] is the transformation matrix obtained from Equation (11).

#### 2.4. Classification Task

#### 2.5. Training Strategy and Performance Evaluation

## 3. Results and Discussions

#### 3.1. Evaluate the Performance of the HSI Model

_{spectrometer}) of the 24-color card. The root-mean-square error (RMSE) was computed to quantify the discrepancy between the two spectra. The calculated average RMSE was 0.0525 (Table S1). Color blocks 13–18 exemplify the variance between the simulated spectrum and the measured spectrum of the 24-color card, as illustrated in Figure 8. These six color blocks were selected for investigation due to their representation of six filters commonly used in chromatic correction. The comparison between the six typical colors demonstrates the correlation between the measured and simulated spectra, indicating that the spectral reproduction algorithm has successfully produced simulation results approximating the measured spectra. Consequently, it can be inferred that the accuracy of the hyperspectral model in converting from real data is notably effective. The simulated spectrum was transformed into the L*a*b* color space for a comparative analysis using the CIEDE 2000 color difference metric (Section S1.5 in Supplementary Materials). This comparison yielded an average color difference of 0.28, as depicted in Table 1. The color difference of the 24-color blocks was illustrated on the CIE 1931 chromaticity diagram for a 2° standard observer, as shown in Figure 9. In this diagram, a black line connects two central points, with “red” indicating simulated colors and “green” representing measured colors. The minimal discrepancy between the simulated and actual colors demonstrated the effectiveness of the HSI algorithm in calibrating the correlation between the camera and the spectrometer.

#### 3.2. Segmentation Model Results

#### 3.3. Classification Task Results

#### 3.4. Comparison with Other Existing Studies

## 4. Conclusions

## Supplementary Materials

_{spectrometer}; Table S1: Root Mean Square Error of S

_{spectrum}and R

_{spectrometer}.

## Author Contributions

## Funding

## Institutional Review Board Statement

## Informed Consent Statement

## Data Availability Statement

## Conflicts of Interest

## References

- Querfeld, C.; Zain, J.; Rosen, S.T. Primary cutaneous T-Cell lymphomas: Mycosis fungoides and Sezary syndrome. T-Cell NK-Cell Lymphomas Biol. Nov. Ther.
**2019**, 176, 225–248. [Google Scholar] - Stoll, J.R.; Willner, J.; Oh, Y.; Pulitzer, M.; Moskowitz, A.; Horwitz, S.; Myskowski, P.; Noor, S.J. Primary cutaneous T-cell lymphomas other than mycosis fungoides and Sézary syndrome. Part I: Clinical and histologic features and diagnosis. J. Am. Acad. Dermatol.
**2021**, 85, 1073–1090. [Google Scholar] [CrossRef] [PubMed] - García-Díaz, N.; Piris, M.Á.; Ortiz-Romero, P.L.; Vaqué, J.P. Mycosis fungoides and sézary syndrome: An integrative review of the pathophysiology, molecular drivers, and targeted therapy. Cancers
**2021**, 13, 1931. [Google Scholar] [CrossRef] [PubMed] - Dobos, G.; Lazaridou, I.; de Masson, A. Mycosis Fungoides and Sézary Syndrome: Microenvironment and Cancer Progression. Cancers
**2023**, 15, 746. [Google Scholar] [CrossRef] [PubMed] - Hsiao, Y.-P.; Wang, H.-C.; Chen, S.-H.; Tsai, C.-H.; Yang, J.-H. Identified early stage mycosis fungoides from psoriasis and atopic dermatitis using non-invasive color contrast enhancement by LEDs lighting. Opt. Quantum Electron.
**2015**, 47, 1599–1611. [Google Scholar] [CrossRef] - Fujimura, T.; Okuyama, R.; Ito, Y.; Aiba, S. Profiles of Foxp3+ regulatory T cells in eczematous dermatitis, psoriasis vulgaris and mycosis fungoides. Br. J. Dermatol.
**2008**, 158, 1256–1263. [Google Scholar] [CrossRef] [PubMed] - Nashan, D.; Faulhaber, D.; Ständer, S.; Luger, T.; Stadler, R. Mycosis fungoides: A dermatological masquerader. Br. J. Dermatol.
**2007**, 156, 1–10. [Google Scholar] [CrossRef] [PubMed] - Cerroni, L. Mycosis fungoides-clinical and histopathologic features, differential diagnosis, and treatment. Semin. Cutan. Med. Surg.
**2018**, 37, 2–10. [Google Scholar] [CrossRef] - Shamim, H.; Johnson, E.F.; Gibson, L.E.; Comfere, N. Mycosis fungoides with spongiosis: A potential diagnostic pitfall. J. Cutan. Pathol.
**2019**, 46, 645–652. [Google Scholar] [CrossRef] - van Santen, S.; Jansen, P.M.; Quint, K.D.; Vermeer, M.H.; Willemze, R. Plaque stage folliculotropic mycosis fungoides: Histopathologic features and prognostic factors in a series of 40 patients. J. Cutan. Pathol.
**2020**, 47, 241–250. [Google Scholar] [CrossRef] - Hodak, E.; Geskin, L.; Guenova, E.; Ortiz-Romero, P.L.; Willemze, R.; Zheng, J.; Cowan, R.; Foss, F.; Mangas, C.; Querfeld, C. Real-Life barriers to diagnosis of early mycosis fungoides: An international expert panel discussion. Am. J. Clin. Dermatol.
**2023**, 24, 5–14. [Google Scholar] [CrossRef] [PubMed] - Roelens, M.; de Masson, A.; Ram-Wolff, C.; Maki, G.; Cayuela, J.M.; Marie-Cardine, A.; Bensussan, A.; Toubert, A.; Bagot, M.; Moins-Teisserenc, H. Revisiting the initial diagnosis and blood staging of mycosis fungoides and Sézary syndrome with the KIR3DL2 marker. Br. J. Dermatol.
**2020**, 182, 1415–1422. [Google Scholar] [CrossRef] [PubMed] - Horna, P.; Wang, S.A.; Wolniak, K.L.; Psarra, K.; Almeida, J.; Illingworth, A.J.; Johansson, U.; Craig, F.E.; Torres, R. Flow cytometric evaluation of peripheral blood for suspected Sézary syndrome or mycosis fungoides: International guidelines for assay characteristics. Cytom. Part B Clin. Cytom.
**2021**, 100, 142–155. [Google Scholar] [CrossRef] [PubMed] - Lyapichev, K.A.; Bah, I.; Huen, A.; Duvic, M.; Routbort, M.J.; Wang, W.; Jorgensen, J.L.; Medeiros, L.J.; Vega, F.; Craig, F.E. Determination of immunophenotypic aberrancies provides better assessment of peripheral blood involvement by mycosis fungoides/Sézary syndrome than quantification of CD26− or CD7− CD4+ T-cells. Cytom. Part B Clin. Cytom.
**2021**, 100, 183–191. [Google Scholar] [CrossRef] [PubMed] - Allen, P.B.; McCook-Veal, A.A.; Switchenko, J.M.; Paulino, D.M.; Niyogusaba, T.; Baird, K.M.; Tarabadkar, E.S.; Lechowicz, M.J. Staging lymph nodes and blood at diagnosis in mycosis fungoides identifies patients at increased risk of progression to advanced stage: A retrospective cohort study. Cancer
**2023**, 129, 541–550. [Google Scholar] [CrossRef] [PubMed] - Guitart, J. Sézary syndrome and mycosis fungoides flow cytometric evaluation: The clinicians’ perspective. Cytom. Part B Clin. Cytom.
**2021**, 100, 129–131. [Google Scholar] [CrossRef] [PubMed] - Vermeer, M.H.; Moins-Teisserenc, H.; Bagot, M.; Quaglino, P.; Whittaker, S. Flow cytometry for the assessment of blood tumour burden in cutaneous T-cell lymphoma: Towards a standardized approach. Br. J. Dermatol.
**2022**, 187, 21–28. [Google Scholar] [CrossRef] [PubMed] - Illingworth, A.; Johansson, U.; Huang, S.; Horna, P.; Wang, S.A.; Almeida, J.; Wolniak, K.L.; Psarra, K.; Torres, R.; Craig, F.E. International guidelines for the flow cytometric evaluation of peripheral blood for suspected Sézary syndrome or mycosis fungoides: Assay development/optimization, validation, and ongoing quality monitors. Cytom. Part B Clin. Cytom.
**2021**, 100, 156–182. [Google Scholar] [CrossRef] - Hodak, E.; Sherman, S.; Papadavid, E.; Bagot, M.; Querfeld, C.; Quaglino, P.; Prince, H.; Ortiz-Romero, P.; Stadler, R.; Knobler, R. Should we be imaging lymph nodes at initial diagnosis of early-stage mycosis fungoides? Results from the PROspective Cutaneous Lymphoma International Prognostic Index (PROCLIPI) international study. Br. J. Dermatol.
**2021**, 184, 524–531. [Google Scholar] [CrossRef] - Calvani, J.; De Masson, A.; de Margerie-Mellon, C.; de Kerviler, É.; Ram-Wolff, C.; Gruber, A.; Meignin, V.; Brice, P.; Sadoux, A.; Mourah, S. Image-guided lymph node core-needle biopsy predicts survival in mycosis fungoides and Sézary syndrome. Br. J. Dermatol.
**2021**, 185, 419–427. [Google Scholar] [CrossRef] - Jonak, C.; Tittes, J.; Brunner, P.M.; Guenova, E. Mycosis fungoides and Sézary syndrome. JDDG J. Der Dtsch. Dermatol. Ges.
**2021**, 19, 1307–1334. [Google Scholar] [CrossRef] [PubMed] - McDonald, E.O.; Amanullah, A.A.; Park, P.S.U.; Song, W.; Werner, T.J.; Alavi, A.; Revheim, M.-E. The role of 18F-FDG PET/CT in primary cutaneous lymphoma: An educational review. Ann. Nucl. Med.
**2023**, 37, 328–348. [Google Scholar] [CrossRef] [PubMed] - Gallamini, A.; Zwarthoed, C.; Borra, A. Positron emission tomography (PET) in oncology. Cancers
**2014**, 6, 1821–1889. [Google Scholar] [CrossRef] [PubMed] - Salem, A.E.; Shah, H.R.; Covington, M.F.; Koppula, B.R.; Fine, G.C.; Wiggins, R.H.; Hoffman, J.M.; Morton, K.A. PET-CT in Clinical Adult Oncology: I. Hematologic Malignancies. Cancers
**2022**, 14, 5941. [Google Scholar] [CrossRef] [PubMed] - Hsiao, Y.-P.; Chiu, C.-W.; Lu, C.-W.; Nguyen, H.T.; Tseng, Y.S.; Hsieh, S.-C.; Wang, H.-C. Identification of skin lesions by using single-step multiframe detector. J. Clin. Med.
**2021**, 10, 144. [Google Scholar] [CrossRef] [PubMed] - Wang, Y.; Wang, Y.; Cai, J.; Lee, T.K.; Miao, C.; Wang, Z.J. Ssd-kd: A self-supervised diverse knowledge distillation method for lightweight skin lesion classification using dermoscopic images. Med. Image Anal.
**2023**, 84, 102693. [Google Scholar] [CrossRef] [PubMed] - Sumithra, R.; Suhil, M.; Guru, D. Segmentation and classification of skin lesions for disease diagnosis. Procedia Comput. Sci.
**2015**, 45, 76–85. [Google Scholar] [CrossRef] - Balaji, V.; Suganthi, S.; Rajadevi, R.; Kumar, V.K.; Balaji, B.S.; Pandiyan, S. Skin disease detection and segmentation using dynamic graph cut algorithm and classification through Naive Bayes classifier. Measurement
**2020**, 163, 107922. [Google Scholar] [CrossRef] - Wu, H.; Chen, S.; Chen, G.; Wang, W.; Lei, B.; Wen, Z. FAT-Net: Feature adaptive transformers for automated skin lesion segmentation. Med. Image Anal.
**2022**, 76, 102327. [Google Scholar] [CrossRef] - Ke, G.; Meng, Q.; Finley, T.; Wang, T.; Chen, W.; Ma, W.; Ye, Q.; Liu, T.-Y. Lightgbm: A highly efficient gradient boosting decision tree. Adv. Neural Inf. Process. Syst.
**2017**, 30, 3149–3157. [Google Scholar] - Chan, I.L.; Cohen, S.; da Cunha, M.G.; Maluf, L.C. Characteristics and management of Asian skin. Int. J. Dermatol.
**2019**, 58, 131–143. [Google Scholar] [CrossRef] [PubMed]

**Figure 1.**The distribution of skin lesions in MF, PsO, and AD is remarkably varied. These lesions can manifest on the upper body with small areas starting from 5 cm

^{2}, extending to larger lesions that spread over extensive regions of the (

**a**) chest, (

**b**) back, (

**c**) arms, flanks, or lower body. The morphological presentation of these lesions ranges from superficial red streaks on the skin to deeper lesions in the epidermal layers, which may be associated with infection and bleeding. These lesions, which appear as red clusters, are relatively challenging to distinguish with the naked eye. As depicted in the figure, the lesions are identified as [(

**a**–

**c**)] MF, [(

**d**,

**e**)] PsO, and (

**f**) AD.

**Figure 2.**The experimental workflow is as follows: Input images were initially segmented in a coarse manner by using Unet-Attention. The exact precision in this segmentation phase is not critical. The primary aim is to segregate the affected skin from the normal skin for processing in the hyperspectral imaging (HSI) model. This step aids in conserving model capacity and enhances accuracy by filtering out noise, such as that from moles and other skin pigments, which are removed during this stage. Subsequently, the output from the HSI model underwent dimensional reduction, and it was classified using the XGBoost algorithm.

**Figure 3.**The data distribution displays the number of images in accordance with their shooting angles, encompassing back-view shots, side-view shots, and close-up shots.

**Figure 4.**A schematic of the proposed method is displayed. Hyperspectral imaging facilitates the conversion between the camera and the spectrometer, employing standard 24-color blocks (X-Rite Classic, 24 Color Checkers) as calibration references. The digital camera captures these color blocks, translating the information into digital images. Concurrently, the spectrometer delineates spectral information by measuring the light from the 24 color blocks. Subsequently, the camera’s images are algorithmically transformed into a spectrum equivalent to that derived from the spectrometer, utilizing hyperspectral imaging techniques.

**Figure 5.**Polynomial regression analysis of the brightness values of these color checker blocks was conducted to represent the gradient of grayscale changes. This analysis comprises (

**a**) first-order (linear) regression, (

**b**) second-order (quadratic) regression, and (

**c**) third-order (cubic) polynomial regression.

**Figure 6.**Cumulative explained variance of the first 12 principal components in the spectrum obtained from the principal component analysis.

**Figure 7.**The t-SNE analysis utilized 5000 data points, sampled from three different skin disease groups. This analysis revealed a clear clustering of features unique to the PsO skin group, setting it apart from the other groups.

**Figure 8.**The correlation between the measured spectrum and the spectrum simulated by the algorithm, representing six basic color filters commonly used in chromatic correction, was assessed. This comparison revealed that the algorithm’s simulated reproduction closely matched the measured spectrum. The largest deviation occurred at longer wavelengths, specifically those exceeding 600 nm.

**Figure 9.**The color difference of 24-color blocks is shown on the CIE 1931 chromaticity diagram 2° standard observer. The black line connecting the center of the two points is “red” to indicate simulated colors and “green” to indicate measured colors.

**Figure 10.**Visualization of the segmentation by the U-Net Attention model illustrates that most lesion areas are effectively isolated from the surrounding normal skin areas. The lesions manifest in various forms: either as close-up images on the arm (

**a**,

**b**), as clusters comprising numerous small, scattered spots (

**c**,

**e**,

**f**), or even as a large area (

**d**) distinctly segmented areas, in contrast to other skin regions.

Color No. | Measured | Simulated | L* (Measured) | a* (Measured) | b* (Measured) | L* (Simulated) | a* (Simulated) | b* (Simulated) | CIEDE 2000 |
---|---|---|---|---|---|---|---|---|---|

1 | 38.94 | 24.35 | 36.89 | 38.81 | 24.92 | 38.23 | 0.52 | ||

2 | 67.93 | 32.09 | 49.08 | 67.95 | 32.03 | 48.94 | 0.05 | ||

3 | 49.45 | 0.07 | 9.87 | 49.34 | −0.66 | 9.68 | 1.01 | ||

4 | 42.33 | −4.05 | 39.45 | 42.59 | −3.97 | 38.03 | 0.57 | ||

5 | 55.89 | 16.17 | 11.32 | 55.88 | 16.12 | 11.33 | 0.04 | ||

6 | 69.10 | −19.39 | 31.48 | 69.10 | −19.46 | 31.45 | 0.05 | ||

7 | 65.32 | 46.33 | 82.11 | 65.37 | 46.43 | 81.81 | 0.14 | ||

8 | 39.09 | 8.89 | −10.49 | 39.23 | 9.36 | −10.40 | 0.49 | ||

9 | 55.11 | 58.96 | 46.87 | 55.06 | 59.03 | 47.19 | 0.14 | ||

10 | 31.22 | 24.81 | 6.34 | 31.78 | 25.36 | 6.05 | 0.57 | ||

11 | 71.87 | −5.83 | 76.56 | 71.87 | −5.85 | 76.51 | 0.02 | ||

12 | 74.87 | 34.26 | 90.13 | 74.84 | 34.16 | 90.25 | 0.08 | ||

13 | 27.87 | 11.77 | −23.10 | 27.60 | 11.34 | −23.26 | 0.46 | ||

14 | 53.97 | −27.31 | 50.50 | 53.95 | −27.22 | 50.74 | 0.11 | ||

15 | 45.76 | 66.20 | 51.52 | 45.76 | 65.96 | 51.01 | 0.17 | ||

16 | 83.76 | 23.22 | 100.15 | 83.76 | 23.26 | 100.18 | 0.02 | ||

17 | 54.10 | 58.28 | 24.01 | 54.11 | 58.30 | 24.01 | 0.01 | ||

18 | 48.03 | −23.70 | −0.08 | 48.08 | −23.35 | −0.02 | 0.19 | ||

19 | 95.47 | 15.52 | 46.28 | 95.48 | 15.52 | 46.28 | 0.00 | ||

20 | 80.98 | 13.58 | 40.13 | 80.95 | 13.62 | 40.11 | 0.04 | ||

21 | 66.39 | 11.35 | 34.00 | 66.52 | 11.47 | 34.18 | 0.14 | ||

22 | 52.19 | 9.40 | 28.14 | 51.90 | 9.29 | 28.26 | 0.31 | ||

23 | 36.45 | 6.84 | 21.33 | 36.46 | 6.79 | 21.44 | 0.09 | ||

24 | 21.36 | 4.87 | 14.76 | 21.05 | 3.68 | 15.29 | 1.55 | ||

average | 0.28 |

**Table 2.**Model training performance was evaluated across different k-fold values (3, 5, and 7) with respect to four performance metrics.

k-Fold | Sensitivity | Specificity | F1-Score | ROC-AUC |
---|---|---|---|---|

k3 | 85.61% | 95.25% | 86.26% | 0.9051 |

k5 | 89.20% | 96.35% | 89.19% | 0.9270 |

k7 | 90.72% | 96.76% | 90.08% | 0.9351 |

Model | Number of Params | Size (MB) | IoU | AUC-ROC | Time Prediction per Image (second) |
---|---|---|---|---|---|

U-Net | 31M | 124 | 0.8003 | 0.8490 | 0.0741 |

U-Net++ | 55M | 220 | 0.8977 | 0.9334 | 0.1182 |

U-Net Attention | 35M | 140 | 0.8521 | 0.9097 | 0.0792 |

DeepLabv3 | 41M | 164 | 0.8730 | 0.9597 | 0.0836 |

Ours (3-folds) | 35M | 148 | 0.8521 | 0.9051 | 0.0810 |

Ours (5-folds) | 35M | 148 | 0.8521 | 0.9270 | 0.0810 |

Ours (7-folds) | 35M | 148 | 0.8521 | 0.9351 | 0.0810 |

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |

© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Huang, H.-Y.; Nguyen, H.-T.; Lin, T.-L.; Saenprasarn, P.; Liu, P.-H.; Wang, H.-C.
Identification of Skin Lesions by Snapshot Hyperspectral Imaging. *Cancers* **2024**, *16*, 217.
https://doi.org/10.3390/cancers16010217

**AMA Style**

Huang H-Y, Nguyen H-T, Lin T-L, Saenprasarn P, Liu P-H, Wang H-C.
Identification of Skin Lesions by Snapshot Hyperspectral Imaging. *Cancers*. 2024; 16(1):217.
https://doi.org/10.3390/cancers16010217

**Chicago/Turabian Style**

Huang, Hung-Yi, Hong-Thai Nguyen, Teng-Li Lin, Penchun Saenprasarn, Ping-Hung Liu, and Hsiang-Chen Wang.
2024. "Identification of Skin Lesions by Snapshot Hyperspectral Imaging" *Cancers* 16, no. 1: 217.
https://doi.org/10.3390/cancers16010217