According to the GLOBOCAN2012 database of World Health Organization （WHO）， in 2015， the number of new cases of cutaneous melanomas （CM） worldwide reached 250 178， including 130 800 males and 119 378 females. There were 60 098 global cutaneous melanoma deaths， including 34 143 males and 25 955 females^［1］. The incidence of cutaneous melanoma has steadily increased over the past 50 years in predominately fair-skinned populations^［2］. Cutaneous melanoma mostly stems from cancerous melanocytes in the epidermis， which mainly occurs in the back， shoulders and lower limbs of human body. In the early stage， it has higher metastatic potential^［3］. The treatment process of cutaneous melanoma is generally divided into diagnosis， treatment， and prognosis. Most cases are diagnosed at an advanced stage， but there is no effective cure for advanced cutaneous melanoma. It is usually only possible to prolong the life of patients by surgical resection and chemotherapy^［4］. Early accurate diagnosis of cutaneous melanoma is of vital significance for timely treatment and recovery.

Detecting melanoma in skin tissue and determining the size of the lesion area are common diagnostic procedures. Researchers at home and abroad have combined numerous modern optical imaging techniques with image processing methods for medical diagnosis. Innovative approach such as dermoscopy can be well applied clinically^［5-6］. In recent years， dermoscopy images processed by deep residual networks can get more accurate skin lesion segmentation^［7］. Based on laser scanning confocal microscope images， an algorithm combined wavelet analysis and regression tree improved the accuracy of melanoma diagnosis ^［8］. However， due to the high degree of visual similarity between melanoma and non-melanoma lesions， the huge intraclass variation of melanomas， and the existence of many artifacts in the images^［9-10］， even well-trained dermatologists may obtain various diagnostic results， and macroscopic digital dermoscopy also have certain limitations ^［11］.

The medical field recognizes that pathological diagnosis remains the gold standard for complex tumor diagnosis. In the treatment of melanoma， the depth of tumor invasion （DoI）， namely Breslow thickness in skin tissue is a crucial prognostic factor ^［12］. Tumor invasion depth is defined as the maximum distance from the top of the granular layer to malignant melanocyte ^［13］. Traditional diagnosis relies on the pathologist's visual inspection of skin biopsies under the microscope. Until 2014， Mokhtari et al. ^［14］ stated that they were the first to apply computer-aided diagnostic algorithms to the measurement of melanoma invasion depth. They used a clustering algorithm and a pre-trained support vector machine （SVM） classifier to extract granular layer and malignant melanocytes， respectively. Another approach is presented by Noroozi et al. for tumor invasion depth， which is based on the entropy of ridges to segment cornified layer and color features to detect cells nuclei ^［15］. In the study ^［16］， multi-threshold segmentation， Bayesian classification， and Hausdorff distance measurement were combined to calculate the melanoma depth of invasion. Immunohistochemical technology provided further supplement and verification for pathological diagnosis ^［17］. Kambiz et al. studied 36 biopsy specimens to determine the difference in vertical thickness of melanoma in hematoxylin and eosin （H&E） and immunohistochemical staining images， they found melanoma-related immunomarkers， the combination of S-100 and Melan-A staining may be sufficient to measure tumor invasion depth ^［18］. In the latest statistics， immunohistochemistry and genetic sequencing have become popular methods for melanoma depth of invasion and lymph node metastasis ^［19］. The existing computer-aided methods almost use 2-D microscopic images of skin biopsy， only spatial information can be referenced for diagnosis. These methods are limited to similar morphology of biological tissues such as melanoma cells and lymphocytes.

The application of hyperspectral imaging technology in biomedical imaging points out a new direction for the identification and diagnosis of cutaneous melanoma ^［20］. Nagaoka et al. proposed an early melanoma screening system by studying melanoma hyperspectral data and obtained a possible melanoma discrimination index based on the characteristics of the pigment molecules. There were not many practical samples of this index， but it was quite successful in distinguishing between melanoma and non-melanoma ^［21］. Zheludev et al. presented a non-invasive method for depicting malignant freckles and malignant freckles-like melanoma. Before the surgical resection， the hyperspectral data of the lesion area and the normal tissue were collected， the spectral features of the pixels in the two regions were extracted， and the internal structure of the spectral data was revealed by a classification algorithm and a diffusion map technique ^［22］. Inspired by Neyman-Pearson lemma， Pardo et al. proposed a reservoir-based non-parametric estimation and classification model with sensitivity tunability， which realized the classification of hyperspectral images of nevi， raised nevi and melanomas ^［23］. In the experiment of detecting melanoma with hyperspectral dermoscopy images， Anna-Marie et al. used artificial intelligence algorithms to analyze， and the overall sensitivity and specificity were 100% and 36%， respectively ^［24］. The achievements illustrate that researchers focus on the use of image algorithms to accurately identify the tumor region of skin surface， but few researchers have used microscopic hyperspectral imaging technique to quantify the depth of invasion of the cutaneous melanoma.

Microscopic hyperspectral imaging （MHSI） technology highlights the abundant spatial and spectral information in the pathological structure of cutaneous melanoma. In this article， to measure the superficial spreading depth of cutaneous melanoma， we use a segmentation approach combined a kernel minimum noise fraction （KMNF） algorithm and two morphological edge detection ways to identify the skin granular layer. Subsequently， a least squares support vector machine based on characteristic spectrum supervision （CSS-LSSVM） segmentation algorithm is applied to separate malignant melanocytes. Comparing the results of different algorithms， our method has realistic significance and may offer a quantitative basis for CM pathological diagnosis.

Hyperspectral images can provide an assessment of tissue pathophysiology based on the spectral characteristics of different tissues^［25-26］. The physical connotation is lattice vibrations of different molecules， atoms， and ions， which cause spectral absorption and transmission at different wavelengths^［27］. In this paper， the main segmentation methods of cutaneous melanoma and melanocytes can be summarized with a flowchart demonstrated in Fig. 1， and the details are introduced in the following parts.

Conventional pathological sections preparation is extremely important in histopathology and is the basis of all pathological analyses. The main steps involved in the production process are tissue sample extraction， fixation， dehydration， waxing， embedding， slicing， baking， dyeing， and sealing ^［28］. Hematoxylin-eosin is a common staining agent for paraffin sections. Hematoxylin staining solution is alkaline， which makes the nuclei， granules， basal layer， small sweat glands， and their ducts and spinous layers appear blue. Eosin is an acid dye that causes the stratum corneum， blood vessels， collagen fibers， apocrine glands and their ducts to be red in the skin tissue.

The structure of skin tissue is complex， and skin melanoma has a wide variety of types. However， under the microscopic field of view， it is not possible to completely contain all the structures of the skin tissue. The cutaneous melanoma samples used in this research are superficial spreading melanoma （SSM）. The lesions are mostly in the epidermis and dermis. The lesion boundaries are obvious， and melanocytes are scattered individually or in-cluster in various layers of the epidermis， with few tumor cell nests and no top-to-bottom maturation. Therefore， we define the tumor depth of invasion within the SSM sample as superficial spreading depth in our experiment. As shown in Fig. 2 （b）， superficial spreading depth refers to the vertical distance from the outer boundary of skin granular layer to the furthest malignant melanocytes， while the structure of each layer of normal skin tissue is intact （Fig. 2（a））. Due to limitation of the microscopic field of view， the layers of skin tissue under low magnification are comprehensive， but the details of cell tissue cannot be expressed， which is contrary to high magnification. Different acquisition fields of view and magnification are usually chosen for different research purposes. According to the field characteristics of sample and research content， the samples studied in this experiment are 100 times images.

The microscopic hyperspectral imaging system as shown in Fig.3 combines hyperspectral technology with microscopic imaging technology ^［29］， which consists of an optical microscope， an acousto-optic tunable filter （AOTF）， a high-sensitivity color charge-coupled device（CCD） camera， a high-density cooled grayscale CCD camera， a high-precision three-dimensional stage， a ring light source， an anti-vibration table and an industrial computer ^［30］. The light source generator adopts a halogen lamp， and combines the characteristics of the system component AOTF hardware to select the spectral range from visible light to near-infrared region. The spectral range selected by the MHSI system is from 550 nm to 1000 nm， and the spectral resolution reaches 2∼6 nm. During the imaging process， after passing through the sample and the objective lens （4×， 10×， or 20×） of the microscope， one way of light enters the color CCD to take ordinary microscope image； the other way enters the AOTF filter， and then imaged on the grayscale CCD to collect hyperspectral image， as shown in Fig. 3 blue arrow. In addition， the green arrow indicates light entering the eyepiece.

In hyperspectral data acquisition， single-wavelength image data are consistent with ordinary two-dimensional image data that contain only spatial information （Fig.4（c））. When acquiring multi-wavelength image data， a plurality of single-wavelength image data are stored together with wavelength information of each pixel （Fig.4（a）） as a hyperspectral image data cube. Figure 4（b） is a data cube of cutaneous melanoma image， containing 60 wavebands of spectral information. Each waveband of the hyperspectral image consists of 1 280 × 1 024 pixels × 16 bits/pixel. Once acquired， the dataset is saved in the band sequential （BSQ） file format.

When the MHSI system is collecting image data， changes in external conditions such as illumination intensity， CCD electrical noise， and sample background have a certain impact on the quality of imaged data. Since the absorption characteristics of biological tissues to the spectrum are much smaller than the transmission and reflection characteristics of slide， the unprocessed sample data exhibit similar spectral characteristics in the same wavelength range. The spectral characteristics of target tissue are weakened， which cannot accurately reflect biochemical features of the sample. Hence samples’ spectrum should be corrected first before data processing.

Lambert-Beer's （LB） law ^［31］ expresses the relationship between absorption intensity of a substance for a single light and its concentration and thickness. It is assumed that the absorbance of a substance is A， the transmittance is T， the parallel monochromatic incident light intensity of substance is I₀， and the transmitted light intensity is I.

A=-logT

T=I0I

In this experiment， we simultaneously acquired hyperspectral images of biological tissues and blank areas of the same slides. The transmittance and absorbance of the cutaneous melanoma samples are：

Tn,m;λ=Dn,m;λ-In,m;λBn,m;λ-In,m;λ

An,m;λ=logBn,m;λ-In,m;λDn,m;λ-In,m;λ

where D（n，m；λ） is the pixel value of the target sample in the n^th row and m^th column in the λ-th wavelength， B（n，m；λ） is the pixel value of the blank sample corresponding to the target sample， and I（n，m；λ） is the image noise acquired by the objective lens between the light source is turned off and on again. On the one hand， because the parameters of the system hardware are stable， the image noise at each pixel location is consistent. On the other hand， hyperspectral imaging mainly relies on spectral characteristics for tissue classification. Previous study has shown that spectral characteristics of the stain are fixed， and the difference in staining basically does not affect the spectral characteristics of biological tissues^［32］. In fact， the process of spectral calibration also purifies the spatial information of the image to a certain extent， and the role of H&E stain after preprocessing will be weakened and negligible. By eliminating the effects of system equipment noise， sample dye inequality， etc.， hyperspectral characteristic curve of actual skin melanoma can be extracted.

The purpose of this section is to achieve segmentation of the skin granular layer. In H&E-staining pathological images， the limitation of RGB information makes it difficult to distinguish different tissues with extremely high similarity. However， in microscopic hyperspectral images， the spectral characteristics of granular layer are different from other tissues. We utilized a KMNF and morphological filtering segmentation method to provide an accurate starting boundary for the calculation of melanoma superficial spreading depth.

Minimum noise fraction （MNF） ^［33］ was proposed by Green et al.， which was first used in the field of remote sensing to obtain image dimensionality reduction and noise fraction. In the study of hyperspectral data of human skin injuries， Randeberg et al. took MNF as a noise reduction method before spectral angle （SAM） processing ^［34］.

KMNF ^［35］ is an improved algorithm based on MNF， which introduces a kernel function in the linear transformation space to make the mapping of samples to nonlinear space. Let the covariance matrix of the original data D（x） be the sum of the signal component covariance matrix C_s and the noise component covariance matrix C_n， then the noise fraction NF_d and the signal to noise ratio SNR are defined as：

NFd=LTCnLLTCL

SNR=LTCsLLTCnL

After linearization， the MNF maximization is：

1NFd=LTCLLTCnL=LTPTPLLTPnTPnL

If a hyperspectral image is regarded as a data set of n pixels and p spectral bands， then P^T is a matrix of n times p， and P_n^T is a similar matrix of noise. Let L be equal to ， then the dual mode of MNF is：

1NFd=ℒTξξTξξTℒℒTξξnTξnξTℒ

Let a map of be R_n， then the maximization of KMNF is

1NFd=ℒTRℛTℛℛTℒℒTℛℛnTℛnℛTℒ=ℒTℝ2ℒℒTℝnℝ2ℒ

where R_n is a two-dimensional matrix， the element of the asymmetric matrix = is the kernel function k（x_i， x_Nj）， i， j=1，···，n， and the average of and kernel matrix columns is zero. An important result of the dual representation is that the dimension of feature space no longer affects calculation. The kernel function is added based on the minimum noise separation transform， and the mapping from original data space to feature space can be completed. The first few dimensions contain a large number of eigenvalues of the image data and a small noise. As the dimension increases， the noise becomes larger and the eigenvalue becomes smaller.

KMNF retains nonlinear characteristics of hyperspectral data effectively. In cutaneous melanoma samples， the granular layer exhibits a black band in hyperspectral grayscale image. The next step is to extract main morphology of the granular layer by morphological filtering. A conventional morphological operator is an N by N （N=3，5，7...） square filter. In this section， in order to better adapt to the shape of the skin granular layer， a multi-size filter operator tending to the diamond shape is designed. We mixed three kinds of filter operators to achieve better morphological extraction results. Namely， the skin granular layer uses a 3 by 3 template， the epidermis layer except the granular layer adopts a 5 by 5 template， and the irregularly shaped fiber structure and malignant melanocytes employ a 7 by 7 template.

The skin granular layer is generally used as starting boundary for the measurement of skin melanoma superficial spreading depth ^［15］. Therefore， after obtaining its morphology， it is necessary to further separate the edge of the granular layer. The image edge is where the gray value changes drastically， reflecting the boundary information between background and outline of the tissue. Image edge detection can filter out a large amount of image information， leaving only important edge structure information.

We studied the image contour extraction method based on level set segmentation ^［36］. The basic idea of the algorithm is to embed the target boundary as a zero-level set into the level set function of the higher dimensional surface. The evolution of high-dimensional closed curve is transformed into the evolution of level set functions ^［37］. The Chan-Vese model^［38］ used in this research is a novel level set segmentation algorithm. The image is divided into a target image area and a background image area and the average gray value is calculated. At this time， the process of image segmentation changes to find out a optimal boundary segmentation curve， so that the gray value error between segmented image region and original image region is minimized. The level set segmentation method is capable of extracting the edge of multiple independent targets， which is advantageous for edge detection of the skin granular layer.

In pathological diagnosis of cutaneous melanoma， the morphology and distribution of malignant melanocytes are one of the main diagnostic criteria. It is also the termination boundary for calculating melanoma superficial spreading depth. In this research， we utilized the least squares support vector machine method based on characteristic spectrum supervision to segment melanocytes.

The least squares support vector machine （LSSVM） is an improved data mining-based machine learning method ^［39］. Aiming at the problems of high dimensional and over-fitting in traditional algorithms， this method follows structural risk minimization criterion， which can solve the local optimal solution and nonlinear problem when the model is established. In specific nonlinear problem-solving process， the kernel function is introduced to replace the complex inner product operation by mapping the sample to the hyperplane ^［40］. The Gaussian kernel function is a relatively simple and efficient radial basis function^［41］， and the expression is as follows：

Kx,xi=exp-x-xi2σ2

The least squares support vector machine based on characteristic spectrum supervision （CSS-LSSVM） segmentation method add characteristic spectra of the target to match in the LSSVM model and segment the LSSVM results again to achieve higher accuracy. The matching way calculates spectral similarity between melanocytes and the pixel of segmentation image one by one and determine the type of the segmented pixel. A spectral angle matching method is typically used for evaluation of spectral similarity. It is known that the number of wavelengths of experimental sample is λ， the spectral characteristic of a single pixel is X=［x₁，x₂，···，x_λ］， and the characteristic spectra of the melanocytes are X’=［x’₁，x’₂，···，x’_λ］， then the spectral similarity ^［42］ of X and X’ is：

α=arccos∑i=1λxi⋅x'i∑i=1λx2i⋅∑i=1λx'2i

The threshold α is usually set artificially， beyond which it is considered not to be a homogeneous biological tissue. The CSS-LSSVM segmentation method includes the following steps：

1） The spectra of region of interest （ROI） are selected as the training set.

2） A feature spectrum library is constructed， and the average spectrum of a plurality of adjacent pixels in the target region is taken as a characteristic spectrum.

3） The ROI data set is trained to establish LSSVM classification model.

The spectra of remaining pixels are sequentially inputted to the LSSVM model. If the model output is determined to be the target class， output is the target class； if not， the pixel’s spectrum is matched with the feature spectral library. If the matching result is a target class， it is determined to be the target class， otherwise， it is not.

The samples used in this section are magnified 100-fold images of cutaneous melanoma， as they contain both granular layers and melanocytes in the field of view. The boundary of granular layers is divided into an inner boundary （near the dermal layer portion） and an outer boundary （near the stratum corneum portion）， and the outer boundary is calculated for superficial spreading depth. For the malignant melanocytes in the samples， the data of the nearer melanocytes are removed with reference to the slope of the outer boundary of granular layer， and the farthest data are retained.

Let the set of outer boundary curve points of the granular layer be D（Φ_i（x，y）|i∈1，2，···，p）， and p is the number of pixels of the outer boundary curve. Let the set of malignant melanocyte boundary points be M（Γ_i（x，y）|j∈1，2，···，q）， q is the number of pixels of malignant melanocytes border. The procedures for calculating superficial spreading depth of skin melanoma are：

1） Since the distance between adjacent pixels is small， in order to reduce the amount of calculation， the curves D and M are resampled. Let the pixel points of sampling distance be d， and the symbol ∟ indicates rounding down. The boundary sets of granular layer and malignant melanocytes are respectively expressed as：

DdΦix,y|i∈1,2,⋯,p/d

MdΓjx,y|j∈1,2,⋯,q/d

2） Let E_kbe a set of coordinate pairs， and the i^th point in the set M_d is Φ_i. Calculating the distance between Φ_i and each point in the set D_din turn， and E_k records the point with the smallest distance， ie：

EkΦix,y,Γjx,y=minΦi-Γj,Φi∈Dd,Γj∈Md

E~Φmx,y,Γmx,y=maxEEk,k∈1,2,⋯,q/d

where ||•|| represents the Euclidean distance between two points， and the coordinate pair indicates the point at which the malignant melanocytes are furthest from granular layer.

3） To improve the accuracy of the tumor superficial spreading depth calculation results， the maximum of the neighborhood is solved by the maximum distance of the previous step. Let Deep be the depth of melanoma， then：

Deep=maxΦix,y-Γjx,y,m-d<i,j<m+d

We operated the microscopic hyperspectral imaging system to capture images of cutaneous melanoma samples at different magnifications. A magnified 100x image contained epidermal layer is displayed in Figs.5（a-b）. The color image Fig.5（a） is ordinary microscopic result and Fig.5（b） is a single-wavelength hyperspectral image at 810nm. In Fig.5（a）， （1） is the entire layer of the epidermis layer， （2） is the stratum corneum， and （3） is the granular layer. From the viewpoint of skin histology， the skin granular layer is located between the stratum corneum and the spinous layer. In the images， the orange arrows point to the blank area of the sample， the red arrows point to the malignant melanocytes， the green arrows point to normal tissue of the epidermis， and the yellow arrows point to fibrous tissue of the dermis. Malignant melanocytes are irregularly distributed in the dermis and epidermal basal layers. As the magnification increases in Figs.5（c-d）， images magnified 200x mainly contains malignant melanocytes in the lesion area. The nucleus and cytoplasm of melanocytes are obvious， and the morphology is clearer. Similar to Fig.5（a）， the blue arrows in Fig.5（c） point to blood vessel in the center， showing that malignant melanocytes are distributed around the blood vessels and begin to invade inward.

After spectral calibration of hyperspectral image， the spectral differences of various pathological tissue components can be relatively truly reflected. Figure 6 is an example of a cutaneous melanoma sample at 200x magnification. The field of view mainly includes the tumor areas and normal fibrous tissues， and the spectra of four regions are selected and compared. Figure 6（a） is a single-wavelength hyperspectral image with a wavelength of 810 nm， and （b） is an enlargement of the red partial region in （a）， wherein the red curve is the nucleus of the malignant melanocytes， the blue curve is the tissue of the lesion region， the orange curve is a blank area of the slide covered by no skin tissue， and the green curve is normal fibrous tissue. Figure 6（c） is the untreated spectra corresponding to the four regions. It can be seen that the spectral characteristics of each type of tissue are similar. As the wavelength increases， significant peaks and troughs appear at the same wavelength， there are only some differences in the range of values. The corrected spectral data are shown in Fig.6（d）， and the spectral transmittance of melanocytes is the lowest. The abundant spectral properties of melanoma samples are a powerful basis for subsequent biological tissue segmentation.

The skin melanoma samples used in the experiment are thick in stratum corneum and have no distinct boundary with granular layer. It is almost difficult to identify and segment the granular layer based on common pathological images. Combining the spectral information and spatial information of hyperspectral images， we used the KMNF algorithm to focus the spectral features of image to the first few wavebands and reduce the image noise. The segmentation process of skin granular layer is demonstrated in Fig.7. Figure 7（b） is still the hyperspectral imaging result with a wavelength of 810 nm. The sixth waveband result obtained after KMNF treatment on the experimental sample is shown in Fig.7（c）. The white area in Fig.7（d） is the granular layer， which is the result of morphological filtering and identifying the largest connected domain. The contour of the skin granular layer is screened by the level set segmentation algorithm. The contour result （e） is superimposed with the original hyperspectral image field of view （b）， and Fig.7（f） is the final segmentation result. Combined with the pathological characteristics of the skin granular layer， the KMNF-based segmentation results are basically consistent with the actual contour.

In this part， three methods for feature extraction of hyperspectral images are studied， and some samples are employed for experiments. Fig.8 shows the comparison results of three methods of PCA （principal component analysis）， MNF， and KMNF. Based on the negative and positive results of disease detection， we introduced a confusion matrix to calculate the accuracy of the algorithms. The accuracy indicates the ability of algorithms to correctly classify the target and non-target for the whole sample. For all test samples， the accuracy of the KMNF-based segmentation method can reach more than 80%， which is higher than the other two algorithms. This method provides a relatively more accurate boundary for the calculation of skin superficial spreading depth. However， it must be acknowledged that the segmentation results of this algorithm depend on the complexity of the cutaneous melanoma samples. The improvement of the algorithm is one of our future research contents.

Skin melanoma is derived from the cancerization of melanocytes. Therefore， malignant melanocytes are the key targets in the diagnosis of cutaneous melanoma. Under the microscopic field of H&E staining samples （Fig. 9（a））， malignant melanocytes have diversified morphological features， and there are subjective errors in human eye observation， which is time consuming and increases the difficulty in rapid pathological diagnosis of cutaneous melanoma. We added a melanocyte characteristic spectral library and used spectral information to assist the least squares support vector machine for segmentation. The malignant melanocyte recognition results are shown in Fig.9. The CSS-LSSVM segmentation method （Fig.9（d）） can identify more cell regions than the SVM segmentation method （Fig.9（c））.

The results of quantitative analysis of four samples are listed in Table 1. In addition to accuracy， the other two commonly used evaluation parameters are sensitivity and specificity. The sensitivity indicates the ratio of correct judgment of the algorithm to the total positive sample； the specificity indicates the ratio of negative sample prediction result to the actual negative sample.

The CSS-LSSVM algorithm has higher segmentation accuracy than the SVM for malignant melanocytes in the samples and can reach more than 85%. Referring to Table 1， the sensitivity of the CSS-LSSVM algorithm is higher than that of the SVM， namely， the probability that the CSS-LSSVM can detect malignant melanocytes is higher. The increase in sensitivity tends to reduce specificity. The specificity of the CSS-LSSVM algorithm in the four samples is lower than that of the SVM， which means the probability that the CSS-LSSVM can detect non-malignant melanocytes is lower. It can be seen that CSS-LSSVM may misidentify a few non-malignant melanocytes while recognizing more malignant melanocytes. However， in view of melanoma is highly metastatic， in diagnosis of cutaneous melanoma， leakage recognition of malignant melanocytes may lead to cancer cell metastasis， which greatly affects the survival rate of patients. In consequence， compared with SVM segmentation method， the CSS-LSSVM segmentation method is more suitable as a recognition algorithm for malignant melanocytes.

Surgical treatment is one of the most effective and preferred treatment methods for cutaneous melanoma. It is to prevent the spread of the tumor by resection of the primary lesion. Whether the resection of the primary lesion is complete is also a direct factor affecting the prognosis. The most important thing in surgical treatment is to determine the boundaries of the tumor. We adopted the CSS-LSSVM algorithm to segment the malignant melanocytes in 100-fold amplified samples， and extracted the skin granular layer structure based on the KMNF algorithm， and retained the outer boundary. During the measurement， the calculated superficial spreading depth denotes the farthest distance between the malignant melanocytes and the granular layer. The yellow arrows in Fig. 10 are the melanoma superficial spreading depth automatically calculated by the computer.

Since the superficial spreading depth is calculated based on pixels， in order to obtain the true physical value corresponding to the pixel size of the image， we conducted the following experiments： at different microscope magnifications， an objective lens calibration ruler was moved to count the distance of pixels in the image. The results are shown in Table 2. It shows that under the same objective magnification， the moving distance of pixels and the scale have a linear growth relationship； under the same scale， the moving distance of the pixel and the objective multiple also have a linear growth relationship. From this， the conversion coefficient between the image size of the pixel and its true physical value at different objective magnifications can be calculated. The experimental data in this paper were collected under the 10x objective lens， hence the conversion coefficient 2 243 （pixel / mm） was selected to measure the true value corresponding to the morphological characteristics. For a sample （a）， the superficial spreading depth is 1 379.5 pixels， approximate 0.61 mm.

To evaluate the depth measurement， we compared the melanoma sample with normal skin tissue， as shown in the Fig.2. Fig.2（b） shows that malignant melanocytes have been produced in the dermis layer. Tumor pathology grading is the key basis for treatment. In the clinical diagnosis of melanoma， Clack level ^［43］ is commonly used to assess the invasion level and Breslow thickness is used to measure the depth of tumor invasion， however， these two indicators are more focused on samples with nests of melanoma that were already evident in the mid to late stages. The sample used in this article is a superficial spreading melanoma. The tissue contains malignant melanocytes， but the cancer nest has not yet formed. Therefore， we defined a superficial spreading depth similar to the Breslow thickness to measure the early spread of the tumor.

The results show the distribution and spread of malignant melanocytes of superficial spreading melanoma. If patients are not treated in time， tumor deterioration and metastasis are likely to occur. we expect to provide a new technique to assist pathological diagnosis. The method proposed in this article is also applicable to advanced skin melanoma samples. However， due to the complexity of the melanoma samples， the morphological and spectral similarity of individual lymphocytes and malignant melanocytes in the melanoma sample may affect the accuracy of identification.

In view of the problem that the traditional pathological diagnosis of cutaneous melanoma cannot quantitatively analyze the tumor tissue， this paper applied microscopic hyperspectral imaging technology to identify and analyze melanoma superficial spreading depth. First， we segmented skin granular layer. Three multi-scale morphological filter operators were designed to extract the main morphological features， and the level set segmentation algorithm was used to obtain the contour of granular layer. By comparing PCA， MNF， and KMNF， the experimental results show that the KMNF-based algorithm has a segmentation accuracy of more than 80%. Second， we explored the segmentation of the malignant melanocytes. After adding the characteristic spectrum supervision of malignant melanocytes， the segmentation accuracy of the LSSVM is greater than 85%， which is better than the SVM algorithm. Finally， on the basis of the above researches， the superficial spreading depth of melanoma was quantitatively calculated. We consider that there is room for improvement in this study. Using microscopic hyperspectral imaging technology to identify and analyze superficial spreading melanoma samples is still a preliminary test， and the experimental results indicate the feasibility of this method. We have found that the thickness of granular layer in different samples is not exactly the same， and the adaptive ability of the filter operator in the experiment is insufficient. Further， we intend to verify on a large number of complex samples， so as to achieve the flexibility of MHSI technology for pathological diagnosis. Improving the accuracy of segmentation and measurement is also the direction of our efforts to developing method.