Method for acquiring skin tumor classification features

Abstract

The invention discloses a skin tumor classification feature acquisition method, and relates to the technical field of image recognition. Comprising the following steps: acquiring near infrared spectrum data and ultrasonic images of skin tumors, constructing a near infrared spectrum feature vector set and an ultrasonic image feature vector set, and generating skin tumor classification indexes and mixed skin tumor classification indexes of different feature vector sets through a regression model; constructing a multi-modal data fusion feature vector set, and screening the multi-modal data fusion feature vector set through a machine learning algorithm to obtain a screened feature set; and performing multi-factor regression analysis on the screening feature set to obtain a consistency analysis result range, and taking the screening features of which the consistency analysis result range is greater than a set threshold value as classification features. According to the method, manual key feature comparison and recording are not needed, and subjectivity and low efficiency of manual feature comparison can be avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method for acquiring classification features of skin tumors. Background Art

[0002] Skin tumors are common dermatological disorders involving cell proliferation in the skin. These neoplasms, which arise within or beneath the skin, vary widely and are clinically categorized as either benign or malignant. Malignant tumors can proliferate, metastasize, and become life-threatening, often referred to as skin cancer. The diagnosis of skin tumors is a systematic, multi-step process that requires comprehensive analysis of the patient's medical history, physical examination, and ancillary test results to arrive at an accurate diagnosis and develop an effective treatment plan. Identifying key features through multiple tests is crucial for skin tumor classification, but this process is extremely time-consuming and labor-intensive.

[0003] Currently, near-infrared spectroscopy (NIR) is used to classify skin tumors by detecting differences in the absorption and scattering properties of tumor tissue for near-infrared light (700-2500 nm), revealing its biochemical composition (such as hemoglobin, water content, and lipids) and structural characteristics. The main steps include data acquisition, preprocessing, feature extraction, and classification. Preprocessing primarily includes denoising, baseline correction, and normalization. Feature extraction methods include feature dimensionality reduction, wavelength selection, and optical parameter extraction. Classification utilizes traditional machine learning models such as support vector machines (SVMs) and random forest models.

[0004] Ultrasound image classification of skin tumors relies on the use of high-frequency ultrasound (20-100 MHz) to reveal the morphological characteristics of skin tumors (boundaries, internal echoes, blood flow signals, etc.). The steps include image preprocessing, feature extraction, and classification. Image preprocessing includes denoising, contrast enhancement, and tumor region of interest (ROI) segmentation. Feature extraction includes morphological features, texture features, and deep learning features. Classification models primarily use traditional methods, such as support vector machines (SVMs) or radio frequency (RF) algorithms combined with handcrafted features.

[0005] The above existing methods require manual key feature comparison and fusion after obtaining the near-infrared spectrum classification results and the ultrasound image classification results. In this process, the functional information of the near-infrared spectrum and the structural information of the ultrasound image need to be considered, and the efficiency of manual comparison and fusion is low. Summary of the Invention

[0006] Based on this, it is necessary to provide a method for obtaining skin tumor classification features to address the above technical issues.

[0007] An embodiment of the present invention provides a method for acquiring skin tumor classification features, comprising: Acquire near-infrared spectroscopy data and ultrasound images of skin tumors; Key features are extracted from the near-infrared spectrum data of the skin tumor to obtain a set of near-infrared spectrum feature vectors; the ultrasound image of the skin tumor is encoded and represented, and key features are extracted to obtain a set of ultrasound image feature vectors; Based on the near-infrared spectrum feature vector set and the ultrasound image feature vector set, a skin tumor classification index of near-infrared features, a skin tumor classification index of ultrasound image features, and a mixed skin tumor classification index of near-infrared features and ultrasound image features were generated through regression models. A multimodal data fusion feature vector set is constructed by vector concatenation based on a near-infrared feature vector set, an ultrasound image feature vector set, a skin tumor classification index of near-infrared features, a skin tumor classification index of ultrasound image features, and a mixed skin tumor classification index of near-infrared features and ultrasound image features. The multimodal data fusion feature vector set was screened using a machine learning algorithm to determine the importance and correlation between each multimodal data fusion feature vector and the skin tumor classification, thereby obtaining a screening feature set. A multifactor regression analysis was performed on the screening feature set to determine the mapping relationship and correlation between each screening feature and the skin tumor classification, thereby obtaining the range of consistency analysis results. The screening features whose consistency analysis result range was greater than the set threshold were used as classification features for skin tumors.

[0008] Optionally, preprocessing the near-infrared spectrum data of the skin tumor to obtain a fitting curve graph of the near-infrared spectrum data; Sort the near-infrared spectral data of skin tumors and use the maximum value in the near-infrared spectral data as the peak feature; The target detection method is used to detect the peak area and trough area in the near-infrared spectrum fitting curve, and the peaks and troughs are counted to obtain the number of peaks, the number of troughs, the maximum value of the peaks, and the minimum value of the troughs; Calculate the difference between the maximum value of each peak and the minimum value of each trough as the difference feature vector of the near-infrared spectrum of the skin tumor; Calculate the curvature of each peak region and trough region in the skin tumor near-infrared spectrum curve as the regional curve curvature feature vector of the skin tumor near-infrared spectrum data; A near-infrared spectrum feature vector set is constructed based on peak characteristics, number of peaks, number of troughs, difference feature vectors and regional curve curvature feature vectors.

[0009] Optionally, preprocessing the near-infrared spectrum data of the skin tumor to obtain a fitting curve graph of the near-infrared spectrum data specifically includes: Baseline correction of near-infrared spectral data was performed using a quadratic polynomial to remove baseline offset or drift; The near-infrared spectral data is subjected to noise removal and smoothing through convolution operation; The near-infrared spectral data were fitted by a long short-term memory network to obtain a smooth and non-overlapping near-infrared spectral curve.

[0010] Optionally, the ultrasound image of the skin tumor is encoded and represented, and key features are extracted to obtain a set of ultrasound image feature vectors, specifically including: Extract skin tumor area and skin tumor boundary contour in ultrasound image through automatic segmentation model; Determining the morphological characteristics of the skin tumor according to the boundary contour of the skin tumor; the morphological characteristics include: regular contour, smooth contour, and infiltrative growth contour; Extract the internal echo characteristics of the skin tumor area to obtain the internal echo uniformity characteristics; Extract the rear echo characteristics of the skin tumor area to obtain the internal echo calcification characteristics and the rear echo attenuation or enhancement characteristics; The blood flow signal characteristics of the skin tumor area are extracted to obtain the blood flow signal richness characteristics, blood flow signal regularity characteristics and blood flow resistance index characteristics; Calculate the depth and size of the skin tumor area to obtain the depth characteristics and tumor size; The skin tumor area is identified through a deep learning model to obtain the characteristics of lymph node enlargement; An ultrasound image feature vector set is constructed based on morphological features, internal echo uniformity, internal echo calcification, rear echo attenuation or enhancement, blood flow signal richness, blood flow signal regularity, blood flow resistance index, depth, tumor size, and lymph node enlargement.

[0011] Optionally, a skin tumor classification index of near-infrared features, a skin tumor classification index of ultrasound image features, and a mixed skin tumor classification index of near-infrared features and ultrasound image features are generated respectively by regression models, specifically including: The near-infrared spectral feature vector set is used as the input of the regression model, and the regression function in the regression model is used to perform regression fitting on different near-infrared spectral features to obtain the skin tumor classification index of the near-infrared features; The ultrasound image feature vector set is used as the input of the regression model, and the regression function in the regression model is used to perform regression fitting on different ultrasound image features to obtain the skin tumor classification index of the ultrasound image features; The near-infrared spectrum feature vector set and the ultrasound image feature vector set are concatenated into a feature vector set, which is used as the input of the regression model. The regression function in the regression model is used to perform regression fitting on the feature vectors in the feature vector set to obtain a hybrid skin tumor classification index of near-infrared features and ultrasound image features.

[0012] Optionally, the multimodal data fusion feature vector set is screened by a machine learning algorithm to obtain a screened feature set, specifically including: The multimodal data fusion feature vector set is input into the machine learning model to determine the importance and correlation between each multimodal data fusion feature vector and skin tumor classification, and the multimodal data fusion feature vectors with importance and correlation greater than the set threshold are used as screening features to construct a screening feature set.

[0013] Optionally, a multivariate regression analysis is performed on the screening feature set based on the following formula to determine the mapping relationship and correlation between each screening feature and the skin tumor classification, and obtain the consistency analysis result range: ICC=(MS R -MS W ) / (MS R +(k-1)MS W ); Among them, MS R is the mean square between groups of the screening feature set, MS W is the within-group mean square of the screening feature set, k is the number of scoring criteria, and ICC is the range of consistency analysis results; When the range of the consistency analysis results is greater than the set threshold, the screening feature has a high contribution to the skin tumor classification; when the range of the consistency analysis results is less than the set threshold, the screening feature has a low contribution to the skin tumor classification.

[0014] The method for obtaining skin tumor classification features provided by the embodiment of the present invention has the following beneficial effects compared with the prior art: The present invention realizes the automation of the entire process from data acquisition to decision output through automated feature extraction and model-driven fusion mechanism; through feature vector concatenation and machine learning model training, the functional information of the near-infrared spectrum is deeply coupled with the structural information of the ultrasound image, eliminating the need for manual key feature comparison and recording, avoiding the subjectivity and inefficiency of manual feature comparison, and improving the representation accuracy of skin tumor classification features, the mapping accuracy of key features of skin tumor classification, and the consistency of selection.

[0015] In addition, the present invention introduces a regression model to generate a multi-dimensional classification index, realizing quantitative feature fusion, which is more objective and repeatable than manual qualitative analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a method for obtaining skin tumor classification features provided in one embodiment; Figure 2This is a flowchart of a method for obtaining skin tumor classification features provided in one embodiment; Figure 3 A schematic diagram of near-infrared spectrum feature vector extraction for a method for acquiring skin tumor classification features provided in one embodiment; Figure 4 A schematic diagram of ultrasound image feature extraction of a method for acquiring skin tumor classification features provided in one embodiment. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] In one embodiment, a method for obtaining skin tumor classification features is provided, such as Figure 1 As shown, the method includes: Acquire near-infrared spectroscopy data and ultrasound images of skin tumors; Key features are extracted from the near-infrared spectrum data of skin tumors to obtain a set of near-infrared spectrum feature vectors. The ultrasound image of the skin tumor is encoded and key features are extracted to obtain a set of ultrasound image feature vectors.

[0019] Based on the near-infrared spectrum feature vector set and the ultrasound image feature vector set, the skin tumor classification index of near-infrared features, the skin tumor classification index of ultrasound image features, and the mixed skin tumor classification index of near-infrared features and ultrasound image features are generated respectively through regression models.

[0020] According to the near-infrared feature vector set, the ultrasound image feature vector set, the skin tumor classification index of the near-infrared feature, the skin tumor classification index of the ultrasound image feature, and the mixed skin tumor classification index of the near-infrared feature and the ultrasound image feature, a multimodal data fusion feature vector set is constructed by vector concatenation.

[0021] A machine learning algorithm was used to filter the multimodal data fusion feature vector set to obtain a screening feature set. Multivariate regression analysis was performed on this screening feature set to determine the mapping relationship and correlation between each screening feature and skin tumor classification. The consistency analysis result range was obtained, and screening features with a consistency analysis result range greater than a set threshold were used as skin tumor classification features.

[0022] like Figure 2 As shown, the specific implementation is: 1. Obtain near-infrared spectral data and ultrasound images of skin tumors.

[0023] 2. Preprocess the near-infrared spectral data of skin tumors and extract key features. Figure 3 As shown, the specific steps are:

[0024] 2.1 Read the collected skin tumor near-infrared spectral data; 2.2 Use quadratic polynomials to perform baseline correction on skin tumor near-infrared spectroscopy data to remove baseline offset or drift caused by instrument and other factors; 2.3 Using convolution operations, remove noise and smooth the near-infrared spectral data of skin tumors; 2.4 Using a long short-term memory (LSTM) network to fit the skin tumor near-infrared spectral data, a smooth, non-overlapping near-infrared spectral curve was obtained; 2.5 Sort the pre-processed skin tumor near-infrared spectral data and obtain the maximum value as the peak feature Peak of the skin tumor near-infrared spectral data IR ; 2.6 Save the near-infrared spectrum curve obtained by preprocessing in step 2.4; 2.7 Use target detection methods (such as the YOLO model) to detect multiple peaks and troughs in the near-infrared spectrum curve of step 2.4, and count the number of peaks and troughs as the peak number feature NP of skin tumors max and trough number characteristic NP min , and simultaneously record the near-infrared spectrum preprocessing data results of each peak area and trough area, as well as the maximum value Peak of each peak max i and the minimum value of the trough Peak min j , i=1,…,NP max , j = 1, ..., NP min ; 2.8 Calculate the maximum value of each peak max i and the minimum value of the trough Peak min j The difference between △P i,j =Peak max i -Peak min j , as the difference feature vector △P of skin tumor near-infrared spectral data; 2.9 Based on the fitted data in step 2.4, calculate the curvature of each peak region and trough region in the skin tumor near-infrared spectrum curve as the regional curve curvature feature vector C of the skin tumor near-infrared spectrum data; 2.10 Through the above steps, we get the near-infrared spectrum feature vector set Peak IR NP max NP min , △P and C.

[0025] 3. Encode the ultrasound image of the skin tumor through a pre-trained model (such as ResNet network, DenseNet network, etc.) and extract key features. Figure 4 As shown, the specific steps are:

[0026] 3.1 Use the automatic segmentation model to extract the skin tumor area in the ultrasound image and the boundary contour of the skin tumor in the ultrasound image; 3.2 Determine the morphological characteristics of the skin tumor boundary contour in the extracted ultrasound image, typically regular or irregular (R), smooth or rough edges (S), and whether there is invasive growth (G); 3.3 Extract the internal echo characteristics of the skin tumor area in the ultrasound image, generally including whether the internal echo is uniform or uneven U and whether there are calcification points So; 3.4 Extract the back echo characteristics of the skin tumor area in the ultrasound image, generally whether the back echo is attenuated or enhanced Ub; 3.5 Extract the blood flow signal characteristics of the skin tumor area in the ultrasound image, generally including whether the blood flow signal is rich (Br), whether the blood flow distribution is regular (Bg), and whether the blood flow resistance index is high or low (Bn); 3.6 Calculate the depth and size of the skin tumor area in the extracted ultrasound image. Generally, compare whether the depth of the skin tumor area is deep or shallow (D) and whether the size of the skin tumor area is consistent (Size); 3.7 Use a deep learning model to determine whether there is lymph node metastasis in the skin tumor area in ultrasound images and automatically find whether there is an enlarged lymph node area Lz; 3.8 Obtain the ultrasound image feature vector set R, S, G, U, So, Ub, Br, Bg, Bn, D, Size, and Lz through steps 3.2-3.7.

[0027] 4. Using the regression model, the features extracted by the feature extraction method in steps 2-3 above are used as input, and the regression function is used to regress and fit the different features to generate skin tumor classification indices for different data. The specific steps are as follows: 4.1 Using the regression model, the near-infrared spectral feature vector set obtained in step 2.10 is mapped into the skin tumor classification index BM of the near-infrared feature through the regression function. IR ; 4.2 Using the regression model, the ultrasound image feature vector set obtained in step 3.8 is mapped into the skin tumor classification index BM of the ultrasound image features through the regression function. UF ; 4.3 Using the regression model, the near-infrared spectrum feature vector set obtained in step 2.10 and the ultrasound image feature vector set obtained in step 3.8 are concatenated into a feature vector set. Through the regression function, they are mapped into a mixed skin tumor classification index BM of near-infrared features and ultrasound image features. IRUF .

[0028] 5. Build a feature screening model. The specific steps are: 5.1 Peak the acquired skin tumor near-infrared feature vector set IR NP max NP min , △P, C, ultrasound image feature vector set R, S, G, U, So, Ub, Br, Bg, Bn, D, Size, Lz and the generated classification index BM IR BM UF BM IRUF The multimodal data fusion feature vector set F={F i} i=1,…,20 ; 5.2 Input the multimodal data fusion feature vector set F into the machine learning model and calculate each multimodal data fusion feature vector F i Importance and relevance to skin tumor classification; 5.3 Fusion feature vector F based on each multimodal data i The top 10 multimodal data fusion feature vectors are obtained by screening based on the dual indicators of importance and relevance, and a screening feature set is constructed. These 10 multimodal data fusion feature vectors are the top 5 features ranked by importance {F i i1 ,F i i2 ,F i i3 ,F i i4 ,F i i5} and the top 5 features with the highest correlation {F i c1 ,F i c2 ,F i c3 ,F i c4 ,F i c5}, and these features are not repeatable, and the feature names in the multimodal data fusion feature vector F are output through the mapping relationship, which is Peak IR NP max NP min ,△P,C,R,S,G,U,So,Ub,Br,Bg,Bn,D,Size,Lz,BM IR BM UF BM IRUF Any 10 feature names in .

[0029] 6. Through the formula ICC=(MS R -MS W ) / (MS R +(k-1)MS W ) Perform multivariate regression analysis on the screening feature set to determine the mapping relationship and correlation between each screening feature and skin tumor classification, obtain the consistency analysis result range, and use the screening features with a consistency analysis result range greater than the set threshold as the classification features of skin tumors. MS R is the mean square between groups of the screening feature set, MS W is the within-group mean square of the screening feature set, k is the number of scoring criteria, and ICC is the range of consistency analysis results.

[0030] Set the ICC threshold for the consistency analysis results, for example, T = 0.75 ± 0.05. If the ICC range between the feature and the skin tumor classification is 0.7-0.8, the feature has a high contribution to the skin tumor classification and is selected. Otherwise, the feature has a low contribution to the skin tumor classification and should be discarded or has no reference value.

[0031] A specific embodiment includes: 1. Obtain near-infrared spectral data and ultrasound images of skin tumors; 2. Preprocess the acquired near-infrared spectral data of skin tumors, including reading the collected near-infrared spectral data; performing baseline correction on the near-infrared spectral data using a quadratic polynomial; removing noise and smoothing the near-infrared spectral data using a convolution operation; fitting the near-infrared spectral data using a long short-term memory network; and saving the fitted curve graph of the preprocessed near-infrared spectral data of skin tumors. 3. Extract the features of the pre-processed near-infrared spectral data of skin tumors, including obtaining the maximum feature Peak of the near-infrared spectral data of skin tumors IR ; Use target detection methods (such as the YOLO model) to detect multiple peaks and troughs in the fitting curve graph, and count the number of peaks and troughs as characteristic NPs of skin tumors max and NPmin , and simultaneously record the near-infrared spectrum preprocessing data results of each peak area and trough area, as well as the maximum value Peak of each peak max i and the minimum value of the trough Peak min j , i=1,…,NP max , j = 1, ..., NP min ; Calculate the maximum value of each peak Peak max i and the minimum value of the trough Peak min j The difference between △P i,j =Peak max i -Peak min j , as the characteristic vector △P of the skin tumor; calculate the curvature of the curve of each peak area and trough area as the characteristic vector C of the skin tumor; obtain the skin tumor near-infrared spectrum characteristic vector Peak IR NP max NP min , △P and C; 4. Encode the ultrasound image of the skin tumor and extract key features, including automatic segmentation and extraction of the skin tumor area and the boundary contour of the skin tumor in the ultrasound image; determine the morphological characteristics of the boundary contour of the skin tumor in the ultrasound image, usually regular or irregular R, smooth or rough edge S, whether there is invasive growth G; extract the internal echo characteristics of the skin tumor area in the ultrasound image, usually uniform or uneven internal echo U, and whether there are calcification points So; extract the rear echo characteristics of the skin tumor area in the ultrasound image, usually whether the rear echo is attenuated or enhanced Ub; extract the skin tumor in the ultrasound image. The blood flow signal characteristics of the tumor area are generally whether the blood flow signal is rich Br, whether the blood flow distribution is regular Bg, and the high or low blood flow resistance index Bn; the depth and size of the skin tumor area in the extracted ultrasound image are calculated, generally comparing whether the depth of the skin tumor area is deep or shallow D, and comparing whether the size of the skin tumor area is consistent Size; using a deep learning model to determine whether there is lymph node metastasis in the skin tumor area in the ultrasound image, and automatically find whether there is an enlarged lymph node area Lz; obtain the skin tumor ultrasound image feature vector R, S, G, U, So, Ub, Br, Bg, Bn, D, Size and Lz; 5. Generate skin tumor classification indexes for different data through regression models, including the skin tumor classification index BM based on near-infrared features IR ; Skin tumor classification index BM based on ultrasound image characteristics UF ; Hybrid skin tumor classification index BM based on near-infrared features and ultrasound image features IRUF ; 6. Construct a feature screening model, including obtaining the skin tumor near-infrared feature vector Peak IR NP max NP min ,△P,C, ultrasonic image feature vector R,S,G,U,So,Ub,Br,Bg,Bn,D,Size,Lz,BM IR BM UF and BM IRUF Constructing a multimodal data fusion feature vector set; inputting the multimodal data fusion feature vector set into a machine learning model, and calculating the importance of each feature and the correlation between the skin tumor classification; screening the top 10 multimodal data fusion feature vectors based on the dual indicators of importance and correlation, and outputting feature names through mapping relationships to obtain a filtered feature set; 7. Through the formula ICC=(MS R -MS W ) / (MS R +(k-1)MS W ) Perform multivariate regression analysis on the screening feature set to determine the mapping relationship and correlation between each screening feature and the skin tumor classification, obtain the range of consistency analysis results, and use the screening features with a consistency analysis result range greater than the set threshold as the classification features of skin tumors.

[0032] The main advantage of this invention is that it acquires near-infrared spectroscopy and ultrasound images non-invasively, alleviating patient pain. Through automated feature extraction and a model-driven fusion mechanism, full-process automation from data acquisition to decision output is achieved. Through feature vector concatenation, machine learning models, and multivariate regression analysis, the functional information of the near-infrared spectrum is deeply coupled with the structural information of the ultrasound image, eliminating the need for manual key feature comparison and recording. This avoids the subjectivity and inefficiency of manual feature comparison, improves the representation accuracy of skin tumor classification features, the mapping accuracy of key features for skin tumor classification, and the consistency of selection.

[0033] In addition, the present invention introduces a regression model to generate a multi-dimensional classification index, realizing quantitative feature fusion, which is more objective and repeatable than manual qualitative analysis.

[0034] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A method for obtaining skin tumor classification features, characterized in that: include: Acquire near-infrared spectroscopy data and ultrasound images of skin tumors; Key features are extracted from the near-infrared spectrum data of the skin tumor to obtain a set of near-infrared spectrum feature vectors; the ultrasound image of the skin tumor is encoded and represented, and key features are extracted to obtain a set of ultrasound image feature vectors; Based on the near-infrared spectrum feature vector set and the ultrasound image feature vector set, a skin tumor classification index of near-infrared features, a skin tumor classification index of ultrasound image features, and a mixed skin tumor classification index of near-infrared features and ultrasound image features were generated through regression models. A multimodal data fusion feature vector set is constructed by vector concatenation based on a near-infrared feature vector set, an ultrasound image feature vector set, a skin tumor classification index of near-infrared features, a skin tumor classification index of ultrasound image features, and a mixed skin tumor classification index of near-infrared features and ultrasound image features. The multimodal data fusion feature vector set was screened using a machine learning algorithm to obtain a screening feature set. Multivariate regression analysis was performed on the screening feature set to determine the mapping relationship and correlation between each screening feature and the skin tumor classification, and the range of consistency analysis results was obtained. The screening features with a consistency analysis result range greater than the set threshold were used as the classification features of skin tumors.

2. The method for obtaining skin tumor classification characteristics according to claim 1, wherein: The key features of the near-infrared spectrum data of the skin tumor are extracted to obtain a set of near-infrared spectrum feature vectors, which specifically includes: Preprocessing the near-infrared spectrum data of skin tumors to obtain a fitting curve graph of the near-infrared spectrum data; Sort the near-infrared spectral data of skin tumors and use the maximum value in the near-infrared spectral data as the peak feature; The target detection method is used to detect the peak area and trough area in the near-infrared spectrum fitting curve, and the peaks and troughs are counted to obtain the number of peaks, the number of troughs, the maximum value of the peaks, and the minimum value of the troughs; Calculate the difference between the maximum value of each peak and the minimum value of each trough as the difference feature vector of the near-infrared spectrum of the skin tumor; Calculate the curvature of each peak region and trough region in the skin tumor near-infrared spectrum curve as the regional curve curvature feature vector of the skin tumor near-infrared spectrum data; A near-infrared spectrum feature vector set is constructed based on peak characteristics, number of peaks, number of troughs, difference feature vectors and regional curve curvature feature vectors.

3. The method for obtaining skin tumor classification characteristics according to claim 2, wherein: The preprocessing of the near-infrared spectrum data of the skin tumor to obtain a fitting curve diagram of the near-infrared spectrum data specifically includes: Baseline correction of near-infrared spectral data was performed using a quadratic polynomial to remove baseline offset or drift; The near-infrared spectral data is subjected to noise removal and smoothing through convolution operation; The near-infrared spectral data were fitted by a long short-term memory network to obtain a smooth and non-overlapping near-infrared spectral curve.

4. The method for obtaining skin tumor classification characteristics according to claim 1, wherein: The ultrasound image of the skin tumor is encoded and represented, and key features are extracted to obtain a set of ultrasound image feature vectors, specifically including: Extract skin tumor area and skin tumor boundary contour in ultrasound image through automatic segmentation model; Determining the morphological characteristics of the skin tumor according to the boundary contour of the skin tumor; the morphological characteristics include: regular contour, smooth contour, and infiltrative growth contour; Extract the internal echo characteristics of the skin tumor area to obtain the internal echo uniformity characteristics; Extract the rear echo characteristics of the skin tumor area to obtain the internal echo calcification characteristics and the rear echo attenuation or enhancement characteristics; The blood flow signal characteristics of the skin tumor area are extracted to obtain the blood flow signal richness characteristics, blood flow signal regularity characteristics and blood flow resistance index characteristics; Calculate the depth and size of the skin tumor area to obtain the depth characteristics and tumor size; The skin tumor area is identified through a deep learning model to obtain the characteristics of lymph node enlargement; An ultrasound image feature vector set is constructed based on morphological features, internal echo uniformity, internal echo calcification, rear echo attenuation or enhancement, blood flow signal richness, blood flow signal regularity, blood flow resistance index, depth, tumor size, and lymph node enlargement.

5. The method for obtaining skin tumor classification characteristics according to claim 1, wherein: The step of generating a skin tumor classification index based on near-infrared features, a skin tumor classification index based on ultrasound image features, and a mixed skin tumor classification index based on near-infrared features and ultrasound image features through the regression model specifically includes: The near-infrared spectral feature vector set is used as the input of the regression model, and the regression function in the regression model is used to perform regression fitting on different near-infrared spectral features to obtain the skin tumor classification index of the near-infrared features; The ultrasound image feature vector set is used as the input of the regression model, and the regression function in the regression model is used to perform regression fitting on different ultrasound image features to obtain the skin tumor classification index of the ultrasound image features; The near-infrared spectrum feature vector set and the ultrasound image feature vector set are concatenated into a feature vector set, which is used as the input of the regression model. The regression function in the regression model is used to perform regression fitting on the feature vectors in the feature vector set to obtain a hybrid skin tumor classification index of near-infrared features and ultrasound image features.

6. The method for obtaining skin tumor classification characteristics according to claim 1, wherein: The multimodal data fusion feature vector set is screened by a machine learning algorithm to obtain a screened feature set, specifically including: The multimodal data fusion feature vector set is input into the machine learning model to determine the importance and correlation between each multimodal data fusion feature vector and skin tumor classification, and the multimodal data fusion feature vectors with importance and correlation greater than the set threshold are used as screening features to construct a screening feature set.

7. The method for obtaining skin tumor classification characteristics according to claim 1, wherein: A multivariate regression analysis was performed on the screening feature set based on the following formula to determine the mapping relationship and correlation between each screening feature and skin tumor classification, and to obtain the range of consistency analysis results: ICC=(MS R -MS W ) / (MS R +(k-1)MS W ); Among them, MS R is the mean square between groups of the screening feature set, MS W is the within-group mean square of the screening feature set, k is the number of scoring criteria, and ICC is the range of consistency analysis results; When the range of the consistency analysis results is greater than the set threshold, the screening feature has a high contribution to the skin tumor classification; when the range of the consistency analysis results is less than the set threshold, the screening feature has a low contribution to the skin tumor classification.

Images