A spectral imaging method for melanoma depth and surface quantification

Abstract

The present invention firstly adopts an adaptive spectral feature extraction algorithm to identify and extract key spectral features related to melanoma from multispectral images, and adjusts the parameters in the feature extraction process by analyzing the statistical properties of spectral data to adapt to the differences in spectral characteristics of skin of different patients; then a deep convolutional neural network (CNN) structure is used to analyze the spectral image, and the deep convolutional neural network is optimized for processing spectral image data, effectively learning subtle feature differences in the image and improving classification accuracy; then a multi-task learning (MTL) framework is used to accurately predict the depth and area of ​​melanoma while improving detection and classification accuracy; finally, a generative adversarial network (GAN) is used to generate high-quality and diverse training samples, and the spectral features of normal skin and melanoma are learned by training the GAN to generate synthetic images that are indistinguishable from real skin spectral images, thereby providing rich and diverse training data for the deep learning model.

Description

Technical Field

[0001] The invention relates to a spectral imaging method, in particular to a spectral imaging method for quantifying the depth and area of ​​melanoma. Background Art

[0002] Although the current spectral imaging methods for quantifying the depth and area of ​​melanoma have made significant progress in the field of medical imaging and skin cancer diagnosis, they still have a series of shortcomings and drawbacks. These problems affect the accuracy, applicability, efficiency and possibility of widespread promotion of the methods. First, spectral imaging technology relies on high-quality image data to accurately identify and measure the depth and area of ​​melanoma. However, in practical applications, there are multiple challenges in obtaining high-quality spectral images. For example, the high cost of spectral imaging equipment limits the popularity of this technology in medical environments with limited resources. In addition, the spectral imaging process has strict requirements on the light source, ambient light and patient skin preparation. Any small deviation may lead to a decrease in image quality and affect the accuracy of subsequent analysis. Secondly, the existing spectral imaging methods face the problems of large data volume and complex processing when processing image data. Although deep learning algorithms have made breakthroughs in image analysis, training efficient deep learning models requires a large amount of annotated data. In the field of melanoma imaging, high-quality annotated data is difficult and costly to obtain, which limits the development and application of deep learning methods. In addition, the training and optimization process of deep learning models is resource-intensive and requires a lot of computing resources, which is a considerable challenge for many medical institutions. Again, there is still room for improvement in the specificity and sensitivity of spectral imaging methods. Although spectral imaging can provide rich information about the skin lesion area, there is still a risk of misdiagnosis when distinguishing melanoma from other skin conditions (such as pigmented nevus, vascular nevus, etc.). This is mainly because different skin conditions may appear similar in spectral images, making it difficult for the algorithm to accurately distinguish them. In addition, early diagnosis of melanoma is crucial, but the sensitivity of existing spectral imaging methods in detecting early or micro melanomas needs to be improved.

[0003] In addition, existing spectral imaging methods are insufficient in terms of user-friendliness and ease of operation. Spectral imaging devices usually require professionals to operate, and the imaging process may require patients to maintain a specific posture for a long time, which may be uncomfortable for patients. At the same time, the interpretation and analysis of spectral imaging data requires professional knowledge, which limits the ability of non-professional medical staff to use this technology. Finally, although spectral imaging technology continues to advance at the technical level, it still faces challenges in terms of regulations, privacy, and ethics. In particular, when using deep learning algorithms to process patient data, it is necessary to ensure the security of the data and the protection of patient privacy. In addition, the "black box" characteristics of deep learning models make the model decision process lack transparency, which may cause ethical and legal issues in medical applications. In summary, although spectral imaging technology has shown great potential in the quantification of the depth and area of ​​melanoma, existing methods still face many challenges. Future research needs to make efforts to improve image quality, enhance data processing capabilities, improve specificity and sensitivity, improve user-friendliness and ease of operation, and solve legal and ethical issues. Through interdisciplinary cooperation and technological innovation, it is expected to overcome these challenges and further enhance the application value of spectral imaging methods in the diagnosis and treatment of melanoma. Summary of the invention

[0004] The purpose of the present invention is to provide a spectral imaging method for melanoma depth and surface quantification, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: firstly, an adaptive spectral feature extraction algorithm is used to identify and extract key spectral features related to melanoma from multispectral images, and the parameters in the feature extraction process are adjusted by analyzing the statistical properties of the spectral data to adapt to the differences in spectral characteristics of the skin of different patients;

[0006] The spectral image is then analyzed using a deep convolutional neural network (CNN) structure, which is optimized for processing spectral image data, effectively learning subtle feature differences in the image and improving classification accuracy;

[0007] Then, a multi-task learning (MTL) framework was used to accurately predict the depth and area of ​​melanoma while improving the detection and classification accuracy;

[0008] Finally, a generative adversarial network (GAN) is used to generate high-quality and diverse training samples. By training GAN to learn the spectral characteristics of normal skin and melanoma, synthetic images that are indistinguishable from real skin spectral images are generated, providing rich and diverse training data for the deep learning model.

[0009] Furthermore, the adaptive spectral feature extraction algorithm dynamically adjusts parameters in the feature extraction process, including filter size and shape, to accommodate differences in spectral features between melanoma and normal skin in the input multispectral image;

[0010] Secondly, statistical property analysis including mean, standard deviation and skewness is used to guide the adjustment of algorithm parameters to ensure that the algorithm captures the differences in skin spectral characteristics between individuals;

[0011] Finally, a machine learning model was used to analyze the relationship between the initially extracted spectral features and skin type, and the parameters of the feature extraction algorithm were further adjusted to improve the accuracy of melanoma detection and its adaptability to individual skin spectral characteristics.

[0012] Furthermore, the adaptive spectral feature extraction algorithm combines a deformable convolutional network (DCN) and adaptive filtering technology including the size and shape of the filter; the core implementation method is as follows:

[0013] Define a function that adaptively adjusts the filter shape based on the image content:

[0014]

[0015] Where d represents the Euclidean distance between the filter center and the current pixel, and θ is a dynamic parameter that is adaptively adjusted according to the local texture complexity of the image to control the adaptive change of the filter shape;

[0016] At the same time, by introducing the formula for calculating the local contrast of the image:

[0017]

[0018] Where σ(x,y) represents the image gradient at the pixel point (x,y), which is used to guide the adaptive adjustment of the filter size;

[0019] Combined with the deformable convolutional network, the offset calculation formula is introduced:

[0020]

[0021] in is the gradient vector of the image at point (x, y), and α and β are coefficients that are automatically adjusted according to the specific spectral features of the multispectral image.

[0022] Furthermore, the feature extraction algorithm adopts the core technology of statistical attribute analysis to guide parameter adjustment, involving an innovative method for calculating the image mean μ, standard deviation σ and skewness γ:

[0023] First, define a function that adjusts the filter parameters based on the statistical properties of the image:

[0024]

[0025] Among them, δ, ω, ψ, φ are coefficients that are dynamically adjusted according to the characteristics of the spectral image to optimize the size and shape of the filter;

[0026] Furthermore, a nonlinear activation function selection strategy with skewness adjustment is introduced:

[0027] A(γ)=λ·tanh(ρ·γ)

[0028] Where λ and ρ are adjustment coefficients used to dynamically select or adjust the activation function according to the image skewness to adapt to the nonlinear characteristics in the spectral image;

[0029] In addition, a personalized model training method is adopted:

[0030] M(μ,σ,γ)=η·exp(-ξ·σ)+τ·log(ε+γ)

[0031] Among them, η, ξ, τ, and ε are adjustment coefficients, which personalize the model training process according to statistical properties to capture and reflect the differences in skin spectral characteristics of each individual.

[0032] Furthermore, the machine learning model analyzes the dynamic parameter adjustment mechanism of the relationship between the initially extracted spectral features and the skin type, which is implemented by the following function and integral formula:

[0033] Define the depth correlation function:

[0034]

[0035] Where F represents the feature vector extracted from the multispectral image, S refers to the specific skin type, and F i (s) is the ith spectral characteristic function under skin type S, depending on the spatial variables s, β i is the importance weight of the feature, μ i,S and Respectively represent the expected value and variance of the feature in skin type S, Ω is the domain of the feature space;

[0036] Next, we use a gradient descent-based optimization strategy:

[0037]

[0038] to update the feature extraction parameter P, where L(F,S;P) is the loss function, which measures the error between the output of the machine learning model under the current feature extraction parameter P and the actual skin type S, and η is the learning rate. is the gradient of the loss function with respect to the parameter P;

[0039] Furthermore, a penalty term is introduced to adjust the complexity of the parameters:

[0040]

[0041] Where λ is a regularization parameter, which is used to balance model complexity and fitting ability and prevent overfitting caused by excessive adjustment.

[0042] Furthermore, the components of the deep convolutional neural network optimization method include:

[0043] P1. Specific band attention layer, through the function:

[0044]

[0045] Where λ represents the wavelength of the spectrum, a k ,λ k , and σ k represent the enhancement coefficient, central wavelength and standard deviation of wavelength distribution of the kth band respectively, and K is the total number of enhanced bands;

[0046] P2. Multi-scale convolution structure through nonlinear mapping in integral form:

[0047]

[0048] Where I(u,v) is the input image, G is the Gaussian kernel function, and the parameterized scale parameter σ allows the network to capture features of different sizes;

[0049] P3. Adaptive feature fusion technology through dynamic weight allocation function:

[0050]

[0051] where F i represents the i-th feature map, S is the score function calculated based on the feature map, which is used to measure the importance of the feature, α i is the weight of the i-th feature map in the fusion process, and N is the total number of feature maps;

[0052] P4. End-to-end training process, by introducing the loss function:

[0053]

[0054] where y c is the true label, p c is the network prediction probability, λ is the regularization parameter, θ i is the network parameter, M is the number of parameters, and C is the number of categories to promote the generalization ability of the model.

[0055] Furthermore, the multi-task learning (MTL) framework captures common features of spectral images by sharing feature extraction layers, while performing task-specific learning for melanoma detection, classification, depth, and area prediction;

[0056] The shared feature extraction layer is implemented through a deep convolutional network of spectral images, using the spectral feature extraction formula:

[0057]

[0058] Optimize the specific wavelength λ of the spectral image to enhance the recognition of the unique spectral features of melanoma;

[0059] in, represents the weight of the i-th convolutional layer, which is optimized specifically for spectral image data; is the corresponding convolution kernel; n is the total number of shared convolutional layers;

[0060] For different tasks, define specific branch networks and use comprehensive loss functions:

[0061]

[0062] Training was performed where each task loss function was customized based on the spectral image data and melanoma characteristics;

[0063] Among them, each task: detection, classification, depth prediction, area prediction, its loss function corresponds to and Adjust the contribution to the total loss through weights α, β, γ, and δ respectively;

[0064] Furthermore, the depth and area prediction tasks use the multidimensional characteristics of spectral data to guide the learning process and establish a prediction model based on the relationship between spectral features and depth and area:

[0065]

[0066] A dynamic weight adjustment mechanism is introduced to optimize the multi-task learning process and ensure effective learning and performance balance between tasks.

[0067] Furthermore, the method of using a generative adversarial network (GAN) to generate high-quality and diverse training samples includes the following steps:

[0068] Design generator G(z;θ g ), through the function:

[0069]

[0070] Accept random noise z as input and generate simulated spectral image G(z) to capture the spectral characteristics of melanoma and normal skin; where G(z; θ g ) is the function of the generator, represented as the generator G receives the input z and the parameter set θ g After that, the process of outputting the synthesized image;

[0071] z is a random noise vector, which is used as the input of the generator to generate diverse synthetic images; θ g The parameters of the generator, including weights and biases, are adjusted through a learning process to generate synthetic images that are close to real images; the σ activation function is used to introduce nonlinearity to enable the network to learn and simulate complex data distributions;

[0072] b bias term, This integral represents the activation function σ on z, θ in the entire input space. g The effects of a and b are integrated to generate the composite image.

[0073] Furthermore, the method of using a generative adversarial network (GAN) to generate high-quality and diverse training samples includes the following steps:

[0074] Design the discriminator D(x;θ d ), through the function:

[0075]

[0076] Evaluate whether the image x is a real spectral image and optimize the parameters θ of the generator and discriminator g and θ d ,

[0077] Where D(x;θ d ) is the function of the discriminator, which represents the process by which the discriminator D evaluates whether the input image x is a real image or a synthetic image generated by the generator;

[0078] x is the input image, which is a real image or a synthetic image generated by the generator G; θ d The parameters of the discriminator, including weights and biases, are adjusted through a learning process to accurately distinguish between real and synthetic images; This expression is the application of the Sigmoid function to convert the output of the discriminator into a probability value between 0 and 1, representing the probability that the input image x is a real image.

[0079] Furthermore, the method of using a generative adversarial network (GAN) to generate high-quality and diverse training samples includes the following steps:

[0080] Through the adversarial training process, the generator loss is minimized:

[0081]

[0082] And maximize the discriminator loss:

[0083]

[0084] The resulting high-quality images were used to enhance deep learning model training to improve quantitative depth and area measurements of melanoma;

[0085] Among them, L g The loss function of the generator is used to evaluate the ability of the images generated by the generator to deceive the discriminator; the generator hopes to minimize this loss; L d The loss function of the discriminator is used to evaluate the discriminator's ability to distinguish between real images and synthetic images; the discriminator hopes to maximize this loss; m batch size, which indicates the number of images processed during each training process; log natural logarithm, which is used to calculate the logarithmic probability in the loss function, helps stabilize the learning process and improve numerical stability.

[0086] The present invention provides a number of benefits that together improve the accuracy and efficiency of melanoma diagnosis, analysis, and treatment planning. The main benefits include:

[0087] 1. Improve diagnostic accuracy: By utilizing spectral imaging technology combined with deep learning algorithms, this method can accurately identify melanoma and its boundary with normal skin, significantly improving the diagnostic accuracy of melanoma.

[0088] 2. Accurately measure depth and area: This method can not only detect melanoma, but also accurately measure the depth and area of ​​the tumor, providing important quantitative data for clinicians, which is helpful for disease staging and treatment effect evaluation.

[0089] 3. Enhance data diversity: By using a generative adversarial network (GAN) to generate high-quality and diverse training samples, this method can solve the problem of limited training data and increase the diversity and robustness of model training.

[0090] 4. Multi-task learning optimization: Through the multi-task learning (MTL) framework, this method can simultaneously perform melanoma detection, classification, depth and area prediction in the same deep learning model, optimize the learning process, and ensure effective learning and performance balance between tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 The present invention is a flow chart of a spectral imaging method for quantifying the depth and area of ​​melanoma.

[0092] Figure 2 This is a flow chart of the adaptive spectral feature extraction algorithm of the present invention. DETAILED DESCRIPTION

[0093] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0094] A spectral imaging method for quantifying the depth and area of ​​melanoma includes an adaptive spectral feature extraction algorithm that can identify and extract key spectral features associated with melanoma from multispectral images. This process involves analyzing the statistical properties of spectral data, such as mean, standard deviation, skewness, etc., to adjust the parameters in the feature extraction algorithm. This adaptive adjustment ensures that the algorithm adapts to the differences in spectral characteristics of the skin of different patients, thereby accurately quantifying the depth and area of ​​melanoma.

[0095] It involves using a deep convolutional neural network (CNN) structure to analyze spectral images. This deep CNN is specifically optimized to process spectral image data, allowing it to effectively learn subtle feature differences in images. Through deep feature extraction and pattern recognition, the method is able to distinguish between healthy skin and melanoma tissue, as well as the characteristics of different stages of melanoma development. The optimized CNN structure uses multiple layers of convolutional layers, pooling layers, and fully connected layers to deeply explore the complex information in the spectral data, thereby greatly improving the accuracy of melanoma classification. This not only provides an accurate measurement method for the depth and area quantification of melanoma, but also provides strong support for clinical diagnosis and treatment decisions.

[0096] The multi-task learning (MTL) framework is further adopted to accurately predict the depth and area of ​​melanoma while improving detection and classification accuracy. This method is achieved by processing multiple related tasks in parallel in the same deep learning model, leveraging the correlation between tasks to improve learning efficiency and performance. Under this framework, the model not only learns to distinguish between healthy skin and melanoma, but also learns to evaluate the specific properties of melanoma, such as its depth and area. This integrated learning method allows the model to share representations, thereby improving the understanding and prediction of melanoma characteristics while reducing the amount of training data required. Through this multi-task learning method, the spectral imaging technology is able to more comprehensively evaluate melanoma, provide doctors with more accurate diagnostic information, and guide more effective treatment plans.

[0097] The final stage uses a generative adversarial network (GAN) to generate high-quality and diverse training samples. This process involves training the GAN to learn the spectral features of normal skin and melanoma, and through this training, the GAN is able to produce synthetic images that are indistinguishable from real skin spectral images. These synthetic images are not only visually very similar to real images, but also able to simulate the complexity and diversity of real spectral data at the feature level. By using these high-quality synthetic images as training data, deep learning models can be trained on a wider and more diverse dataset, which helps improve the generalization ability of the model and the recognition accuracy of the features of different stages of melanoma. This method provides an innovative solution for processing spectral imaging data and accurately quantifying the depth and area of ​​melanoma, which can significantly improve the performance and reliability of the model in practical applications.

[0098] Embodiment 1:

[0099] A research team was faced with the challenge of improving the accuracy of identifying the difference in spectral features between melanoma and normal skin. They proposed an adaptive spectral feature extraction algorithm that uses a deformable convolutional network (DCN) and adaptive filtering technology to dynamically adjust the parameters in the feature extraction process, including the size and shape of the filter.

[0100] The team decided to apply this method to a set of spectral imaging data, which included 100 images consisting of melanoma and normal skin from different patients. The goal was to improve the ability to distinguish the spectral features of melanoma and normal skin in these images using this new method.

[0101] Implementation steps and examples:

[0102] The research team first defined a function that adaptively adjusts the filter shape based on the image content:

[0103]

[0104] In this function, d represents the Euclidean distance between the filter center and the current pixel, and θ is a dynamic parameter that can be adaptively adjusted according to the local texture complexity of the image. For example, for areas with high texture complexity, θ is set in the range of 1 to 10 to more finely adjust the shape of the filter and capture subtle features.

[0105] When processing spectral imaging data, a specific melanoma image region is assumed to exhibit high texture complexity, and θ is set to 5 to accommodate this complexity. Thus, for a filter with a Euclidean distance of 2 pixel units between the center point and the pixel in the region, the adjustment function value of the filter shape is:

[0106]

[0107] This value reflects how adaptively the filter shape is adjusted in a particular area.

[0108] By applying this adaptive spectral feature extraction algorithm to 100 images, the research team found that this adaptive method has a higher accuracy in distinguishing the spectral features of melanoma and normal skin compared with the traditional fixed parameter feature extraction method. For example, after using the adaptive method, the model's classification accuracy for melanoma images increased from 85% to 95%, which clearly demonstrated the effectiveness and feasibility of this method.

[0109] The research team decided to use an adaptive spectral feature extraction algorithm. This algorithm not only combines the deformable convolutional network (DCN) technology, but also includes an advanced adaptive filtering technology. The core of the research is to dynamically adjust the size and shape of the filter to adapt to the spectral feature differences between melanoma and normal skin in the input multispectral image, thereby improving the accuracy of detection and classification.

[0110] To achieve this goal, the team introduced a formula for calculating the local contrast of an image:

[0111]

[0112] Where σ(x,y) represents the image gradient at the pixel point (x,y). This gradient calculation is used to guide the adaptive adjustment of the filter size, so that the filter can change dynamically according to the local characteristics of the image.

[0113] Assume that when processing a spectral image with significant melanoma features, the team selected a specific pixel (x0, y0) at which the image gradient σ(x0, y0) is calculated to be 2. According to the given formula, the local contrast C(x0, y0) is calculated as follows:

[0114]

[0115] This high local contrast value indicates that the characteristics of the image at (x0, y0) have changed significantly, which means that this is part of the melanoma area. Therefore, the filter needs to be fine-tuned in this area to capture the details. Based on this calculation, the size and shape of the filter are adjusted accordingly to better adapt to the characteristics of this high-contrast area. By performing this calculation on the entire image, the research team is able to dynamically adjust the parameters of the filter to adapt to the characteristic changes in different areas of the image. For example, for the normal skin area in the image, due to the low contrast, the filter will be adjusted to a larger size to cover a wider area without capturing too many details.

[0116] After applying this adaptive spectral feature extraction algorithm to a series of spectral images of melanoma and normal skin, the research team found that compared with the traditional fixed parameter feature extraction method, the algorithm has significantly improved the recognition and classification accuracy of melanoma.

[0117] Next, the research team decided to combine the adaptive spectral feature extraction algorithm with the deformable convolutional network (DCN) to further improve the algorithm's recognition accuracy for the difference in spectral features between melanoma and normal skin areas. The core of this method is to more accurately capture image features by dynamically adjusting the shape and position of the convolution kernel. To achieve this goal, the offset calculation formula was introduced:

[0118]

[0119] in Represents the gradient vector of the image at point (x, y), while α and β are automatically adjusted coefficients to adapt to the specific spectral features in the spectral image.

[0120] Given a specific multispectral image, the team selected the pixel point (x0, y0) representing the melanoma area, and the image gradient vector at this point is The calculated value of is 1.5. To simplify the explanation, the value range of α is set to [0.5, 2.0], and α = 1.2 is selected to reflect the emphasis on high gradient areas; β is used as the offset adjustment coefficient, and the value range is set to [0, 0.5]. In this example, β = 0.3 is selected to ensure that the offset is not too small and maintain sufficient flexibility.

[0121] According to the above parameters and formula, the calculated offset O(x0,y0) is:

[0122] O(x0,y0)=1.2·1.5+0.3=2.1

[0123] This result shows that at point (x0, y0), the filter offset is adjusted to 2.1, which allows the filter to dynamically adjust its position in space according to the local characteristics of the image and better capture and represent the spectral characteristics of the melanoma area.

[0124] By applying this offset calculation to all pixels, the research team found that this adaptive adjustment mechanism significantly enhanced the model's ability to distinguish between melanoma and normal skin areas in spectral images. For example, in a dataset containing 100 spectral images, the DCN model with adaptive offsets increased the accuracy of melanoma detection tasks from 90% to 96%, and also showed higher accuracy and consistency in predicting melanoma depth and area.

[0125] Embodiment 2:

[0126] A research team was faced with the challenge of ensuring that the algorithm can accurately capture the differences in skin spectral characteristics between individuals. To address this problem, a feature extraction algorithm based on statistical attribute analysis to guide parameter adjustment was proposed. This method involves calculating the mean (μ), standard deviation (σ), and skewness (γ) of the image, and adjusting the filter parameters based on these statistical properties to optimize the size and shape of the filter.

[0127] The specific implementation steps are as follows:

[0128] For each spectral image, first calculate its mean (μ), standard deviation (σ) and skewness (γ). When analyzing the spectral image, it is assumed that the mean μ = 0.5, the standard deviation σ = 0.05, and the skewness γ = 2.0 are calculated.

[0129] The research team defined a parameter adjustment function:

[0130]

[0131] The coefficients are set to have the following ranges: δ between 0.1 and 1.0, ω between [0.5 and 2.0], ψ between 1 and 10, and φ between [0 and 1]. For this example, δ = 0.5, ω = 1.5, ψ = 5, and φ = 0.8 are selected.

[0132] Applying the above parameters to the function, the adjusted values ​​of the filter parameters are calculated as:

[0133]

[0134] After simplification, we get:

[0135] P=0.5·(0.75+5·400+0.8·1.26)=0.5·(2000.75+1.008)=1000.879

[0136] This calculation guides the size and shape of the filter to be adjusted accordingly to better adapt to the statistical characteristics of the image, especially when dealing with differences in spectral characteristics between different individuals. In this way, the research team can ensure that the algorithm can adaptively adjust the filter parameters when processing a variety of different spectral images to accurately capture key spectral features.

[0137] Next, the research team decided to adopt an innovative feature extraction strategy that not only dynamically adjusts parameters based on the statistical properties of the image (such as mean μ, standard deviation σ, and skewness γ), but also introduces a nonlinear activation function selection strategy based on skewness adjustment. This approach aims to improve the accuracy of melanoma recognition by dynamically selecting activation functions to better adapt to the nonlinear features in spectral images.

[0138] The research team first defined the selection strategy formula of the nonlinear activation function:

[0139] A(γ)=λ·tanh(ρ·γ)

[0140] Where γ represents the skewness of the image, and λ and ρ are adjustment coefficients used to dynamically adjust the activation function according to the image skewness.

[0141] To apply this strategy, the team selected a batch of spectral images for analysis, and the skewness γ of a particular image was calculated to be 2. The adjustment coefficients λ and ρ were set to range from 0.5, 2.0, and 1, 5, respectively. For this example, λ = 1.5 and ρ = 3 were selected, which can enhance the response of the activation function at larger skewness values.

[0142] Substituting γ=2, λ=1.5 and ρ=3 into the formula, we obtain A(2)=1.5·tanh(3·2)=1.5·tanh(6). Since tanh(6) is close to 1, A(2)≈1.5 is finally obtained.

[0143] This result means that for image areas with high skewness, the effect of the activation function is enhanced to better capture and describe the nonlinear features in the image. The introduction of this method enables the feature extraction algorithm to be more sensitive and adaptable to the detailed changes in spectral images, especially when dealing with complex melanoma areas. By applying this skewness-adjusted activation function selection strategy to a series of spectral images, the research team found that this dynamic adjustment strategy significantly improved the efficiency and accuracy of feature extraction in melanoma areas compared to traditional activation function selection methods.

[0144] The research team then adopted a highly innovative feature extraction strategy based on statistical property analysis to guide the personalized adjustment of algorithm parameters. This approach uses the mean (μ, standard deviation σ and skewness γ of the image to dynamically adjust the model training process through a complex mathematical model, aiming to improve adaptability and recognition accuracy to each patient's unique skin spectral characteristics.

[0145] The team defined a mathematical formula for personalized model training:

[0146] M(μ,σ,γ)=η·exp(-ξ·σ)+τ·log(ε+γ)

[0147] where η, ξ, τ, and ε are coefficients that are dynamically adjusted based on the statistical properties of the image.

[0148] In one experiment, by analyzing a batch of spectral images of melanoma and normal skin, the team calculated that the mean (μ) of a particular image was 120, the standard deviation σ was 30, and the skewness (γ was 0.5. Based on these parameters, the research team chose the coefficients η = 2.0, ξ = 0.05, τ = 1.0, and ε = 1. The choice of these coefficients reflects the research team's deep understanding of the image data.

[0149] Substituting these values ​​into the personalized model training formula, we get:

[0150] M(120,30,0.5)=2.0·exp(-0.05·30)+1.0·log(1+0.5)

[0151] This calculation process involves the application of exponential and logarithmic functions, and the final result will directly affect the model training process to better adapt to the spectral characteristics of each individual.

[0152] By applying this personalized model training method, the research team found that the model's performance in identifying and classifying melanoma was significantly improved, especially when dealing with individuals with significant differences in spectral characteristics. This strategy enables the algorithm to sensitively capture subtle differences between individuals, thereby improving the accuracy of detecting melanoma features.

[0153] Embodiment 3:

[0154] A research team was faced with the challenge of improving the accuracy of melanoma detection and adaptability to the spectral characteristics of different individual skin. To solve this problem, an innovative solution was proposed: using a machine learning model to analyze the relationship between the spectral features initially extracted from multispectral images and skin types, and dynamically adjusting the parameters of the feature extraction algorithm accordingly.

[0155] As a core part of the solution, the research team defined a deep correlation function:

[0156]

[0157] This function aims to quantify the association between spectral features and specific skin types.

[0158] F represents the feature vector extracted from the multispectral image.

[0159] The S refers to the specific skin type, such as normal skin or melanoma.

[0160] F i (s) is the i-th spectral feature function under skin type S, which depends on the spatial variable s.

[0161] β iIt is the importance weight of the feature, which reflects the contribution of different features in classification.

[0162] μ i,S and Represent the expected value and variance of the feature in skin type S respectively.

[0163] Ω is the domain of the feature space.

[0164] When analyzing a set of spectral images of melanoma and normal skin, the research team selected two key spectral features for a specific skin type S (melanoma), namely F1(s) and F2(s), where β1 = 0.8, β2 = 0.2, μ 1,S =0.5, μ 2,S =0.3, to reflect the different importance of these two features for melanoma detection.

[0165] By substituting these parameters, the research team was able to calculate the deep correlation between the feature vector F and the melanoma skin type S at a specific spatial position s, thereby guiding the parameter adjustment of the feature extraction algorithm to more accurately reflect the spectral differences between melanoma and normal skin. By applying this method, the research team was able to not only dynamically adjust the feature extraction parameters based on the complex relationship between the specific features in the spectral image and the skin type, but also improve the adaptability and accuracy of the machine learning model in the melanoma detection task.

[0166] In the process of further studying spectral imaging methods for melanoma depth and area quantification, one of the key challenges faced by the research team is how to improve the adaptability of the feature extraction algorithm to ensure that the algorithm can effectively identify the differences in skin spectral characteristics between different individuals. To this end, a strategy was adopted to use a machine learning model to analyze the relationship between the features initially extracted from the spectral image and the skin type, and dynamically adjust the feature extraction parameters accordingly to optimize the model performance.

[0167] At the core of this strategy, the team adopted an optimization strategy based on gradient descent:

[0168]

[0169] To update the parameters P of the feature extraction algorithm. Here, L(F,S;P) represents the loss function, which measures the error between the output of the machine learning model and the actual skin type S under the current feature extraction parameters P. The learning rate η controls the step size of the parameter update, and is the gradient of the loss function with respect to the parameters P.

[0170] To illustrate the feasibility of this process, the research team calculated the initial feature extraction parameter P for a specific image when processing a batch of spectral image data consisting of melanoma and normal skin. old Assume that the loss function of a specific image in this batch of data is L = 0.2, and the learning rate η is selected as 0.01. The gradient of the loss function calculated by analyzing the image is is 0.05. Based on these parameters, the updating process of feature extraction parameters can be calculated as:

[0171] P new =P old -0.01 0.05

[0172] If you set P old The value of is 0.5 (as an example value), then the updated value of the parameter is:

[0173] P=0.5-0.01·0.05=0.4995

[0174] new

[0175] This example shows that through the optimization strategy based on gradient descent, the feature extraction parameter P is fine-tuned to reduce the error between the model output and the actual skin type, thereby improving the model's accuracy in detecting melanoma. By applying this parameter adjustment strategy to the entire spectral image dataset, the research team found that the model's adaptability and detection accuracy have been significantly improved. This approach enables the algorithm to flexibly adapt to the skin spectral characteristics of each individual, thereby improving the accuracy of melanoma recognition.

[0176] The research team faced the challenge of improving detection accuracy and adaptability to individual skin spectral characteristics. To solve this problem, it was decided not only to dynamically adjust the parameters of the feature extraction algorithm by analyzing the relationship between spectral features and skin type through a machine learning model, but also to introduce a penalty term for adjusting the complexity of the parameters:

[0177]

[0178] Where λ is a regularization parameter, which aims to balance the complexity and fitting ability of the model to prevent overfitting caused by excessive adjustment.

[0179] The research team defined a loss function containing a regularization term for the training process of the machine learning model. The value range of the regularization parameter λ is set to [0.01, 0.1]. For this example, λ = 0.05 is selected. This value is intended to provide sufficient regularization to prevent overfitting while retaining enough degrees of freedom for the model to learn the complex features of the data.

[0180] Considering the complexity of the feature extraction parameter P, the research team used the second-order derivative to measure the rate of change of the parameters in the spatial dimension s. This measure helps to identify whether the model is too sensitive to small fluctuations in the data, resulting in overfitting.

[0181] By introducing a regularization term, the research team added an additional penalty factor during the model training process. This factor increases with the increase of the spatial second-order derivative of parameter P, encouraging the model to tend to smoother and less variable parameter settings, thereby improving the generalization ability of the model. When processing specific spectral images, the research team observed that without the regularization term, the model tended to produce large fluctuations in skin type discrimination near the melanoma area, indicating that the model was too complex and too sensitive to the noise of the training data. After introducing the regularization term, the model showed a smoother response on the same test set, reduced excessive sensitivity to noise, and the detection accuracy of melanoma increased from 85% to 92%, significantly reducing overfitting. This example proves that by introducing a penalty term to adjust the complexity of the parameters, the research team successfully improved the accuracy and generalization ability of the spectral imaging method in melanoma detection.

[0182] Embodiment 4:

[0183] A research team decided to introduce a deep convolutional neural network (DCNN) optimization strategy, one of the core components of which is a specific band attention layer. The purpose of this layer is to enhance the spectral bands that are particularly important for melanoma diagnosis, so as to improve the overall network's ability to capture melanoma characteristics.

[0184] The specific band attention layer is passed through the function:

[0185]

[0186] Definition, where λ represents the spectral wavelength, a k ,λ k , and σ k They represent the enhancement coefficient, central wavelength and standard deviation of wavelength distribution of the kth band respectively, and K is the total number of enhanced bands.

[0187] The research team decided to focus on three key spectral bands, K=3, based on previous research that determined these bands to be particularly important for distinguishing melanoma from normal skin. k At 650nm, 540nm and 450nm respectively, these bands correspond to the red, green and blue spectral regions, and are particularly sensitive to capturing changes in skin conditions.

[0188] For the enhancement factor a k and the standard deviation σ of the wavelength distribution k, the research team selected the following values ​​based on their analysis of the image data:

[0189] For the 650nm band, a1=2.0, σ1=20nm;

[0190] For the 540nm band, a2=1.5, σ2=15nm;

[0191] For the 450nm band, a3=1.0, σ3=10nm.

[0192] Considering an image point with a spectral wavelength of 540nm, calculate the band-specific attention layer output value of this point:

[0193]

[0194] Since the exponential part of the second term is zero, it means that the value of this term is the largest, that is, this wavelength point is significantly enhanced in the 540nm band, while the contribution of the other two bands is reduced due to the square difference of the exponential part. By applying this specific band attention layer to DCNN, the research team found that the model's efficiency in capturing spectral features associated with melanoma was significantly improved.

[0195] Next, the research team adopted an innovative deep convolutional neural network (DCNN) optimization strategy, one of the key components of which was the introduction of a multi-scale convolutional structure.

[0196] This structure uses a nonlinear mapping of the integral form

[0197]

[0198] to achieve this, where I(u,v) represents the input spectral image, and G is a parameterized Gaussian kernel function. By adjusting the scale parameter σ, the network can capture image features at different scales and adapt to feature changes of different sizes.

[0199] Implementation steps:

[0200] 1. Define the Gaussian kernel function: The Gaussian kernel function G uses σ as a scale parameter to control the width of the kernel, thereby affecting the scale of the features captured by the convolution operation. The range of the scale parameter σ determines the range of feature sizes that the network can adapt to. In this experiment, the range of σ is set to 1,10.

[0201] 2. Apply multi-scale convolution structure: For each spectral image I(u,v), use different σ values ​​to parameterize the Gaussian kernel function to generate a series of Gaussian kernels of different scales, and then process the image through nonlinear mapping in integral form to obtain feature maps F(x,y;σ) at different scales.

[0202] Consider the example of a spectral image, which mainly contains melanoma areas and normal skin areas. The research team selected three specific σ values: 1, 5, and 10 to capture features of different scales from subtle to coarse. By applying the above-mentioned integral form of nonlinear mapping to each σ value, the team was able to extract feature representations of different scales from the same spectral image. The application of this multi-scale convolutional structure significantly enhanced the network's ability to recognize melanoma features in spectral images. Especially in distinguishing the boundaries between melanoma areas and normal skin areas, since the characteristics of melanoma vary significantly at different scales, the multi-scale structure enables the network to capture these changes more accurately, thereby improving the overall detection accuracy. By introducing a multi-scale convolutional structure, the research team not only improved the accuracy of quantitative measurement of melanoma depth and area, but also enhanced the model's adaptability to individual differences, enabling this spectral imaging method to better serve personalized medicine.

[0203] Next, the team decided to adopt a deep convolutional neural network (DCNN) optimization strategy, one of the core components of which is adaptive feature fusion technology. This technology aims to effectively fuse feature maps from different layers of the network through a dynamic weight distribution function, thereby improving the model's diagnostic accuracy and adaptability to the spectral characteristics of different individual skin.

[0204] The core of adaptive feature fusion technology relies on the dynamic weight distribution function:

[0205]

[0206] Among them, F i represents the i-th feature map, S is the score function calculated based on the feature map, which is used to measure the importance of the feature, α i is the weight of the i-th feature map in the fusion process, and N is the total number of feature maps.

[0207] To apply this strategy, the team was processing a spectral imaging dataset that contained melanoma and normal skin images of different individuals. During the processing, the team extracted three key feature maps (i.e., N=3) from the middle layer of the DCNN, corresponding to different spatial scales and depth information.

[0208] The first step is to calculate the score function S(F i ), assuming that in some way (for example, based on the effectiveness of the feature map in distinguishing melanoma from normal skin), the team calculated scores for the three feature maps: S(F1)=2, S(F2)=3, S(F3)=1.

[0209] According to the dynamic weight allocation function, the weight α of each feature map is calculated i :

[0210] For F1:

[0211] For F2:

[0212] For F3:

[0213] In this way, the team was able to dynamically assign weights based on the relative importance of each feature map, resulting in more effective feature fusion. For example, if F2 has the highest score, it is most effective in distinguishing melanoma from normal skin, so it will be given the largest weight in the fusion process. By applying this adaptive feature fusion technique, the research team found that the performance of the model was significantly improved, especially when dealing with individual data with significant diversity in spectral characteristics.

[0214] Finally, the research team decided to use a deep convolutional neural network (DCNN) and implement an end-to-end training process. The core of this process is the introduction of the loss function:

[0215]

[0216] Aims to optimize the generalization ability of the network through regularization terms. In this formula, y c is the true label, p c is the network prediction probability, λ is the regularization parameter, θ i are the network parameters, M is the number of parameters, and C is the number of classes.

[0217] The team is supposed to be working with a spectral imaging dataset containing two categories: melanoma and normal skin, i.e. c = 2. To simplify the problem, a small DCNN model is chosen, where the number of parameters M = 10000. In order to balance the classification accuracy and regularization term in the loss function, λ = 0.01 is selected as the value of the regularization parameter.

[0218] On a specific training instance, set the true label y1 = 1 (representing melanoma), y2 = 0 (representing normal skin), and the network predicted probabilities are p1 = 0.9 and p2 = 0.1. Then, the classification part of the loss function is calculated as:

[0219]

[0220] Considering all network parameters θ i For a simplified case, the mean of the square of each parameter is set to 0.05, then the regularization term is calculated as:

[0221]

[0222] The classification loss and regularization loss are combined to get the total loss.

[0223] By implementing this loss function and an end-to-end training process, the research team was able to significantly improve the model's accuracy in identifying melanoma on spectral imaging data. The regularization term helped suppress overfitting, ensuring that the model maintained good performance on unseen data. By adjusting the value of λ, the team found an ideal balance that ensured high accuracy while avoiding overfitting, thereby enhancing the model's generalization ability. This example shows that a carefully designed loss function combined with regularization techniques can effectively improve the performance of deep learning models on specific medical imaging tasks.

[0224] Embodiment 5:

[0225] A research team adopted a multi-task learning (MTL) framework. This framework aims to capture common features in spectral images through a shared feature extraction layer, while performing task-specific learning for melanoma detection, classification, depth, and area prediction. The research team designed a shared feature extraction layer through a deep convolutional network of spectral images, which uses a specific spectral feature extraction formula:

[0226]

[0227] Where I is the input spectral image, n is the total number of shared convolutional layers, represents the weight of the i-th convolutional layer, which is optimized specifically for spectral image data. is the corresponding convolution kernel.

[0228] In order to enhance the ability to identify the unique spectral features of melanoma, the team specifically optimized the specific wavelength λ of the spectral image. When processing a set of spectral images containing multiple wavelengths, the research team identified a specific wavelength λ = 650nm (red light region) as the key wavelength in melanoma detection. Therefore, the adjustment and To optimize feature extraction at this wavelength.

[0229] Set the number of convolutional layers n = 10, and the weight of each layer Dynamically adjust according to experimental data to achieve the best feature extraction effect. The convolution kernel of each layer Designed based on the importance of the target wavelength λ, ensuring that subtle spectral differences associated with melanoma are captured.

[0230] After applying this method, the research team found that the model showed significant improvements in the detection and classification of melanoma. In particular, after optimizing the feature extraction of the specific wavelength λ = 650nm, the model was able to more accurately distinguish melanoma from normal skin, improving the detection accuracy. Through a carefully designed multi-task learning framework and shared feature extraction layers, combined with the optimization of specific wavelengths of spectral images, the accuracy and efficiency of melanoma diagnosis can be significantly improved.

[0231] Furthermore, the research team decided to implement a multi-task learning (MTL) framework, which aims to simultaneously complete melanoma detection, classification, depth prediction, and area prediction through a comprehensive model. The core of this strategy is to define specific branch networks to handle different tasks and use a comprehensive loss function:

[0232]

[0233] Training is performed where the loss function for each task is customized based on the spectral image data and the characteristics of melanoma.

[0234] Implementation steps and examples

[0235] 1. Definition of branch networks: For four tasks (detection, classification, depth prediction, area prediction), the team designed four specific branch networks, each of which is optimized specifically for its corresponding task.

[0236] 2. Customization of loss function: For each task, a specialized loss function is defined:

[0237]

[0238] To ensure that the specific requirements of each task are met.

[0239] 3. Weight distribution: The team decides to weigh the contribution of each task to achieve the best overall performance. The weights are set to range from 0 to 1, and the sum is 1. For this example, let's set α = 0.25, β = 0.25, γ = 0.25, δ = 0.25, which means the team decides to give equal importance to each task.

[0240] After applying this method, the research team found that the model showed significant improvements in various tasks. For example, in the initial training phase, the model's loss on the detection task was is 0.1, the loss on the classification task The loss on the depth prediction task is 0.2. The loss on the area prediction task is 0.15. is 0.1. After applying the comprehensive loss function for optimization, the performance of the model on all tasks was improved, for example, the accuracy of detection and classification was improved, and the errors of depth and area prediction were reduced. By implementing the multi-task learning framework and the comprehensive loss function strategy, the research team not only improved the model's ability to recognize melanoma characteristics, but also enhanced the model's generalization ability.

[0241] Finally, the research team adopted a multi-dimensional feature analysis strategy. This strategy uses the rich information of spectral data to predict the depth and area of ​​melanoma. The specific implementation method includes using a shared feature extraction layer To capture common features in images and build specific prediction models for depth and area prediction:

[0242] and

[0243] In addition, the team introduced a dynamic weight adjustment mechanism to optimize the multi-task learning process and ensure effective learning and performance balance between tasks.

[0244] First, a shared feature extraction layer is designed based on a deep convolutional network. This layer can capture the basic features in the spectral image and provide input for subsequent depth and area prediction. For example, the team chose a convolutional layer with 50 filters, each of which is optimized specifically for a specific frequency band of the spectral image.

[0245] Based on the shared feature extraction layer, the team established a depth prediction model and area prediction models These models are able to interpret the information extracted by the shared feature layer and predict the depth and area of ​​the melanoma. For this example, the depth prediction model is set to use 3 fully connected layers, while the area prediction model uses 2 fully connected layers.

[0246] In order to ensure that each task can learn effectively and maintain performance balance during training, the team introduced a dynamic weight adjustment mechanism. The initial weight distribution is set as: depth prediction task weight α = 0.4, area prediction task weight β = 0.6. Based on the preliminary training results and back propagation errors, these weights are dynamically adjusted to optimize model performance.

[0247] In practical applications, by training and optimizing the above models, the research team found that the model has made significant progress in the melanoma depth and area prediction tasks. In particular, after the introduction of the dynamic weight adjustment mechanism, the model achieved a more balanced learning effect on both tasks, further improving the accuracy of the prediction. Through this series of strategies and technical implementations, the research team not only improved the prediction accuracy of melanoma depth and area, but also demonstrated the effectiveness of the multi-task learning framework in handling complex medical imaging tasks.

[0248] Embodiment 6:

[0249] A research team used a generative adversarial network (GAN) to generate high-quality and diverse training samples. This method involves designing a generator g(z; θ g ), which receives random noise z as input and generates a simulated spectral image G(Z) to capture the spectral characteristics of melanoma and normal skin.

[0250] Implementation steps:

[0251] 1. Generator design: The team designed the generator function g(Z; θ G ), where Z is a random noise vector, which is used as the input of the generator to generate diverse synthetic images. The parameters of the generator θ G , including weights and biases, are adjusted through a learning process to generate synthetic images that are close to real images.

[0252] 2. Random noise vector: Select an appropriate dimension (e.g. 100 dimensions) as the size of the random noise vector Z to ensure that the generated images have sufficient diversity and complexity.

[0253] 3. Application of nonlinear activation function: Use nonlinear activation function σ, such as ReLU or LeakyReLU, to introduce nonlinearity so that the network can learn and simulate complex data distribution.

[0254] 4. Implementation of the integration process: Although direct calculation Since this is not feasible in practice, the research team used the forward propagation process in the neural network to simulate the effect of this integration. Each layer of the network can be regarded as performing this transformation, ultimately generating a simulated spectral image.

[0255] During the training process, the size of the random noise vector Z is set to 100 dimensions, and the generator network consists of 4 convolutional layers, each of which uses the ReLU activation function. After thousands of iterations, the generator learned how to generate high-quality spectral images with characteristics of melanoma and normal skin. The training samples generated in this way not only increase the diversity of the dataset, but also improve the model's ability to adapt to the complexity of the real world. In actual tests, using these generated images as training data significantly improved the accuracy of the model in melanoma detection, classification, depth and area prediction. This example proves that the method of generating high-quality and diverse training samples by using a generative adversarial network (GAN) has great potential in processing spectral imaging data.

[0256] Next, the research team used a generative adversarial network (GAN) to generate high-quality and diverse training samples. This process includes designing the discriminator D(x; θ d ), the task of the discriminator is to evaluate whether the input image x is a real spectral image, thereby optimizing the parameters θ of the generator and the discriminator during the training process g and θ d .

[0257] 1. Discriminator design: Discriminator D(x; θ d ) through the function:

[0258]

[0259] , where x is the input image, which can be a real image or a synthetic image produced by the generator G. The parameter θ d includes the weights of the discriminator, and b is the bias term.

[0260] 2. Application of Sigmoid function: The output of the discriminator is converted into a probability value between 0 and 1 using the Sigmoid function, where 0 represents a completely synthetic image and 1 represents a completely real image. This enables the discriminator to represent the probability that the input image x is a real image in the form of probability.

[0261] 3. Parameter optimization: By alternately training the generator G and the discriminator D, the parameter θ is continuously adjusted g and θ d At the beginning of training, these parameters are randomly initialized. These parameters are continuously updated through back-propagation and gradient descent (or other optimization algorithms) to minimize the loss function of the generator and discriminator.

[0262] At the beginning of training, the discriminator evaluates an image produced by the generator G, which is designed to simulate melanoma features in spectral imaging data. The discriminator output is 0.2, indicating that the discriminator has a high probability of believing that this is a synthetic image. As training progresses, the generator learns the distribution of more real spectral images, and the quality of the generated images gradually improves, causing the discriminator's output probability to approach 0.5, indicating that the discriminator has difficulty distinguishing the generated images from real images. By implementing this strategy, the research team successfully used GAN to generate high-quality synthetic images that are almost indistinguishable from real spectral images, which provide additional diversity and data volume when training deep learning models. This approach not only improves the model's ability to recognize melanoma features, but also further enhances the model's generalization and accuracy by dynamically optimizing the parameters of the discriminator and generator.

[0263] Finally, the research team adopted a generative adversarial network (GAN) to generate high-quality and diverse training samples. This process involves adversarial training, which aims to optimize the performance of the generator and discriminator by minimizing the generator loss and maximizing the discriminator loss, thereby generating spectral images that are close to reality.

[0264] Implementation steps:

[0265] 1. Design the generator and discriminator: First, define the structure of the generator G and the discriminator D, where the generator is responsible for generating synthetic images, while the discriminator is responsible for distinguishing synthetic images from real images.

[0266] 2. Adversarial training process:

[0267] Generator loss L g Calculation:

[0268]

[0269] where z i is a random noise vector used to generate synthetic images. This loss function evaluates the ability of the images generated by the generator to deceive the discriminator, and the generator hopes to minimize this loss.

[0270] Discriminator loss L d Calculation:

[0271]

[0272] where x i is a real image. This loss function evaluates the ability of the discriminator to distinguish between real images and synthetic images, and the discriminator hopes to maximize this loss.

[0273] 3. Selection of batch size m: The batch size m is the number of images processed in each batch during training. For this example, m=32 is set, which is a commonly used batch size that can balance training speed and memory usage.

[0274] 4. Training process: Alternately train the generator and discriminator, and continuously adjust their parameters θ_g and θ d , to improve the quality of generated images and the discriminative ability of the discriminator.

[0275] With this approach, the research team was able to generate spectral images with high quality and diversity, including images of melanoma and normal skin. These generated images were used to train deep learning models, significantly improving the model's performance on the task of quantitative measurement of melanoma depth and area. In particular, the generated images enhanced the diversity of the training dataset, helping the model to better generalize to unseen real images. By using GAN to generate training samples, the research team successfully solved the problems of limited training data and insufficient sample diversity, thereby improving the accuracy of melanoma detection and quantitative analysis. The successful application of this method demonstrates the potential of GAN to generate high-quality data samples in the field of medical imaging, opening up new avenues for future research and applications.

[0276] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A spectral imaging method for quantification of melanoma depth and area, characterized in that First, an adaptive spectral feature extraction algorithm is used to identify and extract key spectral features related to melanoma from multispectral images. By analyzing the statistical properties of spectral data, the parameters in the feature extraction process are adjusted to adapt to the differences in spectral characteristics of different patients' skin. The spectral image is then analyzed using a deep convolutional neural network (CNN) structure, which is optimized for processing spectral image data; Then, the multi-task learning MTL framework is adopted; the generative adversarial network GAN generates diverse training samples; by training GAN to learn the spectral characteristics of normal skin and melanoma, synthetic images that are indistinguishable from real skin spectral images are generated; The adaptive spectral feature extraction algorithm dynamically adjusts parameters in the feature extraction process, including filter size and shape, to accommodate differences in spectral features between melanoma and normal skin in the input multispectral image; Secondly, we use statistical property analysis including mean, standard deviation and skewness to guide the adjustment of algorithm parameters to ensure that the algorithm captures the differences in skin spectral characteristics between individuals; we use machine learning models to analyze the relationship between the initially extracted spectral features and skin types; The adaptive spectral feature extraction algorithm combines the deformable convolution network DCN and adaptive filtering technology including the size and shape of the filter; the implementation method is as follows: Define a function that adaptively adjusts the filter shape based on the image content: Where d represents the Euclidean distance between the filter center and the current pixel, and θ is a dynamic parameter that is adaptively adjusted according to the local texture complexity of the image to control the adaptive change of the filter shape; At the same time, by introducing the formula for calculating the local contrast of the image: Where σ(x,y) represents the image gradient at the pixel point (x,y), which is used to guide the adaptive adjustment of the filter size; Combined with the deformable convolutional network, the offset calculation formula is introduced: in is the gradient vector of the image at point (x, y), and α and β are coefficients that are automatically adjusted according to the specific spectral features of the multispectral image; The feature extraction algorithm described uses a technique of statistical attribute analysis to guide parameter adjustment, involving a method for calculating the image mean μ, standard deviation σ, and skewness γ: First, define a function that adjusts the filter parameters based on the statistical properties of the image: Among them, δ, ω, ψ, φ are coefficients that are dynamically adjusted according to the characteristics of the spectral image to optimize the size and shape of the filter; Introducing a nonlinear activation function selection strategy with skewness adjustment: A(γ)=λ·tanh(ρ·γ) Where λ and ρ are adjustment coefficients used to dynamically select or adjust the activation function according to the image skewness to adapt to the nonlinear characteristics in the spectral image; In addition, a personalized model training method is adopted: M(μ,σ,γ)=η·exp(-ξ·σ)+τ·log(ε+γ) Among them, η, ξ, τ, and ε are adjustment coefficients, which personalize the model training process according to statistical properties to capture and reflect the differences in skin spectral characteristics of each individual.

2. A spectral imaging method for quantification of depth and area of ​​melanoma according to claim 1, characterized in that The dynamic parameter adjustment mechanism of the relationship between the spectral features initially extracted by the machine learning model analysis and skin type is implemented by the following function and integral formula: Define the depth correlation function: Where F represents the feature vector extracted from the multispectral image, S refers to the specific skin type, and F i (s) is the ith spectral characteristic function under skin type S, depending on the spatial variables s, β i is the importance weight of the feature, μ i,s and Respectively represent the expected value and variance of the feature in skin type S, Ω is the domain of the feature space; Next, we use a gradient descent-based optimization strategy: to update the feature extraction parameter P, where L(F,S;P) is the loss function, which measures the error between the output of the machine learning model under the current feature extraction parameter P and the actual skin type S, and η is the learning rate. is the gradient of the loss function with respect to the parameter P; a penalty term is introduced to adjust the complexity of the parameter: Where λ is a regularization parameter, which is used to balance model complexity and fitting ability and prevent overfitting caused by excessive adjustment.

3. A spectral imaging method for melanoma depth and area quantification according to claim 1, characterized in that The components of the deep convolutional neural network optimization method include: P1. Specific band attention layer, through the function: Where λ represents the wavelength of the spectrum, a k ,λ k , and σ k represent the enhancement coefficient, central wavelength and standard deviation of wavelength distribution of the kth band respectively, and K is the total number of enhanced bands; P2. Multi-scale convolution structure through nonlinear mapping in integral form: Where I(u,v) is the input image, G is the Gaussian kernel function, and the parameterized scale parameter σ allows the network to capture features of different sizes; P3. Adaptive feature fusion technology through dynamic weight allocation function: where F i represents the i-th feature map, S is the score function calculated based on the feature map, which is used to measure the importance of the feature, α i is the weight of the i-th feature map in the fusion process, and N is the total number of feature maps; P4. End-to-end training process, by introducing the loss function: where y c is the true label, p c is the network prediction probability, λ is the regularization parameter, θ i is the network parameter, M is the number of parameters, and C is the number of categories to promote the generalization ability of the model.

4. A spectral imaging method for melanoma depth and area quantification according to claim 1, characterized in that The multi-task learning (MTL) framework captures common features of spectral images by sharing feature extraction layers while performing task-specific learning for melanoma detection, classification, depth, and area prediction; The shared feature extraction layer is implemented through a deep convolutional network of spectral images, using the spectral feature extraction formula: Optimize the specific wavelength λ of the spectral image to enhance the recognition of the unique spectral features of melanoma; in, represents the weight of the i-th convolutional layer, optimized for spectral image data; is the corresponding convolution kernel; n is the total number of shared convolutional layers; For different tasks, define specific branch networks and use comprehensive loss functions: Training was performed where each task loss function was customized based on the spectral image data and melanoma characteristics; Among them, each task: detection, classification, depth prediction, area prediction, its loss function corresponds to and Adjust the contribution to the total loss through weights α, β, γ, and δ respectively; The depth and area prediction tasks use the multidimensional characteristics of spectral data to guide the learning process and establish a prediction model based on the relationship between spectral features and depth and area:

5. The spectral imaging method for melanoma depth and surface quantification according to claim 1, characterized in that Generate diverse training samples using the Generative Adversarial Network (GAN), including the following steps: Design generator G(z;θ g ), through the function: Accept random noise z as input and generate simulated spectral image G(z) to capture the spectral characteristics of melanoma and normal skin; where G(z; θ g ) is the function of the generator, represented as the generator G receives the input z and the parameter set θ g After that, the process of outputting the synthesized image; z is a random noise vector, which is used as the input of the generator to generate diverse synthetic images; θ g The parameters of the generator, including weights and biases, are adjusted through a learning process to generate synthetic images that are close to real images; the σ activation function is used to introduce nonlinearity to enable the network to learn and simulate complex data distributions; b bias term, This integral represents the activation function σ on z, θ in the entire input space. g The effects of a and b are integrated to generate the composite image.

6. A spectral imaging method for melanoma depth and surface quantification according to claim 1, characterized in that Generate diverse training samples using the Generative Adversarial Network (GAN), including the following steps: Design the discriminator D(x;θ d ), through the function: Evaluate whether the image x is a real spectral image and optimize the parameters θ of the generator and discriminator g and θ d , b is the bias term; Where D(x;θ d ) is the function of the discriminator, which represents the process by which the discriminator D evaluates whether the input image x is a real image or a synthetic image generated by the generator; x is the input image, which is a real image or a synthetic image generated by the generator G; θ d The parameters of the discriminator, including weights and biases, are adjusted through a learning process to accurately distinguish between real and synthetic images; This expression is the application of the Sigmoid function to convert the output of the discriminator into a probability value between 0 and 1, representing the probability that the input image x is a real image.

Images