Port-wine stain segmentation method and device based on multi-color space adaptive fusion

Abstract

A method and device for PWS segmentation based on adaptive fusion of multiple color spaces overcomes the limitations of the single RGB color space and integrates the color features of human visual perception into the deep learning network, thereby improving the detection accuracy and robustness of PWS lesions, enhancing the baseline model's ability to distinguish color and texture, avoiding the mutual interference of multiple color space decisions, and promoting the synergistic effect of multiple color spaces on PWS segmentation. The method includes: (1) introducing multiple color spaces through the multi-color space detection MCSD method to enhance the representation of image color information; (2) using a Unet with dense dilated residual blocks to generate candidate results in different color spaces; (3) obtaining the final segmentation result through adaptive fusion of these candidate results; and (4) measuring the brightness, color, and contrast differences between the predicted lesions and the real lesions through structural similarity loss.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a port-wine stain segmentation method based on multi-color space adaptive fusion, and a port-wine stain segmentation device based on multi-color space adaptive fusion. Background Art

[0002] Port wine stains (PWS) are common vascular malformations of the skin, clinically manifested by a red or purple discoloration of the skin at the site of the lesion, unlike normal skin. They commonly affect the head and neck, but may also occur on the trunk and limbs. The incidence rate in newborns is 0.3% to 0.5%. Lesions darken and thicken with age and are difficult to resolve naturally without treatment. The disfiguring nature of PWS can have long-term consequences for patients' physical and mental health and quality of life, leading to negative emotions such as low self-esteem, anxiety, and depression. Vascular targeted photodynamic therapy (V-PDT) is recognized as a safe and effective treatment for PWS and has been successfully used in China for over 30 years. When treating PWS with V-PDT, physicians assess the vascular distribution based on diagnostic information such as lesion color and size, as well as color fading under pressure, to formulate an effective treatment plan. However, obtaining this diagnostic information currently relies primarily on visual inspection and tactile perception, which are imprecise, unreliable, and difficult to quantify using standardized criteria. To alleviate the above needs, some computer-assisted technologies have been developed to assist in the diagnosis and treatment of PWS. Image analysis technology is used to obtain objective assessment information of lesions, which is of great significance for subsequent diagnosis, treatment planning and disease progression monitoring.

[0003] Isolating PWS lesions from digital images is a crucial step in computer-assisted V-PDT treatment. Currently, research on PWS segmentation is limited and limited to traditional methods such as clustering and thresholding. For example, He et al. proposed a clustering segmentation method based on color similarity and spatial proximity to segment PWS. Kang et al. proposed a PWS segmentation algorithm based on bidirectional perceptual color difference saliency. Because these methods are developed based on small-scale datasets, they suffer from poor robustness and generalization capabilities, making them difficult to apply to clinical research. Furthermore, due to the influence of vascular distribution and blood oxygen content, the morphology and color of PWS lesions vary greatly, posing significant challenges to the lesion segmentation task, such as low contrast between the lesion and the surrounding skin, uneven lesion color, and a similar appearance between the lesion and the background. Therefore, there is an urgent need to develop a PWS lesion segmentation method with high accuracy and good robustness.

[0004] In recent years, deep learning-based image segmentation methods have achieved excellent segmentation performance, providing an alternative solution for skin lesion segmentation. For example, in 2015, Ronnberger et al. proposed a U-net-based architecture for biomedical image segmentation, achieving excellent results. In 2017, Yuan et al. proposed a fully automatic skin lesion segmentation method that does not rely on any prior knowledge of the data and achieves automatic segmentation of skin lesions through an end-to-end trained 19-layer deep convolutional neural network. In 2018, Li et al. proposed a residual learning-based dense deconvolutional network (DDN) for automatic skin lesion segmentation. The network consists of dense deconvolution layers (DDL), chained residual pooling (CRP), and hierarchical supervision (HS). In 2020, Lei et al. proposed an effective generative adversarial network (GAN) for skin lesion segmentation. It uses a Unet with skip connections and dense convolution-based (UNet-SCDC) as the generator and adopts a dual recognition strategy to improve the recognition of genuine inputs. However, the above methods mainly focus on the utilization of texture and shape information without considering the importance of color information, which is important for determining the precise positioning of PWS edges.

[0005] In order to utilize the color information of the image, some literature has enhanced the color texture information of the image by introducing multiple color spaces. For example, in 2017, Yuan et al. proposed an enhanced convolution-deconvolution network with a smaller convolution kernel, and alleviated the overfitting of the model by introducing HSV and LAB color spaces. Tang et al. proposed a separable Unet with random weight averaging for skin lesion segmentation, which greatly captured channel-related contextual features and high-level semantic feature information. Wu et al. proposed a convolutional neural network with an adaptive dual attention module and used the input of multiple color channels to solve the problems of image appearance similarity and low contrast. Hajabdollahi et al. proposed a color channel pruning framework to simplify the complex structure of the network by selecting color channels. These methods improve the segmentation performance of the network by introducing additional color spaces, but they simply combine RGB and other color channels as the input of the network, ignoring the independence of color spaces. Summary of the Invention

[0006] In order to overcome the defects of the existing technology, the technical problem to be solved by the present invention is to provide a port-wine stain segmentation method based on adaptive fusion of multiple color spaces, which makes up for the limitations of the single RGB color space, integrates the color characteristics of human visual perception into the deep learning network, thereby improving the detection accuracy and robustness of PWS lesions, improving the baseline model's ability to distinguish color and texture, avoiding mutual interference of multiple color space decisions, and promoting the synergistic effect of multiple color spaces on PWS segmentation.

[0007] The technical solution of the present invention is: the method for segmenting port-wine stains based on multi-color space adaptive fusion comprises the following steps:

[0008] (1) Multiple color spaces are introduced through the multi-color space detection MCSD method to enhance the representation of image color information;

[0009] (2) Using Unet with dense dilated residual blocks to generate candidate results in different color spaces;

[0010] (3) Obtain the final segmentation result through adaptive fusion of these candidate results;

[0011] (4) Structural similarity loss is used to measure the brightness, color, and contrast differences between the predicted lesions and the true lesions.

[0012] The present invention introduces multiple color spaces through the multi-color space detection MCSD method to enhance the representation of image color information; uses Unet with dense dilated residual blocks to generate candidate results in different color spaces; obtains the final segmentation result through adaptive fusion of these candidate results; measures the brightness, color and contrast differences between predicted lesions and real lesions through structural similarity loss; thereby overcoming the limitations of the single RGB color space, integrating the color features of human visual perception into the deep learning network, and thus improving the detection accuracy and robustness of PWS lesions, enhancing the baseline model's ability to distinguish color and texture, avoiding mutual interference of multi-color space decisions, and promoting the synergistic effect of multiple color spaces on PWS segmentation.

[0013] Also provided is a port-wine stain segmentation device based on multi-color space adaptive fusion, which includes:

[0014] a multi-color detection module configured to introduce multiple color spaces through a multi-color space detection (MCSD) method to enhance the representation of image color information;

[0015] Dense Atrous Residual Module, which is configured to use a Unet with dense dilated residual blocks to generate candidate results in different color spaces;

[0016] An adaptive fusion module, which is configured to obtain the final segmentation result through adaptive fusion of these candidate results;

[0017] The loss module is configured to measure the brightness, color, and contrast differences between the predicted lesions and the true lesions via structural similarity loss. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1The present invention is a flowchart of a method for segmenting port-wine stains based on adaptive fusion of multiple color spaces. DETAILED DESCRIPTION

[0019] like Figure 1 As shown in FIG, the port-wine stain segmentation method based on multi-color space adaptive fusion includes the following steps:

[0020] (1) Multiple color spaces are introduced through the multi-color space detection MCSD method to enhance the representation of image color information;

[0021] (2) Using Unet with dense dilated residual blocks to generate candidate results in different color spaces;

[0022] (3) Obtain the final segmentation result through adaptive fusion of these candidate results;

[0023] (4) Structural similarity loss is used to measure the brightness, color, and contrast differences between the predicted lesions and the true lesions.

[0024] The present invention introduces multiple color spaces through the multi-color space detection MCSD method to enhance the representation of image color information; uses Unet with dense dilated residual blocks to generate candidate results in different color spaces; obtains the final segmentation result through adaptive fusion of these candidate results; measures the brightness, color and contrast differences between predicted lesions and real lesions through structural similarity loss; thereby overcoming the limitations of the single RGB color space, integrating the color features of human visual perception into the deep learning network, and thus improving the detection accuracy and robustness of PWS lesions, enhancing the baseline model's ability to distinguish color and texture, avoiding mutual interference of multi-color space decisions, and promoting the synergistic effect of multiple color spaces on PWS segmentation.

[0025] Preferably, in step (1), the multiple color spaces include: an XYZ color space that is device-independent, an HSV color space that is close to the way humans perceive color, YIQ and YUV color spaces that separate brightness and chromaticity, and a LAB color space with uniform color perception; and a multi-branch parallel approach is adopted to utilize these six branch color spaces to capture different color information contained in the multiple color spaces.

[0026] Preferably, in step (1), MCSD takes the normalized RGB image as input and maps it to the input of the six branches through the color space conversion layer. Assuming X is the input image, the input of the branch corresponding to the color space C is denoted as Xc, then

[0027] X C =T C X,C∈{rgb,hsv,lab,yiq,yuv,xyz} (1)

[0028] Where T C is the transformation operator from RGB color space to C color space. If C = rgb, T C is the identity transformation; if C∈{yiq,yuv,xyz}, T C is a linear transformation; if C∈{hsv,lab}, T C is a nonlinear transformation.

[0029] Preferably, in step (2), a Unet equipped with a DDR module is used to construct a detection model for each branch. All branches adopt the same U-shaped network structure. The model of branch C is recorded as FC. The output of each branch is calculated by formula (2):

[0030]

[0031] A dense atrous residual DDR module is designed to replace the stacked convolutional layers. The DDR module contains 16 basic units, which are improved based on the basic units of ResNetV2. First, the second convolutional layer uses atrous convolution with a dilation rate of 2. The continuous stacking of basic units continuously expands the receptive field of the model, increasing the network depth while improving the model's expression of global information. Then, the stacking of residual units is used to achieve a gradual increase in the number of convolutional layers. Each basic unit only connects the newly generated features to the final output layer, improving the expression of detailed information of the network.

[0032] Preferably, in step (3), the final output image of the model is

[0033] Y pred =sigmoid(∑ C ω c Y C +b) (3)

[0034] where ω c is an adaptive weight coefficient used to control the contribution of each color space to the final prediction result. Its initial value is 1 / 6 and is optimized through model updating. b is a constant term used to balance the systematic error of the model.

[0035] Preferably, in step (4), the loss function is defined as follows:

[0036] L total (x,y true ,y pred )=λL bce (y true ,y pred )+(1-λ)L ssim-final (x,y true ,y pred) (9)

[0037] Where λ is a constant between 0 and 1, which is used to balance the structural similarity loss with the cross entropy loss and is set to 0.5.

[0038] Those skilled in the art will appreciate that all or part of the steps in the above-described embodiment method can be implemented by instructing the relevant hardware through a program. The program can be stored in a computer-readable storage medium. When the program is executed, it includes the steps of the above-described embodiment method. The storage medium can be ROM / RAM, a magnetic disk, an optical disk, a memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a port-wine stain segmentation device based on multi-color space adaptive fusion. The device includes:

[0039] a multi-color detection module configured to introduce multiple color spaces through a multi-color space detection (MCSD) method to enhance the representation of image color information;

[0040] Dense Atrous Residual Module, which is configured to use a Unet with dense dilated residual blocks to generate candidate results in different color spaces;

[0041] An adaptive fusion module, which is configured to obtain the final segmentation result through adaptive fusion of these candidate results;

[0042] The loss module is configured to measure the brightness, color, and contrast differences between the predicted lesions and the true lesions via structural similarity loss.

[0043] Preferably, in the multi-color detection module, the multiple color spaces include: the device-independent XYZ color space, the HSV color space close to the human color perception method, the YIQ and YUV color spaces with separated brightness and chromaticity, and the LAB color space with uniform color perception; a multi-branch parallel approach is adopted to utilize these six branch color spaces to capture the different color information contained in the multiple color spaces.

[0044] Preferably, in the multi-color detection module, MCSD takes the normalized RGB image as input and maps it to the input of the six branches through the color space conversion layer. Assuming X is the input image and the input of the branch corresponding to the color space C is denoted as Xc, then

[0045] X C =T C X,C∈{rgb,hsv,lab,yiq,yuv,xyz} (1)

[0046] Where T C is the transformation operator from RGB color space to C color space. If C = rgb, T Cis the identity transformation; if C∈{yiq,yuv,xyz}, T C is a linear transformation; if C∈{hsv,lab}, T C is a nonlinear transformation.

[0047] Preferably, in the dense hole residual module, a Unet equipped with a DDR module is used to construct a detection model for each branch. All branches adopt the same U-shaped network structure. The model of branch C is recorded as FC. The output of each branch is calculated by formula (2):

[0048]

[0049] A dense atrous residual DDR module is designed to replace the stacked convolutional layers. The DDR module contains 16 basic units, which are improved based on the basic units of ResNetV2. First, the second convolutional layer uses atrous convolution with a dilation rate of 2. The continuous stacking of basic units continuously expands the receptive field of the model, increasing the network depth while improving the model's expression of global information. Then, the stacking of residual units is used to achieve a gradual increase in the number of convolutional layers. Each basic unit only connects the newly generated features to the final output layer, improving the expression of detailed information of the network.

[0050] The present invention is described in more detail below. The method comprises:

[0051] Multi-Color Detection Module (MCDM)

[0052] Since we are interested in processing objects with weak color texture in the dataset, color features are considered first in the model. Since chromaticity and brightness information are mixed together, it is difficult for a simple RGB color space to effectively capture color texture information. Therefore, several other color spaces are introduced to enrich the color features of the image, such as the device-independent XYZ color space, the HSV color space that is close to the way humans perceive color, the YIQ and YUV color spaces that separate brightness and chromaticity, and the LAB color space with uniform color perception. In addition, considering that each color space is independent and there is redundant information in multiple color spaces, we use a multi-branch parallel approach to utilize these color spaces. Specifically, a multi-color space detection module with six branches is constructed to capture the different color information contained in multiple color spaces.

[0053] MCSD takes the normalized RGB image as input and maps it to the input of the six branches through the color space conversion layer. Assume that X is the input image and the input of the branch corresponding to the color space C is recorded as Xc, then

[0054] X C =T CX,C∈{rgb,hsv,lab,yiq,yuv,xyz} (1)

[0055] Where T C is the transformation operator from RGB color space to C color space. If C = rgb, T C is the identity transformation; if C∈{yiq,yuv,xyz}, T C is a linear transformation; if C∈{hsv,lab}, T C is a nonlinear transformation.

[0056] Since its application in segmenting neuronal structures in electron microscopy stacks, Unet has been widely used in medical image segmentation, achieving excellent performance. To achieve better PWS lesion segmentation results, a Unet equipped with a DDR module was used to construct the detection model for each branch. All branches here use the same U-shaped network structure, with the model of branch C denoted as FC. The output of each branch can then be calculated using the following formula:

[0057]

[0058] MCDM requires six branches in parallel, and the model parameters are six times that of a single U-Shape encoding and decoding network. The doubling of the parameters makes the model computationally overburdened, making it difficult to deploy the model on applications with low hardware resources. At the same time, the doubling of the parameters causes the model to overfit, offsetting the performance improvement brought by multiple color spaces. To address the above problems, a dense dilated residual module is designed to replace the stacked convolutional layers. The proposed dense dilated residual module (DDR) contains 16 basic units, which are improved based on the basic units of ResNetV2. First, the second convolutional layer uses a dilated convolution with a dilation rate of 2. By continuously stacking basic units, the receptive field of the model is continuously expanded, increasing the network depth while improving the model's expression of global information; then, drawing on the idea of DenseNet, the stacking of residual units is used to achieve a gradual increase in the convolutional layer. Each basic unit only connects the newly generated features to the final output layer, improving the expression of detailed information of the network. Adaptive Fusion

[0059] These color spaces contain different color and texture information, and their interaction may improve PWS segmentation performance. This phenomenon is called visual complementarity in multiple color spaces. It is difficult to find the edge of the lesion within the yellow ellipse using the color channels of the RGB space because the color texture of the patient's clothing closely matches that of the lesion. However, this is easily achieved in YUV or LAB color spaces. Therefore, the visual complementarity of multiple color spaces may more accurately locate the edges of PWS lesions. To fully utilize this characteristic of multiple color spaces, an adaptive fusion strategy is proposed to fuse the outputs of MCSD.

[0060] The fusion of multiple decision models usually adopts the maximum voting or mean strategy. The former aims to highlight certain well-performing models or features, while weakening the grasp of global information by each sub-model; the latter integrates the characteristics of each sub-model, but ignores the anisotropic characteristics of the model. This paper proposes an adaptive weight strategy for the fusion of detection models of six different color spaces of MCDM, and autonomously learns the weight factors of each model through a preset network. The advantage of this is that there is no need to separately evaluate the contribution of each color space to the final result, nor is there any need to supervise and control each model. Only the final decision layer needs to be supervised. The final output image of the model is

[0061] Y pred =sigmoid(∑ C ω c Y C +b) (3)

[0062] where ω c is an adaptive weight coefficient used to control the contribution of each color space to the final prediction result. Its initial value is 1 / 6 and can be optimized through model updating. b is a constant term used to balance the systematic error of the model.

[0063] Loss Function

[0064] The proposed M-CSAFN is an end-to-end network structure that needs to perform pixel-level classification on the image to predict whether each pixel belongs to the lesion or the background. This is a typical binary classification problem. Binary cross entropy is used to characterize the consistency between the probability distribution of the actual lesion and the predicted result. Its calculation formula is as follows:

[0065]

[0066] where y i ∈{0,1}, N is the total number of pixels, y i and p i are the target value and predicted probability value of pixel i respectively.

[0067] Although cross entropy loss can be used to characterize the consistency between two data distributions, it lacks effective constraints on local details (such as color and contrast). In order to further focus on the details of the image, this paper also introduces structural similarity loss. In 2004, Wang et al. proposed a quantitative image quality evaluation index that conforms to the visual characteristics of the human eye. It evaluates the similarity between the reference image and the distorted image from three aspects: brightness, contrast, and structure:

[0068]

[0069] Where x is the reference image, y is the distorted image, μ x and μ y are the first-order origin moments of the reference image and the distorted image, σ x and σ y are the second-order central moments of the reference image and the distorted image, σ xy is the covariance between the reference image and the distorted image. C1 and C2 are variables used to stabilize the weak denominator division, C1 = (k1L) 2 , C2=(k2L) 2 , where k1 = 0.01, k2 = 0.03, and L = 255. In particular, when x and y are three-channel color images, the structural similarity index of each channel is calculated separately and the average is taken.

[0070] Based on the above structural similarity, the following loss function is defined to supervise the learning process of M-CSAFN using the structural similarity measure between the predicted lesions and the real lesions:

[0071] L ssim (y true ,y pred )=1-SSIM(y true ,y pred ) (6)

[0072] where y true and y pred Represent the gold standard and predicted probability distribution, respectively, both based on grayscale images. In order to make the model pay more attention to the structure and color information of the lesion, constraints are imposed based on the region of interest of the image to increase the network's perception of local details:

[0073] L ssim-color (x,y true ,y pred )=1-SSIM(x⊙y true ,x⊙y pred ) (7)

[0074] Where ⊙ represents the matrix multiplication operation. Taking into account that the training process of M-CSAFN is carried out batch by batch, the loss function is optimized on the batch to accelerate the convergence speed while ensuring the stability of the model. Formula (5) is further improved and the specific calculation is as follows:

[0075] L ssim-final (x,y true ,y pred )=1-mean(SSIM(x⊙y true ,x⊙y pred ))+std(SSIM(x⊙y true ,x⊙y pred )) (8)

[0076] The first two items are structural similarity losses based on the color region of interest, and the last item is used to control the stability of the model.

[0077] In order to better supervise the learning process of M-CSAFN, the final loss function is defined as follows:

[0078] L total (x,y true ,y pred )=λL bce (y true ,y pred )+(1-λ)L ssim-final (x,y true ,y pred ) (9)

[0079] Where λ is a constant between 0 and 1, which is used to balance the structural similarity loss with the cross entropy loss. In this experiment, it is set to 0.5.

[0080] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of protection of the technical solution of the present invention.

Claims

1. A port-wine stain segmentation method based on adaptive fusion of multiple color spaces, characterized by: It includes the following steps: (1) Multiple color spaces are introduced through the multi-color space detection MCSD method to enhance the representation of image color information; (2) Using Unet with dense dilated residual blocks to generate candidate results in different color spaces; (3) Obtain the final segmentation result through adaptive fusion of these candidate results; (4) Structural similarity loss is used to measure the brightness, color, and contrast differences between the predicted lesions and the true lesions.

2. The method for segmenting port-wine stains based on adaptive fusion of multiple color spaces according to claim 1, characterized in that: In step (1), the multiple color spaces include: an XYZ color space that is device-independent, an HSV color space that is close to the way humans perceive color, YIQ and YUV color spaces that separate brightness and chromaticity, and a LAB color space with uniform color perception; a multi-branch parallel approach is adopted to utilize these six branch color spaces to capture different color information contained in the multiple color spaces.

3. The method for segmenting port-wine stains based on adaptive fusion of multiple color spaces according to claim 2, characterized in that: In step (1), MCSD takes the normalized RGB image as input and maps it to the input of the six branches through the color space conversion layer. Assuming X is the input image and the input of the branch corresponding to the color space C is recorded as Xc, then X C =T C X, C∈{rgb, hsv, lab, yiq, yuv, xyz} (1) Where T C is the transformation operator from RGB color space to C color space. If C = rgb, T C is the identity transformation; if C∈{yiq,yuv,xyz}, T C is a linear transformation; if C∈{hsv,lab}, T C is a nonlinear transformation.

4. The method for segmenting port-wine stains based on adaptive fusion of multiple color spaces according to claim 3, characterized in that: In step (2), a Unet equipped with a DDR module is used to construct a detection model for each branch. All branches use the same U-shaped network structure. The model of branch C is denoted as FC. The output of each branch is calculated using formula (2): A dense atrous residual DDR module is designed to replace the stacked convolutional layers. The DDR module contains 16 basic units, which are improved based on the basic units of ResNetV2. First, the second convolutional layer uses atrous convolution with a dilation rate of 2. The continuous stacking of basic units continuously expands the receptive field of the model, increasing the network depth while improving the model's expression of global information. Then, the stacking of residual units is used to achieve a gradual increase in the number of convolutional layers. Each basic unit only connects the newly generated features to the final output layer, improving the expression of detailed information of the network.

5. The method for segmenting port-wine stains based on adaptive fusion of multiple color spaces according to claim 4, characterized in that: In step (3), the final output image of the model is Y pred =sigmoid(∑ C ω c Y C +b) (3) where ω c is an adaptive weight coefficient used to control the contribution of each color space to the final prediction result. Its initial value is 1 / 6 and is optimized through model updating. b is a constant term used to balance the systematic error of the model.

6. The method for segmenting port-wine stains based on adaptive fusion of multiple color spaces according to claim 5, characterized in that: In step (4), the loss function is defined as follows: L total (x,y true ,and pred )=λL bce (and true ,and pred )+(1-λ)L ssim-final (x,y true ,and pred ) (9) Where λ is a constant between 0 and 1, which is used to balance the structural similarity loss with the cross entropy loss and is set to 0.

5.

7. A port-wine stain segmentation device based on adaptive fusion of multiple color spaces, characterized by: It includes: a multi-color detection module configured to introduce multiple color spaces through a multi-color space detection (MCSD) method to enhance the representation of image color information; Dense Atrous Residual Module, which is configured to use a Unet with dense dilated residual blocks to generate candidate results in different color spaces; An adaptive fusion module, which is configured to obtain the final segmentation result through adaptive fusion of these candidate results; The loss module is configured to measure the brightness, color, and contrast differences between the predicted lesions and the true lesions via structural similarity loss.

8. The port-wine stain segmentation device based on multi-color space adaptive fusion according to claim 7, characterized in that: In the multi-color detection module, multiple color spaces include: the device-independent XYZ color space, the HSV color space close to the human color perception method, the YIQ and YUV color spaces that separate brightness and chromaticity, and the LAB color space with uniform color perception; a multi-branch parallel approach is adopted to utilize these six branch color spaces to capture the different color information contained in multiple color spaces.

9. The port-wine stain segmentation device based on multi-color space adaptive fusion according to claim 8, characterized in that: In the multi-color detection module, MCSD takes the normalized RGB image as input and maps it to the input of six branches through the color space conversion layer. Assuming X is the input image and the input of the branch corresponding to the color space C is recorded as Xc, then X C =T C X, C∈{rgb, hsv, lab, yiq, yuv, xyz} (1) Where T C is the transformation operator from RGB color space to C color space. If C = rgb, T C is the identity transformation; if C∈{yiq,yuv,xyz}, T C is a linear transformation; if C∈{hsv,lab}, T C is a nonlinear transformation.

10. The port-wine stain segmentation device based on multi-color space adaptive fusion according to claim 9, characterized in that: In the dense hole residual module, Unet equipped with a DDR module is used to construct the detection model of each branch. All branches adopt the same U-shaped network structure. The model of branch C is recorded as FC. The output of each branch is calculated by formula (2): A dense atrous residual DDR module is designed to replace the stacked convolutional layers. The DDR module contains 16 basic units, which are improved based on the basic units of ResNetV2. First, the second convolutional layer uses atrous convolution with a dilation rate of 2. The continuous stacking of basic units continuously expands the receptive field of the model, increasing the network depth while improving the model's expression of global information. Then, the stacking of residual units is used to achieve a gradual increase in the number of convolutional layers. Each basic unit only connects the newly generated features to the final output layer, improving the expression of detailed information of the network.

Images