Deep learning-based psoriasis skin lesion area evaluation system

Abstract

The invention discloses a deep learning-based psoriasis skin lesion area evaluation system. The system comprises an image acquisition module, a psoriasis skin lesion image preprocessing module, a psoriasis skin lesion image optimization module, a skin lesion area evaluation model design module and a psoriasis skin lesion area evaluation module. The invention belongs to the field of image processing, and particularly relates to a deep learning-based psoriasis skin lesion area evaluation system, which introduces a channel area sensitive factor through an inter-channel interaction weight mechanism, strengthens complementarity of scale and erythema features, reduces leak detection due to fusion of erythema global information into small scales, and improves psoriasis skin lesion area evaluation accuracy. The boundaries of the erythema are clearer by referring to scale textures; by introducing a boundary sensitivity weight, a boundary error is directly associated with an area error; a quadratic term of an area proportion is introduced and standardized to carry out skin damage area feature extraction, so that the distinguishing precision of boundary grades is greatly improved; and constructing a loss function of combination of the gentle probability and the punishment weight so as to obtain a skin lesion area evaluation effect.

Description

Technical Field

[0001] This invention relates to the field of image processing, specifically to a deep learning-based system for assessing the area of ​​psoriasis lesions. Background Technology

[0002] Psoriasis lesion area assessment systems typically use annotated images of psoriasis lesions, extract lesion features, and output assessment results using a specific model. However, these systems often suffer from several problems: scale feature capture imbalances leading to blurred erythema boundaries, spatial feature misalignment and regional fragmentation, and insufficient boundary precision, resulting in low reliability of the final lesion area assessment. Furthermore, they often exhibit issues such as blurred lesion grade boundaries, weak differentiation between lesion grade levels leading to misjudgments, and significant interference from exclusion classes, ultimately resulting in poor lesion area assessment performance. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based psoriasis lesion area assessment system. Addressing the problems of unbalanced scale feature capture leading to blurred erythema boundaries, spatial feature misalignment and regional fragmentation, and insufficient boundary accuracy in general psoriasis lesion area assessment systems, resulting in low reliability of the final lesion area assessment, this solution introduces a channel area-sensitive factor through an inter-channel interactive weighting mechanism. This enhances the complementarity between scale and erythema features. Small scales are less likely to be missed because they incorporate global erythema information, while erythema boundaries are clearer due to reference to scale texture. By introducing boundary sensitivity weights… This approach directly correlates boundary errors with area errors, thereby improving the reliability of lesion area assessment. Addressing the issues of ambiguous lesion grade boundaries, weak differentiation between lesion grade levels leading to misjudgments, and significant interference from exclusion classes resulting in poor lesion area assessment, this solution employs multi-value labeling to avoid cross-grade misjudgments. It introduces a quadratic term of area proportion and standardizes it for lesion area feature extraction, significantly improving the accuracy of boundary grade differentiation. Furthermore, it constructs a loss function combining smooth probability and punitive weights to focus on the accuracy of candidate grades, avoiding misjudgments caused by extreme probabilities, thus improving the overall lesion area assessment performance.

[0004] The technical solution adopted by the present invention is as follows: The psoriasis lesion area assessment system based on deep learning provided by the present invention includes an image acquisition module, a psoriasis lesion image preprocessing module, a psoriasis lesion image optimization module, a lesion area assessment model design module, and a psoriasis lesion area assessment module;

[0005] The image acquisition module acquires historical images of psoriasis lesions and annotates them;

[0006] The psoriasis lesion image preprocessing module performs skin color correction and normalization on the acquired historical psoriasis lesion images;

[0007] The psoriasis lesion image optimization module is designed with a hierarchical coding structure, introduces a channel area sensitive factor, and adopts boundary weighted IoU loss to achieve psoriasis lesion image optimization.

[0008] The skin lesion area assessment model design module is based on the optimized psoriasis skin lesion image, extracts the spatial and statistical features of the skin lesion and performs linear transformation, and trains it with total loss containing smoothing coefficient and punishment coefficient to establish the skin lesion area assessment model.

[0009] The psoriasis lesion area assessment module assesses the lesion area of ​​real-time psoriasis lesion images.

[0010] Furthermore, the image acquisition module acquires historical psoriasis lesion images and performs image annotation, including pixel-level binary masks and ground truth values ​​for lesion area levels.

[0011] Furthermore, the psoriasis lesion image preprocessing module performs skin color correction on the acquired historical psoriasis lesion images using YCbCr color space conversion; hair removal is performed using morphological opening operation and threshold filtering; finally, image standardization and illumination normalization are performed to obtain a set of psoriasis lesion images.

[0012] Furthermore, the psoriasis lesion image optimization module specifically includes:

[0013] Hierarchical coding structure design unit; for psoriasis lesion images, feature extraction is performed by splitting the image into layers according to the lesion scale. The first layer uses a convolutional neural network, and layers 2-5 use a convolutional neural network and a transform neural network in parallel for extraction.

[0014] The design unit for the interaction weighting mechanism between psoriatic skin lesion feature channels is as follows: First, the psoriatic skin lesion channel features are compressed, the mean value of the skin lesion features of each channel is calculated, bidirectional channel attention is generated, and a channel area sensitivity factor is introduced; then, the correlation matrix between psoriatic skin lesion channels is constructed, bidirectional feature projection is performed, and finally, residual feature fusion is performed.

[0015] A unit for integrating spatial dimensional features of psoriatic lesions; first, receptive field adaptation is performed, followed by feature fusion splicing;

[0016] Design unit for psoriasis lesion segmentation decoder; transposed convolution, each layer enlarges the feature map size by 2 times, stride=2, padding=1; skip connection fusion, after upsampling of each layer, channel splicing is performed, and then multi-scale lesion features are fused through double convolutional blocks; the output layer generates a binary mask;

[0017] The loss function is based on IoU loss and introduces boundary sensitivity weights.

[0018] Furthermore, the skin lesion area assessment model design module specifically includes the following:

[0019] Multi-value label creation; the lesion area grading is an ordinal label, and a two-branch tree area grade label conversion is performed, with each area grade mapped to a bi-state encoded sequence; a range-containing label is generated for each binary tree node, and the target sequence is right-shifted to generate the decoded target sequence;

[0020] Area rank label embedding: The label position is associated with the label value to generate a unique identifier, which is then transformed into a vector of the same dimension as the skin lesion feature through the embedding layer and used as the input of the decoder;

[0021] Skin lesion area feature extraction; global average pooling is performed on the final output feature map to obtain spatial feature vector; statistical feature vector is constructed based on area ratio; spatial features and statistical features are concatenated and mapped through a linear transformation layer to obtain the final area feature;

[0022] Design of area level sequence decoder: predict area level bi-state encoded sequence, and introduce mask judgment to suppress the probability contribution of the eliminated area level and focus on the current candidate level; output a set of original predicted values; generate a mask based on the range containing the label in the previous time step and weight the Sigmoid probability; calculate the mean probability and compare it, and iterate 3 times to obtain the complete bi-state encoded sequence;

[0023] Design of area-level loss function; construct smoothing coefficient, introduce penalty coefficient to suppress and exclude level loss, construct penalty coefficient; construct single-time loss, and obtain total loss.

[0024] Furthermore, the psoriasis lesion area assessment module acquires psoriasis lesion images, which are then processed sequentially by the psoriasis lesion image preprocessing module and the psoriasis lesion image optimization module before being input into the lesion area assessment model. The lesion area level output by the model is used as the psoriasis lesion area assessment result. If the lesion area level is 4 or 5, an early warning is issued to the management personnel.

[0025] The beneficial effects achieved by the present invention using the above solution are as follows:

[0026] (1) To address the problem that general psoriasis lesion area assessment systems suffer from scale feature capture imbalance leading to blurred erythema boundaries, spatial feature misalignment and regional fragmentation, and insufficient boundary accuracy, resulting in low reliability of the final lesion area assessment, this solution introduces a channel area sensitive factor through an inter-channel interactive weighting mechanism to enhance the complementarity of scale and erythema features. Small scales are reduced in missed detection due to the integration of global erythema information, and erythema has clearer boundaries due to reference to scale texture. By introducing boundary sensitivity weights, boundary errors are directly correlated with area errors, thereby improving the reliability of lesion area assessment.

[0027] (2) To address the problems of fuzzy lesion level boundaries, weak differentiation of lesion level boundaries leading to misjudgment, and large exclusion interference in general psoriasis lesion area assessment systems, which in turn result in poor lesion area assessment performance, this solution avoids cross-level misjudgment by creating multi-value labels; introduces a quadratic term of area proportion and standardizes it for lesion area feature extraction, which greatly improves the differentiation accuracy of boundary levels; and constructs a loss function that combines smooth probability with punitive weights to focus on the accuracy of candidate levels and avoid misjudgment caused by extreme probabilities, thereby improving the lesion area assessment performance. Attached Figure Description

[0028] Figure 1 A flowchart illustrating the deep learning-based psoriasis lesion area assessment system provided by this invention;

[0029] Figure 2 A flowchart illustrating the design module for the skin lesion area assessment model.

[0030] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0033] Example 1, see Figure 1 The psoriasis lesion area assessment system based on deep learning provided by the present invention includes an image acquisition module, a psoriasis lesion image preprocessing module, a psoriasis lesion image optimization module, a lesion area assessment model design module, and a psoriasis lesion area assessment module.

[0034] The image acquisition module acquires historical images of psoriasis lesions and annotates them; then it sends the data to the psoriasis lesion image preprocessing module.

[0035] The psoriasis lesion image preprocessing module performs skin color correction and normalization on the acquired historical psoriasis lesion images; and sends the data to the psoriasis lesion image optimization module.

[0036] The psoriasis lesion image optimization module is designed with a hierarchical coding structure, introduces a channel area sensitive factor, and adopts boundary weighted IoU loss to optimize the psoriasis lesion image; and sends the data to the lesion area assessment model design module.

[0037] The skin lesion area assessment model design module is based on the optimized psoriasis skin lesion image, extracts the spatial and statistical features of the skin lesion and performs linear transformation, and trains it with total loss containing smoothing coefficient and punishment coefficient to establish a skin lesion area assessment model; and sends the data to the psoriasis skin lesion area assessment module.

[0038] The psoriasis lesion area assessment module assesses the lesion area of ​​real-time psoriasis lesion images.

[0039] Example 2, see Figure 1 This embodiment is based on the above embodiment. The image acquisition module acquires historical psoriasis lesion images, which are RGB images covering different areas, severity levels, and skin tones. Image annotation is performed, including pixel-level binary masks and ground truth values ​​for lesion area levels. In the pixel-level binary mask, 1 represents the lesion area, and 0 represents normal skin, hair, or background. Lesion area level classification: the percentage of lesion area is calculated. And mapped to ordinal category labels, represented as: ;in, It is a classification of the area of ​​skin lesions.

[0040] Example 3, see Figure 1 This embodiment is based on the above embodiment. The psoriasis lesion image preprocessing module performs skin color correction on the acquired historical psoriasis lesion images, and uses YCbCr color space conversion to convert the RGB image to YCbCr space, as shown below: ; ; Where Y is the luminance component in the YCbCr space; Cb and Cr are the blue and red components in the YCbCr space, respectively; R, G, and B are the red, green, and blue components of each pixel in the psoriatic lesion image, respectively. Hair removal is performed using morphological opening operations and thresholding. First, hair and lesions are initially separated by opening operations of 1×5 vertical structuring elements. Then, based on the large gray-level difference of hair pixels, Sobel edge detection is used to filter hair areas. Finally, the hair areas are filled by interpolation of surrounding normal skin pixels to avoid hair being misjudged as scales. Finally, image standardization is performed to unify the image size to 512×512, and illumination normalization is performed. CLAHE is used to enhance contrast within 8×8 local blocks to avoid large areas of dark lesions being missed due to insufficient illumination. A set of psoriatic lesion images is obtained.

[0041] Example 4, see Figure 1 This embodiment is based on the above embodiment, and the psoriasis lesion image optimization module specifically includes:

[0042] A hierarchical coding structure design unit is used to address the issue that psoriatic lesions exhibit both small-scale details, including punctate scaling, and large-scale distributions, such as extensive erythema on the trunk. A hierarchical encoder is designed to extract features from psoriatic lesion images by splitting them into layers according to lesion scale. The first layer uses a convolutional neural network (CNN) to lay the foundation for local details. Layers 2-5 simultaneously use a CNN and a transform neural network (VLT) for parallel extraction, ensuring that scaling details and global erythema features are generated synchronously at each scale level, providing multi-scale lesion feature support for subsequent fusion. The CNN backbone uses ResNet18, and the VLT backbone uses ViT-Base. The hierarchical feature extraction process is as follows: Layer 1 (input layer) uses only ResNet18's Conv1+MaxPool to convert the preprocessed image into a 128×128×64 basic local feature map, capturing the initial edges of the lesions. Layers 2-5 (multi-scale layers) simultaneously execute a ResNet18 block (outputting CNN features) and a ViT-Base block (outputting VLT features).

[0043] The encoder layer number and feature map size correspond to the scale of psoriatic skin lesions. Convolutional neural networks emphasize this aspect, while transform neural networks emphasize this aspect.

[0044] The second layer consists of 64×64 small scales with scale edge texture and clustered distribution of small scales.

[0045] Layer 3, 32×32 medium-sized plaque, plaque boundary details, association between plaque and surrounding skin lesions;

[0046] Layer 4: 16×16 large area of ​​erythema, internal infiltration texture of erythema, global extent of erythema;

[0047] Layer 5, 8×8 full-map skin lesion distribution, core lesion area features, and correlation of skin lesions in multiple locations;

[0048] Design unit for the interaction weighting mechanism between channels of psoriatic skin lesions; The channel features of psoriatic skin lesions exhibit significant differences. Through channel compression, bidirectional attention, feature projection, and residual connection feature integration, the scaling channel of the convolutional neural network and the erythema channel of the transform neural network are bidirectionally associated, screening highly correlated features, strengthening the complementarity of local details and global context, and avoiding missed detection of small scales and blurred erythema boundaries; First, the channel features of psoriatic skin lesions are compressed, and the mean value of the lesion features in each channel is calculated to avoid interference from single pixel noise, represented as: ; Where P is the original lesion feature map extracted from the psoriasis lesion image by the convolutional neural network; V is the original lesion feature map extracted by the transform neural network. It is a channel-level average pooling function; and This is the result of average pooling; bidirectional channel attention generation allows the scaling channel of the convolutional neural network to perceive the dimension of the erythema channel of the neural network, and vice versa, generating attention weights that can filter highly relevant channels, and introducing a channel area sensitivity factor. For each channel, the average area ratio of its corresponding lesion in the training set is statistically analyzed as a sensitivity factor, expressed as: ; ; ; ;in, and These are the channel attention weight vectors on the convolutional neural network side and the transform neural network side, respectively; and These are the sensitivity factors on the convolutional neural network side and the transform neural network side, respectively. and These are the average areas of the skin lesions corresponding to the convolutional channel p and the transform channel v, respectively. It is the Sigmoid activation function; and These are the weight matrices for increasing and decreasing the dimensionality of the convolutional neural network, respectively. It is the ReLU activation function; and These are the weight matrices for dimensionality increase and dimensionality reduction on the neural network side, respectively; and a correlation matrix between channels in psoriatic lesions is constructed to quantify the correlation between the scaling channel and the erythematous channel, expressed as: Bidirectional feature projection integrates convolutional neural network features with global erythema information to avoid missing scales at the erythema edge, and transform neural network features with detailed scale information to avoid blurring of erythema boundaries. This is represented as: ; Finally, residual feature fusion is performed, represented as: ; Where Q is the inter-channel correlation matrix; It is the projection result of the features of the convolutional neural network onto the channel space of the transform neural network; The spatial dimension of P is preserved by marking; It is the Softmax normalization function; It is the projection result of the transformed neural network features onto the channel space of the convolutional neural network; The spatial dimension of V is preserved; and These are the fused convolutional neural network feature maps and transform neural network feature maps, respectively.

[0049] A unit for integrating spatial features of psoriatic lesions; the spatial distribution of psoriatic lesions exhibits local-global misalignment. Through receptive field adaptation, iterative fusion, and parameter control, the spatial features of scaling from the convolutional neural network are aligned with the spatial features of erythema from the transform neural network, ensuring the continuity of the lesion region and providing a complete and unbroken lesion mask for area statistics; firstly, receptive field adaptation is performed, represented as: ; Subsequently, feature fusion and stitching are performed to enhance the commonly identified lesion areas, represented as follows: ;in, and These are the fused convolutional neural network feature maps and the transformed neural network feature maps, respectively. It is a two-dimensional convolution operation; and It is a convolution kernel; This is the final output feature map; It is feature splicing;

[0050] The design unit for the psoriasis lesion segmentation decoder is as follows: The boundary accuracy of psoriasis lesions directly affects area assessment. A U-Net-like architecture of transposed convolution + double convolution is adopted, combined with skip connections in the spatial dimension feature fusion unit of psoriasis lesions, to fuse multi-scale lesion features and ultimately generate a binary segmentation mask of the same size as the input image. Transposed convolution enlarges the feature map size by a factor of 2 in each layer, with a stride of 2 and padding of 1, ensuring accurate lesion size recovery. Skip connection fusion, after upsampling in each layer, is combined with... Channel concatenation is performed, followed by a double convolutional block of 3×3 convolution + ReLU + 3×3 convolution + ReLU to fuse multi-scale skin lesion features; the output layer generates a binary mask, represented as: ;in, It is the value of the final output binary segmentation mask at pixel (x,y), where 1 indicates that the pixel belongs to the psoriasis lesion area and 0 indicates that the pixel does not belong to the psoriasis lesion area; It is a 1×1 convolution kernel; First to Perform convolution to obtain a single-channel feature map, then apply the sigmoid function to this single-channel feature map to obtain a probability map of each pixel belonging to skin lesions. This represents the pixel value corresponding to the coordinate in the probability graph;

[0051] The loss function is based on IoU loss and introduces boundary sensitivity weights. This links the segmentation loss of the boundary region with the area error, strengthening the impact of the boundary on area assessment, as shown below: ; ;in, It is a boundary-weighted loss; It is an IoU loss; It is a truly labeled binary segmentation mask; It is the boundary enhancement coefficient; It is an erosion operation; training uses the AdamW optimizer, and the learning rate scheduling strategy uses cosine annealing.

[0052] By performing the above operations, this solution addresses the problems of unbalanced scale feature capture leading to blurred erythema boundaries, spatial feature misalignment and regional fragmentation, and insufficient boundary accuracy in general psoriasis lesion area assessment systems, resulting in low reliability of the final lesion area assessment. This solution introduces a channel area-sensitive factor through an inter-channel interactive weighting mechanism, enhancing the complementarity of scale and erythema features. Small scales are less likely to be missed due to the integration of global erythema information, while erythema boundaries are clearer by referencing scale texture. By introducing boundary sensitivity weights, boundary errors are directly correlated with area errors, thereby improving the reliability of lesion area assessment.

[0053] Example 5, see Figure 1 and Figure 2 This embodiment is based on the above embodiment, and the skin lesion area assessment model design module specifically includes the following:

[0054] Multi-value marker creation; lesion area grading These are ordinal labels. Direct prediction can lead to misjudgments due to blurred boundaries between levels. Therefore, a binary tree area level label conversion is performed. Assuming the binary tree height is d=3, each area level is mapped to a bi-state encoded sequence, represented as: ; represents the path from the root node to the leaf node, where 0 in binary indicates the left branch and 1 indicates the right branch. Each leaf node uniquely corresponds to a skin lesion area level; generates a range label for each binary tree node, indicating the positive class included / excluded negative class of the current node, represented as: The target sequence is then shifted right, starting with the initial query marker 's', to generate the decoded target sequence. , represented as: ;in, It is a bi-state encoded sequence label at the area level; It is the multi-hot vector at time t, where 0 corresponds to the left branch and 1 corresponds to the right branch, and t∈{1,2,3}; The scope includes the tag; It is the multi-heat vector at time t, which is a binary vector of length 6. Each time corresponds to a level, and 1 indicates that the current level node contains the level of skin lesion area.

[0055] Area rank label embedding; binary labels at different levels have ambiguous meanings, so the label position is associated with the label value to generate a unique identifier, which is then transformed into a vector of the same dimension as the skin lesion features through an embedding layer and used as the input to the decoder; the unique identifier is represented as: Decoder input Represented as: ;in, This is the position-value association encoding function for the area marker; k is the position index of the element in the target sequence, where k=0 corresponds to the initial marker s, and k=1 corresponds to... ; It is the value of the element at the k-th position in the target sequence; It is an embedded layer for area markers;

[0056] Skin lesion area feature extraction; for Perform global average pooling to obtain spatial feature vectors. Constructing statistical feature vectors based on area proportion This includes the percentage and the quadratic term, improving the differentiation of boundary levels, and is expressed as: Spatial and statistical features are combined, and the final area feature is obtained through a linear transformation layer mapping. , represented as: ;in, and These are all samples in the training set. The mean and variance; It is a linear transformation layer mapping;

[0057] Design of area-level sequence decoder; Inputting skin lesion features, with As the label embedding input, predict the area level bi-state encoded sequence, and simultaneously introduce a mask decision to suppress the probability contribution of the already eliminated area levels, focusing on the current candidate level; let the decoder input at time t be... During training During testing, the embedding of the predicted label from the previous time step is used; after processing with multi-head self-attention and cross-attention, a set of original predicted values ​​is output. , represented as: Based on the range containing the label from the previous time step. Generate a mask, suppress the weights of excluded categories, and weight the Sigmoid probability, represented as: ; ;in, It is an area-level sequence decoder; It is the mask vector at time t. It is a vector element; It is a suppressor factor; The range at time t-1 includes the label; Let i be the label of level i in the multi-hot vector at time t-1; let the area level ranges of the left and right subtrees at level t of the binary tree be [l,m] and [m+1,r], respectively. Calculate the mean probability and compare them, expressed as: ; ; The complete two-state encoding sequence is obtained by iterating 3 times. ;in, and These are the mean probabilities of the left and right branches, respectively; l, m, and r are boundary parameters. It is the mask weighted probability of level i at time t; It is the decision result at time t; , and These are the decision results at three different moments;

[0058] Area-level loss function design; constructing a smoothing coefficient to avoid probability extremes, expressed as: A penalty coefficient is introduced to suppress the loss caused by excluding certain levels, prioritizing the prediction accuracy of candidate levels, and constructing a penalty coefficient system. , represented as: Single-moment loss Represented as: The assessment of psoriasis area requires a step-by-step screening process at three different times, hence the total loss... Represented as: ;in, It is the smoothness coefficient at time t; T is the temperature range coefficient; It is the penalty coefficient; the AdamW optimizer is used, and the learning rate is dynamically adjusted using a cosine annealing strategy; after each round of training, the model is evaluated using a validation set, and the rank accuracy is used as the core indicator to judge the model performance.

[0059] By performing the above operations, this solution addresses the problems of general psoriasis lesion area assessment systems, such as blurred lesion grade boundaries, weak differentiation of lesion grade boundaries leading to misjudgments, and large exclusion of class interference, resulting in poor lesion area assessment performance. This solution avoids cross-grade misjudgments by creating multi-valued labels; it introduces a quadratic term of area proportion and standardizes it for lesion area feature extraction, significantly improving the accuracy of boundary grade differentiation; and it constructs a loss function combining smooth probability and punitive weights to focus on the accuracy of candidate grades, avoiding misjudgments caused by extreme probabilities. These improvements ultimately enhance the effectiveness of lesion area assessment.

[0060] Example 6, see Figure 1 This embodiment is based on the above embodiment. The psoriasis lesion area assessment module acquires psoriasis lesion images, which are then processed by the psoriasis lesion image preprocessing module and the psoriasis lesion image optimization module in sequence before being input into the lesion area assessment model. The lesion area level output by the model is used as the psoriasis lesion area assessment result. If the lesion area level is 4 or 5, an early warning is issued to the management personnel.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0062] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A deep learning-based Psoriasis Area and Severity Index system, characterized by: The system comprises an image acquisition module, a psoriasis lesion image preprocessing module, a psoriasis lesion image optimization module, a lesion area evaluation model design module and a psoriasis lesion area evaluation module. The image acquisition module acquires historical psoriasis lesion images and performs image labeling. The psoriasis lesion image preprocessing module performs skin color correction and normalization processing on the acquired historical psoriasis lesion images. The psoriasis lesion image optimization module designs a hierarchical coding structure, introduces a channel area sensitive factor, and adopts a boundary weighted IoU loss to realize the optimization of the psoriasis lesion image. The lesion area evaluation model design module extracts lesion spatial and statistical features and linearly transforms them based on the optimized psoriasis lesion image to train a total loss containing a smoothness coefficient and a penalty coefficient, and establishes a lesion area evaluation model. The psoriasis lesion area evaluation module evaluates the lesion area of real-time psoriasis lesion images.

2. The deep learning-based Psoriasis Area and Severity Index system of claim 1, wherein: The psoriasis lesion image optimization module specifically comprises: A hierarchical coding structure design unit; for psoriasis lesion images, feature extraction is split according to lesion scale levels, the first layer uses a convolutional neural network, and the 2-5 layers use a convolutional neural network + transformation neural network in parallel for feature extraction; A psoriasis lesion feature channel interaction weight mechanism design unit; A psoriasis lesion spatial dimension feature intersection integration unit; first, the perception field is adapted, and then the fused features are spliced; A psoriasis lesion segmentation decoder design unit; transpose convolution, each layer enlarges the feature map size by 2 times, step = 2, padding = 1; skip connection fusion, after upsampling each layer, the channel is spliced, and then the multi-scale lesion features are fused through a double convolution block; the output layer generates a binary mask; The loss function is based on the IoU loss and introduces a boundary sensitivity weight.

3. The deep learning-based Psoriasis Area and Severity Index system of claim 2, wherein: The psoriasis lesion feature channel interaction weight mechanism design unit first performs psoriasis lesion channel feature compression, calculates the lesion feature mean value of each channel, generates a bidirectional channel attention, and introduces a channel area sensitive factor; A psoriasis lesion channel correlation matrix construction is performed, bidirectional feature projection is performed, and finally residual feature fusion is performed.

4. The deep learning-based Psoriasis Area and Severity Index system of claim 3, wherein: The lesion area evaluation model design module specifically comprises the following contents: Multi-value label creation; lesion area grading is ordinal label, double-branch tree area level label conversion is performed, each area level is mapped to a binary encoding sequence; a range-inclusive label is generated for each binary tree node, and the target sequence is right-shifted to generate a decoding target sequence; Area level label embedding; the label position is associated with the label value to generate a unique identifier, which is then converted into a vector with the same dimension as the lesion feature through an embedding layer and used as the decoder input; Lesion area feature extraction; Global average pooling is performed on the final output feature map to obtain a spatial feature vector; a statistical feature vector is constructed based on the area ratio; the spatial feature and the statistical feature are spliced The spatial feature and the statistical feature are mapped to the final area feature through a linear transformation layer; Area level sequence decoder design; Area level loss function design.

5. The deep learning-based Psoriasis Area and Severity Index system of claim 4, wherein: The area level sequence decoder design is to predict the area level binary encoding sequence, while introducing a mask decision to suppress the probability contribution of the eliminated area level, focusing on the current selected level; output a set of original prediction values; generate a mask based on the previous time range containing label, and weight the Sigmoid probability; calculate the mean probability and compare it, and iterate 3 times to get the complete binary encoding sequence.

6. The deep learning-based Psoriasis Area and Severity Index system of claim 5, wherein: The area level loss function design is to construct a gentle coefficient, introduce a penalty coefficient to suppress the excluded level loss, and construct a penalty coefficient; construct a single time loss to get the total loss.

7. The deep learning-based Psoriasis Area and Severity Index system of claim 6, wherein: The psoriasis lesion image preprocessing module performs skin color correction on the collected historical psoriasis lesion image, adopts YCbCr color space conversion; and performs hair removal, adopts morphological opening operation and threshold screening; finally performs image standardization and illumination normalization; obtain a set of psoriasis lesion images.

8. The deep learning-based Psoriasis Area and Severity Index system of claim 7, wherein: The psoriasis lesion area evaluation module obtains the psoriasis lesion image, sequentially processes the psoriasis lesion image through the psoriasis lesion image preprocessing module and the psoriasis lesion image optimization module, and inputs the processed image into the lesion area evaluation model; the lesion area level output by the model is taken as the psoriasis lesion area evaluation result; if the lesion area level is 4 or 5, the management personnel are warned.

Images