Intelligent skin disease image classification method and system based on multi-scale boundary guidance
The intelligent classification method for skin disease images guided by multi-scale boundaries utilizes multi-scale convolutional gated linear units and boundary collaborative enhancement modules to solve the problems of insufficient feature extraction and susceptibility to interference in existing technologies, and achieves high-precision and interpretable skin disease image classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HENAN UNIV OF SCI & TECH
- Filing Date
- 2025-11-19
- Publication Date
- 2026-04-10
AI Technical Summary
Existing dermatology assistance systems based on machine learning and computer vision suffer from problems such as low accuracy in diagnosing complex and rare diseases, difficulty in data acquisition, and insufficient model interpretability when processing skin lesion images. They also struggle to effectively extract and selectively fuse multi-scale features and are easily affected by interference information such as hair and shadows, leading to the loss of key features and misjudgment.
A multi-scale boundary-guided intelligent classification method for skin disease images is adopted. Through multi-scale convolutional gated linear unit modules and boundary collaborative enhancement modules, dynamic fusion and screening of multi-scale lesion features are achieved, and edge and subtle lesion features are enhanced. The method combines residual blocks and depth-separable attention blocks for feature extraction and fusion, and outputs high-level semantic features.
It significantly improves the accuracy, robustness, and interpretability of skin disease image classification, and can effectively overcome the problems of single receptive field, loss of detail, and insufficient anti-interference ability of traditional methods without the need for pre-segmentation, thus achieving high-quality image acquisition and accurate diagnosis.
Smart Images
Figure CN121837697A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of image processing technology in computer vision, and in particular to a skin disease image intelligent classification method and system based on multi-scale boundary guidance. BACKGROUND
[0002] Skin is the most complex structure and the most comprehensive function organ of human body, and has sharp feeling and multiple barrier effect, which makes the types and forms of skin diseases diverse. The pathogenesis of skin diseases is jointly regulated by internal factors such as genetics, immunity, endocrine, and psychology, and external factors such as environmental stimulation, chemical substances, and pathogen infection, showing a high heterogeneity. The traditional diagnosis of skin diseases relies on the clinical experience of doctors, and inexperienced doctors are easy to misdiagnose because of the similarity of early symptoms. With the continuous growth of the number of skin disease patients, the traditional diagnosis mode has been difficult to cope with the increasing medical burden, and it is particularly urgent to develop efficient and accurate computer-aided diagnosis technology.
[0003] The clinical diagnosis and treatment of skin diseases faces challenges such as difficult diagnosis, uneven treatment effect and easy recurrence, and uneven distribution of medical resources. Skin imaging technology has evolved from early expectation diagnosis to digital means such as dermatoscope, ultrasound and CT, laying a foundation for intelligent diagnosis. However, the existing auxiliary system based on machine learning and computer vision still has significant bottlenecks in processing skin lesion images: for example, low diagnosis accuracy for complex and rare diseases, difficulty in data acquisition, and insufficient model interpretability; common diseases and rare diseases have high similarity in appearance, and the lesion scale span is large, traditional methods are difficult to effectively extract and selectively fuse multi-scale features, and are easily affected by interference information such as hair and shadow, resulting in loss of key features and misjudgment. Therefore, it is urgent to build a model that can accurately capture multi-scale lesion information and dynamically fuse features to improve classification accuracy and diagnosis reliability, and promote the development of skin disease auxiliary diagnosis to a more intelligent and reliable direction. SUMMARY
[0004] The technical problem to be solved by the present application is to provide a skin disease image intelligent classification method and system based on multi-scale boundary guidance, which can effectively solve the problems of poor image quality during image acquisition and insufficient feature extraction in processing skin disease images, without segmentation steps, directly focusing on lesion edge features, extracting different lesion area features, to realize stronger feature extraction capability, improve the clarity of lesion boundary and classification accuracy.
[0005] In order to achieve the above purpose, on the one hand, the present application provides a skin disease image intelligent classification method based on multi-scale boundary guidance, comprising the following steps: S1, collecting skin disease image data conforming to a predetermined range; S2, initial feature extraction and down-sampling processing are performed on the skin disease image data to obtain a basic feature map; S3, the basic feature map is input into a first residual block group, and key features are strengthened through a plurality of first residual blocks, a parallel spatial fusion attention module and an adaptive boundary enhancement module, and residual fusion is performed on two output results to generate a first fusion feature; S4, the first fusion feature is input into a multi-scale convolution gate linear unit module, the first fusion feature is first split into two branch features, then multi-scale deep convolution spatial feature extraction and fusion are performed on the first branch feature to obtain a multi-scale fusion feature, and a channel weight vector is generated based on the multi-scale fusion feature, the second branch feature is weighted by using the channel weight vector, and an optimized feature is output; S5, the optimized feature is respectively input into a boundary cooperative enhancement module and a second residual block group, the boundary cooperative enhancement block extracts boundary enhancement features, and the second residual block group extracts high-level semantic features in combination with a depth separable attention residual block; After fusion, a high-level semantic feature is obtained; the high-level semantic feature and the boundary enhancement feature are size-aligned and channel-connected, and channel-unified through a convolution layer to obtain a cross-layer fusion feature; S6, the cross-layer fusion feature is input into a third residual block group for deep semantic extraction, and then processed through a classification head, and the skin disease classification result is output by sequentially performing average pooling, flattening, full connection and Softmax activation on the feature.
[0006] Further, the step S4 specifically comprises: S41, the input feature is flattened into a feature sequence, and is split into a first branch feature and a second branch feature through a first linear layer; S42, multi-scale deep convolution spatial feature extraction is performed on the first branch feature, including a 3*3 convolution path and a 5*5 convolution path, and adaptive dynamic weights are generated through a Softmax function to weight and fuse the outputs of the two paths to obtain a multi-scale fusion feature; S43, the multi-scale fusion feature is input into a double linear dynamic gate, sequentially passes through a second linear layer, a GELU activation function, a third linear layer and a Sigmoid function, and outputs a channel weight vector; S44, the channel weight vector and the second branch feature are multiplied element by element to realize channel selective enhancement, and then the fourth linear layer is adjusted to adjust the feature dimension to output an optimized feature conforming to the input dimension of the next stage; Further, in the step S42, the process of the multi-scale deep convolution spatial feature extraction is: converting the first branch feature into a two-dimensional feature map format, inputting into a 3x3 deep convolution layer and a 5x5 deep convolution layer respectively for parallel feature extraction; using a Softmax function to generate dynamic weights of two paths, and weighting and summing the convolution output according to the dynamic weights.
[0007] Further, in the step S43, the specific process of the bilinear dynamic gate is: the channel number of the second linear layer is compressed from 256 to 64, after the non-linear representation ability is enhanced by a GELU activation function, the channel number of the third linear layer is restored to 256, and the weight is output by a Sigmoid function.
[0008] Further, in the step S44, the expression of the optimized feature is: wherein v represents the second branch feature, u represents the multi-scale fusion feature, g represents linear variation, [u;v] represents feature splicing, σ is a Sigmoid function, and ⨂ represents element-wise multiplication; when the weight calculated by GELU(u)⨂σ(g([u;v])) is close to 1, GELU(u) is preferentially retained, which is applicable to images with clear lesions; when the weight calculated by v⨂(1-σ(g([u;v])) is close to 0, the second branch feature is preferentially retained, which avoids the loss of complex image information.
[0009] Further, in the step S1, the collecting of the skin disease image data conforming to the preset range specifically includes: acquiring an original image of a skin disease area through an image acquisition module; performing angle backtracking calibration and polarization shadowless imaging processing on the original image in sequence, wherein the angle backtracking calibration is used to compensate for device jitter in the acquisition process, and the polarization shadowless imaging is used to eliminate reflections and local shadows on the image surface; and finally obtaining the skin disease image data conforming to the preset range.
[0010] Further, in the step S2, the initial feature extraction and down-sampling processing specifically includes: performing preliminary feature mapping by using a 7x7 convolution kernel, then performing batch normalization and activation function processing in sequence, and down-sampling the image size from 224x224 to 56x56 through a 3x3 max-pooling operation, and increasing the channel number to 64 to retain color features and texture information.
[0011] Further, in the step S5, the boundary cooperative enhancement module includes four parallel branches, and the channel numbers of the branches are evenly distributed in a ratio of 1:1:1:1; 1x1 convolution is used in each branch for dimension reduction processing, and then multi-scale feature extraction is performed on the branch features respectively; the module is applicable to skin disease image analysis, and can capture full-scale features from subtle lesions to large area skin lesions.
[0012] The application also provides a skin disease image classification system based on a multi-scale guided boundary network, comprising: The data acquisition unit is composed of an image acquisition module, an angle backtracking calibration module, a polarization shadowless imaging module and a data sending module connected in sequence, and is used for collecting skin disease image data meeting a preset range; The image classification unit performs enhancement operations such as random rotation, color jitter, contrast adjustment and affine transformation on the input image in the data preprocessing stage; in the feature extraction stage, a multi-scale convolution network architecture is adopted, and feature extraction is performed through 7*7 convolution layers in sequence, feature strengthening is performed in combination with residual blocks, spatial fusion attention and adaptive boundary enhancement convolution, advanced feature fusion is realized by using multi-scale convolution gate, boundary collaborative enhancement and deep separable residual block; finally, classification output is completed through average pooling, feature flattening, full connection layer and Softmax activation, and an interpretable diagnosis result is provided.
[0013] According to the above technical solution, the application has the following advantages: The application realizes dynamic fusion and screening of multi-scale lesion features by introducing a multi-scale convolution gate linear unit module, and strengthens edge and subtle lesion features in combination with a boundary collaborative enhancement module, thereby effectively overcoming the problems of single receptive field, detail loss and insufficient anti-interference ability of traditional methods without pre-segmentation, significantly improving the accuracy, robustness and interpretability of skin disease image classification, and realizing high-quality image acquisition without shaking, removing reflections and local shadows through a handheld device. BRIEF DESCRIPTION OF DRAWINGS
[0014] Figure 1 The flowchart of the application is shown in the figure; Figure 2 The model framework diagram of the application is shown in the figure; Figure 3 The schematic diagram of the multi-scale convolution gate linear unit module of the application is shown in the figure; Figure 4 The schematic diagram of the device in Example 2 is shown in the figure; Figure 5 The left view of the device in Example 2 is shown in the figure; Figure 6 The visualization effect diagram of the eczema feature map is shown in the figure; Figure 7 The visualization effect diagram of the miliaria papulosa feature map is shown in the figure. DETAILED DESCRIPTION
[0015] The application will be further described in detail below in combination with the drawings and specific embodiments.
[0016] The structure, proportion, size, etc. shown in the drawings of the specification are only used to cooperate with the content disclosed in the specification, to be understood and read by those skilled in the art, and do not have technical substantial significance, and any modification of the structure, change of the proportion relationship or adjustment of the size, without affecting the effect and purpose that can be achieved by the present application, should still fall within the scope covered by the disclosed technology.
[0017] Embodiment 1 As Figures 1-3 shown, the present application proposes a multi-scale boundary guided skin disease image intelligent classification method, comprising the following steps: S1, collect skin disease image data conforming to a preset range.
[0018] Specifically, the original image of the skin disease area is obtained through the image acquisition module; the original image is sequentially subjected to angle backtracking calibration and polarization shadowless imaging processing, wherein the angle backtracking calibration is used to compensate for the device jitter in the acquisition process, and the polarization shadowless imaging is used to eliminate the reflection and local shadow on the image surface; finally, the skin disease image data conforming to the preset range is obtained.
[0019] S2, initial feature extraction and down-sampling processing are performed on the skin disease image data to obtain a basic feature map.
[0020] Wherein, the initial feature extraction and down-sampling processing specifically includes: adopting a 7x7 convolution kernel for preliminary feature mapping, then sequentially performing batch normalization, activation function processing, and through a 3x3 maximum pooling operation, the image size is down-sampled from 224x224 to 56x56, and the channel number is increased to 64, so as to retain color features and texture information.
[0021] Specifically, first receive the input image , wherein I represents the input image set, B represents the batch size, C represents the channel number of the input image, H and W represent the height and width of the input image respectively. The input image is preliminarily mapped by 7x7 convolution, combined with batch normalization, activation and 3x3 pooling down-sampling, the size is reduced from 224x224 to 56x56, and the channel number is increased from 3 to 64. The basic color features (such as erythema, pigmentation) and preliminary texture information (such as rough surface, scale distribution) of skin lesions are effectively retained, which provides a clear lesion area positioning basis for subsequent deep analysis, and helps to identify skin diseases such as psoriasis, eczema, etc. with typical color and texture features.
[0022] S3, input the base feature map into the first residual block group, strengthen the features through a plurality of first residual blocks, and strengthen the key features through a parallel spatial fusion attention (SFA) module and an adaptive boundary enhancement (ABEC) module, and perform residual fusion on two output results to generate a first fused feature.
[0023] Among them, the SFA module generates an adaptive weight distribution map of the lesion area and the background area through separate coordinate information encoding and double-channel pooling operations, and reweights the features according to the distribution map to suppress the feature response of irrelevant backgrounds such as normal skin and light interference in the image, so as to accurately focus on the lesion position and optimize the feature expression.
[0024] The ABEC module constructs a parallel decoupling architecture of convolution and self-attention, wherein the convolution branch adopts 3x3 convolution to capture the edge gradient details of small lesions, and the self-attention branch focuses on the boundary area and models its global association with the lesion through adaptive average pooling; the outputs of the two branches are dynamically fused through gated connection to balance detail preservation and interference suppression, and a cross-layer fidelity channel from shallow fine boundary features to deep network is constructed through residual connection, so that the high-fidelity retention of complex boundary details, effective suppression of normal texture interference and complete transmission of cross-layer features are realized at the same time while reducing the computational overhead.
[0025] Specifically, in the first residual block group Layer1, the initial features are strengthened through the residual block Bottleneck. Layer1 is composed of 3 residual blocks, each residual block retains key features through convolution, batch normalization and residual connection, and the output channel number is 64xblock.expansion (expansion is 256).
[0026] The SFA module and the ABEC module strengthen the output of Layer1 respectively to obtain strengthened features and . The two strengthened features are element-wise added and fused to obtain the first fused feature .
[0027] S4, input the first fusion feature into a multi-scale convolutional gated linear unit (MCGLU) module, first split the first fusion feature into two branch features, then perform multi-scale deep convolution spatial feature extraction and fusion on the first branch feature to obtain a multi-scale fusion feature, generate a channel weight vector based on the multi-scale fusion feature, use the channel weight vector to perform channel weighting on the second branch feature, and output an optimized feature.
[0028] S41, flatten the input feature into a feature sequence, and split it into a first branch feature and a second branch feature through a first linear layer; S42, perform multi-scale deep convolution spatial feature extraction on the first branch feature, convert the first branch feature into a two-dimensional feature map format, and input it into a 3x3 deep convolution layer and a 5x5 deep convolution layer respectively for parallel feature extraction; generate dynamic weights of two paths using a Softmax function, and perform weighted fusion on the outputs of the two paths to obtain multi-scale fusion features.
[0029] S43, input the multi-scale fusion feature into a bilinear dynamic gate, sequentially pass through a second linear layer, a GELU activation function, a third linear layer, and a Sigmoid function, and output a channel weight vector.
[0030] Among them, the specific process of the bilinear dynamic gate is: the channel number of the second linear layer is compressed from 256 to 64, after the non-linear representation ability is enhanced by the GELU activation function, the channel number of the third linear layer is restored to 256, and the weight is output through the Sigmoid function.
[0031] S44, multiply the channel weight vector and the second branch feature element by element to realize channel selective enhancement, and then adjust the feature dimension through a fourth linear layer to output an optimized feature conforming to the input dimension of the next stage.
[0032] Specifically, the first fusion feature is input into the MCGLU module for feature optimization. First, the first fusion feature is flattened into a sequence feature , realizing the format adaptation from "image feature" to "sequence feature", laying a foundation for subsequent processing.
[0033] By splitting into two branches x and v through the first linear layer, the x branch is first converted into a 2D feature map format, and then multi-scale spatial features are extracted in parallel through 3x3 and 5x5 depth convolutions (dwconv3 / dwconv5). Dynamic weights are generated through the Softmax function to adaptively adjust the contribution of the two convolutions, and the feature map u is obtained by weighted fusion. u is processed by a bilinear dynamic gate, specifically: the channel number of the second linear layer is compressed from 256 to 64, and then the channel number of the third linear layer is restored to 256 after the non-linear representation ability is enhanced by the GELU activation function, and the weight is output through the Sigmoid function . Then u is multiplied element by element with the output weight , giving high weight to the lesion-related channels and low weight to the interference channels. Finally, the feature dimension is adjusted through the fourth linear layer to match the input dimension of the next stage.
[0034] The expression of the output MCGLU optimized feature is: where v represents the second branch feature, u represents the multi-scale fusion feature, g represents the linear change, represents feature concatenation, is the Sigmoid function, represents element-wise multiplication; when the calculated weight is close to 1, the is preferentially retained, suitable for images with clear lesions; when the calculated weight is close to 0, the second branch feature is preferentially retained to avoid loss of complex image information.
[0035] S5, the optimized feature is input into a Boundary Synergistic Enhancement Module (BSEM) and a second residual block group, the BSEM extracts boundary enhancement features, and the second residual block group extracts high-level semantic features in combination with a Depthwise Attentive Residual Block (DARB); after fusion, high-level semantic features are obtained; the size of the high-level semantic features is aligned with the boundary enhancement features, and the channels are connected, and the channels are unified through a convolution layer to obtain cross-layer fusion features.
[0036] where the BSEM module includes four parallel branches, and the channel number of each branch is evenly distributed in a ratio of 1:1:1:1; each branch is reduced in dimension by 1x1 convolution, and then multi-scale feature extraction is performed on each branch feature; the module is suitable for skin disease image analysis and can capture full-scale features from subtle lesions to large-area skin lesions.
[0037] The DARB module fuses grouped convolution and global normalization, uses a grouped convolution unit to capture features, and then uses a global normalization unit to suppress environmental interference and strengthen the consistency of lesion features, thereby providing more stable semantic representations for intelligent diagnosis of skin diseases and improving computational efficiency.
[0038] Specifically, the optimized features are input into the BSEM module to obtain through feature enhancement. Meanwhile , the features are also input into a second residual block group Layer2 (consisting of 6 residual blocks with a stride of 2), in which the MCGLU module output features are extracted and enhanced for high-level semantics, the output channel is expanded to 256x256, and the feature map shape is changed to . Then, the features are strengthened through the DARB module, and the output is . This step focuses on high-level semantic feature extraction while reducing computational complexity through lightweight convolution.
[0039] Subsequently, the boundary features and high-level semantic features are fused across layers. The size is adjusted to match (H / 8xW / 8) through bilinear interpolation, obtaining . After channel dimension concatenation, the channel number is unified to 1024 through a 1x1 convolution layer, and the batch normalization and activation output the cross-layer fusion features .
[0040] S6, the cross-layer fusion features are input into a third residual block group for deep semantic extraction, and then processed through a classification head. The classification result of skin diseases is output by sequentially performing average pooling, flattening, full connection, and Softmax activation on the features.
[0041] Specifically, the classification result is finally output by the third residual block group Layer3 and the classification head. The cross-layer fusion features are input into Layer3 (3 residual blocks with a stride of 2 for downsampling) to extract high-level semantic features . Through the classification head, the features are compressed into a vector through global average pooling, and the class probability is output through the full connection layer (fc).
[0042] Another aspect of the present application provides a multi-scale boundary guided intelligent classification system for skin disease images, comprising: A data acquisition unit is composed of an image acquisition module, an angle backtracking calibration module, a polarization shadowless imaging module, and a data sending module connected in sequence, which is used to collect skin disease image data within a predetermined range; The image classification unit performs enhancement operations such as random rotation, color jitter, contrast adjustment, and affine transformation on the input image in the data preprocessing stage; in the feature extraction stage, a multi-scale convolutional network architecture is adopted, and the features are extracted through a 7x7 convolutional layer in sequence, combined with residual blocks, spatial fusion attention, and adaptive boundary enhancement convolution for feature enhancement, and then multi-scale convolution gating, boundary collaborative enhancement, and deep separable residual blocks are used to realize high-level feature fusion; The result output unit completes class prediction through average pooling, feature flattening, a fully connected layer, and a Softmax activation function, generates a diagnostic conclusion with explainability, and outputs to a display screen.
[0043] Embodiment 2 As Figures 4-5 shown, the application also includes a device for implementing the skin disease image classification method and system based on the multi-scale guided boundary network described in embodiment 1.
[0044] The device includes a camera module 1, an angle backtracking calibration component 2, a polarized shadowless imaging component 3, a communication component 4, a computing unit 5, a storage component 6, a display screen 7, and a focusing button 8; the modules are connected through electrical signals or data buses, and cooperatively realize stable collection, processing, and output result display of skin lesions.
[0045] The camera module 1 is located on the right side of the device, used to start the imaging function and preliminarily capture the lesion area picture. The focusing button 8 is arranged at the bottom of the device for easy operation, and is connected with the camera module 1, used to control the camera focusing, and ensure clear imaging of the lesion area.
[0046] The angle backtracking calibration component 2 cooperates with the camera module 1, and its built-in high-sensitivity inertial measurement unit is used to detect the device shaking angle in real time, and compare with the reference stable angle. When the angle deviation exceeds the set threshold, the module drives the miniature anti-shake adjusting mechanism to make reverse compensation, so that the shooting angle backtracks to the stable state before shaking.
[0047] The polarized shadowless imaging component 3 is started synchronously with the angle backtracking calibration component 2, including a polarized filter and a ring-shaped shadowless lamp. The polarized filter is used to filter the stray light reflection caused by oil or keratin on the skin surface; the ring-shaped shadowless lamp is arranged around the camera module 1, providing multi-directional uniform illumination, eliminating local shadows, ensuring stable and distortion-free imaging light, and clearly presenting the pathological features such as subcutaneous pigments and blood vessels.
[0048] The communication component 4 is connected with the camera module 1, the polarized shadowless imaging component 3, and the angle backtracking calibration component 2, used to transmit the standardized image data processed by anti-shake, polarization, and light enhancement to the computing unit 5 and the storage component 6 in real time.
[0049] The computing unit 5 receives the image data from the communication component 4 and performs the MGBNet proposed in the application to perform lesion image classification and generate a classification result.
[0050] The storage component 6 is used for storing the collected image data and the diagnosis result, and is a pluggable memory card arranged in a card slot on the side of the device, facilitating data export and backup.
[0051] The display screen 7 is arranged on the left side of the device housing and is used for displaying images and diagnosis results. The display screen 7 includes an image display area 701 and a classification result display area 702. The image display area 701 is used for displaying the collected lesion images in real time, and the classification result display area 702 is used for displaying the classification result output by the computing unit 5, such as lesion type, confidence, and the like.
[0052] The application will be further demonstrated below in combination with ablation experiments and comparative experiments.
[0053] The experimental data of the application includes two parts: one is a self-built original lesion data set, which is obtained from the dermatology database of a cooperative hospital. All images are in the original state and have not been subjected to any lesion segmentation preprocessing. To ensure data quality, invalid samples with blurred lesions and serious occlusion are removed through manual screening, and finally an effective data set covering twenty different types of skin lesions is formed. The second is a comparative data set. Based on existing public data resources, the ISIC2018 dermatoscope data set commonly used in the field is selected as a control to more comprehensively verify the performance of the model on different types of skin images.
[0054] The two data sets are trained using the skin disease image intelligent classification method and system based on the multi-scale boundary guide proposed in the application, and the best model is saved and the training result is output through distributed training, mixed precision calculation, various optimizations, and a learning rate scheduler.
[0055] When testing the trained model, ACC (accuracy), Precision (precision), Recall (recall), and F1-score (F1 score) are used as verification indicators to evaluate the classification performance of the model on the test set.
[0056] ACC is used to calculate the proportion of all correctly predicted samples to the total samples; Precision is used to calculate the proportion of samples that the model predicts as "positive" to the actual "positive" samples; Recall is used to calculate the proportion of samples that the model successfully predicts as "positive" to the actual "positive" samples; and F1-score is used to calculate the harmonic mean of Precision and Recall, which is used to comprehensively evaluate the performance of the two.
[0057] Table 1 shows the effects of the self-built data set on the six comparative models and the improved model, and it can be seen that the effects of the skin disease image intelligent classification method and system based on multi-scale boundary guidance proposed in the application are the best in the four evaluation indexes, and the Recall is even 33 percentage points higher than ConvNext.
[0058] Table 1 Comparison experiment of self-built skin lesion data set Table 2 shows the effects of the ISIC2018 data set on the six models and the improved model, and it can be seen that the effects of the skin disease image intelligent classification method and system based on multi-scale boundary guidance proposed in the application are the best in the four evaluation indexes, and the ACC is even 34 percentage points higher than InceptionNext.
[0059] Table 2 Comparison experiment of ISIC2018 data set Finally, the heat map visualization results of the test set images are shown, and the pictures of the test set are randomly extracted for visualization using the trained weights, which shows the area of interest of the lesion in the skin disease image intelligent classification method and system based on multi-scale boundary guidance proposed in the application. The original image is the input skin lesion image, the heatmap is the heat map generated according to the weight, and the superimposed image is the visualization of the heat map superimposed on the original image.
[0060] Figure 6 The visualization of eczema is shown, Figure 7 The visualization of miliaria is shown. From the visualization results, it can be seen that the constructed model can accurately locate the lesion area, which provides help for subsequent auxiliary diagnosis and treatment.
[0061] In summary, the application realizes dynamic fusion and screening of multi-scale lesion features by introducing a multi-scale convolution gate linear unit module, and strengthens edge and subtle lesion features by combining a boundary cooperative enhancement module. Without pre-segmentation, the application effectively overcomes the problems of single receptive field, detail loss and insufficient anti-interference ability of traditional methods, significantly improves the accuracy, robustness and explainability of skin disease image classification, and realizes high-quality image acquisition of anti-shake, removal of reflection and local shadow through a handheld device.
[0062] It should be noted that the above embodiments are only used to illustrate the application, but the application is not limited to the above embodiments, and any simple modification, equivalent change and modification of the above embodiments according to the technical essence of the application all fall within the protection scope of the application.
Claims
1. A method for intelligent classification of skin disease images based on multi-scale boundary guidance, characterized in that, Includes the following steps: S1. Collect skin disease image data that falls within the preset range; S2. Perform initial feature extraction and downsampling on the skin disease image data to obtain the basic feature map; S3. Input the basic feature map into the first residual block group, perform feature enhancement through multiple first residual blocks, enhance key features through parallel spatial fusion attention module and adaptive boundary enhancement module, and perform residual fusion of the two output results to generate the first fused feature; S4. Input the first fusion feature into the multi-scale convolutional gated linear unit module. First, split the first fusion feature into two branch features. Then, perform multi-scale deep convolutional spatial feature extraction and fusion on the first branch feature to obtain the multi-scale fusion feature. Then, generate a channel weight vector based on the multi-scale fusion feature. Use the channel weight vector to perform channel weighting on the second branch feature and output the optimized feature. S5. Input the optimized features into the boundary collaborative enhancement module and the second residual block group respectively. The boundary collaborative enhancement block extracts the boundary enhancement features, and the second residual block group combines the depth-separable attention residual block to extract high-level semantic features. The two are fused to obtain high-level semantic features; the high-level semantic features are then sized and channel-connected with the boundary enhancement features, and channel unification is achieved through a convolutional layer to obtain cross-layer fused features; S6. Input the cross-layer fusion features into the third residual block group for deep semantic extraction, and then process them through the classification head. By sequentially performing average pooling, flattening, full connection and softmax activation on the features, the skin disease classification results are output.
2. The intelligent classification method for skin disease images based on multi-scale boundary guidance according to claim 1, characterized in that, Step S4 specifically includes: S41. Flatten the input features into a feature sequence, and split it into a first branch feature and a second branch feature through the first linear layer; S42. Multi-scale deep convolutional spatial feature extraction is performed on the first branch features, including 3×3 convolutional path and 5×5 convolutional path, and adaptive dynamic weights are generated through the Softmax function. The outputs of the two paths are weighted and fused to obtain multi-scale fused features. S43. Input the multi-scale fused features into a bilinear dynamic gate, and pass them sequentially through a second linear layer, a GELU activation function, a third linear layer, and a Sigmoid function to output a channel weight vector. S44. Multiply the channel weight vector element-wise with the second branch feature to achieve channel selectivity enhancement. Then, adjust the feature dimension through the fourth linear layer to output optimized features that meet the input dimension of the next stage.
3. The intelligent classification method for skin disease images based on multi-scale boundary guidance according to claim 2, characterized in that, In step S42, the process of multi-scale deep convolution spatial feature extraction is as follows: the first branch features are converted into a two-dimensional feature map format and input into a 3×3 deep convolutional layer and a 5×5 deep convolutional layer for parallel feature extraction; the dynamic weights of the two paths are generated using the Softmax function, and the convolution output is weighted and summed accordingly.
4. The intelligent classification method for skin disease images based on multi-scale boundary guidance according to claim 2, characterized in that, In step S43, the specific process of the bilinear dynamic gate is as follows: the number of channels in the second linear layer is compressed from 256 to 64, and after the nonlinear representation capability is enhanced by the GELU activation function, the number of channels in the third linear layer is restored to 256, and the weights are output through the Sigmoid function.
5. The intelligent classification method for skin disease images based on multi-scale boundary guidance according to claim 2, characterized in that, In step S44, the expression for the optimized feature is: Where v represents the second branch feature, u represents the multi-scale fusion feature, g represents linear transformation, [u;v] represents feature concatenation, σ is the Sigmoid function, and ⨂ represents element-wise multiplication. When the weight calculated by GELU(u)⨂σ(g([u;v])) is close to 1, GELU(u) is retained first, which is suitable for images with clear lesions. When the weight calculated by v⨂(1-σ(g([u;v])) is close to 0, the second branch feature is retained first to avoid the loss of complex image information.
6. The intelligent classification method for skin disease images based on multi-scale boundary guidance according to claim 1, characterized in that, In step S1, collecting skin disease image data within a preset range specifically includes: acquiring the original image of the skin disease area through the image acquisition module; performing angle backtracking calibration and polarization shadowless imaging processing on the original image in sequence, wherein angle backtracking calibration is used to compensate for equipment jitter during the acquisition process, and polarization shadowless imaging is used to eliminate reflections and local shadows on the image surface; finally, skin disease image data within a preset range is obtained.
7. The intelligent classification method for skin disease images based on multi-scale boundary guidance according to claim 1, characterized in that, In step S2, the initial feature extraction and downsampling process specifically includes: using a 7×7 convolution kernel for preliminary feature mapping, followed by batch normalization and activation function processing, and then downsampling the image size from 224×224 to 56×56 and increasing the number of channels to 64 through a 3×3 max pooling operation to preserve color features and texture information.
8. The intelligent classification method for skin disease images based on multi-scale boundary guidance according to claim 1, characterized in that, In step S5, the boundary collaborative enhancement module includes four parallel branches, with the number of channels in each branch evenly distributed in a ratio of 1:1:1:1; each branch uses 1×1 convolution for dimensionality reduction, and then multi-scale feature extraction is performed on the features of each branch; the module is suitable for skin disease image analysis and can simultaneously capture full-scale features from subtle lesions to large-area skin lesions.
9. A multi-scale boundary-guided intelligent classification system for skin disease images for implementing the method of any one of claims 1 to 8, characterized in that, include: The data acquisition unit consists of an image acquisition module, an angle backtracking calibration module, a polarization shadowless imaging module, and a data transmission module connected in sequence. It is used to collect skin disease image data that meets the preset range. In the image classification unit, the data preprocessing stage performs enhancement operations on the input image, such as random rotation, color jitter, contrast adjustment, and affine transformation. In the feature extraction stage, a multi-scale convolutional network architecture is adopted, which extracts features sequentially through 7×7 convolutional layers. It then combines residual blocks, spatial fusion attention, and adaptive boundary enhancement convolution to enhance features. Finally, it utilizes multi-scale convolutional gating, boundary co-enhancement, and depth-separable residual blocks to achieve advanced feature fusion. The output unit performs category prediction through average pooling, feature flattening, fully connected layers, and the Softmax activation function, generates interpretable diagnostic conclusions, and outputs them to the display screen.