Improved GAN and multi-feature fusion skin LC-OCT image segmentation method

Abstract

This application discloses a method for skin LC-OCT image segmentation that improves GAN and multi-feature fusion. The method includes using an LC-OCT device to collect skin B-scan samples and store them as images; constructing a generator that integrates reflection filling, multi-scale encoding-decoding, self-attention, and residual blocks, and coordinating it with a multi-scale fusion discriminator and a weight-sharing twin network to achieve high-fidelity denoising and enhancement through joint optimization of perceptual loss, multi-scale SSIM loss, and Wasserstein adversarial loss; locating and normalizing skin regions based on threshold or morphological preprocessing; extracting gradients from the denoised image, and combining the minimum gradient difference criterion with random forests for dynamic path search; and after adaptive sampling and spline fitting of the paths, generating candidates through random forests and correcting for breaks or misalignments in the lesion area. This method can achieve accurate segmentation of skin LC-OCT images in high-noise environments, construct a high-quality enhanced dataset, and provide a reliable basis for the early diagnosis of skin diseases.

Description

Technical Field

[0001] The present application relates to the field of medical image processing technology, and in particular to a skin LC-OCT image segmentation method using an improved GAN and multi-feature fusion. Background Art

[0002] Optical coherence tomography (OCT) is a noninvasive optical tomographic imaging technique based on the principle of low-coherence interferometry. Since its initial introduction in 1991, various implementations have been developed, including time-domain OCT (TD-OCT) and frequency-domain OCT (FD-OCT). OCT utilizes a broadband, low-coherence light source to illuminate biological tissue. By interfering the backscattered light from the sample with light from a reference arm, the reflected intensity distribution at different depths is measured, allowing the reconstruction of two-dimensional or three-dimensional tomographic images of the tissue interior with micron-level axial resolution (approximately 1–10 μm). Compared to traditional dermatoscopy and CT imaging, OCT can reveal subtle morphological changes in the epidermis, dermis, and accessory structures (such as hair follicles and sweat glands) in real time and noninvasively, with greater sensitivity to tissue boundaries and heterogeneity. In recent years, numerous studies have applied OCT to the early detection of skin lesions, lesion localization, and intraoperative navigation, demonstrating its superior imaging depth and resolution, providing crucial structural and functional information for clinical diagnosis and treatment decision-making.

[0003] Skin segmentation is a crucial technology for diagnosing skin diseases. Its core objective is to accurately segment the skin's various layers from optical coherence tomography (OCT) images. With the widespread adoption of high-resolution LC-OCT (line-field confocal optical coherence tomography), the richness of image detail has significantly increased, but interference issues such as speckle noise and motion artifacts have also become more prominent. Existing methods fall into two main categories: traditional image processing techniques and deep learning methods, both of which suffer from significant drawbacks. Traditional methods, based on edge detection (such as the Sobel and Canny operators) and path-finding algorithms (such as dynamic programming), while not requiring labeled data, are extremely sensitive to noise and rely on manual selection of initial points, making them difficult to address with blurred layer boundaries or lesions (such as distortion and hemorrhage). Furthermore, while traditional denoising algorithms (such as median filtering and wavelet denoising) can suppress noise, they can blur the fine structure between skin layers, resulting in reduced segmentation accuracy. These limitations are particularly pronounced in complex clinical scenarios, failing to meet the requirements for fully automated and highly robust performance.

[0004] In recent years, deep learning-based segmentation methods (such as U-Net and Transformer) have performed well on public datasets, but their implementation faces significant challenges. On the one hand, deep learning models require a large number of pixel-level annotated LC-OCT images for training, but medical data annotation is expensive, and consistency in annotations between different physicians is difficult to ensure. On the other hand, such models are computationally complex and rely on GPU acceleration, making them difficult to deploy on low-cost portable medical devices. More importantly, the "black box" nature of deep learning models results in a lack of explainability in their decision-making process, making it difficult for physicians to trust their output, hindering their practical clinical application.

[0005] Existing technologies attempt to combine traditional methods with deep learning (such as using convolutional networks to optimize gradient calculations), but there are still significant shortcomings. Denoising and segmentation are designed independently in stages, resulting in error accumulation at each level; the model has a large number of parameters and cannot meet real-time requirements; and there is a lack of a fusion mechanism for multiple features (gradients, textures, spatial positions), making it difficult to adaptively adjust the segmentation strategy. Summary of the Invention

[0006] The present application aims to solve one of the technical problems in the related art at least to a certain extent.

[0007] To this end, the purpose of this application is to propose an improved skin LC-OCT image segmentation method that combines GAN with multi-feature fusion, which can not only effectively suppress noise interference, but also achieve fully automatic segmentation through a lightweight model while retaining the decision logic that doctors can understand.

[0008] To achieve the above objectives, this application proposes a skin LC-OCT image segmentation method based on an improved GAN and multi-feature fusion method, comprising the following steps:

[0009] S1. Use LC-OCT equipment to collect skin B-scan samples from different subjects and save them as image data;

[0010] S2. Select and classify the collected images to ensure that the image information is discernible and representative;

[0011] S3, LC-OCT image denoising: An improved generative adversarial network is used for image denoising, with noisy images as input and real images as targets for training. The generator captures structural information through multi-scale feature extraction and self-attention mechanism, and integrates perceptual loss, MSSIM loss, and adversarial loss to optimize generation quality.

[0012] S4. Perform segmentation preprocessing based on thresholding or morphological operations, and limit the pixel value range to [0, 1] to eliminate illumination differences;

[0013] S5, dynamic path search based on Sobel operator and random forest multi-features, combined with adaptive spline fitting to achieve image segmentation;

[0014] S6. Output the skin LC-OCT image with segmentation results to achieve accurate segmentation effect, and compare and confirm with the original image.

[0015] An apparatus for improving a skin LC-OCT image segmentation method using GAN and multi-feature fusion includes a computer comprising a processor, a memory, and a communication interface; an LC-OCT device for collecting skin B-scan samples; and a computer program stored in the memory.

[0016] In addition, the skin LC-OCT image segmentation method based on the improved GAN and multi-feature fusion proposed in the application may also have the following additional technical features:

[0017] Furthermore, step S3 includes the following steps: performing reflection filling on the input image, the generator extracts features through multi-scale downsampling and processes them through the residual block and self-attention module, and then fuses features of different scales through upsampling, the discriminator uses multi-scale discrimination to evaluate the authenticity of the image, and the twin network calculates the similarity probability.

[0018] Furthermore, step S5 includes the following steps: using the Sobel operator to calculate the gradient and calibrate the initial search point, combining gradient features, texture features, and position features to perform random forest path optimization, dynamically adjusting the sampling points according to the complexity to perform spline fitting, and correcting the lesion area through random forest.

[0019] Furthermore, in the multi-scale feature extraction of the generator, downsampling uses a 3×3 convolution kernel with a step size of 2, the number of channels is doubled successively, the resolution is reduced from 450×900 to 56×112, and upsampling restores the resolution through transposed convolution and is spliced ​​and fused with the corresponding features.

[0020] Furthermore, the self-attention module converts the feature map into Q, K, and V matrices to calculate the global dependency weights. The discriminator uses four 4×4 convolution downsampling combined with LeakyReLU activation, and the twin network extracts features through a 5-layer CNN to calculate the L1 distance.

[0021] Furthermore, in the random forest multi-features, the gradient feature is calculated by the Sobel operator, the texture feature is extracted using the LBP descriptor, and the position feature includes pixel coordinate information. The random forest predicts the optimal segmentation path by extracting gradient, curvature, and texture features.

[0022] Furthermore, the adaptive spline fitting dynamically adjusts the sampling density according to the path complexity: sparse sampling is used in flat areas to reduce the amount of calculation, dense sampling is used in complex areas to improve the accuracy, and spline fitting smoothing processing is used to reduce noise interference.

[0023] Compared to existing technologies, this application demonstrates the following advantages: By integrating an improved generative adversarial network (reflection filling, multi-scale encoding-decoding, self-attention, and residual blocks) with dynamic path optimization (Sobel gradient difference criterion, random forest evaluation, adaptive spline fitting, and offset correction), it achieves high-fidelity denoising and accurate segmentation of LC-OCT B-scan skin images. The joint optimization of a triple loss (VGGMSE perceptual loss, multi-scale SSIM loss, and Wasserstein adversarial loss) effectively preserves both macro-texture and micro-detail in the denoised image. Furthermore, dynamic path search and random forest correction not only automatically track inter-layer boundaries in complex lesions but also maintain layer continuity in central depressions, providing a reliable foundation for constructing high-quality enhanced datasets.

[0024] Compared with the traditional threshold method or single morphological segmentation method, this application can stably extract the skin layer structure under high noise and lesion interference without manual parameter adjustment; compared with the existing simple GAN denoising technology, this method additionally introduces the structural consistency constraint of the twin network and the cross-layer feature fusion of the multi-scale discriminator, which significantly improves the segmentation accuracy and robustness; compared with manual segmentation, this application significantly reduces the time cost and avoids subjective bias.

[0025] Furthermore, on a moderately configured computer, the processing time for each 450×900 OCT image in this application can be controlled within a few seconds. Verified by Hausdorff distance and Dice coefficient, the segmentation results are highly consistent with the manual gold standard, with zero segmentation in the lesion area. This method has been validated in both clinical retrospective data and simulated high-noise scenarios, demonstrating excellent versatility and potential for widespread adoption, providing an efficient and reliable technical means for the quantitative analysis and auxiliary diagnosis of skin diseases.

[0026] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0028] Figure 1 This is the overall flow chart of this application;

[0029] Figure 2Optimizing the GAN network denoising flow chart for skin LC-OCT images in this application;

[0030] Figure 3 Flowchart of the skin LC-OCT segmentation algorithm after denoising in this application. DETAILED DESCRIPTION

[0031] The specific embodiments of this application involve the entire process of denoising and segmenting LC-OCT skin segmentation images, which cross multiple fields and cover a wide range of topics. Therefore, unless otherwise expressly stated herein, all technical and scientific terms shall have the meanings commonly understood by those skilled in the art. As used herein, the words "including," "having," and any variations thereof are intended to refer to non-exclusive inclusions, as set forth in the claims and accompanying drawings.

[0032] The following describes the skin LC-OCT image segmentation method of the improved GAN and multi-feature fusion embodiment of the present application with reference to the accompanying drawings.

[0033] like Figure 1 As shown, the skin LC-OCT image segmentation method of the improved GAN and multi-feature fusion embodiment of the present application includes the following steps:

[0034] S1. Data collection:

[0035] Using an LC-OCT device to perform OCT data acquisition, collecting skin data as samples, and saving images; the acquisition includes skin B-scan samples of different subjects;

[0036] S2, and select and classify to ensure that the image information is discernible and representative;

[0037] S3. LC-OCT image denoising:

[0038] The image takes a noisy image as input and a real image as the target. This application performs improved generative adversarial network (GAN) training: the generator generates an output close to the target image from the noisy image, and its authenticity is determined by the discriminator, and then fed back to the generator. At the same time, the similarity between the generated image and the target image is evaluated by the twin network and also fed back to the generator. In the joint training, the generator gradually generates high-quality images through multi-task optimization such as perceptual loss, structural similarity, and discriminator feedback, achieving the effects of denoising and image enhancement, and thus outputting a denoised image. The important parts of the improved generative adversarial network are described in detail below, including related modules such as reflection filling, multi-scale fusion, and self-attention, as well as the network structure design of the generator, discriminator, and twin network, as well as the implementation methods of perceptual loss, multi-scale structural similarity, and Wasserstein loss function.

[0039] S3.1. Related modules

[0040] S3.1.1 Reflection Fill

[0041] In the convolution operation, if the image edge is directly truncated or zero-filled, the features at the boundary will be lost or distorted, thus affecting the response of the subsequent convolution kernel to these areas. Reflection filling expands the image by "mirroring" the adjacent pixels at each edge, which not only keeps the input and output sizes consistent (if the original image height and width are and , the number of channels is , fill the edges rows of pixels, the size after padding is , can also maximize the preservation of edge information, reduce artifacts and enhance the model's ability to perceive edge features. Specifically, for the filled coordinates , the reflection mapping will map it to the original image coordinates, where and They are obtained by the following formulas:

[0042] (2-1)

[0043] (2-2)

[0044] By reflecting and padding 3 pixels around the input image (450×900), the size can be expanded to 456×906, thus maintaining the same height and width as the original image before and after convolution and deconvolution. Reflection padding does not fill in a constant, but mirrors the original pixels along the edge; this not only expands the convolution receptive field of the boundary, but also avoids the discontinuity artifacts caused by zero padding. Position after padding It will be mapped back to the real coordinates in the original image through symmetric mapping, ensuring the natural connection between edge features and internal features, and can significantly improve the detail preservation effect in high-resolution image generation and edge perception tasks.

[0045] S3.1.2 Multi-scale Fusion Design

[0046] Multi-scale feature fusion aims to address the difficulty of balancing global structure and detail at a single resolution. In the generator, after the initial convolution, the image enters a series of downsampling layers with a stride of 2, each time reducing the spatial dimension and increasing the number of channels to gradually capture macroscopic semantics. At the minimum resolution, residual blocks further deepen the feature representation; then, transposed convolution layers restore the spatial size in reverse order. At each upsampling stage, the current feature map is concatenated with the features saved in the corresponding downsampling stage along the channel dimension. Then, through 3×3 convolution, batch normalization, and ReLU activation, it strengthens global consistency while preserving local details.

[0047] The discriminator also employs a multi-scale design: the input is continuously downsampled to extract feature maps at different resolutions. These coarse and fine features are then fused at the same scale, allowing the discriminant network to simultaneously focus on large-scale textures and subtle edges. This cross-layer connection enables the discriminator to more accurately distinguish between genuine and fake images, improving overall discrimination performance.

[0048] S3.1.3 Self-attention module design

[0049] In traditional generators, the convolution kernel size and network depth jointly limit the receptive field, while restoring resolution through deconvolution easily introduces blurring and edge artifacts. To address these issues, we embed a self-attention module in the generator to globally model long-range dependencies between pixels. This module first maps the feature map into query, key, and value. It then calculates the correlation between any two points using dot-product attention and then takes a weighted sum of the values ​​to generate a new global feature representation.

[0050] To balance training stability and attention effectiveness, the module output is added to the original features via a residual connection with a learnable scalar γ. Initially, γ ≈ 0, making the network behave like a convolutional architecture without attention. As training progresses, γ increases autonomously, and the attention mechanism gradually takes effect, accelerating convergence and effectively suppressing mode collapse.

[0051] S3.2 Network Structure

[0052] This network employs both generative adversarial and Siamese architectures: It takes noisy images from the raw folder as input and real images from the average folder as targets. Guided by an adversarial discriminator, the generator continuously approximates the distribution of real images. Simultaneously, the Siamese network calculates the feature distance between the generated output and the target, using structural similarity as feedback. During training, the generator integrates perceptual loss, structural similarity loss, and adversarial loss, gradually improving output quality through a multi-task approach, achieving effective denoising and enhancement.

[0053] S3.2.1. Generator Model

[0054] The generator starts with reflection padding and a 7×7 convolution kernel, followed by batch normalization and ReLU activation to extract initial edge and texture features. Three 3×3 convolutions with a stride of 2 are then performed to downsample the spatial resolution to 1 / 8 the original and expand the number of channels to nf×2, nf×4, and nf×8, respectively. Each convolution is followed by batch normalization and ReLU to ensure training stability and enhance nonlinear representation.

[0055] At the minimum resolution, a series of residual blocks (ResNetBlock) work in conjunction with a self-attention module: residual connections enable each module to retain shallow information from the input, while self-attention globally models long-range dependencies, deepening the semantic representation. The generator then gradually restores the spatial size through three transposed convolutions. Each level of upsampling is concatenated and fused with the corresponding downsampled feature map in the channel dimension. The generator then undergoes another 3×3 convolution, BatchNormalization, and ReLU processing, balancing low-level details and high-level semantics. Finally, the feature map undergoes another 7×7 convolution and uses a tanh activation to map it to the target channel range. The generated result is then residually added to the input image, ensuring the output is within the [−1, 1] interval and further improving image consistency and detail preservation.

[0056] S3.2.2 Discriminator Module

[0057] The discriminator first performs a preliminary downsampling of the input image using a 4×4 convolution kernel (stride 2). Batch normalization is not used in this layer to avoid instability in the early stages of training. It is then activated with a LeakyReLU. The network then undergoes further downsampling using three identical 4×4 convolution blocks (stride 2), each followed by BatchNormalization and LeakyReLU. This reduces the spatial resolution while extracting deeper image features. The extracted multi-scale features are fused in a 3×3 convolution layer, integrating both detailed and global information. A 4×4 convolution with stride 1 then maps the features to a single-channel output, which is then flattened and fed into two fully connected layers: the first layer contains 1024 neurons and uses a tanh activation, while the second layer outputs a single scalar, mapped to [0, 1] using a sigmoid function, representing the probability that the image is a real sample. A discriminant value closer to 1 indicates a more likely real image; a value closer to 0 indicates a more likely generated image.

[0058] S3.2.3 Siamese Network Model

[0059] The two input images are first encoded using two identical, weight-sharing convolutional subnetworks. Each subnetwork consists of five convolutional layers with kernel sizes of 10×10, 7×7, 4×4, 4×4, and 4×4, respectively. Activation functions and pooling layers are interspersed to extract hierarchical features, ultimately mapping each image into a 4096-dimensional vector. The absolute difference between elements at corresponding positions is then calculated to obtain an L1 distance vector. This vector is then passed through a fully connected layer and activated with a sigmoid function, and the result is normalized to [0, 1] to serve as the similarity score between the two images. This structure ensures consistent feature extraction and determines similarity by directly measuring feature differences.

[0060] S3.3 Loss Function

[0061] S3.3.1. Perceptual Loss

[0062] Perceptual loss aims to measure the difference between the generated image and the real image in the high-level semantic space. It uses the pre-trained VGG network to extract feature maps at a certain intermediate layer and calculates the mean squared error between the two in this feature domain (Equation (2-3)). This avoids direct comparison at the pixel level and focuses more on the content consistency of the image. The formula is: (2-3)

[0063] Among them is Real images, is to generate an image, represents the feature map extracted by the VGG network, is the number of pixels in the image.

[0064] S3.3.2 Multi-scale structural similarity index loss

[0065] The multi-scale structural similarity (MSSIM) loss combines the perceptual loss with the structural similarity loss. At multiple resolutions, SSIM evaluates the local consistency of image patches by using three components: brightness, contrast, and structure (Formulas (2-4) and (2-5)). and Balancing the two weights makes the generated image close to the real sample in terms of overall structure while retaining the detailed texture. and The formula is as follows:

[0066] (2-4)

[0067] (2-5)

[0068] in, is the perceptual loss, is the structural similarity loss, and is a hyperparameter used to balance the weights of the two loss terms. and They are images and The mean of and is an image and The standard deviation of is their covariance, and is a constant used to stabilize the calculation.

[0069] S3.3.3 Wasserstein loss

[0070] Wasserstein loss (Equation (2-6)) approximates the earth mover distance by the difference between the output of the discriminator for real samples and generated samples, and is used to measure the distance between the two distributions. Compared with the binary classification loss of traditional GAN, Wasserstein loss is smoother and more stable, and can effectively alleviate the problems of gradient disappearance and mode collapse. Wasserstein loss The formula is as follows:

[0071] (2-6)

[0072] in, is the discriminator's response to the real data The output, is the image generated by the generator, and They are data distribution and generated data distribution respectively.

[0073] S4. Perform segmentation preprocessing based on thresholding or morphological operations, and limit the pixel value range to [0, 1] to eliminate illumination differences;

[0074] S5. OCT image segmentation algorithm design:

[0075] This application designs a new skin LC-OCT B-Scan image segmentation method to identify representative structural features in skin images, thereby achieving accurate segmentation of skin layers and constructing an enhanced dataset. The details are as follows:

[0076] S5.1. Input denoised image:

[0077] The original skin LC-OCT images were denoised using an improved GAN model. After denoising, the image size was maintained at 450×900 and the image was segmented.

[0078] S5.2. Gradient calculation and initial path search:

[0079] First, a starting pixel is manually selected on a representative layer of the skin image as the reference point for the segmentation path. Then, Sobel horizontal and vertical operators are applied to the entire image, calculating the gradient vector of each pixel to reveal its edge strength and direction. Using the gradient information of this reference point as a reference, a progressive search is performed in adjacent columns in both the left and right directions. The pixel with the smallest gradient difference from the current point is selected as the next node. This process is repeated until the image edge is reached, ultimately constructing a smooth and continuous skin segmentation path.

[0080] The dynamic path update strategy adopted by this application will determine the next path node according to the gradient difference between the current point and the adjacent column candidate points in each search step, thereby adaptively tracking the skin layer structure. Take the right search as an example: for position Search 5 pixels above and below the adjacent column to the right ,in . Calculate the difference between them and the current point gradient for:

[0081] (2-7)

[0082] in, is the gradient of 5 pixels above and below adjacent columns, is the gradient of the current point. The pixel that minimizes this gradient is selected as the new starting point. This iterative process continues, extending along the direction of the smoothest gradient until it reaches the edge of the image, ultimately generating a smooth, continuous segmentation path that closely adheres to the natural layers of the skin.

[0083] S5.3, Random Forest Dynamic Path Optimization:

[0084] A random forest model is introduced into the path search to optimize candidate point selection. Multi-dimensional features of candidate points are extracted: gradient amplitude, neighborhood gradient directional consistency (5×5 window histogram), lateral position offset, and local texture feature (LBP). During the training phase, a random forest classifier is constructed using labeled data (positive samples: correct path points, negative samples: deviated points) to predict the probability that a candidate point belongs to the correct path. . Comprehensive gradient difference and random forest probability score during online search ,The candidate point with the highest score is selected to enhance the path ,robustness to noise and diseased areas.

[0085] S5.4. Segmentation Region Division and Dual-Path Fitting:

[0086] To solve the problem of overlapping layers caused by skin lesions, the image is divided into Divide into left and right independent areas. The left and right areas perform path search from the initial point and departure point respectively to avoid cross-area interference. , Perform adaptive spline fitting: dynamically adjust the number of sampling points according to the path curvature (10 points for flat areas and 30 points for complex areas), and generate subsets through linear interpolation , and optimize the smoothing spline function ,in The spline function is The function value at the sampling point, For the sampling points Coordinates (observations), is the smoothing parameter, is the square of the second-order derivative, which measures the curvature of the curve.

[0087] S5.5, Random Forest Assisted Path Correction

[0088] Candidate offset paths (±3 pixels) are generated from the initial fitted path. Multidimensional features of the candidate paths are extracted: average gradient, path curvature (second-order derivative), and surrounding texture consistency (gray-level co-occurrence matrix). A random forest model predicts the confidence level of each candidate path, and the path with the highest confidence level is selected as the final segmentation boundary. This step effectively corrects path deviations caused by noise or local structural anomalies, improving inter-layer discrimination in complex areas.

[0089] S5.6. Result Verification and Quantitative Evaluation

[0090] Segmentation performance was evaluated through visual comparison and quantitative metrics: the Hausdorff distance (measure of maximum local deviation) between the segmented path and the manually annotated gold standard, the Dice coefficient (measure of regional overlap), and the average gradient consistency of the segmented edges were calculated. For lesion images, the continuity of the segmented path within the macular depression (number of breakpoints) was also statistically analyzed.

[0091] S6. Output the segmentation result image with the segmentation result to achieve the segmentation effect, and compare and confirm it with the original image.

[0092] This application constructs an end-to-end LC-OCT B-Scan skin segmentation processing framework by deeply integrating the improved generative adversarial network and the intelligent path optimization algorithm. This framework integrates denoising, enhancement and segmentation, and can run stably under high noise and complex lesion backgrounds without relying on tedious manual parameter adjustment; at the same time, multi-scale feature extraction, self-attention mechanism and structural consistency constraints of the twin network effectively take into account global texture and microscopic details, and the output quality far exceeds traditional methods. The overall process can be processed in seconds on a medium-configuration computer, providing efficient, accurate and repeatable technical guarantees for large-scale clinical review analysis and data set construction.

[0093] For image denoising, the generator in this application uses reflection padding to initiate a multi-scale encoding-decoding process. It extracts and reconstructs features through successive downsampling and upsampling layers. A self-attention module with residual blocks and a learnable scalar γ is incorporated at the minimum resolution to enhance the ability to capture long-range dependencies. During training, the perceptual loss (VGGMSE), multi-scale SSIM loss, and Wasserstein adversarial loss are collaboratively optimized, enabling the network to suppress noise while maximally preserving skin layering and texture details. The interactive feedback between the multi-scale discriminator and the Siamese network further enhances the accuracy of the real structure, resulting in denoised images that are both smooth and natural, yet rich in detail.

[0094] For skin layer segmentation, this application introduces a dynamic path search based on Sobel gradient differences and random forest scoring. Starting from a starting point, it compares adjacent candidate pixels column by column along the left and right sides, automatically outlining a smooth and coherent preliminary boundary. Subsequently, it adaptively adjusts the sampling density based on curvature and uses linear interpolation and spline fitting to balance path accuracy and smoothness. Finally, it corrects lesion misalignment by offsetting candidates by ±3 pixels and re-evaluating with random forests. This segmentation algorithm approaches the manual gold standard in multiple quantitative metrics, including Hausdorff distance, Dice coefficient, and edge gradient consistency, achieving a harmonious combination of high precision and high robustness.

[0095] In summary, the skin LC-OCT image segmentation method proposed in this application, based on an improved generative adversarial network (GAN) and multi-feature fusion, significantly outperforms existing technologies in many aspects: First, a deep GAN integrating reflection filling, multi-scale encoding-decoding, self-attention, and residual blocks can precisely reconstruct the microstructure of each layer of skin while suppressing complex speckle noise; second, the structural consistency constraint of the siamese network and the cross-layer feature fusion of the multi-scale discriminator ensure that the denoised image fully preserves both global texture and local details. Furthermore, through dynamic path search based on Sobel gradient and random forest scores, adaptive spline fitting, and offset correction, it can automatically track complex boundaries such as lesions, maintaining interlayer continuity and structural integrity. The overall algorithm can achieve processing in seconds on a medium-spec computer, requires low computing resources, and has good interpretability and stability, providing an efficient and reliable solution for the construction and automated quantitative analysis of large-scale clinical datasets.

[0096] Compared with conventional GAN ​​denoising methods, this method additionally introduces multi-objective joint losses (VGGMSE perceptual loss, multi-scale SSIM loss, Wasserstein adversarial loss) and L1 similarity feedback of the twin network, so that network training focuses on both semantic consistency and structural fidelity; compared with manual segmentation operations, it not only greatly reduces subjective errors and time costs, but also achieves segmentation accuracy comparable to the manual gold standard, showing excellent robustness.

[0097] Obviously, the specific implementation methods described above are only some embodiments of the present application, not all implementation methods. The contents shown in the drawings are descriptions of the preferred embodiments of the present application, but do not limit the scope of the present application. All equivalent structures covered by the description and drawings of this application, or equivalent improvement schemes directly or indirectly applied in the relevant technical field, should be included in the scope of protection of this application. The present application can be implemented in various forms, and the embodiments provided are only used to fully illustrate the contents of this application. Technicians in the technical field can make corresponding changes or replacements to the specific implementation methods, and these changes and replacements should be deemed to fall within the scope of protection of this application.

Claims

1. Improved GAN and multi-feature fusion skin LC-OCT image segmentation method, characterized by: The following steps are involved: S1. Use LC-OCT equipment to collect skin B-scan samples from different subjects and save them as image data; S2. Select and classify the collected images to ensure that the image information is discernible and representative; S3, LC-OCT image denoising: An improved generative adversarial network is used for image denoising, with noisy images as input and real images as targets for training; The generator captures structural information through multi-scale feature extraction and self-attention mechanism, the discriminator evaluates the authenticity of the image, and the Siamese network calculates the similarity probability. The perceptual loss, MSSIM loss and adversarial loss are integrated to optimize the generation quality. S4. Perform segmentation preprocessing based on thresholding or morphological operations, and limit the pixel value range to [0, 1] to eliminate illumination differences; S5, dynamic path search based on Sobel operator and random forest multi-features, combined with adaptive spline fitting to achieve image segmentation; S6. Output the skin LC-OCT image with segmentation results to achieve accurate segmentation effect, and compare and confirm with the original image.

2. The skin LC-OCT image segmentation method based on the improved GAN and multi-feature fusion according to claim 1 is characterized in that: The step S3 includes the following steps: performing reflection filling on the input image, the generator extracts features through multi-scale downsampling and processes them through the residual block and self-attention module, and then fuses features of different scales through upsampling, the discriminator uses multi-scale discrimination to evaluate the authenticity of the image, and the twin network calculates the similarity probability.

3. The skin LC-OCT image segmentation method based on the improved GAN and multi-feature fusion according to claim 1 is characterized in that: The step S5 includes the following steps: using the Sobel operator to calculate the gradient and calibrate the initial search point, combining the gradient features, texture features, and position features to perform random forest path optimization, dynamically adjusting the sampling points according to the complexity to perform spline fitting, and correcting the lesion area through random forest.

4. The skin LC-OCT image segmentation method based on the improved GAN and multi-feature fusion according to claim 2 is characterized in that: In the multi-scale feature extraction of the generator, downsampling uses a 3×3 convolution kernel with a step size of 2, the number of channels is doubled successively, and the resolution is reduced from 450×900 to 56×112. Upsampling restores the resolution through transposed convolution and is spliced ​​and fused with the corresponding features.

5. The skin LC-OCT image segmentation method based on the improved GAN and multi-feature fusion according to claim 4 is characterized in that: The self-attention module converts the feature map into Q, K, and V matrices to calculate the global dependency weights. The discriminator uses four 4×4 convolution downsamplings with LeakyReLU activation. The twin network extracts features through a 5-layer CNN to calculate the L1 distance.

6. The skin LC-OCT image segmentation method based on the improved GAN and multi-feature fusion according to claim 3 is characterized in that: In the random forest multi-features, the gradient feature is calculated by the Sobel operator, the texture feature is extracted using the LBP descriptor, and the position feature includes pixel coordinate information. The random forest predicts the optimal segmentation path by extracting gradient, curvature, and texture features.

7. The skin LC-OCT image segmentation method based on the improved GAN and multi-feature fusion according to claim 6 is characterized in that: The adaptive spline fitting dynamically adjusts the sampling density according to the path complexity: sparse sampling is used in flat areas to reduce the amount of calculation, dense sampling is used in complex areas to improve the accuracy, and spline fitting smoothing processing is used to reduce noise interference.

8. An apparatus for improving the skin LC-OCT image segmentation method by combining GAN and multiple features, used in the skin LC-OCT image segmentation method by combining GAN and multiple features according to any one of claims 1 to 7, characterized in that: The system comprises a computer including a processor, a memory and a communication interface; an LC-OCT device for collecting skin B-scan samples; and a computer program stored in the memory.

Images