Skin LC-OCT image segmentation method based on improved GAN and multi-feature fusion

Abstract

The invention discloses a skin LC-OCT image segmentation method based on improved GAN and multi-feature fusion. The skin LC-OCT image segmentation method comprises the following steps: collecting a skin B-scan sample by adopting LC-OCT equipment and storing the skin B-scan sample as an image; the method comprises the following steps: constructing a generator fusing reflection filling, multi-scale coding-decoding, self-attention and residual blocks, cooperating with a multi-scale fusion discriminator and a weight sharing twin network, and realizing high-fidelity denoising and enhancement through joint optimization of perception loss, multi-scale SSIM loss and Wasserstein adversarial loss; preprocessing and positioning a skin area based on a threshold value or morphology and normalizing the skin area; gradient is extracted from the de-noised image, and dynamic path search is carried out in combination with a gradient difference minimum criterion and a random forest; and after adaptive sampling and spline fitting are carried out on the path, candidates are generated through a random forest, and fracture or dislocation of a focus area is corrected. Therefore, accurate segmentation of the LC-OCT image of the skin layer can be realized in a high-noise environment, a high-quality enhanced data set can be constructed, and a reliable basis is provided for early diagnosis of skin diseases.

Description

Technical Field

[0001] This application relates to the field of medical image processing technologies, and in particular, to a skin LC-OCT image segmentation method that improves GAN and multi-feature fusion. Background Art

[0002] Optical coherence tomography (OCT) is a non-invasive optical tomography technology based on the principle of low-coherence interferometry. Since its first proposal in 1991, various implementation methods such as time-domain OCT (TD-OCT) and frequency-domain OCT (FD-OCT) have been developed. OCT uses a broadband low-coherence light source to irradiate biological tissues. By interfering the backscattered light of the sample with the light in the reference arm, the reflection intensity distribution at different depths is measured, thereby reconstructing two-dimensional or three-dimensional tomographic images inside the tissue with a micron-level axial resolution (about 1–10 μm). Compared with traditional dermoscopy and CT imaging, OCT can reveal the subtle morphological changes of the epidermis, dermis, and appendage structures (such as hair follicles, sweat glands) in real time and non-invasively, and has higher sensitivity to tissue boundaries and heterogeneity. In recent years, a large number of studies have applied OCT to the early detection, lesion localization, and intraoperative navigation of skin diseases, demonstrating excellent imaging depth and resolution advantages, and providing important structural and functional information for clinical diagnosis and treatment decisions.

[0003] Skin segmentation is an important technology for skin disease diagnosis, and its core is to accurately segment the structures of each layer of the skin from optical coherence tomography (OCT) images. With the wide application of high-resolution LC-OCT (line-field confocal optical coherence tomography), the richness of image details has been significantly improved, but interference problems such as speckle noise and motion artifacts have become more prominent. Existing methods are mainly divided into two categories: traditional image processing technologies and deep learning methods, but both have obvious defects. Traditional methods are based on edge detection (such as Sobel and Canny operators) and path search algorithms (such as dynamic programming). Although they do not require labeled data, they are extremely sensitive to noise and rely on manual selection of initial points, making it difficult to handle layer boundary blur or lesion areas (such as distortion, bleeding). In addition, although traditional denoising algorithms (such as median filtering and wavelet denoising) can suppress noise, they will blur the fine structures between skin layers, resulting in a decrease in segmentation accuracy. The limitations of such methods are particularly significant in complex clinical scenarios and cannot meet the requirements of full automation and high robustness.

[0004] In recent years, segmentation methods based on deep learning (such as U-Net, Transformer) have shown excellent performance on public datasets, but their implementation faces severe challenges. On the one hand, deep learning models require a large number of LC-OCT images with pixel-level annotations for training, while the cost of medical data annotation is high, and it is difficult to ensure the consistency of annotations by different physicians. On the other hand, such models have high computational complexity and rely on GPU acceleration, making it difficult to deploy them to low-cost portable medical devices. More importantly, the "black box" characteristic of deep learning models leads to the lack of interpretability in their decision-making process, making it difficult for doctors to trust their output results and hindering clinical practical applications.

[0005] Existing technologies have tried to combine traditional methods with deep learning (such as using convolutional networks to optimize gradient calculation), but there are still significant deficiencies. Denoising and segmentation are designed independently in stages, resulting in the cumulative error; the number of model parameters is huge and cannot meet the real-time requirements; and there is a lack of a fusion mechanism for multiple features (gradient, texture, spatial position), making it difficult to adaptively adjust the segmentation strategy. Summary of the Invention

[0006] This application aims to solve at least one of the technical problems in the related technologies to some extent.

[0007] To this end, the purpose of this application is to propose a skin LC-OCT image segmentation method that improves GAN and fuses multiple features, which can not only effectively suppress noise interference, but also achieve full-automatic segmentation through a lightweight model, while retaining the decision-making logic that doctors can understand.

[0008] To achieve the above object, this application proposes a skin LC-OCT image segmentation method that improves GAN and fuses multiple features, including the following steps:

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

[0010] S2. Select and classify the collected images to ensure that the image information has discriminability and representativeness;

[0011] S3. Denoise the LC-OCT image: Use an improved generative adversarial network to denoise the image, and train with the noise image as the input and the real image as the target; the generator captures structural information through multi-scale feature extraction and self-attention mechanism, and fuses perceptual loss, MSSIM loss and adversarial loss to optimize the generation quality;

[0012] S4. Perform segmentation preprocessing based on the threshold method or morphological operation, and limit the pixel value range to [0,1] to eliminate the illumination difference;

[0013] S5. Perform dynamic path search based on the Sobel operator and multiple features of the random forest, and combine adaptive spline fitting to achieve image segmentation;

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

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

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

[0017] Further, the step S3 includes the following steps: perform reflection padding on the input image, the generator extracts features through multi-scale downsampling, processes them through residual blocks and self-attention modules, then fuses different-scale features through upsampling, the discriminator uses multi-scale discrimination to evaluate the authenticity of the image, and the siamese network calculates the similarity probability.

[0018] Further, the step S5 includes the following steps: use the Sobel operator to calculate the gradient and calibrate the initial search point, optimize the random forest path by combining gradient features, texture features, and position features, dynamically adjust the sampling points according to the complexity for spline fitting, and correct the lesion area through random forest processing.

[0019] Further, in the multi-scale feature extraction of the generator, the downsampling uses a 3×3 convolution kernel with a stride of 2, the number of channels doubles in turn, and the resolution is reduced from 450×900 to 56×112. The upsampling restores the resolution through transposed convolution and splices and fuses it with the corresponding features.

[0020] Further, the self-attention module converts the feature map into Q, K, V matrices to calculate the global dependence weight, the discriminator uses four 4×4 convolution downsamplings in cooperation with the LeakyReLU activation, and the siamese network extracts features through 5 layers of CNNs to calculate the L1 distance.

[0021] Further, in the random forest multiple features, the gradient feature is calculated by the Sobel operator, the texture feature is extracted by the LBP descriptor, the position feature includes pixel coordinate information, and 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 adopted in flat areas to reduce the computational load, dense sampling is adopted in complex areas to improve the accuracy, and the spline fitting is smoothed to reduce noise interference.

[0023] Compared with the prior art, the present application has the following advantages: by integrating and improving the generative adversarial network (reflection filling, multi-scale encoding-decoding, self-attention and residual blocks) and dynamic path optimization (Sobel gradient difference criterion, random forest evaluation, adaptive spline fitting and offset correction), high-fidelity denoising and accurate segmentation of B-Scan skin images of LC-OCT are achieved. On the one hand, the joint optimization of triple losses (VGGMSE perceptual loss, multi-scale SSIM loss, Wasserstein adversarial loss) effectively preserves the denoised image in both macroscopic texture and microscopic details; on the other hand, the dynamic path search and random forest correction not only automatically track the interlayer boundaries of complex lesion areas, but also maintain the lamination continuity at the fovea centralis, providing a reliable basis for constructing a high-quality enhanced dataset.

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

[0025] In addition, on a medium-configured computer, the processing time of each 450×900 OCT image of the present application can be controlled within seconds; verified by the Hausdorff distance and Dice coefficient, the segmentation results are highly consistent with the manual gold standard and there are zero breaks in the lesion areas. This method has been verified in clinical retrospective data and simulated high-noise scenarios, demonstrating excellent versatility and promotion value, 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 following description, become apparent in part from the following description, or be learned through the 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 be readily understood from the following description of the embodiments in conjunction with the drawings, wherein:

[0028] Figure 1 is the overall flowchart of the present application;

[0029] Figure 2Flowchart of the optimized GAN network for denoising skin LC-OCT images in this application;

[0030] Figure 3 Flowchart of the segmentation algorithm for skin LC-OCT after denoising in this application. Detailed implementation manners

[0031] The detailed implementation manners of this application involve the entire process of denoising and segmenting LC-OCT skin segmentation images, with multiple fields intersecting and a wide coverage. Therefore, unless otherwise clearly stated herein, the meanings of all technical and scientific terms should conform to the general understanding of those skilled in the art. The terms "including", "having" and any variants thereof used herein are intended to refer to a non-exclusive inclusion relationship, as shown in the claims and the drawings.

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

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

[0034] S1. Data acquisition:

[0035] Use an LC-OCT device to collect OCT data, collect skin data as samples, and save the images; the collection is, for example, skin B-scan samples of different subjects;

[0036] S2. And perform selection and classification to ensure that the image information has discriminability and representativeness;

[0037] S3. LC-OCT image denoising:

[0038] The image takes the noisy image as the input and the real image as the target. This application performs improved generative adversarial network training (GAN): the generator generates an output close to the target image from the noisy image, and its authenticity is judged 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 through a siamese 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 the denoised image. The following will detail the important parts in the improved generative adversarial network, including related modules such as reflection padding, multi-scale fusion, and self-attention, as well as the network structure design of the generator, discriminator, and siamese network, and the implementation methods of perceptual loss, multi-scale structural similarity, and Wasserstein loss function.

[0039] S3.1. Related modules

[0040] S3.1.1. Reflection Padding

[0041] In the convolution operation, if the image edges are directly truncated or zero padding is used, the features at the boundaries will be lost or distorted, thus affecting the response of the subsequent convolutional kernels to these regions. Reflection padding extends the image by "mirroring" and replicating adjacent pixels on each edge, not only maintaining the same size of the input and output (if the original image has a height of and , a channel number of , and rows of pixels are padded on each edge, the size after padding is ), but also maximizing the retention of edge information, reducing artifacts, and enhancing the model's perception ability of edge features. Specifically, for the padded coordinate , the reflection mapping will map it to the original image coordinates, where and are obtained by the following formulas respectively:

[0042] (2-1)

[0043] (2-2)

[0044] By reflecting and padding 3 pixels on each side of the input image (450×900), the size can be extended 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 constants but mirrors and replicates the original pixels along the edges; this not only expands the convolutional receptive field of the boundary but also avoids the discontinuous artifacts brought by zero padding. The position after padding 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 significantly improving the detail retention 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 make up for the deficiency that it is difficult to balance global structure and detail performance at a single resolution. In the generator, the image enters a series of downsampling layers with a stride of 2 after the initial convolution, each time reducing the spatial dimension and increasing the number of channels to gradually capture the macroscopic semantics. At the minimum resolution, the residual blocks further deepen the feature expression; subsequently, the transposed convolution layers restore the spatial size in the 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, and then through 3×3 convolution, batch normalization, and ReLU activation, both the global consistency is strengthened and the local details are retained.

[0047] The discriminator also adopts a multi-scale design: the input is continuously downsampled to extract feature maps at different resolutions, and then these coarse and fine features are fused at the same scale, enabling the discriminative network to simultaneously focus on large-scale textures and tiny edges. Through this cross-layer connection, the discriminator can more accurately distinguish real and fake images and improve the overall discriminative performance.

[0048] S3.1.3, Self-attention Module Design

[0049] In traditional generators, the size of the convolutional kernel and the depth of the network jointly limit the receptive field, and restoring the resolution through transposed convolution is prone to introducing blurring and edge artifacts. To address these issues, we embed a self-attention module in the generator to model long-range dependencies between pixels globally. This module first maps the feature map into Query, Key, and Value, calculates the correlation between any two points through dot-product attention, and then weights and sums the Value to generate a new global feature representation.

[0050] To balance training stability and attention effect, the module output is added to the original feature through a residual connection with a learnable scalar γ. Initially, γ≈0, making the network behavior equivalent to a convolution structure without attention; as training progresses, γ increases autonomously, and the attention mechanism gradually comes into play, thereby accelerating convergence and effectively suppressing mode collapse.

[0051] S3.2, Network Structure

[0052] This network simultaneously adopts a generative adversarial and siamese structure: using the noisy images in the raw folder as the input and the real images in the average folder as the target. The generator continuously approaches the real image distribution under the guidance of the adversarial discriminator; at the same time, the siamese network calculates the feature distance between the generated result and the target, with the structural similarity as the feedback. During training, the generator comprehensively perceives the perceptual loss, structural similarity loss, and adversarial loss, and gradually improves the output quality under the drive of multiple tasks, achieving effective denoising and enhancement.

[0053] S3.2.1, Generator Model

[0054] The generator starts with reflection padding and a 7×7 convolutional kernel, followed by batch normalization and ReLU activation to extract the initial edge and texture features. Subsequently, through three 3×3 convolutions with a stride of 2, the spatial resolution is downsampled to 1 / 8 of the original and the number of channels is successively expanded to ngf×2, ngf×4, and ngf×8. After each convolution, BatchNormalization and ReLU are used to ensure training stability and enhance the non-linear expression.

[0055] At the minimum resolution, a series of residual blocks (ResNetBlock) work jointly with the self-attention module: the residual connection enables each module to retain the shallow information of the input, while self-attention models long-range dependencies globally to deepen the semantic representation. Subsequently, the generator gradually restores the spatial size through three transposed convolutions. After each upsampling stage, it concatenates and fuses with the corresponding feature map obtained from downsampling in the channel dimension, and is further processed by a 3×3 convolution, Batch Normalization, and ReLU, thus taking into account both low-level details and high-level semantics. Finally, the feature map passes through another 7×7 convolution and uses the tanh activation to map to the target channel range. The generated result is then added to the input image residually to ensure that the output is within the range of [−1,1] and further enhance image consistency and detail retention.

[0056] S3.2.2. Discriminator Module

[0057] The discriminator first uses a 4×4 convolutional kernel (stride 2) to perform preliminary downsampling on the input image. Batch normalization is not used in this layer to avoid instability in the early stage of training, followed by LeakyReLU activation. Subsequently, the network further downsamples through three 4×4 convolutional blocks with the same configuration (stride 2). After each block, Batch Normalization and LeakyReLU are connected to extract deeper image features while continuously compressing the spatial resolution. The extracted multi-scale features are fused in a 3×3 convolutional layer to integrate details and global information. Then, a 4×4 convolution with a stride of 1 maps the features to a single-channel output, which is then flattened and input into two fully connected layers: the first layer contains 1024 neurons and uses tanh activation, and the second layer outputs a single scalar and is mapped to [0,1] through sigmoid, representing the probability that the image is a real sample. The closer the discriminant value is to 1, the more likely it is a real image; the closer it is to 0, the more likely it is a generated result.

[0058] S3.2.3. Siamese Network Model

[0059] Two input images are first encoded separately through two identical and weight-sharing convolutional sub-networks. Each sub-network consists of five convolutional layers with kernel sizes of 10×10, 7×7, 4×4, 4×4, and 4×4 in sequence, interleaved with activation functions and pooling layers to extract hierarchical features. Finally, each image is mapped to a 4096-dimensional vector. Subsequently, the absolute difference of the corresponding elements is calculated to obtain the L1 distance vector, which is then mapped through a fully connected layer and activated by sigmoid to normalize the result to [0,1] as the similarity score of the two images. This structure ensures the consistency of feature extraction and completes similarity judgment by directly measuring feature differences.

[0060] S3.3. Loss Function

[0061] S3.3.1. Perceptual Loss

[0062] The perceptual loss aims to measure the difference between the generated image and the real image in the high-level semantic space. It extracts feature maps at an intermediate layer with the pre-trained VGG network and calculates the mean square error between the two in this feature domain (Equation (2-3)), avoiding direct comparison at the pixel level, thus paying more attention to the content consistency of the images. The perceptual loss The formula is: (2-3)

[0063] where is the real image, is the generated image, represents the feature map extracted by the VGG network, and 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 and the structural similarity loss. At multiple resolutions, SSIM evaluates the local consistency of image patches through three components: luminance, contrast, and structure (Equations (2-4) and (2-5)), and then uses hyperparameters and to balance the weights of the two terms, so that the generated image is both close to the real sample in the overall structure and can retain the detailed texture. The MSSIM loss and The formulas are as follows:

[0066] (2-4)

[0067] (2-5)

[0068] where, is the perceptual loss, is the structural similarity loss, and are hyperparameters used to balance the weights of the two loss terms. and are the means of the images and respectively, and are the standard deviations of the images and respectively, is their covariance, and are constants used to stabilize the calculation.

[0069] S3.3.3, Wasserstein Loss

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

[0071] (2-6)

[0072] where is the output of the discriminator for real data and is the image generated by the generator, and are the data distribution and the generated data distribution respectively.

[0073] S4. Perform segmentation preprocessing based on the threshold method 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 B-Scan image segmentation method for skin LC-OCT to identify representative structural features in skin images, so as to achieve precise segmentation of skin layers and construct an enhanced dataset. Specifically as follows:

[0076] S5.1. Input denoised image:

[0077] The original skin LC-OCT image is denoised using an improved GAN model, and the size of the denoised image remains 450×900, and the image is segmented.

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

[0079] First, a starting pixel is manually selected on a typical layer of the skin image as the reference point for the segmentation path. Subsequently, the Sobel horizontal and vertical operators are applied to the entire image to calculate the gradient vector of each pixel to reveal its edge strength and direction. Referring to the gradient information of this reference point, a progressive search is carried out in the adjacent columns in the left and right directions, and the pixel with the smallest gradient difference from the current point is selected as the next node, and so on until the image edge is reached, and finally a smooth and continuous skin segmentation path is constructed.

[0080] The dynamic path update strategy adopted in this application determines the next path node based on the gradient difference between the current point and the candidate points in the adjacent column during each step of the search, thereby adaptively tracking the skin layer structure. Taking the rightward search as an example: For the position Search for 5 pixels above and below the adjacent column to the right , among which . Calculate the difference between their gradients and the current point as:

[0081] (2 - 7)

[0082] Among them, is the gradient of 5 pixels above and below the adjacent column, is the gradient of the current point. And select the pixel with the smallest value as the new starting point. Iterate in this way, continuously expand along the direction with the smoothest gradient until reaching the image edge, and finally generate a smooth, continuous and highly fitting segmentation path for the natural layers of the skin.

[0083] S5.3. Random forest dynamic path optimization:

[0084] Introduce a random forest model in the path search to optimize the candidate point selection. Extract multi-dimensional features of the candidate points: gradient amplitude, neighborhood gradient direction consistency (5×5 window histogram), horizontal position offset, local texture feature (LBP). Use labeled data (positive samples: correct path points, negative samples: deviated points) to construct a random forest classifier during the training phase, and predict the probability that the candidate point belongs to the correct path . During the online search, comprehensively consider the gradient difference and the random forest probability score , and select the candidate point with the highest score to enhance the path robustness in the noise and lesion areas.

[0085] S5.4. Segmentation region division and double-path fitting:

[0086] Aiming at the problem of interlayer overlap caused by skin lesions, divide the image into left and right independent regions according to the demarcation point . Perform path searches from the initial points in the left and right regions respectively, avoiding cross-region interference. For the left and right path point sets , , perform adaptive spline fitting: Dynamically adjust the number of sampling points according to the path curvature (take 10 points in the flat area and 30 points in the complex area), generate a subset through linear interpolation , and optimize the smoothing spline function , where is the function value of the spline function at the th sampling point, is the th sampling point's Coordinates (observed values), is the smoothing parameter, is the square of the second derivative, which measures the degree of curve bending.

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

[0088] Generate candidate offset paths (±3 pixels) for the initially fitted path, and extract multi-dimensional features of the candidate paths: average gradient, path curvature (second derivative), and surrounding texture consistency (gray-level co-occurrence matrix). The random forest model predicts the confidence of each candidate path, and selects the path with the highest confidence as the final segmentation boundary. This step can effectively correct path deviations caused by noise or local structural anomalies, and improve the inter-layer discrimination in complex regions.

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

[0090] Evaluate the segmentation effect through visual comparison and quantitative metrics: Calculate the Hausdorff distance (measuring the maximum local deviation), Dice coefficient (measuring the degree of regional overlap), and average gradient consistency of the segmentation edge between the segmented path and the gold standard of manual annotation. For lesion images, additionally count the continuity (number of break points) of the segmented path in the fovea region.

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

[0092] This application constructs an end-to-end B-Scan skin segmentation processing framework for LC-OCT by deeply integrating an improved generative adversarial network and an intelligent path optimization algorithm. This framework integrates denoising, enhancement, and segmentation, and can operate stably under high-noise and complex lesion backgrounds without relying on cumbersome manual parameter tuning; at the same time, multi-scale feature extraction, self-attention mechanism, and structural consistency constraints of the siamese network effectively take into account global texture and microscopic details, and the quality of the output results far exceeds traditional methods. The overall process can achieve second-level processing on a computer with medium configuration, providing efficient, accurate, and reproducible technical support for large-scale clinical retrospective analysis and dataset construction.

[0093] In terms of image denoising, the generator of this application adopts reflection padding to initiate a multi-scale encoding-decoding process. It extracts and reconstructs features through consecutive downsampling and upsampling layers, and incorporates a residual block and a self-attention module with a learnable scalar γ at the minimum resolution to enhance the ability to capture long-range dependencies. During the training process, the perceptual loss (VGGMSE), multi-scale SSIM loss, and Wasserstein adversarial loss are jointly optimized, enabling the network to suppress noise while maximizing the retention of skin layering and texture details. Supplementary interactive feedback from the multi-scale discriminator and the Siamese network further improves the restoration degree of the real structure, making the denoised image both smooth and natural, as well as rich in details.

[0094] In terms of skin layer segmentation, this application introduces dynamic path search based on Sobel gradient difference and random forest scoring. Starting from the starting point, it compares adjacent candidate pixels column by column on both the left and right sides to automatically outline a smooth and coherent preliminary boundary. Subsequently, it adaptively adjusts the sampling density according to the curvature and uses linear interpolation and spline fitting to balance path accuracy and smoothness. Finally, through ±3 pixel offset candidates and random forest re-evaluation, the misalignment of the lesion is corrected. This segmentation algorithm is close to the manual gold standard in multiple quantitative indicators such as Hausdorff distance, Dice coefficient, and edge gradient consistency, achieving an organic unity of high precision and high robustness.

[0095] In summary, the skin LC-OCT image segmentation method based on the improved generative adversarial network (GAN) and multi-feature fusion proposed in this application is significantly superior to the prior art in many aspects: on the one hand, the deep GAN integrating reflection padding, multi-scale encoding-decoding, self-attention, and residual blocks can finely reconstruct the microstructures of each skin layer while suppressing complex speckle noise; on the other hand, the structural consistency constraint of the Siamese network and the cross-layer feature fusion of the multi-scale discriminator are fused, enabling the denoised image to fully retain both global texture and local details. Through dynamic path search based on Sobel gradient and random forest scoring, adaptive spline fitting, and offset correction, it can automatically track complex boundaries such as the lesion area, maintaining interlayer continuity and structural integrity. The overall algorithm can achieve second-level processing on a computer with medium configuration, has low computational resource requirements, and has good interpretability and stability, providing an efficient and reliable solution for large-scale clinical dataset construction and automated quantitative analysis.

[0096] Compared with the conventional GAN denoising method, this method additionally introduces a multi-objective joint loss (VGGMSE perceptual loss, multi-scale SSIM loss, Wasserstein adversarial loss) and the L1 similarity feedback of the siamese network, enabling the network training to focus on both semantic consistency and structural fidelity; compared with manual segmentation operations, it not only significantly reduces the subjective error and time cost, but also achieves a segmentation accuracy comparable to that of the manual gold standard, demonstrating excellent robustness.

[0097] Obviously, the specific embodiments described above are only partial embodiments of this application, not all implementation schemes. The content shown in the drawings is an illustration of the preferred embodiments of this application, but does not limit the scope of this application. Any equivalent structure covered by the specification and drawings of this application, or any equivalent improvement scheme directly or indirectly applied in the relevant technical field, should be included in the protection scope of this application. This application can be implemented in various forms. The provided embodiments are only used to comprehensively elaborate on the content of this application. Those skilled in the art can make corresponding changes or substitutions to the specific embodiments, and these changes and substitutions should be regarded as belonging to the protection scope of this application.

Claims

1. An improved skin LC-OCT image segmentation method integrating GAN and multi-features, characterized in that It includes the following steps: S1. Use an LC-OCT device to collect skin B-scan samples of different subjects and save them as image data; S2. Select and classify the collected images to ensure that the image information is discriminative and representative; S3. Denoise the LC-OCT images: Use an improved generative adversarial network for image denoising, and train with the noisy image as the input and the real image as the target; The generator captures structural information through multi-scale feature extraction and self-attention mechanism, and optimizes the generation quality by fusing perceptual loss, MSSIM loss, and adversarial loss; S4. Perform segmentation preprocessing based on the threshold method or morphological operation, and limit the pixel value range to [0,1] to eliminate illumination differences; S5. Perform dynamic path search based on multi-features of the Sobel operator and random forest, and combine adaptive spline fitting to achieve image segmentation; S6. Output the skin LC-OCT image with the segmentation result, achieve an accurate segmentation effect, and compare and confirm it with the original image.

2. The improved GAN and multi-feature fusion-based skin LC-OCT image segmentation method according to claim 1, wherein The step S3 includes the following steps: Perform reflection padding on the input image. After the generator extracts features through multi-scale downsampling, it is processed by residual blocks and self-attention modules, and then fuses different-scale features through upsampling. The discriminator uses multi-scale discrimination to evaluate the authenticity of the image, and the siamese network calculates the similarity probability.

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

4. The improved GAN and multi-feature fusion-based skin LC-OCT image segmentation method according to claim 2, characterized in that, In the multi-scale feature extraction of the generator, the downsampling uses a 3×3 convolutional kernel with a stride of 2, the number of channels doubles in turn, and the resolution is reduced from 450×900 to 56×112. The upsampling restores the resolution through transposed convolution and splices and fuses with the corresponding features.

5. The improved GAN and multi-feature fusion-based skin LC-OCT image segmentation method according to claim 4, wherein, The self-attention module converts the feature map into Q, K, V matrices to calculate the global dependence weight. The discriminator uses four times of 4×4 convolutional downsampling with LeakyReLU activation, and the siamese network extracts features through 5 layers of CNN to calculate the L1 distance.

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

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

8. An apparatus for an improved GAN and multi-feature fusion-based skin LC-OCT image segmentation method, which is used for the improved GAN and multi-feature fusion-based skin LC-OCT image segmentation method according to any one of claims 1-7, characterized in that, It includes a computer, which includes a processor, a memory, and a communication interface; an LC-OCT device for collecting skin B-scan samples; and a computer program is stored in the memory.

Images