Feature sensitive image segmentation method

Abstract

The application relates to the technical field of computer vision and image processing, and discloses a feature-sensitive image segmentation method, which comprises the following steps: firstly, converting a medical image to a CIELAB space, constructing a five-dimensional feature vector, and clustering to generate a superpixel image block; then, calculating local statistical features such as the mean value of pixels in the block, information entropy and roughness; for a pixel point, based on the gradient continuity to distinguish the region attribute, combining the local feature to calculate an adaptive fractional differential order, so as to distinguish the edge enhancement and noise smoothing operation; meanwhile, a global background reference value is used to generate a saliency image; on this basis, a level set evolution equation containing a double data fitting term and an adaptive fractional length regularization term is constructed, and the zero level set contour is evolved to a target boundary. Through dynamic adjustment of the differential strategy, the application effectively overcomes the interference of uneven gray and weak boundary, and significantly improves the segmentation precision and robustness of complex medical images.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a feature-sensitive image segmentation method. Background Technology

[0002] Medical image segmentation is a crucial step in computer-aided diagnosis and treatment, and its accuracy directly impacts the reliability of clinical decisions. Currently, medical image segmentation techniques mainly encompass three categories of methods: threshold-based, deep learning-based, and deformation model-based approaches.

[0003] Threshold-based methods embody an intuitive segmentation concept: classifying image pixels based on specific pixel values. For example, the Otsu thresholding method assumes the image contains foreground and background, achieving segmentation by searching for a threshold that minimizes intra-class variance and maximizes inter-class variance; it can even be extended to multi-threshold segmentation. However, these methods are limited by the difficulty in selecting and optimizing the threshold, which is highly susceptible to factors such as image content complexity, uneven pixel distribution, and noise interference, leading to inaccurate segmentation boundaries or misclassification of regions.

[0004] Deep learning methods typically employ supervised or semi-supervised learning mechanisms to construct artificial neural networks that simulate the human visual system. Typical examples include the U-Net model, which utilizes symmetrical downsampling and upsampling structures for pixel localization and segmentation, and the Inf-Net model, which identifies infected regions from CT images through parallel partial decoders and attention mechanisms. While these "data-driven" methods excel in semantic-level segmentation, their performance is highly dependent on massive amounts of labeled training samples and high-performance computing resources. More importantly, after multiple downsampling operations, the network model often loses high-frequency information such as edges and fine structures, resulting in insufficient detail preservation in the segmentation results.

[0005] Deformation-based image segmentation methods, such as active contour models and level set methods, drive curve evolution by constructing specific energy functions. These methods can adaptively find target regions and handle topological changes, possessing a unified mathematical expression and strong generalization potential. Although users can customize energy functions according to their needs, existing level set models still face challenges when processing medical images. Due to the prevalence of uneven grayscale, blurred target edges (weak boundaries), and severe noise interference in medical images, traditional integer-order differential operators or energy functions with fixed parameters struggle to retain weak edge features while suppressing noise. This can easily lead to leakage or local minima in the evolution curve at weak boundaries, making it difficult to achieve high-quality segmentation of complex lesion regions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a feature-sensitive image segmentation method that solves the problems of decreased segmentation accuracy caused by uneven grayscale, weak boundaries, and noise interference in complex medical images.

[0007] To achieve the above objectives, the present invention provides a feature-sensitive image segmentation method, the main technical solutions of which include: First, the medical image data to be segmented is acquired, and a color space conversion process is performed to transform the image data from the RGB color space to the CIELAB color space. In the CIELAB space, the luminance components, color components, and spatial coordinate values ​​of each pixel are extracted to construct a five-dimensional feature vector. This five-dimensional feature vector is then used to perform a clustering operation on the image, dividing it into multiple non-overlapping superpixel image patches. This achieves dimensionality reduction of the image data while maintaining the consistency of local features.

[0008] Secondly, for each generated superpixel image block, the gray-level distribution of pixels within the block is statistically analyzed, and statistical features reflecting local texture and structural characteristics are calculated. These local statistical features specifically include the pixel mean representing brightness, the pixel variance representing dispersion, the image information entropy representing the degree of texture disorder, and the roughness representing the relative gray-level offset. These features provide a data foundation for subsequent adaptive parameter adjustment.

[0009] In the parameter determination phase, this method introduces a gradient continuity discrimination mechanism for each pixel in the image. By calculating the difference in gradient magnitude between the current pixel and its eight neighboring pixels, it identifies whether the pixel is located in an edge region or a noise region. Subsequently, by combining the image information entropy and roughness of the superpixel image block to which the pixel belongs, an adaptive fractional derivative order matching the local features of the pixel is calculated. This step enables the dynamic adjustment of the derivative order according to the local features of the image, that is, different derivative strategies are used in different regions.

[0010] Meanwhile, to overcome the problem of insufficient local contrast, this method constructs a saliency image. A global background reference value is obtained by calculating the arithmetic mean of the pixel values ​​of all superpixel image blocks. This global background reference value is then differenced from the Gaussian-smoothed image data to generate the saliency image. This saliency image highlights target regions in the image that are more salient to the global background.

[0011] Finally, a level set evolution equation is constructed, comprising a dual data fitting term and an adaptive fractional-order length regularization term. The dual data fitting term integrates information from the original image and salient image information, providing stronger evolutionary driving force in regions with uneven grayscale or weak edges. The fractional-order length regularization term utilizes the previously calculated adaptive fractional-order derivative order to perform fractional-order differentiation on the level set function. This level set evolution equation drives the zero-level set contour to iteratively evolve towards the target boundary until a convergence condition based on the overall change is met, thus outputting the final image segmentation result.

[0012] In one specific embodiment of the present invention, clustering is initiated during superpixel segmentation by uniformly generating initial seed points in the image plane. The distance between a pixel and a seed point in the five-dimensional feature space is calculated using a weighted distance formula, and the pixel is assigned to the nearest cluster. The cluster centers are continuously updated until convergence. This five-dimensional feature-based clustering ensures the compactness of superpixel blocks in both color and spatial dimensions.

[0013] In terms of quantification of local statistical features, image information entropy is calculated by statistically analyzing the frequency and probability of gray values, and then summing the product of the probability and its logarithmic value to accurately quantify the complexity of the texture. Roughness is calculated based on pixel variance, and a normalized value that is positively correlated with the variance is constructed to reflect the smoothness of the region.

[0014] For the core adaptive fractional derivative order calculation, this invention employs a strategy based on region attribute discrimination. When a pixel is determined to be located in an edge region by comparing gradient differences, the calculated derivative order is positive, causing the subsequent evolution process to perform fractional derivative enhancement operations to sharpen the edges. When a pixel is determined to be located in a noise region, the calculated derivative order is negative, causing the subsequent evolution process to perform fractional integral smoothing operations to suppress noise. The specific absolute value of the derivative order is determined by constructing an exponential function model, based on a weighted sum of the gradient magnitude, image information entropy, and roughness, achieving fine-grained parameter control.

[0015] In constructing the level set evolution equation, this invention introduces several constraint mechanisms. A dual data fitting term divides the image domain into internal and external regions using a regularized Heaviside function, and calculates the weighted gray-level mean of the original image and the salient image in each region, thereby constructing the evolutionary force. The edge indicator function is constructed based on the gradient magnitude of the smoothed image; its value approaches zero where the gradient is large, preventing boundary leakage. The distance regularization term maintains the signed distance property of the level set function, eliminating the computational burden of re-initializing the level set function during evolution.

[0016] Beneficial effects 1. This invention constructs an adaptive fractional derivative model based on superpixel local features, dynamically adjusting the derivative order according to the gradient continuity differences of image pixels. It automatically matches positive derivatives in edge regions to enhance weak boundary signals and automatically matches negative integrals in flat or noisy regions to smooth interference. It also constructs a dual data fitting term by combining a saliency image generated from the global mean and the smoothed image. This mechanism effectively overcomes common problems in medical images such as uneven gray levels, blurred boundaries, and noise interference, improves the ability to identify weak features, and thus obtains high-quality segmentation results.

[0017] 2. This invention utilizes superpixel clustering to transform image data from pixel-level to region-level, achieving effective dimensionality reduction and lowering the complexity of feature calculation. Simultaneously, this method supports optimization of the computational process: for structurally similar medical sequence images, the superpixel partitioning and differential order parameters of the first frame image can be reused, avoiding repeated calculations frame by frame; furthermore, core steps such as adaptive order calculation and level set evolution are independent, making them suitable for acceleration using parallel computing architectures such as CUDA, thereby shortening algorithm execution time and meeting the timeliness requirements of clinical applications.

[0018] 3. This invention introduces a distance regularization term into the level set evolution equation, which can automatically maintain the sign distance attribute of the level set function, eliminating the re-initialization operation that must be frequently performed in traditional methods, further reducing the computation time. Combined with the edge indicator function based on Gaussian smooth gradient, it effectively prevents leakage of the evolution curve at weak boundaries. In addition, a dynamic threshold related to the total number of image pixels is used as the iteration termination condition to ensure that the algorithm can accurately determine the convergence state when processing images of different resolutions, thus ensuring the stability of the segmentation results. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the image superpixel segmentation and local feature extraction process of the present invention; Figure 3 This is a schematic diagram illustrating the iterative process of the level set evolution and numerical implementation of the present invention. Detailed Implementation

[0020] See attached document Figure 1 This invention provides a feature-sensitive image segmentation method, which is executed by a computer device or image processing terminal, and mainly includes the following steps: First, the medical image data to be segmented is acquired. This medical image data can be a two-dimensional grayscale image or a color image containing weakly salient features, blurred boundaries, or noise interference. After acquiring the image, superpixel segmentation preprocessing is performed on the medical image data. Specifically, the medical image data is converted from the original RGB color space to the CIELAB color space. In the CIELAB color space, each pixel has a luminance component. Color components and color components For each pixel in the medical image data, its color component value and its two-dimensional spatial coordinate value in the image plane are extracted. The color component value and spatial coordinate value are combined to construct a five-dimensional feature vector for each pixel.

[0021] Next, a predetermined number of initial seed points are generated on the image plane of the medical image data. Based on the five-dimensional feature vector, the vector distance between each pixel in the medical image data and each initial seed point is calculated. According to the calculated vector distance, each pixel is assigned to the cluster containing the nearest seed point. By iteratively updating the cluster centers and reallocating pixels, until the clustering convergence condition is met, the medical image data is divided into multiple non-overlapping superpixel image patches. Each superpixel image patch contains several pixels with similar features. This division can aggregate local features of the image and reduce the amount of data required for subsequent calculations.

[0022] Subsequently, for each superpixel image block, the local statistical features within that block are calculated. These local statistical features include the pixel mean, pixel variance, image information entropy, and roughness. The total number of pixels within each superpixel image block is counted. The pixel mean is the arithmetic mean of the grayscale values ​​of all pixels within the superpixel image block. The pixel variance is a measure of the dispersion of the grayscale values ​​of all pixels within the superpixel image block relative to the pixel mean.

[0023] The image information entropy is further calculated. Image information entropy is used to characterize the smoothness or disorder of texture within a superpixel image patch. Specifically, the ratio of the frequency of a specific gray value occurring within a superpixel image patch to the total number of pixels in that patch is calculated to obtain the probability of that specific gray value, and the image information entropy is calculated based on the probabilities of all gray values. When the superpixel image patch is located in a smooth region of the image, the image information entropy value is relatively small; when the superpixel image patch is located at the image edge or in a region with rich texture, the image information entropy value is relatively large.

[0024] Simultaneously, roughness is calculated based on pixel variance. Roughness characterizes the relative offset of pixel grayscale values. The roughness value is positively correlated with the pixel variance. In smooth regions of the image, the calculated roughness value is smaller due to the smaller pixel variance; in textured regions, the calculated roughness value is larger due to the larger pixel variance. The pixel mean, image information entropy, and roughness will be used to subsequently determine the order of the fractional derivative.

[0025] After obtaining the aforementioned statistical characteristics, the corresponding fractional derivative order is determined for each pixel within a superpixel image block. First, the gradient magnitude of the pixel is calculated, and it is determined whether the pixel belongs to an edge region or a noise region. Then, using the gradient magnitude, image entropy, and roughness, a variable fractional derivative order is calculated through an adaptive function. Pixels identified as edge regions are assigned a positive fractional derivative order to enhance the signal; pixels identified as noise regions are assigned a negative fractional derivative order to suppress noise.

[0026] Finally, image segmentation is performed using fractional derivative order. First, a saliency image is generated based on the pixel mean, reflecting the difference between local regions and the global background. Next, a level set function is initialized, and a level set evolution equation is constructed, including data fitting, length regularization, and distance regularization terms. In the level set evolution equation, a fractional derivative operator based on fractional derivative order replaces the traditional integer gradient operator. Based on information from the saliency image and the original image, the value of the level set function is iteratively updated, driving the zero-level set contour to move towards the target boundary. The iteration stops when the change in the level set function between two consecutive iterations is less than a preset termination threshold, and the image segmentation result is output based on the final level set function. See attached document Figure 2 In the initial stage of image segmentation processing, to reduce the complexity of subsequent calculations and enhance the consistency of local features, this embodiment first employs superpixel segmentation technology to preprocess the input medical image. This process aims to aggregate pixel-level data into perceptually meaningful region-level data, i.e., superpixel image patches.

[0027] First, color space conversion and feature vector construction are performed. Considering the perceptual non-uniformity of the RGB color space, the acquired raw image data is linearly or non-linearly converted to the CIELAB color space. In the CIELAB space, the image consists of a luminance component. And two contrasting color components and Composition. For each pixel on the image plane. Extract its corresponding color component values And combine this color information with spatial location information By combining them, a five-dimensional feature vector is constructed.

[0028] ; By introducing spatial coordinates and The subsequent clustering process can simultaneously constrain the similarity of pixels in color and their spatial proximity, thereby ensuring the compactness of the generated superpixel blocks at the visual boundary.

[0029] Next, Simple Linear Iterative Clustering (SLIC) is performed based on the five-dimensional feature vector. Multiple initial seed points are uniformly generated within the image domain, let... The image is divided into predefined superpixel patches. Within a search window centered on each seed point, the distance between neighboring pixels and the seed point in the five-dimensional feature space is calculated. This distance metric combines color distance and spatial distance. Each pixel is assigned to the cluster containing its nearest seed point. Subsequently, the average of the five-dimensional feature vectors of all pixels within each cluster is calculated, and this average is used as the new cluster center. The above assignment and update steps are repeated until the change in the position of the cluster center is less than a preset threshold or the maximum number of iterations is reached, thus completing the superpixel segmentation of the image. At this point, the original image is divided into... Let the nth non-overlapping superpixel image patch be denoted as . Each superpixel image block is .

[0030] After completing the block division, for each superpixel image block We will calculate the local statistical characteristics within each pixel, including pixel mean, pixel variance, image information entropy, and roughness. These characteristics will serve as the physical basis for subsequently determining the fractional-order differential order.

[0031] First, statistics Superpixel image blocks The total number of pixels contained therein is denoted as . Calculate the average pixel value of this image patch. and pixel variance Pixel average Defined as the arithmetic mean of the gray values ​​of all pixels within an image block, its calculation formula is as follows: ; in, Represents pixel coordinates, This represents the pixel grayscale value at that coordinate. The pixel mean reflects the overall brightness level of the superpixel block.

[0032] Pixel variance Defined as the dispersion of pixel grayscale values ​​within an image patch relative to the mean, its calculation formula is as follows: ; Pixel variance measures the degree of grayscale variation within an image patch. The larger the variance, the more pronounced the texture or edge features within the patch.

[0033] Secondly, calculate the information entropy of the image patch. This is used to quantify the texture smoothness of an image patch. For an image patch... Statistical analysis of grayscale values ​​of specific pixels The frequency of occurrence, divided by the total number of pixels. Get the probability of this grayscale value appearing. Information entropy The calculation formula is as follows: ; in, Represents image blocks The set of all grayscale values ​​appearing in the image. In smooth regions of the image, the grayscale distribution is concentrated, and the information entropy... The value is relatively small; however, in areas with dense edges or complex textures, the grayscale distribution is discrete, and the information entropy is relatively small. The value is relatively large. This feature can effectively distinguish between smooth backgrounds and regions containing structural information.

[0034] Finally, the roughness of the image patch is calculated. Roughness is a dimensionless feature derived from pixel variance, used to characterize the relative offset of pixel grayscale values. Its calculation formula is as follows: ; As can be seen from the formula, roughness With variance They exhibit a positively correlated nonlinear mapping relationship. In smooth regions of the image, Approaching 0, then Approaching 0; in areas with rich texture or high noise, If it is larger, then Approaching 1. By calculating the roughness, the variance can be mapped to a normalized interval, facilitating subsequent adjustment of the fractional order as a parameter input into the exponential function. Through the above steps, the original medical image is transformed into a series of images with rich local statistical features ( The superpixel image patches provide an accurate data foundation for subsequent feature-sensitive fractional derivative calculations.

[0035] After completing superpixel segmentation and local feature extraction, this embodiment further constructs adaptive fractional derivative parameters for each pixel in the image. The core purpose of this process is to resolve the contradiction that traditional integer-order derivative operators often over-amplify noise when enhancing high-frequency edge information in images, while easily blurring weak texture details when performing smoothing and denoising. By introducing a variable-order fractional derivative, this method can dynamically switch and balance between derivative enhancement and integral smoothing according to the local environmental characteristics of the pixel.

[0036] First, a mechanism for distinguishing between edges and noise is established. In medical images, edge points and noise points often exhibit high-frequency abrupt changes in grayscale, making it difficult to effectively distinguish them based on a single gradient magnitude. This embodiment utilizes the continuity of the gradient within a local neighborhood for discrimination. For any central pixel in the image... The gradient magnitude of the center pixel is calculated, and the gradient magnitudes of all pixels in its eight neighborhoods are also obtained. Then, the absolute values ​​of the differences between the gradient magnitude of the center pixel and the gradient magnitudes of each neighboring pixel are calculated, and the minimum value is found among these differences.

[0037] Define a discriminant indicator variable This is used to identify the attributes of the current pixel. A judgment threshold is set. When the minimum value of the above gradient difference is less than or equal to the threshold When the gradient change of the center pixel and its neighboring pixels is continuous or similar, it indicates that the pixel belongs to the image edge or texture region. Therefore, let... Conversely, when the minimum value of the above gradient difference is greater than the threshold... When this occurs, it indicates that the gradient change of that pixel is isolated and abrupt, and the pixel is determined to belong to noise interference. In this case, let... This discriminant logic can be described by a mathematical expression: ; in, This represents the gradient magnitude of the center pixel. This represents the gradient magnitude of neighboring pixels.

[0038] After determining the attributes of the pixels, an adaptive fractional order is constructed by further combining the local statistical features calculated in the previous stage. The computational model aims to implement the following control logic: for edge and texture regions, a positive fractional order is assigned ( This involves using fractional differential operators to enhance the intensity of high-frequency signals, thereby sharpening blurred lesion boundaries; for noisy regions, a negative fractional order is assigned (…). This essentially constitutes a fractional integral with this order as a parameter, which smooths and suppresses noise.

[0039] Specifically, this embodiment uses an exponential nonlinear function to fit this order variation relationship. This function comprehensively considers the gradient magnitude of the pixels. Information entropy of the superpixel block and roughness The calculation formula is as follows: ; In the above formula, and This is an adjustment constant used to control the overall range of the order values ​​and the reference offset. These are the weighting coefficients for gradient magnitude, information entropy, and roughness, respectively.

[0040] The physical significance of this computational model is as follows: First, by introducing The term directly uses the aforementioned discrimination results to control the sign of the order. When identified as an edge ( When the coefficient is positive, differential enhancement is performed; when it is identified as noise ( When the coefficient is negative, integral denoising is performed.

[0041] Second, the exponential term aggregates multidimensional features using a weighted summation method. Gradient magnitude Directly reflects the steepness of the edge; information entropy Reflects the richness of local texture; roughness This reflects the fluctuation of grayscale. The larger these three values ​​are, the larger the calculated exponential value, meaning that for significant edges or complex textures, the algorithm will apply a stronger differential enhancement (a larger positive value). Alternatively, apply stronger smoothing force (smaller negative) to areas with extremely high noise levels. Depending on the specific parameters (As determined).

[0042] Third, the introduction of the exponential function makes the order... It has a non-linear response to changes in features, meaning it is highly sensitive to subtle changes in features. This allows it to better capture and enhance the features of weakly salient targets (such as ground-glass opacities) commonly found in medical images, preventing them from being missed during segmentation.

[0043] Through the above construction process, each pixel in the image obtains a fractional order that precisely matches its local structure and statistical properties. This provides key adaptive parameters for calculating fractional gradients in subsequent level set evolution.

[0044] After enhancing the distinction between edges and noise through adaptive fractional-order differential parameters, this embodiment introduces a saliency image information extraction step to further address the problems of low contrast and blurred boundaries (e.g., ground-glass opacities) between lesion areas and background tissue in medical images. This step aims to construct a saliency map that reflects the structural features of the image, and introduce it as an independent data item into the subsequent level set evolution equation to compensate for the shortcomings of segmentation based solely on the gray-level gradient of the original image.

[0045] The construction of the saliency image is based on a measure of the difference between global statistical information and local structural information. First, the global background reference value of the image is calculated. To avoid pixel-level noise interfering with the global statistics, this embodiment utilizes the superpixel block pixel mean value calculated in the preceding steps. As the basic data, iterate through all the images. For each superpixel image patch, the arithmetic mean of the pixel values ​​of these patches is calculated again to obtain the global mean of the image. The global mean This represents the macroscopic brightness benchmark of the entire image after removing local texture details, and its calculation expression is as follows: ; By using this dual averaging method (first averaging within blocks, then averaging globally), a more robust background estimate can be obtained than by directly calculating the total pixel mean, effectively suppressing the influence of outlier noise.

[0046] Secondly, local smoothing features of the image are obtained. A Gaussian filter is used to smooth the original input image. Perform convolution operations to extract low-frequency structural information from the image. Define a standard deviation as... Gaussian kernel function The Gaussian kernel is convolved with the original image to obtain a smoothed image.

[0047] ; Here, * represents the convolution operation. This step filters out high-frequency details and minute noise in the image, preserving the image's main morphological and structural features.

[0048] Finally, based on the global mean With smooth image Construct a saliency image of the differences between them For each pixel in the image Calculate its smoothed grayscale value and global mean. The absolute value of the difference between them (i.e., Euclidean distance): ; In this salient image In this context, the pixel value directly reflects the saliency of that location relative to the global background. For lesion areas (such as tumors or inflammation) in medical images, their grayscale values ​​typically deviate statistically from the surrounding normal tissue (background). Through the above calculations, the background region, because its grayscale value is close to the global mean, [is considered less significant]. In the image, it appears as a low grayscale value (approaching black); while the target lesion area, due to its significant difference from the global mean, is... The image shows high grayscale values ​​(approaching white), which significantly highlights the location information of the target area.

[0049] This saliency map, generated based on the concept of frequency domain tuning, not only enhances the contrast of weakly saliency targets, but also provides a constraint field independent of the original gray-level gradient for subsequent level set segmentation. This allows the evolution curve to continue converging towards the true edge based on saliency differences when facing fuzzy boundaries, rather than getting trapped in local minima.

[0050] After obtaining the aforementioned adaptive fractional derivative parameters and saliency image information, this embodiment employs an improved level set evolution model to perform the final image segmentation. Traditional level set segmentation methods typically rely solely on the grayscale gradient information of the original image to construct the edge stop function, i.e., using the magnitude of the integer gradient to control the speed of the evolution curve. However, when processing medical images with weak saliency features or blurred boundaries, relying solely on the integer gradient is insufficient to capture subtle edge changes, easily leading to the evolution curve exceeding the true boundary (i.e., boundary leakage), and the integer derivative tends to amplify high-frequency noise in the image.

[0051] To address the aforementioned technical issues, this embodiment constructs a level set evolution model with specific configurations and improvements in the construction of the energy functional. Instead of solely relying on the grayscale information of the original image, the model incorporates the saliency image generated in the preceding steps as an independent data constraint term into the energy functional. By combining the grayscale distribution information of the original image with the structured difference information of the saliency image, a dual data fitting term is constructed. This dual constraint mechanism ensures that in regions with low grayscale contrast in the original image, the evolution curve maintains the correct evolution direction based on the difference information between the foreground and background in the saliency image, thereby locking in the blurred target boundary.

[0052] Furthermore, in this embodiment, a variable-order fractional differential operator is used instead of the traditional integer-order gradient operator in the length term of the level set evolution. Specifically, the adaptive fractional order calculated for each pixel in the aforementioned steps is used. For level set functions Perform fractional differentiation and calculate its fractional gradient magnitude. .

[0053] The introduction of this fractional-order differential operator alters how the evolution curve responds to image features. In the edge regions of the image, due to the increased order... When set to a positive value, fractional-order differentiation can non-linearly enhance the intensity of high-frequency signals, making previously blurred edges numerically steeper, thus strengthening the edge stopping function's ability to truncate weak boundaries; in smooth or noisy regions of an image, due to the order... When set to a negative value, fractional-order differentiation is actually transformed into fractional-order integration, smoothing the local region and thus automatically suppressing noise interference on the curve motion during evolution. This mechanism, based on adaptive adjustment of the differentiation order according to local features, enables the level set model to maintain its adaptability to complex topological structures while significantly improving its robustness to weakly textured and noisy data.

[0054] This embodiment updates the level set function by iteratively solving partial differential equations. The value of the level set function drives the evolution of the zero-level set contour (i.e., the segmentation curve) in the image domain and ultimately converges to the target boundary. It is a high-dimensional function defined on the image domain, whose zero level set This represents the current segmentation curve. To achieve accurate segmentation sensitive to features, this embodiment constructs a level set evolution equation that includes a dual data fitting term, a fractional-order length regularization term, and a distance regularization term. The equation is defined as follows: ; The specific technical meanings and calculation methods of each term in the above equations are as follows: Dual data fitting terms based on weighted regions: The first four terms of the equation constitute the external energy force driving the evolution of the curve. Among them, The first two terms are the variable fractional derivative order values. Based on the original image The grayscale information, the last two items Based on saliency image Structural information.

[0055] These are preset positive weighting coefficients used to balance the contribution ratios of original grayscale information and saliency information in segmentation.

[0056] and These represent the original images. Within the current level set contour region and external areas The weighted gray mean; and These represent saliency images. Within the current level set contour region and external areas The weighted gray mean. These means vary with the level set function. It is dynamically updated based on evolution, and its computation utilizes a regularized Heaviside function. Divide the area into regions: ; ; ; ; By introducing the mean difference of the saliency image and When the grayscale contrast of the original image is insufficient, the algorithm uses the high contrast features in the saliency map to continuously provide evolutionary force, pulling the curve to move towards the target boundary.

[0057] Variables in the equation of the edge indicator function This is an edge indicator function used to slow down the evolution rate at image edges, preventing the curve from crossing the true boundary. This function is calculated based on the gradient of the Gaussian-smoothed image. ; in, It is the squared magnitude of the gradient of the Gaussian-smoothed image. This represents the pixel grayscale value at that coordinate. The standard deviation is Gaussian kernel function, This represents the gradient operator. In smooth regions, the gradient is smaller. The gradient approaches 1; in the edge region, the gradient is larger. Approaching 0, thus suppressing the influence of data items and length items at the edges.

[0058] The fifth term of a fractional-order length regular equation div Used to control the smoothness and length of the evolution curve. Among them... These are weighting coefficients.

[0059] This is a regularized Dirac function. Unlike traditional methods, the divergence operator in this term... Acting on fractional gradient Instead of integer gradients, fractional gradients. order The adaptive fractional-order parameters are determined by the steps described above. When the curve is located in a weakly textured region... The value of allows the fractional differential operator to capture minute changes that integer operators ignore, thus maintaining sensitivity to the geometry of weak edges. This term minimizes the weighted arc length, preserving boundary smoothness while avoiding the loss of small structural features.

[0060] Distance regularization term: The sixth term of the equation div This is a distance regularization term used to maintain the level set function during evolution. The signed distance property, i.e., maintaining... .

[0061] in The regularization coefficient is . Let be the potential function. The existence of this term avoids the frequent re-initialization operations that must be performed in the traditional level set method, thus improving the stability and efficiency of numerical computation.

[0062] In practical implementation, the level set function The function is initialized as a binary function, with values ​​set to 1 at the initial superpixel cluster edges and 0 at other locations. Subsequently, the partial differential equation is discretized and solved using the finite difference method. At each time step... Internally, according to the current Calculate and update all values Repeat this process until the evolution termination condition is met.

[0063] See attached document Figure 3 After determining the specific form of the level set evolution equation, this embodiment uses a numerical iterative method to solve the partial differential equation to obtain the final image segmentation result. This process mainly involves the initialization of the level set function, the approximation of the regularization function, finite difference iteration, and the determination of the convergence termination condition.

[0064] First, the level set function is initialized. Based on the initial region segmentation results obtained through superpixel segmentation in the previous steps, the initial level set function is constructed. The boundaries of the superpixel blocks are used as the initial zero level set contour. For regions inside the contour, the level set function is assigned a positive constant; for regions outside the contour, the level set function is assigned a negative constant or zero. To ensure the stability of numerical computation and prevent shocks or singularities from occurring in the level set function during evolution, this embodiment introduces a regularized Dirac function. This is used to approximate the derivative term of the Heaviside function. This regularization smooths out numerical variations near the contour of the level set, and is defined as follows: ; in, The smoothing parameter is set to 1.5 in this embodiment. When the variable... The absolute value is less than or equal to When the variable is in a smooth transition, a cosine function is used to construct a smooth transition; when the variable is in a smooth transition, a cosine function is used to construct a smooth transition. The absolute value is greater than When the function value is zero, the function value is 0. This local support property ensures that the computation of level set evolution is concentrated only in a narrow band region near the zero level set, thereby reducing computational redundancy.

[0065] Next, the discretization and iterative update stage begins. The finite difference method is used to discretize the continuous time and space variables. In the... In this iteration, the level set function value at the current moment is used. Based on the aforementioned evolutionary equation, which includes a dual data fitting term, a fractional length term, and a distance regularization term, the time derivative of the level set function is calculated. Set the time step. The level set function value for the next time step is updated as follows:

[0066] ; in, This represents the numerical approximation of all operators on the right-hand side of the evolutionary equation at the current time step. In this calculation, the fractional-order differential operations involved are numerically implemented using either a frequency-domain algorithm based on Fourier transform or a time-domain difference algorithm based on the Grünwald-Letnikov definition to ensure accuracy regarding the fractional-order differential operations. A precise response.

[0067] Subsequently, after each iteration, the evolution termination condition is determined. To automatically determine the optimal segmentation time, this embodiment monitors the change in the level set function between adjacent iterations. The total number of pixels is defined as... Set a dynamic threshold based on image size. When the overall change in the level set function is less than this threshold, the evolution curve is considered to have converged and stabilized at the target boundary, and iteration stops at this point. The mathematical expression for the termination condition is as follows: ; in, express The level set function value at time (current iteration step). express The level set function value at time (next iteration step).

[0068] Finally, based on the level set function at the end of the iteration. Output the segmentation results. Extract the segments that meet the criteria. The set of pixels is used as the final target contour line, or the set of pixels that satisfy the condition is extracted. The region is used as the target segmentation region (mask). In addition, in order to improve the processing efficiency of large-scale medical image data, the matrix operations in the above numerical calculation process, especially the convolution operation and the difference operation, are configured to be executed on parallel computing units (such as the CUDA core of the GPU). The update value of each pixel in the image domain is calculated simultaneously by parallel threads, thereby shortening the time consumption of segmentation processing.

Claims

1. A feature-sensitive image segmentation method, characterized in that, Includes the following steps: The medical image data to be segmented is obtained, the medical image data is converted from RGB color space to CIELAB color space, a five-dimensional feature vector containing color component values ​​and spatial coordinate values ​​is constructed based on CIELAB color space, and a clustering operation is performed using the five-dimensional feature vector to divide the medical image data into multiple non-overlapping superpixel image blocks. For each superpixel image block, local statistical features within the block are calculated by statistically analyzing the pixel grayscale distribution within the superpixel image block. These local statistical features include: pixel mean representing brightness, pixel variance representing dispersion, image information entropy representing texture disorder, and roughness representing relative grayscale offset. For each pixel in the medical image data, the gradient continuity of each pixel determines whether each pixel belongs to an edge region or a noise region. An adaptive fractional derivative order that matches the local features of each pixel is calculated by combining the image information entropy and roughness of the superpixel image block to which each pixel belongs. A global background reference value is calculated by taking the arithmetic mean of the pixel mean of each superpixel image block, and a saliency image is generated based on the difference between the global background reference value and the smoothed image data. A level set evolution equation is constructed, which introduces the saliency image to construct a dual data fitting term and introduces the fractional derivative order to construct an adaptive fractional length regularization term. The level set function is iteratively updated using the level set evolution equation, driving the zero level set contour to evolve towards the target boundary until a preset termination condition is met. The image segmentation result is then output based on the final level set function.

2. The feature-sensitive image segmentation method according to claim 1, characterized in that, The process of dividing the medical image data into multiple non-overlapping superpixel image blocks using the five-dimensional feature vector includes: A predetermined number of initial seed points are uniformly generated on the image plane of the medical image data; For each pixel, extract the brightness component, the first color component, the second color component, and the spatial coordinates in the horizontal and vertical dimensions, and combine them to construct the five-dimensional feature vector. The weighted vector distance between each pixel and each initial seed point in the five-dimensional feature space is calculated using a weighted distance formula. Based on the weighted vector distance, each pixel is assigned to the cluster of the nearest seed point. The new cluster center is calculated by taking the arithmetic mean of the five-dimensional feature vectors of all pixels in the cluster. The allocation and update steps are iteratively executed until the clustering converges, and the superpixel image block is obtained.

3. The feature-sensitive image segmentation method according to claim 1, characterized in that, The image information entropy and roughness in the local statistical features are calculated as follows: The image information entropy is calculated as follows: the probability of the current gray value is calculated by counting the frequency of the current gray value in the current superpixel image block; the probability of the current gray value is multiplied by the logarithm of the probability of the current gray value; and the image information entropy is calculated by summing the product of all gray values ​​in the superpixel image block and taking the opposite number. The roughness is calculated as follows: an intermediate variable is calculated by taking the reciprocal of the sum of the number 1 and the pixel variance, and the roughness is calculated by subtracting the intermediate variable from the number 1; the value of the roughness is normalized to the range of zero to one, and is positively correlated with the pixel variance.

4. The feature-sensitive image segmentation method according to claim 1, characterized in that, The calculation process for the adaptive fractional derivative order includes: The gradient magnitude of the current pixel and the gradient magnitudes of all pixels in the eight-neighborhood of the current pixel are calculated using the gradient operator. The absolute value of the difference between the gradient magnitude of the current pixel and the gradient magnitude of each neighboring pixel is calculated by subtraction operation, and the minimum value among the absolute values ​​is found. The region attribute is determined by comparing the minimum value with a preset determination threshold: if the minimum value is less than or equal to the determination threshold, the current pixel is determined to be an edge region, and the determination indicator variable is set to a first value; if the minimum value is greater than the determination threshold, the current pixel is determined to be a noise region, and the determination indicator variable is set to a second value. A calculation model for the order of the fractional derivative is constructed, and the sign of the fractional derivative is determined by the discrimination indicator variable. The absolute value of the fractional derivative is determined by weighted summation of the gradient magnitude, the image information entropy, and the roughness.

5. The feature-sensitive image segmentation method according to claim 4, characterized in that: When the discrimination indicator variable corresponds to the edge region, the fractional derivative order is calculated to be a positive value, which is used to indicate that a fractional derivative enhancement operation will be performed in subsequent evolution; When the discrimination indicator variable corresponds to a noise region, the fractional derivative order is calculated to be negative, which is used to indicate that a fractional integral smoothing operation is performed in subsequent evolution.

6. The feature-sensitive image segmentation method according to claim 1, characterized in that, The process of generating a saliency image includes: The global background reference value is calculated by taking the arithmetic mean of the pixel mean of all superpixel image blocks, and is used as the macroscopic brightness benchmark for the entire image. The medical image data is convolved using a Gaussian kernel function to obtain a smooth image with high-frequency details filtered out. The gray value of a pixel in the salient image is calculated by taking the absolute value of the difference between the gray value of each pixel in the smoothed image and the global background reference value, where a high gray value represents a salient region and a low gray value represents a background region.

7. The feature-sensitive image segmentation method according to claim 1, characterized in that, The level set evolution equation defines the rate of change of the level set function over time, and the constituent terms of the level set evolution equation include: The dual data fitting term includes a weighted mean difference squared term based on the original image and a weighted mean difference squared term based on the saliency image; the saliency image is used to provide additional evolutionary driving force when the contrast of the original image is low. The fractional length regularization term calculates the fractional gradient by performing fractional differentiation on the level set function using the adaptive fractional derivative order, and calculates the normalized divergence of the fractional gradient. Distance regularization term: It is used to maintain the sign distance property of the level set function and avoid re-initialization operations during the evolution process.

8. The feature-sensitive image segmentation method according to claim 7, characterized in that, The level set evolution equation also introduces an edge indicator function to adjust the weights of the dual data fitting term and the fractional-order length regularization term: The edge indicator function is calculated based on the gradient magnitude of the original image after Gaussian smoothing; the calculation logic of the edge indicator function is as follows: the value of the edge indicator function is calculated by taking the reciprocal of the sum of the number 1 and the square of the gradient magnitude of the smoothed image; In regions with large image gradients, the value of the edge indicator function approaches zero, reducing the influence of the dual data fitting term and the fractional-order length regularization term on the evolution curve and preventing boundary leakage.

9. The feature-sensitive image segmentation method according to claim 7, characterized in that, The calculation of the weighted mean difference squared term involved in the dual data fitting term depends on the region division. The region division uses a regularized Heaviside function to process the current level set function, dividing the image domain into an inner region and an outer region of the level set contour. The weighted gray mean of the original image and the salient image in the inner region and the outer region are calculated by integral operation, respectively.

10. The feature-sensitive image segmentation method according to claim 1, characterized in that, The termination condition is: The overall change in the level set function value obtained from two adjacent iterations is calculated by subtraction. The iteration stops when the overall change is less than a preset dynamic threshold. The dynamic threshold is calculated by multiplying the total number of pixels in the medical image data by a preset constant.

Images