A surgical area lesion image segmentation method based on a multi-scale dynamic segmentation kernel

Abstract

This invention discloses a method for surgical lesion image segmentation based on a multi-scale dynamic segmentation kernel, belonging to the field of image data processing technology. It includes: S1, acquiring raw frames of medical images, performing preprocessing, dividing the image into scales and segmentation units, and acquiring the local feature parameter set of each segmentation unit; S2, quantifying the lesion-sensitive features of each segmentation unit, screening out sub-centimeter-level lesion risk areas, and constructing a medical image segmentation model; S3, for sub-centimeter-level lesion risk areas, performing structural feature evaluation and orientation consistency quantification, generating dynamic segmentation kernel output results, achieving adaptive segmentation of the lesion area, and screening out abnormal areas; S4, for abnormal areas, performing multi-scale inference and boundary accuracy evaluation, achieving smoothing and blur correction of sub-centimeter-level lesion edges. This invention solves the problem that sub-centimeter-level lesion signals are weak, have blurred boundaries, and are easily missed, making it difficult to perform highly sensitive multi-scale fine segmentation of lesion images.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a method for segmenting surgical lesion images based on a multi-scale dynamic segmentation kernel. Background Technology

[0002] Surgical lesion image segmentation is a core direction in the intelligent development of medical imaging. It aims to accurately identify and differentiate different tissues and lesion structures from multimodal images of the surgical area, providing intuitive quantitative evidence for intraoperative navigation, lesion localization, and precision treatment. With the increase in medical image resolution and data complexity, surgical images exhibit high-dimensionality, dynamics, and heterogeneity, placing higher demands on the accuracy, stability, and real-time performance of segmentation results. Continuous development in this field is driving the evolution of medical imaging towards intelligence, visualization, and personalization, providing crucial support for precision medicine and surgical decision-making.

[0003] For example, invention patent CN120672776A discloses a method for lesion segmentation by fusing PET and CT dual-modal images, relating to the field of computer technology. The method includes: acquiring lesion image data, separating the lesion image data to obtain CT data and PET data, and inputting the CT data and PET data into a dual-modal medical image segmentation model to obtain preliminary segmentation results; generating difference regions based on the preliminary segmentation results, and performing difference region detection on the difference regions to generate click signal data; the difference regions are areas not marked as lesions in the preliminary segmentation results; training the dual-modal medical image segmentation model a preset number of times using the click signal data, CT data, and PET data to obtain segmentation results; and determining the segmentation result generated by the last training as the lesion image segmentation result. This method can improve the accuracy of lesion segmentation in medical images.

[0004] For example, invention patent CN116309647B discloses a method for constructing a segmentation model for craniocerebral lesions, an image segmentation method, and an apparatus, relating to the field of image processing technology. Specifically, it provides a method for constructing a segmentation model for craniocerebral lesions, an image segmentation method, and an apparatus. The method for constructing a segmentation model for craniocerebral lesions includes: acquiring multiple images of craniocerebral lesions; preprocessing each image; converting the processed images to grayscale; constructing a gradient histogram of the grayscale images and extracting lesion feature vectors; annotating each image with lesion features based on the corresponding lesion feature vectors and constructing a craniocerebral image dataset; and training an initial model using the craniocerebral image dataset to obtain a segmentation model for craniocerebral lesions. The technical solution of this invention addresses the complex characteristics of craniocerebral lesions by constructing and training a model, thereby improving the segmentation accuracy of the craniocerebral lesion image segmentation model for craniocerebral lesions.

[0005] However, sub-centimeter lesions in the surgical area, such as early-stage tumors and microaneurysms, have extremely weak signals and are often overlooked by the main segmentation nucleus, resulting in unclear boundaries and easy missed detection. Achieving highly sensitive multi-scale dynamic detection and fine segmentation of tiny targets while maintaining global segmentation efficiency presents an extremely high technical challenge.

[0006] Therefore, in order to address the above problems, there is an urgent need for a method for segmenting surgical lesion images based on multi-scale dynamic segmentation kernels. Summary of the Invention

[0007] To address the technical problem in existing technologies where weak signals, blurred boundaries, and easy missed detection of sub-centimeter lesions in the surgical area hinder highly sensitive multi-scale fine segmentation of lesion images, this invention provides a surgical lesion image segmentation method based on a multi-scale dynamic segmentation kernel. The technical solution is as follows:

[0008] A method for surgical lesion image segmentation based on a multi-scale dynamic segmentation kernel is disclosed. The method includes: S1, real-time acquisition of raw medical images of the surgical area; uniform preprocessing of the raw medical images; generation of multi-scale resolution images based on the preprocessed raw medical images; division of the images into segmentation units according to spatial coordinates; acquisition of local feature parameter sets for each segmentation unit; and establishment of a spatial mapping relationship between the multi-scale resolution images and the segmentation units; S2, quantification of lesion-sensitive features for each segmentation unit at each scale based on the local feature parameter sets; determination of region type based on the quantification results; selection of sub-centimeter-level lesion risk regions; construction of a medical image segmentation model; generation of basic parameters for the segmentation kernel; and realization of differentiated segmentation. S3, for sub-centimeter-level lesion risk areas, combining the basic parameters of the segmentation kernel and the quantification results of lesion sensitive features, structural feature evaluation and orientation consistency quantification are performed based on the local feature parameter set. The combined results of structural feature evaluation and orientation consistency quantification are used to generate dynamic segmentation kernel output results. Adaptive segmentation of lesion areas is achieved based on the dynamic segmentation kernel output results, and abnormal areas are screened out. S4, for abnormal areas, multi-scale inference is performed using a medical image segmentation model. Boundary accuracy is evaluated based on the multi-scale inference results, and adaptive smoothing and blur correction of sub-centimeter-level lesion edges are achieved. At the same time, based on the lesion sensitive feature quantification results and boundary accuracy evaluation results, visualization feedback and closed-loop optimization are achieved.

[0009] Optionally, the process of acquiring raw medical images of the surgical area in real time, performing unified preprocessing on the raw medical images, generating multi-scale resolution images based on the preprocessed raw medical images, dividing the images into segments according to spatial coordinates, acquiring the local feature parameter set of each segment, and establishing the spatial mapping relationship between the multi-scale resolution images and the segmentation units is as follows: Real-time acquisition of raw medical images, synchronously acquiring the spatial resolution information, device acquisition parameters, image time series labels, and spatial coordinates of each frame, and acquiring the pixel grayscale matrix of the raw medical images; performing unified normalization processing on the acquired raw medical images, including pixel grayscale normalization, noise filtering, and contrast enhancement; downsampling the normalized raw medical images at different scaling ratios using the spatial pyramid method to generate multi-scale resolution images and obtain the number of scale layers; at each scale, dividing the corresponding resolution image into pixels based on spatial coordinates. The hierarchical segmentation is divided into multiple segmentation units. The number of segmentation units is obtained, and the spatial coordinates and normalized pixel gray values ​​of each segmentation unit are recorded synchronously. Based on the normalized pixel gray values, the gray-level co-occurrence matrix algorithm is used to statistically analyze the joint distribution of gray-level pairs within a fixed neighborhood sliding window centered on the spatial coordinates. This yields texture features including contrast, correlation, energy, and local structural entropy. Based on the spatial coordinates, normalized pixel gray values, and texture features of each segmentation unit, a local feature parameter set is constructed. All scales, all segmentation units, and the local feature parameter set are uniformly mapped to establish a spatial correspondence between segmentation units and the original medical image frame and multi-scale resolution images. An image segmentation database is established, storing the original medical image frame, the preprocessed original medical image frame, multi-scale resolution images, segmentation units, their corresponding local feature parameter sets, and spatial mapping relationships.

[0010] Optionally, the specific process of quantifying lesion-sensitive features of segmentation units at each scale based on local feature parameter sets is as follows: Obtain the number of scale layers, the number of segmentation units, multi-scale resolution images, segmentation units, their corresponding local feature parameter sets, and spatial mapping relationships. Input these into a deep convolutional feature extraction network integrating a multi-scale pyramid structure and an attention mechanism. Employ a backbone network and a multi-scale parallel branch structure to perform deep feature representation on segmentation units at each scale, outputting the corresponding multi-scale feature tensor. For each segmentation unit at each scale, use principal component analysis to reduce the dimensionality of the multi-scale feature tensor, taking the first principal component score as the feature response value. Simultaneously, statistically analyze the feature response values ​​of all segmentation units at each scale, calculate and output the corresponding... The mean and standard deviation of the feature response are calculated. For each segmentation unit at each scale, the absolute difference between the feature response value and the mean feature response is calculated and divided by the sum of the standard deviation of the feature response and the minimum constant value to obtain the normalized feature anomaly. The normalized feature anomaly is multiplied by the corresponding scale-adaptive weighting factor to obtain the weighted feature anomaly. The absolute difference between the feature response value of the segmentation unit at the same location in the current scale and the previous scale is calculated and multiplied by the cross-scale amplification weighting factor to obtain the multi-scale differential enhancement value. The weighted feature anomaly is added to the multi-scale differential enhancement value to obtain the single anomaly significance value of the segmentation unit. The single anomaly significance values ​​of the segmentation units at the same location in all scales are summed and divided by the number of scale layers to obtain the sub-centimeter lesion sensitivity value.

[0011] Optionally, based on the quantification results of lesion sensitivity features, region type determination is performed to screen out sub-centimeter-level lesion risk regions, and a medical image segmentation model is constructed to generate basic parameters for the segmentation kernel. The specific process for achieving differentiated segmentation and annotation is as follows: Calculate the sub-centimeter-level lesion sensitivity values ​​of all segmentation units at all scales, write them into the image segmentation database, and compare the sub-centimeter-level lesion sensitivity value of each segmentation unit with the sensitivity threshold; when the sub-centimeter-level lesion sensitivity value is less than the sensitivity threshold, it is determined to be a regular region. Using the BraTS public medical image segmentation benchmark dataset as input, the medical image segmentation model is trained using a convolutional neural network algorithm. Using multi-scale resolution images as input, forward inference is performed on the regular region. By statistically analyzing the mean of the convolution output response values ​​of the medical image segmentation model in the regular region, the basic value of the segmentation kernel is obtained, and the basic segmentation kernel parameters are output; the basic segmentation kernel parameters are combined with the medical image segmentation model to segment the regular region; when the sub-centimeter-level lesion sensitivity value is greater than or equal to the sensitivity threshold, it is determined to be a sub-centimeter-level lesion risk region, and the dynamic segmentation kernel generation process begins.

[0012] Optionally, for sub-centimeter-level lesion risk areas, the specific process of structural feature evaluation and directional consistency quantification based on local feature parameter sets, combining the basic parameters of the segmentation kernel and the quantification results of lesion sensitivity features, is as follows: Receive the basic value of the segmentation kernel and the sub-centimeter-level lesion sensitivity value of the sub-centimeter-level lesion risk area, and execute the dynamic segmentation kernel generation process: For each sub-centimeter-level lesion risk area, obtain the local structural entropy. Simultaneously, with the current sub-centimeter-level lesion risk area as the center, select all segmentation units within a fixed neighborhood sliding window, and use the Sobel operator to calculate the gradient values ​​of the normalized pixel grayscale values ​​of all segmentation units within the window in the horizontal and vertical directions. Combine all gradient values ​​into a gradient vector set of window pixels. Use principal component analysis to perform principal component decomposition on the gradient vector set, extract the first principal component direction as the principal direction component, and obtain the principal direction angle through the polar angle of the principal component vector. Multiply the sub-centimeter-level lesion sensitivity value of the sub-centimeter-level lesion risk area with the corresponding local structural entropy to obtain the structural anomaly response value. Calculate the product of the principal direction component magnitude and the cosine of the corresponding principal direction angle, and add the stabilization constant value to obtain the structural correction value.

[0013] Optionally, the specific process for generating the dynamic segmentation kernel output result by integrating the structural feature evaluation and directional consistency quantification results is as follows: divide the structural anomaly response value by the structural correction value, and take the opposite number as the exponent for natural exponentiation. Add the result of natural exponentiation to a constant to obtain the dynamic segmentation kernel suppression term. Multiply the reciprocal of the dynamic segmentation kernel suppression term by the cross-scale weighting factor and add a constant to obtain the segmentation kernel adaptive adjustment value. Multiply the basic value of the segmentation kernel by the adaptive adjustment value of the segmentation kernel to obtain the dynamic segmentation kernel output value.

[0014] Optionally, the specific process of adaptive segmentation of lesion regions based on the output results of the dynamic segmentation kernel and screening out abnormal regions is as follows: Calculate the dynamic segmentation kernel output value of all sub-centimeter-level lesion risk regions, assign the dynamic segmentation kernel output value to the corresponding spatial position, and use it as the local convolution kernel parameter of the convolutional neural network algorithm in the sub-centimeter-level lesion risk regions. Use the medical image segmentation model to perform adaptive segmentation processing on the sub-centimeter-level lesion risk regions; calculate the average value of the dynamic segmentation kernel output value of all sub-centimeter-level lesion risk regions, and screen out the sub-centimeter-level lesion risk regions whose dynamic segmentation kernel output value is higher than the average value of the dynamic segmentation kernel output value, and mark them as abnormal sub-centimeter-level lesion risk regions; write all dynamic segmentation kernel output values ​​and adaptive segmentation processing results into the image segmentation database.

[0015] Optionally, for abnormal regions, the specific process of multi-scale inference using the medical image segmentation model is as follows: The dynamic segmentation kernel output value and adaptive segmentation processing result of the abnormal sub-centimeter lesion risk region are received; the main segmentation network in the medical image segmentation model is called to perform forward inference on the multi-scale resolution image, outputting the main branch segmentation probability value for each abnormal sub-centimeter lesion risk region; the auxiliary segmentation network in the medical image segmentation model is called in parallel to output the auxiliary segmentation probability value; using the Sobel operator, convolution calculations are performed on the main branch segmentation probability value in the horizontal and vertical directions respectively to obtain the horizontal and vertical gradients of the main branch segmentation probability value; for each abnormal sub-centimeter lesion risk region, the gradient magnitude of the main branch segmentation probability value is calculated based on the horizontal and vertical gradients; using the main branch segmentation probability value and the auxiliary segmentation probability value as input, the structural similarity index algorithm is used to calculate the probability distribution consistency value.

[0016] Optionally, the specific process for boundary accuracy assessment based on multi-scale inference results is as follows: The edge sharpness value is obtained by exponentially calculating the edge sharpening weight factor as the exponent and the gradient magnitude of the main branch segmentation probability value as the base; for the same sub-centimeter-level lesion risk region, the absolute difference between the main branch segmentation probability value and the auxiliary segmentation probability value is calculated, and exponentially calculated using the consistency amplification weight factor as the exponent and the absolute difference as the base; the result of the exponential calculation is multiplied by the consistency suppression weight factor and added to a constant to obtain the probability consistency suppression value; the negative of the ratio of the probability distribution consistency value to the structural similarity weight factor is used as the exponent, and natural exponential calculation is performed to obtain the structural similarity correction value; the ratio of the edge sharpness value to the probability consistency suppression value is multiplied by the structural similarity correction value to obtain the sub-centimeter-level lesion boundary confidence value.

[0017] Optionally, adaptive smoothing and blur correction of sub-centimeter lesion edges are implemented. Simultaneously, based on the lesion sensitivity feature quantification results and boundary accuracy evaluation results, the specific process of achieving visual feedback and closed-loop optimization is as follows: Calculate the sub-centimeter lesion boundary confidence value for all abnormal sub-centimeter lesion risk regions, write it into the image segmentation database, and compare it with a confidence threshold for segmentation correction. For abnormal sub-centimeter lesion risk regions with a sub-centimeter lesion boundary confidence value higher than the confidence threshold, no fine segmentation is required; the adaptive segmentation processing result is directly output. For abnormal sub-centimeter lesion risk regions with a sub-centimeter lesion boundary confidence value less than or equal to the confidence threshold... For lesion risk areas, an adaptive boundary smoothing method is used for edge denoising and blur correction. The abnormal sub-centimeter lesion risk area image after edge denoising and blur correction is used as input, and the medical image segmentation model is called to perform local secondary inference for fine segmentation. Based on the sub-centimeter lesion sensitivity value and the sub-centimeter lesion boundary confidence value, a spatial heat map is generated for all segmentation units of each frame of medical image. The sub-centimeter lesion sensitivity value, the sub-centimeter lesion boundary confidence value and all corresponding segmentation results are analyzed, and the Bayesian optimization method is used to optimize the medical image segmentation model, convolution kernel parameters, sensitivity threshold and confidence threshold.

[0018] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0019] (1) This invention generates multi-scale resolution images based on unified normalization and spatial pyramid, and establishes spatial mapping relationships between image frames, segmentation units and local feature parameters, thereby realizing a unified multi-scale expression of surgical area image data. This avoids the problems of feature separation and inconsistent mapping at different scales in traditional methods, and enables subsequent lesion feature quantification and dynamic segmentation kernel generation to have a unified spatial reference system, thereby improving the continuity and stability of the overall segmentation.

[0020] (2) This invention constructs a multi-scale feature anomaly degree and differential enhancement mechanism to quantify the lesion-sensitive features of segmentation units at each scale, and uses the quantification results to distinguish between conventional areas and sub-centimeter-level lesion risk areas. Through this mechanism, small, low-contrast lesion areas can be automatically identified while maintaining global segmentation efficiency, realizing differentiated detection and adaptive modeling of different types of lesions, and significantly improving the identification sensitivity and segmentation accuracy of small lesions.

[0021] (3) This invention constructs a dynamic segmentation kernel generation mechanism by combining local structural entropy, gradient direction, and principal component features, thereby achieving adaptive adjustment of the segmentation kernel sensitivity. When the local structure is complex or the directionality is unstable, the dynamic segmentation kernel automatically enhances the segmentation weights, improving the model's response to fine structures and fuzzy boundaries; when the structure is simple, the segmentation kernel reverts to the basic value, maintaining model stability. This enables the segmentation process to dynamically adjust the convolution kernel parameters according to local structural features, effectively improving the oversegmentation or undersegmentation problems of traditional models under complex structures.

[0022] (4) This invention quantifies and evaluates the boundary accuracy of sub-centimeter lesions through main and auxiliary branch probabilistic reasoning and structural similarity index calculation, and automatically determines whether to perform refined secondary segmentation based on the boundary confidence threshold. When the confidence is high, the result is directly output; when the confidence is low, adaptive smoothing and fuzzy correction are triggered, achieving a dynamic balance between segmentation accuracy and computational efficiency. At the same time, Bayesian optimization is used to continuously learn and update the segmentation model, convolution kernel parameters, and thresholds, forming a closed-loop optimization mechanism for segmentation performance, thereby achieving self-evolution of segmentation accuracy and long-term stability improvement. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a surgical lesion image segmentation method based on a multi-scale dynamic segmentation kernel provided by an embodiment of the present invention;

[0025] Figure 2 This is a heatmap of sub-centimeter-level lesion sensitivity values ​​provided in an embodiment of the present invention;

[0026] Figure 3 This is a flowchart of the surgical lesion image segmentation method based on multi-scale dynamic segmentation kernel provided in the embodiments of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0028] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0029] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0030] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0031] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0032] This invention provides a method for segmenting surgical lesion images based on a multi-scale dynamic segmentation kernel, such as... Figure 1The flowchart illustrates a surgical lesion image segmentation method based on a multi-scale dynamic segmentation kernel. The processing flow of this method includes the following steps: S1, real-time acquisition of raw medical images of the surgical area; uniform preprocessing of the raw medical images; generation of multi-scale resolution images based on the preprocessed raw medical images; division into segmentation units according to spatial coordinates; acquisition of local feature parameter sets for each segmentation unit; and establishment of a spatial mapping relationship between the multi-scale resolution image and the segmentation units; S2, quantification of lesion-sensitive features for each segmentation unit at each scale based on the local feature parameter sets; determination of region type based on the lesion-sensitive feature quantification results; screening of sub-centimeter-level lesion risk regions; and construction of a medical image segmentation model to generate the segmentation kernel basis. S3, for sub-centimeter-level lesion risk areas, combines the basic parameters of the segmentation kernel with the quantification results of lesion sensitive features, performs structural feature evaluation and orientation consistency quantification based on the local feature parameter set, and generates dynamic segmentation kernel output results by integrating the structural feature evaluation and orientation consistency quantification results. Based on the dynamic segmentation kernel output results, adaptive segmentation of lesion areas is achieved, and abnormal areas are screened out; S4, for abnormal areas, multi-scale inference is performed using the medical image segmentation model, boundary accuracy is evaluated based on the multi-scale inference results, and adaptive smoothing and blur correction of sub-centimeter-level lesion edges are achieved; at the same time, based on the lesion sensitive feature quantification results and boundary accuracy evaluation results, visualization feedback and closed-loop optimization are achieved.

[0033] Optionally, the process of acquiring raw frames of medical images of the surgical area in real time, performing unified preprocessing on the raw frames, generating multi-scale resolution images based on the preprocessed raw frames, dividing them into segments according to spatial coordinates, acquiring the local feature parameter set of each segment, and establishing the spatial mapping relationship between the multi-scale resolution images and the segmentation units is as follows: acquiring raw frames of medical images in real time, synchronously acquiring the spatial resolution information, device acquisition parameters, image time series labels, and spatial coordinates of each frame, and acquiring the pixel grayscale matrix of the raw frames of medical images; wherein, the spatial resolution information includes the actual physical size represented by each pixel, the device acquisition parameters include the scanner type, imaging mode, and exposure time, the image time series labels are used to locate the acquisition time point, and the spatial coordinates use the image two-dimensional or three-dimensional coordinate system to record the specific position of each pixel. The acquired raw medical image frames undergo uniform normalization processing, including pixel grayscale normalization using linear scaling, noise filtering using Gaussian filtering, and contrast enhancement using histogram equalization. The normalized raw medical image frames are then downsampled at different scaling ratios using the spatial pyramid method to generate multi-scale resolution images and determine the number of scale layers. The spatial pyramid downsampling uses a Gaussian pyramid, with configurable scaling ratios for each level (e.g., 1, 1 / 2, 1 / 4, 1 / 8). The number of scale layers is typically 3 to 5, with different levels corresponding to different resolutions, flexibly set according to actual needs. The resolution of the newly generated image at each level and its mapping relationship to the original frame must be recorded. At each scale, the corresponding resolution image is divided into multiple segmentation units at the pixel level based on spatial coordinates. Each pixel is an independent segmentation unit. The number of segmentation units is obtained, and the spatial coordinates and normalized pixel grayscale values ​​are recorded simultaneously for each segmentation unit. Based on the normalized pixel grayscale values, the Gray-Level Co-occurrence Matrix (GLCM) algorithm is used to statistically analyze the joint distribution of grayscale pairs within a fixed neighborhood sliding window centered on the spatial coordinates. This yields texture features including contrast, correlation, energy, and local structure entropy. The fixed neighborhood sliding window size is preferably 5×5 pixels with a step size of 1 pixel, and the boundaries are processed using mirror filling. In the GLCM algorithm, the preferred orientation parameters are 0°, 45°, 90°, and 135°, and the preferred distance parameter is 1 pixel. These can be adjusted appropriately according to the actual image resolution to ensure that the texture features cover all main directions. During the statistical analysis, the GLCM in four directions is extracted using the center pixel of the sliding window as a reference. Contrast, correlation, energy, and local structure entropy are calculated separately, and their mean or maximum values ​​are used as the texture feature representation of the segmentation unit.Based on the spatial coordinates, normalized pixel grayscale values, and texture features of each segmentation unit, a local feature parameter set is constructed. A hash table is used to uniformly map all scales, all segmentation units, and the local feature parameter set, establishing a spatial correspondence between segmentation units and the original medical image frame and multi-scale resolution images. This ensures that any segmentation unit at each scale can be traced back to the original pixel and the corresponding multi-scale region. An image segmentation database is established, storing the original medical image frame, the preprocessed original medical image frame, the multi-scale resolution images, the segmentation units, the corresponding local feature parameter sets, and the spatial mapping relationships into the image segmentation database.

[0034] This implementation scheme achieves pixel-level detailed management of surgical area medical images through multi-source acquisition, standardized preprocessing, and multi-scale resolution construction. The spatial pyramid and gray-level co-occurrence matrix algorithms are used to extract multi-directional texture features of segmentation units, significantly improving the ability to identify minute structures and blurred boundaries. Data at each scale and in each unit is uniformly managed through hash mapping, ensuring traceability and data integrity. The overall scheme lays a solid data foundation for subsequent intelligent segmentation and anomaly detection, effectively enhancing the sensitivity of small lesion detection and the ability to implement the scheme.

[0035] Optionally, the specific process of quantifying lesion-sensitive features of segmentation units at various scales based on local feature parameter sets is as follows: Obtain the number of scale layers, the number of segmentation units, multi-scale resolution images, segmentation units, and their corresponding local feature parameter sets and spatial mapping relationships. Input this data into a deep convolutional feature extraction network integrating a multi-scale pyramid structure and an attention mechanism. The deep convolutional feature extraction network adopts the mainstream U-Net segmentation network framework and embeds a self-attention mechanism in each scale branch to enhance the response capability to small regions and weak contrast targets. A backbone network and a multi-scale parallel branch structure are used to perform deep feature representation of segmentation units at various scales, outputting the corresponding multi-scale feature tensors. The feature tensor is a multidimensional matrix, with each dimension corresponding to the spatial coordinates of the segmentation unit, the channel dimension, and the scale index, respectively. For each segmentation unit at each scale, principal component analysis (PCA) is used to reduce the dimensionality of the multi-scale feature tensor. During PCA dimensionality reduction, the high-dimensional feature vector extracted for each segmentation unit is projected onto the direction of the first principal component to reflect the dominant feature strength of the segmentation unit at this scale. The score of the first principal component is taken as the feature response value. Simultaneously, the feature response values ​​of all segmentation units at each scale are statistically analyzed, and the corresponding mean and standard deviation of the feature response are calculated and output. The mean and standard deviation are used to measure the distribution of the segmentation unit features at this scale, facilitating anomaly sensitivity normalization. For each segmentation unit at each scale, the absolute difference between the feature response value and the mean feature response is calculated and divided by the sum of the standard deviation of the feature response and the minimum constant value to obtain the normalized feature anomaly degree. The minimum constant value is set to a value of 1. To avoid instability caused by a zero denominator, the weighted feature anomaly is obtained by multiplying the normalized feature anomaly by the corresponding scale-adaptive weighting factor. The absolute difference between the feature response values ​​of the segmentation unit at the same location in the current scale and the previous scale is calculated and multiplied by the cross-scale amplification weighting factor to obtain the multi-scale differential enhancement value; if it is the first layer, the multi-scale differential enhancement value is set to zero. The weighted feature anomaly and the multi-scale differential enhancement value are added to obtain the single anomaly significance value of the segmentation unit, which reflects the comprehensive strength of the segmentation unit's anomaly response and multi-scale mutation in this layer. The single anomaly significance values ​​of the segmentation units at the same location in all scales are summed and divided by the number of scale layers to obtain the sub-centimeter lesion sensitivity value. That is, the final sensitivity value of each pixel integrates the multi-scale anomaly representation, significantly improving the recognition sensitivity of small lesions, especially sub-centimeter lesions.

[0036] The specific formula for the sensitivity value of sub-centimeter lesions is as follows:

[0037] ;

[0038] In the formula, It represents the sub-centimeter-level lesion sensitivity value, which is used to measure the abnormal sensitivity of each pixel (x, y) to small, weak signals and lesions that are easily missed by the backbone network under multi-scale fusion, thereby improving the intelligent segmentation capability for blurred boundaries and extremely small abnormal structures. Indicates the number of scale layers; This represents the feature response value at the current scale; This represents the mean of the characteristic response, used to reflect the global normal level at this scale; The standard deviation of the characteristic response reflects the overall dispersion of the characteristic response at this scale. This represents the feature response value at the next higher scale. Represents a very small constant value, taking values ​​of ; The scale-adaptive weighting factor is automatically optimized by an attention mechanism algorithm based on the local feature parameter set of each segmentation unit, and its value range is (0,1]. This represents the cross-scale amplification weighting factor, which is obtained by constructing a validation set and fitting it using a Bayesian optimization algorithm based on the statistical distribution of the differences in the mean values ​​of feature response values ​​between scales. The value range is (0,2). It represents the normalized feature anomaly degree, which describes the degree of deviation of the segmentation unit from the normal level at this scale, and normalizes it with the standard deviation to eliminate the influence of distribution differences at different scales and make the anomaly intensity comparable. It represents the multi-scale differential enhancement value, which measures the degree of abrupt change in the feature response of the same spatial location between the current scale and the previous scale. It amplifies the anomalous boundaries at multiple scales and helps to enhance the response of edges, microstructures and weak contrast targets.

[0039] In this embodiment, Table 1 is a data table of sensitivity values ​​for sub-centimeter lesions. The cross-scale magnification weighting factor is set to 0.6, the scale adaptation weighting factor is 0.8, and the number of scale layers is 3. The mean value of the feature response corresponding to the first scale is 0.49, and the standard deviation of the feature response is 0.037; the mean value of the feature response corresponding to the second scale is 0.57, and the standard deviation of the feature response is 0.044; the mean value of the feature response corresponding to the third scale is 0.59, and the standard deviation of the feature response is 0.029. The table records in detail the feature response values ​​of the five segmentation units at the first scale, the second scale, the third scale, and the sensitivity value of sub-centimeter lesions. Among them, the feature response value of segmentation unit 1 at the first scale is 0.54, the feature response value at the second scale is 0.63, the feature response value at the third scale is 0.67, and the sensitivity value of sub-centimeter lesions is 1.486. Segmentation unit 2 at the first scale... The feature response value of segmentation unit 1 at the first scale is 0.56, at the second scale it is 0.59, at the third scale it is 0.65, and the sensitivity value for sub-centimeter lesions is 1.195; the feature response value of segmentation unit 2 at the first scale is 0.55, at the second scale it is 0.61, at the third scale it is 0.63, and the sensitivity value for sub-centimeter lesions is 1.059; the feature response value of segmentation unit 3 at the first scale is 0.52, at the second scale it is 0.62, at the third scale it is 0.60, and the sensitivity value for sub-centimeter lesions is 0.324; the feature response value of segmentation unit 4 at the first scale is 0.58, at the second scale it is 0.66, at the third scale it is 0.71, and the sensitivity value for sub-centimeter lesions is 2.324.

[0040] Table 1. Sensitivity Values ​​for Subcentimeter-Level Lesions

[0041]

[0042] like Figure 2 The image shows a heatmap of sensitivity values ​​for sub-centimeter lesions. The color intensity represents the sensitivity value of the sub-centimeter lesion; the color gradually changes from dark red (low value) to yellow and white (high value); each square represents the sensitivity value of a sub-centimeter lesion in its corresponding spatial coordinates. (Based on Table 1 and...) Figure 2As can be seen, the four segmentation units with sensitivity values ​​of 1.49, 1.20, 1.06, and 2.32 for sub-centimeter lesions have relatively high values, appearing as significantly bright areas on the heatmap. This indicates that lesions at these locations have high sensitivity and pose a potential risk of microlesions or require close monitoring. The value of 2.32 is the highest, displayed as the brightest white, indicating an extremely significant risk of lesions at this location. The segmentation unit corresponding to 0.32 and other segmentation units are located in the darker areas of the heatmap, appearing as normal areas. This suggests that the lesions corresponding to these segmentation units have low sensitivity and are less likely to show microlesions.

[0043] In this implementation scheme, a deep convolutional feature extraction network integrating a multi-scale pyramid structure and an attention mechanism is used, combined with principal component analysis and adaptive weighting factors, to achieve accurate quantification of sensitive features of segmentation units at different scales. Each segmentation unit obtains normalized and enhanced abnormal responses under multi-scale fusion, significantly improving the detection rate and boundary segmentation accuracy for sub-centimeter-level micro-lesions. This effectively overcomes challenges such as weak signals, structural ambiguity, and the tendency for the backbone network to miss detections, ensuring the system possesses highly sensitive and robust automatic segmentation capabilities, providing strong technical support for intelligent detection and precise localization of lesions in surgical areas.

[0044] Optionally, based on the quantification results of lesion sensitivity features, region type determination is performed to screen out sub-centimeter-level lesion risk regions. A medical image segmentation model is then constructed to generate basic segmentation kernel parameters. The specific process for achieving differentiated segmentation and annotation is as follows: The sub-centimeter-level lesion sensitivity values ​​of all segmentation units at all scales are calculated and written into the image segmentation database for subsequent traceability and global statistical analysis. The sub-centimeter-level lesion sensitivity value of each segmentation unit is compared with a sensitivity threshold. When the sub-centimeter-level lesion sensitivity value is less than the sensitivity threshold, it is determined to be a regular region. Using the BraTS publicly available medical image segmentation benchmark dataset as input, the medical image segmentation model is trained using the U-Net convolutional neural network algorithm. Using multi-scale resolution images as input, forward inference is performed on regular regions. By statistically analyzing the mean of the convolutional output response values ​​of the medical image segmentation model in regular regions, the basic segmentation kernel value is obtained, and the basic segmentation kernel parameters are output. The basic segmentation kernel parameters are combined with the medical image segmentation model to segment regular regions. When the sub-centimeter-level lesion sensitivity value is greater than or equal to the sensitivity threshold, it is determined to be a sub-centimeter-level lesion risk region, and the dynamic segmentation kernel generation process begins.

[0045] This implementation scheme achieves automatic and accurate classification and differentiated processing of routine areas and sub-centimeter-level lesion risk areas in medical images. By quantifying pixel-level sensitive features and determining thresholds, it effectively screens out small lesions that are easily missed, and adaptively configures segmentation parameters for different areas. Routine areas are segmented efficiently in batches using deep segmentation networks such as U-Net, improving overall processing efficiency; risk areas are processed using a dynamic segmentation kernel, further enhancing sensitivity to abnormal structures and boundary recognition capabilities. This method not only ensures the consistency and efficiency of global segmentation but also significantly improves the detection and fine segmentation capabilities of sub-centimeter-level lesions, meeting the clinical needs for highly sensitive, multi-scale, and intelligent segmentation of lesions in surgical areas.

[0046] Optionally, for sub-centimeter-level lesion risk areas, the specific process of structural feature evaluation and orientation consistency quantification based on local feature parameter sets, combining the basic parameters of the segmentation kernel and the lesion sensitivity feature quantification results, is as follows: Receive the basic values ​​of the segmentation kernel and the sub-centimeter-level lesion sensitivity values ​​of the sub-centimeter-level lesion risk areas, and execute the dynamic segmentation kernel generation process: For each sub-centimeter-level lesion risk area, obtain the local structural entropy, which reflects the local tissue complexity of the segmentation unit; simultaneously, with the current sub-centimeter-level lesion risk area as the center, select all segmentation units within a fixed neighborhood sliding window, with a sliding window step size of 1 pixel, and use mirror filling at the boundaries. Calculate all segmentation units within the window using the Sobel operator. The gradient values ​​of normalized pixel grayscale values ​​in the horizontal and vertical directions are combined into a gradient vector set of window pixels. Principal component analysis is used to decompose the gradient vector set into principal components, and the direction of the first principal component is extracted as the principal direction component. The principal direction component is the direction of the maximum variance, and the polar angle of the principal component vector is calculated by the arctan2 function to obtain the principal direction angle. The sub-centimeter lesion sensitivity value of the sub-centimeter lesion risk area is multiplied by the corresponding local structural entropy to obtain the structural anomaly response value. The product of the magnitude of the principal direction component and the cosine of the corresponding principal direction angle is calculated, where the magnitude of the principal direction component represents the intensity of the dominant direction, and a stabilization constant is added to obtain the structural correction value.

[0047] This implementation scheme achieves dynamic adaptive generation and adjustment of the segmentation kernel for sub-centimeter lesion risk areas by integrating basic segmentation kernel parameters, sub-centimeter lesion sensitivity values, and local structural features. This scheme can accurately capture abnormalities and complex structures in lesion areas, improve the sensitivity of boundary detection and the specificity of segmentation, and significantly enhance the automatic segmentation capability for complex, small, and low-contrast lesions, providing strong data support for high-precision medical image segmentation and intelligent assisted diagnosis.

[0048] Optionally, the specific process for generating the dynamic segmentation kernel output result based on the combined structural feature assessment and directional consistency quantification results is as follows: Divide the structural anomaly response value by the structural correction value, and take the opposite as the exponent for natural exponentiation. Add the result of the natural exponentiation to a constant to obtain the dynamic segmentation kernel suppression term. The structural anomaly response value is obtained by multiplying the sub-centimeter lesion sensitivity value by the local structural entropy. The structural correction value is obtained by multiplying the principal direction component magnitude by the cosine of the principal direction angle, plus a stabilization constant, ensuring that the denominator is not zero and that there is sufficient adjustment capability for structures protruding in the principal direction. The natural exponentiation is an exponentiation operation with base e, and the constant is used to balance the numerical range of the dynamic segmentation kernel suppression term. Multiply the reciprocal of the dynamic segmentation kernel suppression term by the cross-scale weighting factor and add the constant to obtain the segmentation kernel adaptive adjustment value. Multiply the segmentation kernel base value by the segmentation kernel adaptive adjustment value to obtain the dynamic segmentation kernel output value. The base value of the segmentation kernel reflects the average level of convolutional features in the conventional region, while the output value of the dynamic segmentation kernel serves as the local convolutional parameter within the sub-centimeter-level lesion risk region. This parameter guides the adaptive adjustment of the medical image segmentation model, enabling it to respond to the segmentation of small and complex lesions.

[0049] The specific formula for the output value of the dynamic segmentation kernel is as follows:

[0050] ;

[0051] In the formula, This represents the output value of the dynamic segmentation kernel, which is used to dynamically generate adaptive segmentation kernel values ​​for sub-centimeter-level lesion risk areas. Based on the salience of local abnormalities, structural complexity, and principal direction information, it adaptively adjusts the sensitivity of the segmentation kernel for sub-centimeter-level lesion risk areas, enhancing the segmentation response for small, vaguely defined, and complex lesions, and improving the precision and targeting of automatic segmentation. This represents the basic value of the segmentation kernel, which serves as the benchmark for the adaptive adjustment of the entire segmentation kernel. It reflects the segmentation capability under conventional structures and provides a basic reference for dynamic adjustment. This represents the sensitivity value for sub-centimeter-level lesions, characterizing the abnormal intensity of multi-scale features at the current location. It is used to identify small, weak-contrast lesions that are easily overlooked by the backbone network. The larger the value, the more abnormal it is. Adjusting the segmentation kernel enhances sensitivity. It represents the local structural entropy, which measures the texture complexity and organizational uncertainty of the segmentation unit. The larger the value, the higher the local organizational complexity, the greater the segmentation difficulty, and the more the sensitivity of the segmentation kernel needs to be improved. It represents the magnitude of the principal direction component, reflecting the dominant directional strength of the local structure; the larger the magnitude, the more prominent the directionality. The cosine value of the principal direction angle is used to adjust the projection weight of the principal direction in the horizontal direction, reflecting the consistency of the principal direction; This represents a stabilization constant, with a value of 0.01. This represents the structural anomaly response value, capturing the coupling between the salience of the anomaly and the complexity of the local structure. The larger the value, the more anomaly and complexity there is, and the more sensitive the segmentation kernel needs to be. This represents the structural correction value, reflecting the significance of the principal direction and the weight after correction for the principal direction angle. A larger value indicates a dominant structural direction. The segmentation kernel can appropriately suppress the hypersensitive response. Used to balance the dimensions of the principal direction component and the angle term, preventing the denominator from being zero or too small; The cross-scale weighted factor is initially set to 0.5. In subsequent optimization, based on the original frames of medical images in the validation set and the true segmentation labels of sub-centimeter lesion risk areas, the Dice coefficient is maximized as the sub-centimeter lesion segmentation performance index. The cross-scale weighted factor is automatically searched through the Bayesian optimization algorithm, and the cross-scale weighted factor corresponding to the maximum value of the Dice coefficient is selected as the optimal weight factor, with a value range of (0,2).

[0052] In this implementation scheme, by combining structural anomaly response with orientation consistency correction, segmentation kernel parameters for each sub-centimeter-level lesion risk region are dynamically generated. This not only adaptively adjusts segmentation sensitivity and effectively suppresses noise and artifact responses, but also highlights the true characteristics of small, poorly defined, and complex lesions. This significantly improves the detection rate and segmentation accuracy of medical image segmentation methods for extremely small lesions, providing solid data support and technical assurance for intelligent identification of early lesions in the surgical area and personalized, refined surgical planning.

[0053] Optionally, the process of adaptively segmenting lesion regions and filtering out abnormal regions based on the output of the dynamic segmentation kernel is as follows: Calculate the dynamic segmentation kernel output value for all sub-centimeter-level lesion risk regions, and assign the dynamic segmentation kernel output value to the corresponding spatial location as the local convolutional kernel parameter of the convolutional neural network algorithm within the sub-centimeter-level lesion risk region. The dynamic segmentation kernel output value is directly mapped to the convolutional kernel amplitude adjustment factor of the corresponding pixel using an amplitude scaling method. Only the amplitude is adjusted at each spatial location, without changing the spatial shape or offset of the convolutional kernel. The amplitude scaling relationship is achieved through a linear normalization function, ensuring a one-to-one correspondence between the dynamic segmentation kernel output value and the convolutional kernel response. For medical image segmentation models integrating deformable convolution mechanisms, the dynamic segmentation kernel output result can also be directly mapped to the spatial offset of the convolutional kernel through a preset mapping function. This allows the convolutional kernel to flexibly adapt to local deformations and abnormal regions based on local structural features, achieving high-precision capture of heterogeneous regions, complex boundaries, and small lesions, ensuring a unique local convolutional kernel configuration for each segmentation unit location. Adaptive segmentation processing was performed on sub-centimeter lesion risk areas using a medical image segmentation model. The average dynamic segmentation kernel output value of all sub-centimeter lesion risk areas was calculated. Sub-centimeter lesion risk areas with dynamic segmentation kernel output values ​​higher than the average dynamic segmentation kernel output value were selected. These areas indicated that the local structure, degree of abnormality, or texture complexity was significantly higher than that of conventional risk areas, and there was a higher demand for anomaly detection and segmentation difficulty. These areas were marked as abnormal sub-centimeter lesion risk areas. All dynamic segmentation kernel output values ​​and adaptive segmentation processing results were written into the image segmentation database.

[0054] In this implementation scheme, spatial adaptive segmentation of sub-centimeter lesion regions is performed using dynamic segmentation kernel output values. This allows for dynamic adjustment of the convolution kernel's amplitude, shape, and spatial offset to address local structural differences and abnormal features in different regions, achieving refined responses to heterogeneity, complex boundaries, and minute lesions. By filtering regions where the dynamic segmentation kernel output value is higher than the global average, high-risk lesions with complex structures and high degrees of abnormality are automatically identified, enabling intelligent focusing and subsequent refined processing of key areas. This improves the segmentation sensitivity and boundary accuracy for early-stage minute lesions.

[0055] Optionally, for abnormal regions, the specific process of multi-scale inference using the medical image segmentation model is as follows: The dynamic segmentation kernel output value and adaptive segmentation processing result of the abnormal sub-centimeter-level lesion risk region are received; the main segmentation network in the medical image segmentation model is called to perform forward inference on the multi-scale resolution image, outputting the main branch segmentation probability value for each abnormal sub-centimeter-level lesion risk region; and the auxiliary segmentation network in the medical image segmentation model is called in parallel to output the auxiliary segmentation probability value. The main segmentation network adopts a U-Net-based architecture, with a ResNet34 backbone encoder and bilinear upsampling decoder. The auxiliary segmentation network uses a single-layer shallow convolutional structure, i.e., a 3×3 convolutional kernel and ReLU activation, and inputs the same feature tensor as the main branch, focusing on microstructures and edge perception. The features of the main branch and auxiliary branch are concatenated along the channel dimension, and then fused using a 1×1 convolution. The main branch uses the Dice loss function during training, and the auxiliary branch uses the binary cross-entropy loss function. The fused output of the main and auxiliary branches introduces an L1 consistency loss to constrain the consistency of the prediction results between the main branch and the auxiliary branch, and the fused result is used for subsequent inference. Using the Sobel operator, convolution calculations are performed on the main branch segmentation probability values ​​in both the horizontal and vertical directions to obtain the horizontal and vertical gradients of the main branch segmentation probability values. For each abnormal sub-centimeter lesion risk region, the gradient magnitude of the main branch segmentation probability value is calculated based on the horizontal and vertical gradients. Using the main branch segmentation probability values ​​and auxiliary segmentation probability values ​​as inputs, the structural similarity index algorithm is used to calculate the probability distribution consistency value. The window size of the structural similarity index algorithm is set to 5×5, which is the standard implementation method.

[0056] In this implementation scheme, a multi-scale deep segmentation network with coordinated main and auxiliary branches is used. This network combines the main branch U-Net with auxiliary shallow edge-aware convolutions to achieve high-resolution feature extraction and structural boundary refinement for abnormal sub-centimeter lesion regions. Main-auxiliary feature fusion and L1 consistency loss constraints effectively improve the model's segmentation consistency and robustness for small and complex lesions. The Sobel operator and structural similarity index are employed to further enhance the sharpness of segmentation probability boundaries and the spatial consistency of main-auxiliary outputs. The overall scheme improves the accuracy of identifying and segmenting extremely small lesions, ensuring more sensitive boundary localization of abnormal regions and more reliable segmentation results, providing a solid technical guarantee for highly sensitive lesion detection and automatic annotation in surgical areas.

[0057] Optionally, the specific process for boundary accuracy assessment based on multi-scale inference results is as follows: The edge sharpness value is obtained by exponentially calculating the edge sharpening weight factor as the exponent and the gradient magnitude of the main branch segmentation probability value as the base. The gradient magnitude of the main branch segmentation probability value refers to the first derivative calculated in the horizontal and vertical directions for each pixel within a 5×5 neighborhood, implemented using a Sobel filter. For the same sub-centimeter-level lesion risk region, the absolute difference between the main branch segmentation probability value and the auxiliary segmentation probability value is calculated. This difference is then exponentially calculated using the consistency amplification weight factor as the exponent and the absolute difference as the base. The result of the exponential calculation is multiplied by the consistency suppression weight factor and added to a constant to obtain the probability consistency suppression value. The consistency amplification weight factor and the consistency suppression weight factor are used to amplify the prediction difference between the main and auxiliary branches and suppress over-response, respectively. The negative of the ratio of the probability distribution consistency value to the structural similarity weight factor is used as the exponent for natural exponential calculation to obtain the structural similarity correction value. The ratio of the edge sharpness value to the probability consistency suppression value is multiplied by the structural similarity correction value to obtain the sub-centimeter-level lesion boundary confidence value. All calculations are performed on a unit-by-unit basis, and can be processed in batches and in parallel under spatial coordinate mapping. Sub-centimeter-level lesion boundary confidence values ​​are used to ultimately determine whether the boundary needs fine-tuning, enabling segmentation optimization for blurred, weak signals, and complex structural edges.

[0058] The specific formula for the confidence value of the sub-centimeter lesion boundary is as follows:

[0059] ;

[0060] In the formula, It represents the confidence value of sub-centimeter lesion boundaries, used for the refined discrimination of small lesions and their boundaries. It not only needs to accurately distinguish between lesions and normal tissue, but also needs to achieve sharp enhancement of boundaries, correction of primary and secondary consistency, and confirmation of multimodal structures in the edge areas where there are blurred, weak signals, and errors that are easily missed by the main segmentation network. This represents the probability value of splitting the main branch; This represents the auxiliary segmentation probability value; It represents the gradient magnitude and describes the sharpness of the boundary. It represents the consistency value of the probability distribution, reflecting the degree of consistency between the main and auxiliary branches in the spatial structure; The edge sharpening weight factor is represented by the probability value of the main branch segmentation and gradient distribution statistics of the risk area of ​​abnormal sub-centimeter lesions. The Bayesian optimization algorithm is used to evaluate the boundary clarity index of the segmentation results for different edge sharpening weight factors and fit to obtain the optimal edge sharpening weight factor, with a value range of [0.5,3]. The consistency amplification weight factor is obtained by combining the statistical characteristics of the distribution of the absolute difference between the main branch segmentation probability value and the auxiliary segmentation probability value with cross-validation and Bayesian optimization algorithm, and fitting the consistency penalty effect on the boundary ambiguity area and error area. The value range is [1,5]. The consistency suppression weight factor is defined as the statistical data of the difference between the normalized gray pixel values ​​corresponding to the main branch segmentation probability value and the auxiliary segmentation probability value. The grid search algorithm is used to traverse different consistency suppression weight factors to evaluate the false negative rate and boundary stability of the segmentation results in the main and auxiliary branch consistency region. The optimal consistency suppression weight factor is obtained by fitting, and the value range is [0.1,2]. The structural similarity weight factor is represented by the probability distribution consistency value. Using the Bayesian optimization algorithm, the reliability of the segmentation results in regions with low structural similarity is compared for different structural similarity weight factors, and the optimal structural similarity weight factor is automatically fitted. The value range is [1,5].

[0061] In this implementation scheme, by jointly quantifying the edge sharpness, consistency, and structural similarity indices of the multi-scale segmentation probability map, the fuzzy or complex boundaries of sub-centimeter lesions can be automatically identified, achieving accurate assessment of segmentation reliability. Utilizing the synergy of main and auxiliary branches and the regulation of multiple weighting factors, the sensitivity and reliability of the backbone segmentation model in boundary detection of small lesions and difficult-to-define regions are significantly improved, effectively reducing the risk of false detections and missed detections at the edges. This provides solid data support for subsequent segmentation correction and refinement, comprehensively enhancing the automatic detection and high-precision segmentation capabilities for early-stage small lesions.

[0062] Optionally, adaptive smoothing and blur correction of sub-centimeter lesion edges are implemented. Simultaneously, based on the quantification results of lesion sensitive features and boundary accuracy evaluation results, the specific process of achieving visual feedback and closed-loop optimization is as follows: The confidence value of the sub-centimeter lesion boundary in all abnormal sub-centimeter lesion risk areas is calculated, written into the image segmentation database, and compared with a confidence threshold for segmentation correction. For abnormal sub-centimeter lesion risk areas with a confidence value higher than the confidence threshold, no fine segmentation is required, and the adaptive segmentation result is directly output. For abnormal sub-centimeter lesion risk areas with a confidence value less than or equal to the confidence threshold, an adaptive smoothing method is used for edge denoising and blur correction. The boundary adaptive smoothing method is based on a fully variational regularized denoising algorithm, with parameters automatically adapted to regional features to preserve the main boundary structure and eliminate high-frequency artifacts. Blur correction can be achieved using morphological reconstruction to avoid excessive smoothing leading to the loss of the true structure. Using images of aberrant sub-centimeter lesion risk areas, processed by edge denoising and blur correction, as input, a medical image segmentation model is invoked for local secondary inference. During local secondary inference, the medical image segmentation model automatically adjusts the inference window and convolution kernel weights based on the spatial distribution of the corrected region, achieving accurate boundary reconstruction for residual blurred regions; fine-grained segmentation is then performed; based on sub-centimeter lesion sensitivity values ​​and sub-centimeter lesion boundary confidence values, a spatial heatmap is generated for all segmentation units in each frame of the medical image; the sub-centimeter lesion sensitivity values, sub-centimeter lesion boundary confidence values, and all corresponding segmentation results are analyzed, defining... Using the Dice coefficient as the comprehensive evaluation index as the optimization objective, a Bayesian optimization method is employed to jointly and adaptively search the structural hyperparameters of the medical image segmentation model, such as backbone network depth and number of channels, convolution kernel parameters such as dynamic kernel amplitude and kernel shape adjustment factor, as well as the sensitivity threshold and confidence threshold. In each iteration, the segmentation process is automatically run on the validation set, recording the correspondence between parameter combinations and optimization objectives. The Bayesian optimization module fits the parameter performance function, selects the optimal parameter configuration, and finally optimizes and pushes the medical image segmentation model, convolution kernel parameters, sensitivity threshold, and confidence threshold, achieving automatic closed-loop optimization and continuous performance improvement in medical image segmentation.

[0063] like Figure 3The diagram shows the workflow of a surgical lesion image segmentation method based on a multi-scale dynamic segmentation kernel. First, surgical area medical images are acquired and preprocessed. Multi-scale images are constructed using a spatial pyramid, and local texture features are extracted. Then, lesion sensitivity quantification is performed on the multi-scale features, and the results of comparing sensitivity values ​​with thresholds distinguish between regular areas and sub-centimeter-level lesion risk areas. For risk areas, a dynamic segmentation kernel is generated and adjusted to achieve adaptive segmentation, and abnormal areas are filtered based on the output values. Next, multi-scale inference is performed through a master-slave network to quantify edge sharpness, probability consistency, and structural similarity, calculate boundary credibility, and perform refined segmentation. Finally, a closed-loop update of model parameters and thresholds is implemented, thereby achieving highly sensitive and refined segmentation of small lesions.

[0064] In this implementation scheme, high-precision segmentation of ambiguous and complex boundaries is achieved by implementing adaptive smoothing, fuzzy correction, and local secondary inference in the risk areas of abnormal sub-centimeter lesions. This effectively suppresses noise while preserving true structural features. Combined with pixel-level heatmaps and multi-dimensional risk indicators, it assists doctors in accurately locating high-risk lesion areas. Simultaneously, a closed-loop feedback mechanism based on Bayesian optimization continuously optimizes the segmentation model and parameter configuration, ensuring dynamic adaptation and long-term stable improvement in segmentation results, significantly enhancing the intelligent detection and boundary discrimination capabilities of sub-centimeter lesions.

[0065] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0066] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0067] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0068] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0069] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0070] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0071] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0072] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0073] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0074] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for segmenting a surgical site lesion image based on a multi-scale dynamic segmentation kernel, characterized in that, The method comprises: S1, real-time acquisition of medical image original frames in a surgical area, uniform preprocessing of the medical image original frames, generation of multi-scale resolution images based on the preprocessed medical image original frames, division of segmentation units according to spatial coordinates, acquisition of local feature parameter sets of each segmentation unit, and establishment of a spatial mapping relationship between the multi-scale resolution images and the segmentation units; S2, based on the local feature parameter sets, conducting lesion sensitive feature quantization on the segmentation units at each scale, determining the region type according to the lesion sensitive feature quantization results, screening out sub-centimeter lesion risk regions, constructing a medical image segmentation model, generating segmentation kernel basic parameters, and realizing differentiated segmentation and labeling; S3, for the sub-centimeter lesion risk region, combining the segmentation kernel basic parameters and the lesion sensitive feature quantization results, based on the local feature parameter sets, conducting structure feature evaluation and direction consistency quantization, comprehensively analyzing the structure feature evaluation and direction consistency quantization results, generating dynamic segmentation kernel output results, realizing adaptive segmentation of the lesion region according to the dynamic segmentation kernel output results, and screening out abnormal regions; S4, for the abnormal region, using the medical image segmentation model to conduct multi-scale reasoning, according to the multi-scale reasoning results, conducting boundary precision evaluation, and realizing adaptive smoothing and fuzzy correction of the sub-centimeter lesion edge; at the same time, based on the lesion sensitive feature quantization results and the boundary precision evaluation results, realizing visual feedback and closed-loop optimization.

2. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The specific process of real-time acquisition of medical image original frames, uniform preprocessing of the medical image original frames, generation of multi-scale resolution images based on the preprocessed medical image original frames, division of segmentation units according to spatial coordinates, acquisition of local feature parameter sets of each segmentation unit, and establishment of a spatial mapping relationship between the multi-scale resolution images and the segmentation units is as follows: Real-time acquisition of medical image original frames, synchronous acquisition of spatial resolution information, device acquisition parameters, image timing labels and spatial coordinates of each frame, and acquisition of pixel gray matrix of the medical image original frames; Uniform normalization processing of the acquired medical image original frames, including pixel gray normalization, noise filtering processing and contrast enhancement; Down-sampling of the normalized medical image original frames at different scaling ratios by the spatial pyramid method to generate multi-scale resolution images and obtain the number of scale layers; At each scale, the corresponding resolution image is pixel-level segmented into multiple segmentation units based on the spatial coordinates, the number of segmentation units is obtained, and the spatial coordinates and normalized pixel gray values of each segmentation unit are recorded synchronously, and based on the normalized pixel gray values, the joint distribution of gray pairs in the segmentation unit is counted in a fixed neighborhood sliding window with the spatial coordinates as the center by using the gray co-occurrence matrix algorithm, and texture features including contrast, correlation, energy and local structure entropy are calculated; based on the spatial coordinates, normalized pixel gray values and texture features of each segmentation unit, a local feature parameter set is constructed; All scales, all segmentation units and local feature parameter sets are uniformly mapped to establish a spatial correspondence between the segmentation units and the medical image original frames and the multi-scale resolution images; The image segmentation database is established, and the medical image original frame, the preprocessed medical image original frame, the multi-scale resolution image, the segmentation unit, and the corresponding local feature parameter set and spatial mapping relationship are stored in the image segmentation database.

3. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The specific process of carrying out lesion sensitive feature quantization on the segmentation unit under each scale based on the local feature parameter set is as follows: The scale layer number, the segmentation unit number, the multi-scale resolution image, the segmentation unit, and the corresponding local feature parameter set and spatial mapping relationship are obtained, and a deep convolution feature extraction network integrated with a multi-scale pyramid structure and an attention mechanism is inputted, a main network and a multi-scale parallel branch structure are adopted to perform deep-level feature expression on the segmentation unit under each scale, and a corresponding multi-scale feature tensor is outputted; For each segmentation unit of each scale, the multi-scale feature tensor is processed by dimension reduction by using principal component analysis, and the first principal component score is taken as a feature response value; meanwhile, the feature response values of all segmentation units under each scale are counted, and a corresponding feature response mean and feature response standard deviation are calculated and outputted; For each segmentation unit of each scale, the absolute difference value between the feature response value and the feature response mean is calculated, and then divided by the sum of the feature response standard deviation and an extremely small constant value to obtain a normalized feature abnormality degree, and the normalized feature abnormality degree is multiplied by a corresponding scale adaptive weight factor to obtain a weighted feature abnormality degree; The absolute difference value between the feature response values of the segmentation unit at the same position in the current scale and the previous scale is calculated, and then multiplied by a cross-scale amplification weight factor to obtain a multi-scale difference enhancement value; the weighted feature abnormality degree and the multi-scale difference enhancement value are added to obtain a single abnormality saliency value of the segmentation unit; The single abnormality saliency values of the segmentation unit at the same position of all scales are summed and divided by the scale layer number to obtain a sub-centimeter level lesion sensitive value.

4. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The specific process of performing region type determination according to the lesion sensitive feature quantization result, screening out a sub-centimeter level lesion risk region, constructing a medical image segmentation model, generating segmentation kernel basic parameters, and realizing differentiated segmentation and labeling is as follows: The sub-centimeter level lesion sensitive values of all segmentation units under all scales are calculated, written into the image segmentation database, and compared with a sensitive threshold value; When the sub-centimeter level lesion sensitive value is less than the sensitive threshold value, it is determined as a regular region, the BraTS public medical image segmentation benchmark dataset is taken as input, a convolutional neural network algorithm is used to train a medical image segmentation model, a multi-scale resolution image is taken as input, a forward inference is performed on the regular region, the mean of the convolution output response value of the medical image segmentation model in the regular region is obtained to obtain a segmentation kernel basic value, and the basic segmentation kernel parameters are outputted; the basic segmentation kernel parameters are used in combination with the medical image segmentation model to segment the regular region; When the sub-centimeter level lesion sensitive value is greater than or equal to the sensitive threshold value, it is determined as a sub-centimeter level lesion risk region, and a dynamic segmentation kernel generation process is entered.

5. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The specific process of performing structure feature evaluation and direction consistency quantization based on the local feature parameter set for the sub-centimeter level lesion risk region in combination with the segmentation kernel basic parameters and the lesion sensitive feature quantization result is as follows: The sub-centimeter lesion sensitive value of the sub-centimeter lesion risk area is multiplied by the corresponding local structure entropy to obtain a structure anomaly response value. The product of the main direction component module length and the cosine value of the corresponding main direction angle is calculated, and a stabilization constant value is added to obtain a structure correction value. The specific process of the comprehensive structure feature evaluation and the direction consistency quantization result to generate the dynamic segmentation kernel output result is as follows:

6. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The structure anomaly response value is divided by the structure correction value, and the reciprocal is taken as the exponential power to perform natural exponential operation, and the natural exponential operation result is added to constant one to obtain a dynamic segmentation kernel inhibition term; the reciprocal of the dynamic segmentation kernel inhibition term is multiplied by the cross-scale weighted weight factor, and constant one is added to obtain a segmentation kernel adaptive adjustment value; the segmentation kernel basic value is multiplied by the segmentation kernel adaptive adjustment value to obtain a dynamic segmentation kernel output value. The specific process of realizing adaptive segmentation of the lesion area according to the dynamic segmentation kernel output result and screening the abnormal area is as follows:

7. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The dynamic segmentation kernel output values of all sub-centimeter lesion risk areas are calculated, and the dynamic segmentation kernel output values are respectively assigned to the corresponding spatial positions as the local convolution kernel parameters of the convolutional neural network algorithm in the sub-centimeter lesion risk area, and the medical image segmentation model is used to implement adaptive segmentation processing on the sub-centimeter lesion risk area; The average value of the dynamic segmentation kernel output values of all sub-centimeter lesion risk areas is calculated, and the sub-centimeter lesion risk area with a dynamic segmentation kernel output value higher than the average value of the dynamic segmentation kernel output value is selected and marked as an abnormal sub-centimeter lesion risk area; all dynamic segmentation kernel output values and adaptive segmentation processing results are written into an image segmentation database. The specific process of using the medical image segmentation model to perform multi-scale reasoning on the abnormal area is as follows:

8. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The dynamic segmentation kernel output value and the adaptive segmentation processing result of the abnormal sub-centimeter lesion risk area are received, the main segmentation network in the medical image segmentation model is called, forward reasoning is performed on the multi-scale resolution image, and the main branch segmentation probability value of each abnormal sub-centimeter lesion risk area is output, and the auxiliary segmentation network in the medical image segmentation model is called in parallel to output the auxiliary segmentation probability value; ​ The Sobel operator is used to perform convolution calculation on the main branch segmentation probability value in the horizontal direction and the vertical direction respectively, to obtain the horizontal direction gradient and the vertical direction gradient of the main branch segmentation probability value, and for each abnormal sub-millimeter lesion risk area, the gradient module length of the main branch segmentation probability value is calculated based on the horizontal direction gradient and the vertical direction gradient; The structural similarity index algorithm is used to calculate the probability distribution consistency value based on the main branch segmentation probability value and the auxiliary segmentation probability value.

9. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The specific process of boundary precision evaluation according to the multi-scale inference result is: The edge sharpness weight factor is used as the exponential power, and the main branch segmentation probability value gradient module length is used as the base number to perform exponential operation to obtain the edge sharpness value; For the same abnormal sub-millimeter lesion risk area, the absolute difference value of the main branch segmentation probability value and the auxiliary segmentation probability value is calculated, the consistency amplification weight factor is used as the exponential power, and the absolute difference value is used as the base number to perform exponential operation, the exponential operation result is multiplied by the consistency suppression weight factor, and a constant one is added to obtain the probability consistency suppression value; The inverse of the ratio of the probability distribution consistency value and the structural similarity weight factor is used as the exponential power to perform natural exponential operation to obtain the structural similarity correction value; The sub-millimeter lesion boundary credibility value is obtained by multiplying the ratio of the edge sharpness value and the probability consistency suppression value by the structural similarity correction value.

10. The multi-scale dynamic split kernel based surgical site lesion image segmentation method of claim 1, wherein, The specific process of realizing adaptive smoothing and fuzzy correction of sub-millimeter lesion edge, and realizing visual feedback and closed-loop optimization based on lesion sensitive feature quantization result and boundary precision evaluation result is: The sub-millimeter lesion boundary credibility value of all abnormal sub-millimeter lesion risk areas is calculated and written into the image segmentation database, and compared with the credibility threshold to perform segmentation correction: for the abnormal sub-millimeter lesion risk area with sub-millimeter lesion boundary credibility value higher than the credibility threshold, no fine segmentation is needed, and the adaptive segmentation processing result is directly output; For the abnormal sub-millimeter lesion risk area with sub-millimeter lesion boundary credibility value less than or equal to the credibility threshold, edge denoising and fuzzy correction are performed by using boundary adaptive smoothing, and the abnormal sub-millimeter lesion risk area image after edge denoising and fuzzy correction is taken as input to call the medical image segmentation model to perform local secondary inference and fine segmentation; Based on the sub-millimeter lesion sensitivity value and the sub-millimeter lesion boundary credibility value, a spatial heat map is generated for all segmentation units of each frame of medical image; The Bayesian optimization method is used to optimize the medical image segmentation model, the convolution kernel parameter, the sensitivity threshold and the credibility threshold based on the sub-millimeter lesion sensitivity value, the sub-millimeter lesion boundary credibility value and all corresponding segmentation results.

Images