A feature optimization method and device for colonoscopy image segmentation

Abstract

The application discloses a feature optimization method and device for colonoscopy image segmentation, and belongs to the technical field of medical image processing and intelligent analysis. The method comprises the following steps: extracting structure features reflecting structure changes and geometric features reflecting morphological changes from colonoscopy images or intermediate features; calculating a consistency index based on the spatial correlation or change consistency between the structure features and the geometric features; simultaneously obtaining a reflection feature for reflecting reflection interference; generating a gating weight according to the consistency index and the reflection feature, and weighting and optimizing the structure features by using the gating weight, and fusing the structure features with original features; and using the fused features in the training or reasoning process of a colonoscopy image segmentation model. In the above manner, the application can effectively reduce the influence of reflection interference on the colonoscopy image segmentation result, improve the accuracy and stability of the segmentation result, and does not depend on a specific network structure, and has good universality and practical value.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and intelligent analysis technology, specifically to a feature optimization method and apparatus for colonoscopy image segmentation. Background Technology

[0002] Colonoscopy is an important tool for screening and diagnosing colorectal diseases. By observing images of the colonic wall, it helps doctors detect abnormal tissues such as polyps and lesions. With the continuous growth of medical imaging data, computer-based automatic segmentation technology for colonoscopy images has gradually become an important research direction for improving diagnostic efficiency and accuracy.

[0003] Existing colonoscopy image segmentation methods are typically based on deep learning models. These models extract, fuse, and decode the images to reconstruct the segmented target region. However, colonoscopy images are often subject to complex interference factors during imaging, such as specular reflection, mucus highlights, tissue folds, and uneven illumination. These factors can easily create strong local brightness variations or pseudo-structural boundaries in the images, thus adversely affecting the segmentation results.

[0004] In practical applications, reflective or highlighted areas often resemble real tissue boundaries in terms of pixel intensity and local gradient. Existing segmentation methods are prone to misidentifying such reflective interference as target boundaries during feature extraction and enhancement, leading to false positives, broken boundaries, or contour shifts in the segmentation results. Furthermore, for polyp regions with blurred boundaries, low contrast, or relatively gentle morphological changes, existing methods struggle to reliably distinguish between real tissue structures and background regions at the feature level, further impacting the reliability of segmentation.

[0005] To improve segmentation results, some existing technologies attempt to optimize features by introducing edge enhancement, attention mechanisms, multi-branch feature extraction, or reflection removal preprocessing. However, most of these methods focus on enhancing feature saliency or suppressing noise interference, failing to effectively distinguish between real tissue structural changes and reflection interference at the feature level. Furthermore, they lack the ability to characterize the intrinsic consistency between structural information and tissue morphological changes, thus still resulting in segmentation instability in complex colonoscopy imaging scenarios.

[0006] Therefore, there is an urgent need for a feature optimization method for colonoscopy image segmentation that can comprehensively analyze structural changes, tissue morphology changes, and reflective interference at the feature level. By optimizing the features in a targeted manner, the influence of reflective interference on the segmentation results can be suppressed, thereby improving the accuracy and stability of colonoscopy image segmentation. Summary of the Invention

[0007] The purpose of this invention is to provide a feature optimization method and apparatus for colonoscopy image segmentation. By jointly analyzing structural features, geometric features and reflective features at the feature level, constructing a structure-geometric consistency index, and generating gating weights by combining reflective interference information, the features are optimized in a targeted manner, thereby suppressing reflective false boundaries, enhancing real tissue boundary features, and improving the accuracy and stability of colonoscopy image segmentation results.

[0008] To achieve the above objectives, the present invention employs the following technical solution:

[0009] This invention provides a feature optimization method for colonoscopy image segmentation, comprising:

[0010] Obtain the features to be optimized from colonoscopy images;

[0011] Structural features are extracted from the features to be optimized to characterize tissue boundaries, texture changes, or local intensity changes in the image.

[0012] Geometric features are extracted from the features to be optimized to characterize tissue morphology, surface changes, or regional change trends in the image.

[0013] Calculate the structural-geometric consistency index based on the structural and geometric features;

[0014] Acquire reflective features to characterize bright interference areas in colonoscopy images;

[0015] Gating weights are generated based on the structure-geometry consistency index and the reflective features;

[0016] The structural features are weighted and optimized based on the gating weights to obtain optimized structural features, which are then fused with the original features to be optimized to form a comprehensive feature representation, which is used in the training or inference process of the colonoscopy image segmentation model.

[0017] Preferably, the features to be optimized for acquiring colonoscopy images include:

[0018] The intermediate features output from the encoding stage of colonoscopy images or segmentation models are used as the features to be optimized, represented as follows: , in, Indicates the features to be optimized. This represents the size of the feature map in the vertical direction. This represents the dimension of the feature map in the horizontal direction. This indicates the number of channels in the feature map.

[0019] Preferably, the features to be optimized for acquiring colonoscopy images further include:

[0020] Normalize the mean and standard deviation of each channel of the feature to be optimized to obtain the normalized feature tensor. The normalized feature tensor As a feature to be optimized.

[0021] Preferably, structural features are extracted from the features to be optimized, and are represented as follows:

[0022] ,

[0023] in, Indicates multi-channel structural features, This represents the structural feature extraction function. The number of channels for structural features, satisfying ;

[0024] The structural feature extraction function can be implemented using any of the following methods:

[0025] Local feature extraction based on convolution operations, local intensity change detection based on gradient or difference operators, and edge enhancement based on learnable filters.

[0026] Preferably, geometric features are extracted from the features to be optimized, as follows:

[0027] ,

[0028] in, Represents multi-channel geometric features, Represents the geometric feature extraction function. The number of channels for geometric features, satisfying ;

[0029] The geometric feature extraction function can be implemented using any of the following methods:

[0030] Multi-scale convolution, dilated convolution, second-order difference filtering, and feature calculation methods for local structure tensors.

[0031] Preferably, the structure-geometric consistency index is calculated based on the structural and geometric features, including:

[0032] By enhancing the local structural changes of the structural features using a first-order gradient operator, a structural strength map is obtained, as follows:

[0033] ,

[0034] ,

[0035] ,

[0036] in, To reflect spatial location Structural strength diagram showing local structural changes. This represents the mapping of structural features in spatial location. The response value at that location, Indicates the spatial location of the feature to be optimized. The feature vector at that location, and These represent the structural convolution kernels along the horizontal and vertical directions, respectively, with symbols... Represents convolution operation, operator and These represent first-order transformation operations performed along the horizontal and vertical directions, respectively.

[0037] The local structural changes of the geometric features described by the second-order difference approximation operator are used to obtain the geometric intensity map, as follows:

[0038] ,

[0039] ,

[0040] ,

[0041] in, Indicates spatial location Geometric intensity map at the location, operator and These represent second-order transformation operations performed along the horizontal and vertical directions, respectively. Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The eigenvector at that location;

[0042] The structure-geometric consistency index is defined as follows:

[0043] ,

[0044] in, Spatial location The structure-geometric consistency index at the location, A positive constant set to prevent the denominator from being zero;

[0045] Will Perform a linear mapping to convert it into a normalized structure-geometric consistency index on the interval [0,1]. :

[0046] .

[0047] Preferably, the acquisition of reflective features for characterizing the bright interference region in the colonoscopy image is expressed as follows:

[0048] ,

[0049] in, This is the reflective feature extraction function. Indicates spatial location The probability value that the location belongs to the reflective interference area;

[0050] The reflective feature extraction function can be implemented using any of the following methods:

[0051] Rule-based estimation based on brightness and saturation;

[0052] A reflection detection subnetwork is constructed based on a convolutional neural network, and the reflection probability map is automatically predicted through training.

[0053] Highlight region detection based on local statistical features or threshold segmentation.

[0054] Preferably, a gating weight is generated based on the structure-geometric consistency index and the reflective features, expressed as follows:

[0055] ,

[0056] in, Spatial location Gating weights at the location, and These are the adjustment coefficients for the consistency term and the reflectivity term, respectively. It is a non-linear mapping function.

[0057] Preferably, the structural features are weighted and optimized based on the gating weights to obtain optimized structural features, which are then fused with the original features to be optimized, including:

[0058] Based on the gating weight The structural features are weighted pixel by pixel, as follows:

[0059] ,

[0060] in, For structural features in spatial location First Feature values ​​on each channel express Optimized eigenvalues;

[0061] The optimized structural features are then fused with the original features to be optimized to obtain the fused optimized features. ;

[0062] The This can be achieved in the following way:

[0063] ,

[0064] or

[0065] ,

[0066] in, This represents the optimized structural features. This represents a feature concatenation operation along the channel dimension. For channel-aligned linear transformation operators, These are scalar weighting coefficients.

[0067] The present invention also provides a feature optimization apparatus for colonoscopy image segmentation, used to implement the above-mentioned feature optimization method for colonoscopy image segmentation, the apparatus comprising:

[0068] The original feature acquisition module is used to acquire the features to be optimized from colonoscopy images;

[0069] The structural feature extraction module is used to extract structural features from the features to be optimized, which are used to characterize tissue boundaries, texture changes or local intensity changes in the image.

[0070] The geometric feature extraction module is used to extract geometric features from the features to be optimized, which are used to characterize tissue morphology, surface changes or regional change trends in the image.

[0071] The first calculation module is used to calculate the structure-geometric consistency index based on the structural features and geometric features;

[0072] The second calculation module is used to obtain reflective features that characterize the bright interference areas in colonoscopy images.

[0073] The third calculation module is used to generate gating weights based on the structure-geometric consistency index and the reflective features;

[0074] The optimization and fusion module is used to perform weighted optimization on the structural features based on the gating weights to obtain optimized structural features, and then fuse them with the original features to be optimized to form a comprehensive feature representation, which is used in the training or inference process of the colonoscopy image segmentation model.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] (1) This invention introduces joint modeling of structural features and geometric features at the feature level, and uses the structure-geometric consistency index to characterize the intrinsic difference between the real tissue boundary and the background area, which helps to distinguish the real boundary from the pseudo-structural response caused by noise or illumination changes.

[0077] (2) The present invention explicitly constructs reflective features for characterizing specular reflection and high-brightness interference regions, and combines reflective features with structure-geometric consistency index during the gating weight generation process. By suppressing the structural response of high reflective regions, the influence of reflective pseudo-boundaries on the segmentation results is effectively reduced.

[0078] (3) The present invention optimizes the structural features element by element by gating weights, so that the optimized structural features are enhanced in areas with high structural-geometric consistency and weak reflective interference, and suppressed in areas with low consistency or high reflective probability, thereby improving the expression quality of target boundaries in colonoscopy images.

[0079] (4) The feature optimization method of the present invention takes feature tensors as input and output, does not depend on specific segmentation network structure, loss function form or training strategy, and can be embedded as an independent module in various colonoscopy image segmentation models, with good versatility and engineering portability.

[0080] (5) By comprehensively analyzing and optimizing structural changes, geometric changes and reflective interference at the feature level, the present invention can effectively reduce false boundaries, false positives and boundary fragmentation in complex colonoscopy imaging scenarios, significantly improve the accuracy and stability of segmentation results, and thus help improve the overall performance of the colonoscopy-assisted diagnostic system. Attached Figure Description

[0081] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below. Obviously, the description in the drawings is 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.

[0082] Figure 1 This is a schematic diagram of the overall process of the feature optimization method for colonoscopy image segmentation provided in an embodiment of the present invention;

[0083] Figure 2 This is a schematic diagram illustrating the generation process of structural features, geometric features, consistency indicators, and reflective features in an embodiment of the present invention.

[0084] Figure 3 This is a schematic diagram illustrating the process of feature gating optimization and output feature fusion in an embodiment of the present invention;

[0085] Figure 4 This is a schematic diagram illustrating the experimental results of the segmentation task on the Kvasir-SEG public dataset using the method of the present invention in this embodiment of the invention;

[0086] Figure 5 This is a schematic diagram illustrating the experimental results of a segmentation task on the CVC-ClinicDB public dataset using the method of this invention in an embodiment of the invention. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0088] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0089] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0090] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0091] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0092] It should be emphasized here that the step markers mentioned below are not a limitation on the order of the steps, but should be understood as meaning that the steps can be executed in the order mentioned in the embodiments, or in a different order than in the embodiments, or several steps can be executed simultaneously.

[0093] This invention provides a feature optimization method for colonoscopy image segmentation, applicable to the automatic segmentation and auxiliary analysis of lesion regions, tissue structures, or target regions in colonoscopy images. This method can be embedded as an independent module into existing colonoscopy image segmentation models to optimize features generated during the model's encoding or intermediate stages, without depending on the specific network structure.

[0094] See Figure 1 The feature optimization method for colonoscopy image segmentation provided by this invention includes the following steps:

[0095] Acquire colonoscopy images or intermediate features generated during the coding stage of existing colonoscopy image segmentation models, and use them as features to be optimized;

[0096] Structural features are extracted from the acquired features to be optimized to characterize tissue boundaries, texture changes, or local intensity changes in the image.

[0097] Geometric features are extracted from the acquired features to be optimized to characterize tissue morphology, surface changes, or regional change trends in the image.

[0098] Calculate the structure-geometric consistency index based on the correlation or consistency of change between the extracted structural features and geometric features at their corresponding spatial locations.

[0099] Acquire reflective features to characterize specular reflections or bright interference areas in colonoscopy images;

[0100] Based on the above structure-geometric consistency index and reflectivity features, a gating weight is generated, and the structural features are weighted and optimized based on the gating weight to obtain the optimized structural features;

[0101] The optimized structural features are fused with the original features to be optimized to form a comprehensive feature representation containing more stable boundary information, which is used in the training or inference process of the colonoscopy image segmentation model to obtain the segmentation result.

[0102] In a preferred embodiment, the present invention uniformly represents the intermediate features output during the encoding stage of colonoscopy images or segmentation models as a multi-channel feature tensor:

[0103] ,

[0104] in, Represents the characteristic tensor; This indicates the size of the feature map in the vertical direction; This indicates the dimension of the feature map in the horizontal direction; This indicates the number of channels in the feature map.

[0105] Feature tensor in spatial location First The eigenvalues ​​of each channel can be expressed as:

[0106] .

[0107] To enhance numerical stability, the feature tensor can be normalized in some implementations. For example, the mean and standard deviation of each channel can be normalized to obtain a normalized feature tensor, which can be expressed as:

[0108] ,

[0109] in, Indicates the first The mean value of the feature value corresponding to each channel can be statistically analyzed in the current feature map or a small batch of data. Indicates the first The standard deviation of the eigenvalues ​​corresponding to each channel; The preferred range of values ​​for a positive constant set to prevent the denominator from being zero is as follows: .

[0110] After the above preprocessing, the normalized feature tensor is obtained. It can be used as a specific implementation of the feature to be optimized and input into subsequent steps for extracting structural, geometric, and reflective features; unless otherwise specified, the feature tensors mentioned in the following steps... Both refer to the features to be optimized, which can be either the original feature tensor or the feature tensor after normalization. .

[0111] See Figure 2 In one embodiment of the present invention, a structural feature extraction unit extracts structure-related information from the feature tensor to characterize local texture changes and potential boundary responses. The structural features can be represented as:

[0112] ,

[0113] in, Indicates multi-channel structural features, The structural feature extraction function can be represented in this invention using one or more of the following methods, including but not limited to: local feature extraction based on convolution operations, local intensity change detection based on gradient or difference operators, edge enhancement based on learnable filters, etc. The specific design is not limited to a fixed structure, as long as it can highlight the local structural change information. The number of channels for structural features, satisfying .

[0114] In a preferred embodiment, local variations can be enhanced using a first-order gradient operator. Taking the typical Sobel operator as an example, its convolution kernels in the horizontal and vertical directions are as follows:

[0115] ,

[0116] Structural feature extraction is performed on the feature tensor to be optimized, in order to characterize the local structural changes of the features in the spatial dimension. Let... Represents the spatial location of the feature tensor to be optimized. The feature vector at that location should be noted; it should be noted that the feature tensor to be optimized can be the original feature tensor. It can also be a normalized feature tensor. .

[0117] By performing convolution operations on the feature vectors using a predefined structural convolution operator, the changes in structural features in two orthogonal directions can be obtained, defined as follows:

[0118] ,

[0119] ,

[0120] in, Represents the characteristic tensor The structural features obtained after structural feature extraction are mapped to spatial locations. The response value at the location; and These represent the structural convolution kernels along the horizontal and vertical directions, respectively; symbols Represents the convolution operation. Operator and These represent the first-order transformation operations performed on the structural feature mapping along the horizontal and vertical directions, respectively, used to characterize the changing trend of the structural features in the corresponding directions.

[0121] Furthermore, in order to obtain an overall characterization of the structural change magnitude, the present invention defines a structural strength diagram as follows:

[0122] ,

[0123] in, Used to reflect spatial location The strength of local structural changes is indicated by its numerical value; a larger value suggests that the location is more likely to correspond to a potential boundary or a texture transition region with significant structural changes. In other embodiments, the structural feature extraction function... It can also be implemented using multi-layer convolution, edge enhancement filters, or learnable feature extraction modules; this invention does not limit this to any particular approach.

[0124] In one embodiment of the present invention, geometric features are used to describe the morphological change trend of the target region on a larger spatial scale, such as gentle undulations, bulges, or depressions. This feature can compensate for the instability that may arise from relying solely on local gradient information. Geometric features can be defined as:

[0125] ,

[0126] in, Represents multi-channel geometric features, The geometric feature extraction function includes, but is not limited to, multi-scale convolution, dilated convolution, second-order difference filtering, and feature calculation based on local structure tensors, to capture morphological change information and surface change trends within a large receptive field. This invention does not limit the specific implementation method. The number of channels for geometric features, satisfying .

[0127] In a preferred embodiment, to characterize the curvature of a local shape, a second-order difference approximation of the curvature information can be used. The second-order differences along the horizontal and vertical directions are defined as follows:

[0128] ,

[0129] ,

[0130] Based on the above second-order difference, the geometric intensity map is defined as follows:

[0131] ,

[0132] in, Indicates spatial location The degree of geometrical change at a given location; the larger the value, the more pronounced the morphological undulations or structural transitions within that region at a local scale. Operator and These represent second-order transformation operations performed along the horizontal and vertical directions, respectively. Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The eigenvector at that location.

[0133] From an intuitive point of view, It is more inclined to indicate "whether a mutation boundary exists", while It is more inclined to indicate whether "the area is undergoing overall structural change". Both will be used together in subsequent consistency calculations to distinguish between true tissue boundaries and pseudo-boundaries caused by factors such as reflection.

[0134] In other embodiments, It can also be replaced with a more complex learnable geometric modeling module, such as a structural trend extraction network that combines convolutions at different scales; this invention does not limit this.

[0135] In some preferred embodiments, to reduce the computational complexity of multi-channel features during the consistency calculation stage and to highlight the overall trend of change in the spatial dimension, multi-channel structural features and geometric features can be compressed into single-channel intermediate response maps for subsequent consistency analysis. Specifically, for multi-channel structural features... It can be accessed through Convolution or other channel compression methods are used to weight and aggregate the responses of each channel to obtain the structured aggregate response, defined as:

[0136] ,

[0137] Similarly, for multi-channel geometric features The geometrical polymerization response can be obtained:

[0138] ,

[0139] in, Indicates the spatial location of multi-channel structural features The structural polymerization response obtained from polymerization at the site; This represents the geometric aggregation response obtained by aggregating multi-channel geometric features; and Representing structural features respectively With geometric features In spatial location First Feature values ​​on each channel; and These represent the number of channels for structural features and geometric features, respectively. Weighting coefficients. and It can be a preset constant or a learnable parameter that can be optimized during training. This invention does not limit its specific form.

[0140] It should be noted that the above-mentioned aggregation response and It serves as an intermediate representation for subsequent structural-geometric consistency calculations; it is consistent with the structural strength diagram defined earlier. With geometric intensity diagram They differ in function and purpose.

[0141] See Figure 2In one embodiment of the present invention, in order to distinguish between true tissue boundaries and false boundaries caused by reflection or noise, a consistency analysis is performed on the changing trends of the structural strength map and the geometric strength map, thereby constructing a joint discrimination criterion at the structure-morphology level. First, the structural strength map... With geometric intensity diagram Calculate its gradient in the horizontal and vertical directions respectively:

[0142] ,

[0143] ,

[0144] in and These represent discrete difference or gradient operators in the horizontal and vertical directions, respectively. For example, first-order forward difference, central difference, or gradient approximation based on the Sobel operator can be used. and It is a two-dimensional vector used to describe The direction and magnitude of local changes in structural strength and geometric strength at the location.

[0145] The gradient vectors described above describe the dominant directions of local structural changes and geometric changes, respectively.

[0146] Based on gradient direction similarity, this invention defines a structure-geometric consistency index as follows:

[0147] ,

[0148] The theoretical range of values ​​is ,when When the value is close to 1, it indicates that the direction of structural change is basically consistent with the direction of geometric change, and is more likely to correspond to the actual tissue boundary; when... When the correlation is close to 0, it indicates a weak correlation between the two, which may be texture noise or a non-significant structural region; when... When the values ​​are negative, the two change in opposite directions, which is more likely to correspond to the erroneous boundary response caused by reflection or pseudostructure.

[0149] To prevent the denominator from being zero, a positive constant is usually adopted. Any value within the range; when the gradient magnitude is small or close to zero, the above equation still applies. Ensure numerical stability.

[0150] To facilitate joint modeling with other features in subsequent gating processes, this invention further... Perform a linear mapping to transform it into a normalized structure-geometric consistency index over the interval [0,1].

[0151] .

[0152] Normalized structure-geometric consistency index It retains the high response at the true boundary and provides a unified numerical scale for the subsequent weight generation module, which is more conducive to stable modeling.

[0153] In other embodiments, the consistency index can also be calculated using correlation coefficient, covariance measure, angle difference or other similarity function. This invention does not limit the specific form of consistency calculation, as long as it can reflect the consistency of changes between structural features and geometric features, it can be used.

[0154] See Figure 2 During colonoscopy imaging, the light source at the lens tip directly illuminates the moist mucosal surface, often producing locally bright or even saturated specular reflective areas. These areas may resemble real tissue boundaries in intensity and gradient characteristics, but they do not correspond to actual structural changes. If not distinguished, they can easily lead to false boundaries and false positive areas in the segmentation results.

[0155] Therefore, this invention constructs reflective features in a feature space to characterize high-brightness interference areas in colonoscopy images caused by factors such as specular reflection, mucus coverage, or uneven illumination. The reflective features are represented in a spatially location-dependent form and are defined as follows:

[0156] ,

[0157] in, This is the reflective feature extraction function. Represents the spatial location of the feature tensor to be optimized. The eigenvector at that location; Indicates spatial location The probability value of a location belonging to a reflective interference area. The larger the value, the more likely the location is to correspond to a reflective or bright interference area.

[0158] Reflection feature extraction function The specific implementation methods may include, but are not limited to, the following forms:

[0159] (1) Rule-based estimation methods based on brightness and saturation, such as assigning a higher reflectivity probability to areas with high brightness and low saturation in RGB images or feature maps;

[0160] (2) A reflective detection subnetwork based on a convolutional neural network automatically learns the discriminative features of reflective areas through training and outputs the corresponding reflective probability map;

[0161] (3) A bright area detection method based on local statistical features or threshold segmentation is used to identify potential reflective interference areas.

[0162] The above reflective characteristics As an important input for subsequent gating weight generation, it is used to adaptively suppress or enhance structural information at the feature level, thereby reducing the impact of reflective interference on the segmentation results.

[0163] In a preferred embodiment, considering that reflective areas typically exhibit imaging characteristics of "high brightness and low saturation," the brightness component and saturation component can be combined for modeling. Represents the luminance component. If we denote the saturation component, then the reflective response can be expressed as:

[0164] ,

[0165] in Indicates spatial location Reflective response at the location, and These are non-negative weighting coefficients used to control the contribution of brightness and saturation to reflectance estimation. This indicates that the input quantity is normalized so that its value range falls within the range [0,1].

[0166] Furthermore, to give the reflective features probabilistic meaning, this invention transforms the above response into a reflective probability over the interval [0,1] using a nonlinear mapping function:

[0167] ,

[0168] in, The Sigmoid function is defined as follows:

[0169] .

[0170] thus, The closer the value is to 1, the more likely the current location is to be in a reflective interference area; conversely, the closer the value is to 0, the lower the likelihood of reflection.

[0171] In other embodiments, Alternatively, convolutional neural networks or shallow classification models can be used to learn and model the feature maps to obtain a reflectivity probability map. This invention does not limit the specific implementation method.

[0172] See Figure 3 After obtaining the structure-geometric consistency index and reflectivity probability features, this invention further designs a gated weight generation module to adaptively adjust the contribution of structural features to the final segmentation result at the feature level.

[0173] Gating weights are used to characterize whether the structural information at the current location is worth preserving or enhancing. The basic idea is that when the structural response is consistent with the geometry and the probability of reflection is low, the structure should be enhanced; conversely, if the area is more likely to be caused by reflection or noise, its effect should be suppressed.

[0174] To achieve the above objectives, the present invention constructs the following gating weights:

[0175] ,

[0176] in, and These are the adjustment coefficients for the consistency term and the reflectivity term, respectively, used to control their relative importance in the weight generation process. It is a nonlinear mapping function, preferably the sigmoid function or other function that maps real numbers to... A monotonically increasing function over an interval.

[0177] From the above definition, we can obtain:

[0178] .

[0179] when Larger and When the value is small, it indicates that the structure at that location is reliable and the reflective interference is weak. When the value is close to 1, structural features are preserved or enhanced; when Smaller or A larger value indicates structural instability or potential susceptibility to reflection. When both are close to 0, structural features are explicitly suppressed; when both are at moderate levels, By taking the middle value, flexible adjustment is achieved instead of binary clipping.

[0180] In practical applications, parameters and The optimal value range can be set based on the statistical results of the training set or experience. This allows for flexible adjustment of the consistency index and reflection interference intensity under different colonoscopy imaging conditions. For example, when reflection interference is significant, a larger value can be selected. A larger value can be selected to enhance the reflection suppression effect; when morphological change information is more important, a larger value can be selected. Values ​​are used to enhance the impact of structure-geometric consistency on gating.

[0181] It is worth noting that the gating weights only adjust the structural information at the feature level and do not change the overall architecture of the network. Therefore, they can be flexibly embedded into different segmentation models as pluggable modules.

[0182] See Figure 3 After obtaining the gating weights Subsequently, the present invention performs pixel-by-pixel weighting on structural features, thereby achieving adaptive suppression and enhancement of structural information.

[0183] Structural features in spatial location The corresponding number Each channel value can be denoted as The present invention performs weighted optimization of structural features in the following manner:

[0184] ,

[0185] in, This represents the optimized structural feature tensor.

[0186] It can be seen that the present invention adopts a weighting method of "pixel-by-pixel, cross-channel consistency", that is, the same spatial position All channels share the same weight This ensures consistent semantic control of each pixel location without disrupting its internal channel structure.

[0187] This operation can be written in tensor form:

[0188] ,

[0189] Among them, symbols This indicates element-wise multiplication, which can be automatically matched in terms of dimensions through a broadcast mechanism at the implementation level.

[0190] Through the above weighted optimization process, the present invention achieves the following effects:

[0191] (1) When the structural-geometric consistency is high at a certain location ( Larger) and lower reflectivity ( When the gating weight is relatively small, Approximately 1, thus preserving the structural features of that location, which helps to highlight the true organizational boundary;

[0192] (2) When the structural-geometric consistency at a certain location is low ( (smaller) or higher reflectivity ( When the gating weight is relatively large, Approaching 0 suppresses the structural response at that location, reduces the impact of reflective pseudo-boundaries on the segmentation results, and helps avoid breakage or jaggedness at the segmentation boundaries.

[0193] Therefore, this invention effectively improves the reliability of structural information through gating adjustment within the feature space.

[0194] See Figure 3 After the structural features are weighted and optimized, the present invention fuses the optimized structural features with the original feature tensor to form a comprehensive feature representation containing more stable boundary information.

[0195] The fusion method can be represented as:

[0196] ,

[0197] or

[0198] ,

[0199] in, This represents a feature concatenation operation along the channel dimension. This indicates the optimized features after fusion. Can be a channel-aligned linear transformation operator, such as convolution, This is a scalar weighting coefficient used to adjust the injection intensity of structural residuals.

[0200] In one implementation, The input is fed into the decoding stage or subsequent prediction module of the segmentation model to generate the final segmentation result. This process does not change the backbone of the original network structure, but only inserts optimization modules in the intermediate feature layers, thus exhibiting good compatibility and portability.

[0201] During the training phase, the method of this invention can be jointly optimized with common segmentation loss functions; during the inference phase, this method can also be enabled independently to directly enhance existing models. This makes this invention applicable not only to new model design but also to upgrading the functionality of existing segmentation systems.

[0202] The feature optimization method of the present invention can be embedded in different types of colonoscopy image segmentation models in the form of modules, including but not limited to convolutional neural network models, attention-based segmentation models, encoder-decoder structure models, etc.

[0203] In one typical embodiment, the feature optimization module can be applied to intermediate layers of the encoding stage of the segmentation model or to skip-connection features between encoding and decoding, in order to improve the reliability of boundary region features while maintaining multi-scale information. In another embodiment, the feature optimization module can also be applied to high-resolution features in the decoding stage to suppress the reconstruction of reflective false boundaries. Regardless of the specific embedding location, this invention optimizes features in a targeted manner by jointly analyzing structural features, geometric features, and reflective features, and using the structure-geometric consistency index and reflective features to generate gating weights, thereby improving the robustness of the colonoscopy image segmentation model to reflective interference and weak boundary regions at the feature level.

[0204] The proposed feature optimization method for colonoscopy image segmentation is validated on the Kvasir-SEG and CVC-ClinicDB datasets. After obtaining comprehensive features based on the method of this invention, the results are input into the colonoscopy image segmentation model, as shown below. Figure 4 and Figure 5 As shown. In Figure 4 In the Kvasir-SEG dataset shown, it can be observed that in the presence of obvious specular reflections and highlighted areas, after applying the feature optimization method proposed in this invention, structure-geometric consistency and reflectivity probability are jointly used to constrain the weights of structural features. This effectively suppresses the false boundaries of reflective areas, while the true tissue contours are preserved more clearly. The comparison results in the figure show that the optimized segmentation results are more stable in terms of boundary coherence and target region integrity, helping to reduce misjudgments caused by changes in illumination.

[0205] Similarly, in Figure 5 In the CVC-ClinicDB dataset shown, the method of this invention consistently demonstrates improved segmentation across polyp samples with varying acquisition conditions, morphologies, and scales. Even in low-contrast, blurred-edge scenarios, the model can more accurately identify the true boundary range by introducing a structure-geometric consistency criterion, while avoiding over-segmentation of non-target regions. This indicates that the invention is not only effective on specific samples but also universally adaptable to common endoscopic scenarios involving lighting variations, glare interference, and weak boundary issues, thereby improving the stability and reliability of colonoscopy image segmentation results while maintaining a simple model structure.

[0206] Based on the above-described inventive concept, the present invention also provides a feature optimization apparatus for colonoscopy image segmentation, used to implement the feature optimization method for colonoscopy image segmentation described in the above embodiments. The apparatus includes:

[0207] The original feature acquisition module is used to acquire the features to be optimized from colonoscopy images;

[0208] The structural feature extraction module is used to extract structural features from the features to be optimized, which are used to characterize tissue boundaries, texture changes or local intensity changes in the image.

[0209] The geometric feature extraction module is used to extract geometric features from the features to be optimized, which are used to characterize tissue morphology, surface changes or regional change trends in the image.

[0210] The first calculation module is used to calculate the structure-geometric consistency index based on the structural features and geometric features;

[0211] The second calculation module is used to obtain reflective features that characterize the bright interference areas in colonoscopy images.

[0212] The third calculation module is used to generate gating weights based on the structure-geometric consistency index and the reflective features;

[0213] The optimization and fusion module is used to perform weighted optimization on the structural features based on the gating weights to obtain optimized structural features, and then fuse them with the original features to be optimized for use in the training or inference process of the colonoscopy image segmentation model.

[0214] It is worth noting that this device embodiment corresponds to the above method embodiment. The implementation methods of the above method embodiments are all applicable to this device embodiment and can achieve the same or similar technical effects, so they will not be described in detail here.

[0215] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0216] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0217] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0218] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0219] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A feature optimization method for colonoscopy image segmentation, characterized in that, include: Obtain the features to be optimized from colonoscopy images; Structural features are extracted from the features to be optimized to characterize tissue boundaries, texture changes, or local intensity changes in the image. Geometric features are extracted from the features to be optimized to characterize tissue morphology, surface changes, or regional change trends in the image. Calculate the structural-geometric consistency index based on the structural and geometric features; Acquire reflective features to characterize bright interference areas in colonoscopy images; Gating weights are generated based on the structure-geometry consistency index and the reflective features; The structural features are weighted and optimized based on the gating weights to obtain optimized structural features, which are then fused with the original features to be optimized to form a comprehensive feature representation, which is used in the training or inference process of the colonoscopy image segmentation model.

2. The feature optimization method for colonoscopy image segmentation according to claim 1, characterized in that, The features to be optimized for acquiring colonoscopy images include: The intermediate features output from the encoding stage of colonoscopy images or segmentation models are used as the features to be optimized, represented as follows: ， in, Indicates the features to be optimized. This represents the size of the feature map in the vertical direction. This represents the dimension of the feature map in the horizontal direction. This indicates the number of channels in the feature map.

3. The feature optimization method for colonoscopy image segmentation according to claim 2, characterized in that, The features to be optimized for acquiring colonoscopy images also include: Normalize the mean and standard deviation of each channel of the feature to be optimized to obtain the normalized feature tensor. The normalized feature tensor As a feature to be optimized.

4. The feature optimization method for colonoscopy image segmentation according to claim 2, characterized in that, Structural features are extracted from the features to be optimized, and are represented as follows: ， in, Indicates multi-channel structural features, This represents the structural feature extraction function. The number of channels for structural features, satisfying ; The structural feature extraction function can be implemented using any of the following methods: Local feature extraction based on convolution operations, local intensity change detection based on gradient or difference operators, and edge enhancement based on learnable filters.

5. The feature optimization method for colonoscopy image segmentation according to claim 4, characterized in that, Geometric features are extracted from the features to be optimized, and are represented as follows: ， in, Represents multi-channel geometric features, Represents the geometric feature extraction function. The number of channels for geometric features, satisfying ; The geometric feature extraction function can be implemented using any of the following methods: Multi-scale convolution, dilated convolution, second-order difference filtering, and feature calculation methods for local structure tensors.

6. The feature optimization method for colonoscopy image segmentation according to claim 5, characterized in that, The structure-geometric consistency index is calculated based on the structural and geometric features, including: By enhancing the local structural changes of the structural features using a first-order gradient operator, a structural strength map is obtained, as follows: ， ， ， in, To reflect spatial location Structural strength diagram showing local structural changes. This represents the mapping of structural features in spatial location. The response value at that location, Indicates the spatial location of the feature to be optimized. The feature vector at that location, and These represent the structural convolution kernels along the horizontal and vertical directions, respectively, with symbols... Represents convolution operation, operator and These represent first-order transformation operations performed along the horizontal and vertical directions, respectively. The local structural changes of the geometric features described by the second-order difference approximation operator are used to obtain the geometric intensity map, as follows: ， ， ， in, Indicates spatial location Geometric intensity map at the location, operator and These represent second-order transformation operations performed along the horizontal and vertical directions, respectively. Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The feature vector at that location, Indicates the spatial location of the feature to be optimized. The eigenvector at that location; The structure-geometric consistency index is defined as follows: ， in, Spatial location The structure-geometric consistency index at the location, A positive constant set to prevent the denominator from being zero; Will Perform a linear mapping to convert it into a normalized structure-geometric consistency index on the interval [0,1]. : 。 7. The feature optimization method for colonoscopy image segmentation according to claim 6, characterized in that, The acquisition of reflective features used to characterize the bright interference areas in colonoscopy images is expressed as follows: ， in, This is the reflective feature extraction function. Indicates spatial location The probability value that the location belongs to the reflective interference area; The reflective feature extraction function can be implemented using any of the following methods: Rule-based estimation based on brightness and saturation; A reflection detection subnetwork is constructed based on a convolutional neural network, and the reflection probability map is automatically predicted through training. Highlight region detection based on local statistical features or threshold segmentation.

8. The feature optimization method for colonoscopy image segmentation according to claim 7, characterized in that, The gating weights are generated based on the structure-geometric consistency index and the reflective features, and are expressed as follows: ， in, Spatial location Gating weights at the location, and These are the adjustment coefficients for the consistency term and the reflectivity term, respectively. It is a non-linear mapping function.

9. The feature optimization method for colonoscopy image segmentation according to claim 8, characterized in that, The structural features are weighted and optimized based on the gating weights to obtain optimized structural features, which are then fused with the original features to be optimized, including: Based on the gating weight The structural features are weighted pixel by pixel, as follows: ， in, For structural features in spatial location First Feature values ​​on each channel express Optimized eigenvalues; The optimized structural features are then fused with the original features to be optimized to obtain the fused optimized features. ; The This can be achieved in the following way: ， or ， in, This represents the optimized structural features. This represents a feature concatenation operation along the channel dimension. For channel-aligned linear transformation operators, These are scalar weighting coefficients.

10. A feature optimization device for colonoscopy image segmentation, characterized in that, The apparatus for implementing the feature optimization method for colonoscopy image segmentation as described in claim 1 includes: The original feature acquisition module is used to acquire the features to be optimized from colonoscopy images; The structural feature extraction module is used to extract structural features from the features to be optimized, which are used to characterize tissue boundaries, texture changes or local intensity changes in the image. The geometric feature extraction module is used to extract geometric features from the features to be optimized, which are used to characterize tissue morphology, surface changes or regional change trends in the image. The first calculation module is used to calculate the structure-geometric consistency index based on the structural features and geometric features; The second calculation module is used to obtain reflective features that characterize the bright interference areas in colonoscopy images. The third calculation module is used to generate gating weights based on the structure-geometric consistency index and the reflective features; The optimization and fusion module is used to perform weighted optimization on the structural features based on the gating weights to obtain optimized structural features, and then fuse them with the original features to be optimized to form a comprehensive feature representation, which is used in the training or inference process of the colonoscopy image segmentation model.

Images