Cervical cancer detection system, method, device and medium based on hyperspectral imaging

Abstract

The present invention belongs to the field of image processing technology. In order to solve the problem of inaccurate detection in existing hyperspectral detection, a cervical cancer detection system, method, equipment and medium based on hyperspectral imaging are proposed. The long-range dependence of hyperspectral cervical images in the spatial dimension is captured through spatial state modeling, so as to capture the contextual features of the lesion structure and improve the modeling ability of the integrity of the lesion structure. In addition, the spectral band response of the lesion area may span multiple adjacent bands. GCN mapping can model the nonlinear connection between the bands and enhance the perception of weak spectral variations. The importance of spatial and spectral features to anomaly detection varies in different samples. The gated fusion mechanism is used to dynamically balance the spatial spectrum contribution, realize automatic adjustment in different scenarios, improve robustness, and reduce overfitting of specific feature channels.

Description

Technical Field

[0001] The present invention belongs to the technical field related to image processing, and in particular relates to a cervical cancer detection system, method, equipment and medium based on hyperspectral imaging. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Traditional cervical diagnostic and treatment equipment often uses single-band white light imaging, which is unable to obtain rich spectral information about lesions, making it difficult to identify subtle abnormalities in the early stages. Hyperspectral imaging technology is based on image data from multiple narrow bands. It combines imaging and spectroscopy to detect the two-dimensional geometric space and one-dimensional spectral information of the target, obtaining continuous, narrow-band image data with high spectral resolution. Existing research has attempted to use hyperspectral imaging, but it is limited by single-point scanning or uneven lighting, resulting in low acquisition efficiency and the introduction of artifacts.

[0004] Most existing hyperspectral anomaly detection algorithms are based on PCA, ICA or 3D-CNN. Existing methods find it difficult to simultaneously capture spatial fine-grainedness and cross-band non-Euclidean relationships, and are insufficient in modeling complex noise and cross-band correlations, resulting in a high false detection rate. Summary of the Invention

[0005] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a cervical cancer detection system, method, device and medium based on hyperspectral imaging, which can amplify the error response of the abnormal area while maintaining the accurate reconstruction of the background area, thereby improving the significance of the lesion.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a cervical cancer detection system based on hyperspectral imaging, comprising:

[0008] an acquisition unit configured to acquire a hyperspectral cervical image;

[0009] The first extraction unit is configured to capture the long-range dependency of the hyperspectral cervical image in the spatial dimension using the visual state space according to the acquired hyperspectral cervical image, and obtain a high-order spatial spectral feature;

[0010] The second extraction unit is configured to map the hyperspectral cervical image using a graph convolutional network to capture cross-band relationships and obtain spectral features;

[0011] a dynamic fusion unit configured to dynamically fuse the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image;

[0012] The detection unit is configured to realize auxiliary detection of cervical cancer based on the reconstructed background of the hyperspectral cervical image.

[0013] In a second aspect, the present invention provides a method for detecting cervical cancer based on hyperspectral imaging, comprising:

[0014] Acquire hyperspectral cervical images;

[0015] According to the acquired hyperspectral cervical image, the visual state space is used to capture the long-range dependence of the hyperspectral cervical image in the spatial dimension and obtain the high-order spatial spectral features;

[0016] Graph convolutional networks are used to map hyperspectral cervical images and capture cross-band relationships to obtain spectral features.

[0017] Dynamically fusing the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image;

[0018] Based on the reconstructed background of hyperspectral cervical images, auxiliary detection of cervical cancer is achieved.

[0019] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed by the processor, wherein when the computer instructions are executed by the processor, the following steps are performed:

[0020] Acquire hyperspectral cervical images;

[0021] According to the acquired hyperspectral cervical image, the visual state space is used to capture the long-range dependence of the hyperspectral cervical image in the spatial dimension and obtain the high-order spatial spectral features;

[0022] Graph convolutional networks are used to map hyperspectral cervical images and capture cross-band relationships to obtain spectral features.

[0023] Dynamically fusing the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image;

[0024] Based on the reconstructed background of hyperspectral cervical images, auxiliary detection of cervical cancer is achieved.

[0025] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the following steps are performed:

[0026] Acquire hyperspectral cervical images;

[0027] According to the acquired hyperspectral cervical image, the visual state space is used to capture the long-range dependence of the hyperspectral cervical image in the spatial dimension and obtain the high-order spatial spectral features;

[0028] Graph convolutional networks are used to map hyperspectral cervical images and capture cross-band relationships to obtain spectral features.

[0029] Dynamically fusing the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image;

[0030] Based on the reconstructed background of hyperspectral cervical images, auxiliary detection of cervical cancer is achieved.

[0031] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program performs the following steps:

[0032] Acquire hyperspectral cervical images;

[0033] According to the acquired hyperspectral cervical image, the visual state space is used to capture the long-range dependence of the hyperspectral cervical image in the spatial dimension and obtain the high-order spatial spectral features;

[0034] Graph convolutional networks are used to map hyperspectral cervical images and capture cross-band relationships to obtain spectral features.

[0035] Dynamically fusing the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image;

[0036] Based on the reconstructed background of hyperspectral cervical images, auxiliary detection of cervical cancer is achieved.

[0037] One or more of the above technical solutions have the following beneficial effects:

[0038] In the present invention, the long-range dependence of hyperspectral cervical images in the spatial dimension is captured through spatial state modeling, the contextual features of the lesion structure are captured, and the modeling capability of the integrity of the lesion structure is improved, while maintaining linear computational complexity, which is conducive to lightweight deployment; in addition, the spectral band response of the lesion area may span multiple adjacent bands, and GCN mapping can model the nonlinear relationship between bands, enhance the perception of weak spectral variations, and help graded judgment; the importance of spatial and spectral features to anomaly detection in different samples is different, and the gated fusion mechanism is used to dynamically balance the spatial spectrum contribution, realize automatic adjustment in different scenarios, improve robustness, and reduce overfitting of specific feature channels.

[0039] The solution of the present invention can amplify the error response of the abnormal area while maintaining the accuracy of background area reconstruction, thereby improving the significance of the lesion.

[0040] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0042] Figure 1 Schematic diagram of uniform illumination of the MMF array in Example 1 of the present invention;

[0043] Figure 2 This is a schematic diagram of a cervical cancer detection process based on hyperspectral imaging in Example 1 of the present invention;

[0044] Figure 3 This is a diagram of a visual state branch network structure in the first embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of anomaly detection in Example 1 of the present invention;

[0046] Figure 5 This is a schematic diagram of the gating mechanism fusion in Example 1 of the present invention. DETAILED DESCRIPTION

[0047] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0048] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0049] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0050] Example 1

[0051] This embodiment discloses a cervical cancer detection system based on hyperspectral imaging, comprising:

[0052] an acquisition unit configured to acquire a hyperspectral cervical image;

[0053] The first extraction unit is configured to capture the long-range dependency of the hyperspectral cervical image in the spatial dimension using the visual state space according to the acquired hyperspectral cervical image, and obtain a high-order spatial spectral feature;

[0054] The second extraction unit is configured to map the hyperspectral cervical image using a graph convolutional network to capture cross-band relationships and obtain spectral features;

[0055] a dynamic fusion unit configured to dynamically fuse the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image;

[0056] The detection unit is configured to realize auxiliary detection of cervical cancer based on the reconstructed background of the hyperspectral cervical image.

[0057] This embodiment captures the long-range dependency of hyperspectral cervical images in the spatial dimension through spatial state modeling, achieves the capture of contextual features of lesion structures, improves the modeling capability of lesion structural integrity, and maintains linear computational complexity, which is conducive to lightweight deployment. In addition, the spectral band response of the lesion area may span multiple adjacent bands. GCN mapping can model the nonlinear relationship between bands, enhance the perception of weak spectral variations, and help with graded judgment. The importance of spatial and spectral features to anomaly detection varies in different samples. The gated fusion mechanism is used to dynamically balance the spatial spectrum contribution, achieve automatic adjustment in different scenarios, improve robustness, and reduce overfitting of specific feature channels. While maintaining accurate reconstruction of the background area, the error response of the abnormal area can be amplified, thereby improving the significance of the lesion.

[0058] First, the hardware involved in this embodiment is described, including: a real-time imaging module and a hyperspectral imaging module. The real-time imaging module uses an LED ring light source and a CCA module to achieve positioning and navigation. The hyperspectral imaging module uses a xenon lamp with a wavelength of 220–2000 nm and a typical operating range of 400–900 nm. The ROI is illuminated by a 5×5 MMF array, and a hyperspectral line scan camera with a telephoto lens is used for data acquisition.

[0059] Figure 1 Schematic diagram of uniform illumination of MMF array. The optical parameters and formulas are:

[0060] The output angle of a single optical fiber is θ = arcsin(NA); the spot diameter is d_spot = 2htanθ.

[0061] When the numerical aperture NA of the optical fiber is 0.22 and the distance from the optical fiber end face to the tissue surface is h = 5mm, the irradiation spot diameter d_spot ≈ 2.26mm.

[0062] 25 optical fibers cover an area of ​​10 mm × 10 mm, with uniformity U = E_min / E_max ≥ 0.9.

[0063] Among them, E_min and E_max represent the minimum irradiance and maximum irradiance respectively, and the unit is W / m 2

[0064] In response to the current situation where the spectral characteristics of early lesions are weak, the reconstruction residual amplifies the abnormality: in hyperspectral images, the spectral response of the diseased tissue is usually slightly different from that of the surrounding normal tissue, and direct segmentation is difficult to accurately define; while maintaining the accurate reconstruction of the background area, the error response of the abnormal area can be amplified, thereby improving the significance of the lesion.

[0065] The following combination Figure 1 The cervical cancer detection system based on hyperspectral imaging proposed in this embodiment is described in detail:

[0066] In the acquisition unit, the hyperspectral cervical image is acquired and the bands are extracted. The band selection algorithm is used to select the optimal 64 bands to reduce the dimension and noise; To the selected data .

[0067] Hyperspectral cervical images are large in raw size, typically containing over 300 bands, and are complex to process. The differences between normal and cancerous cervical tissue typically center on the strong hemoglobin absorption peaks at 415±10nm and 525±10nm. Band selection prioritizes bands of strong biological relevance and removes noise bands before and after data acquisition due to insufficient spectral resolution. Then, based on the correlation of information across bands, a subset of bands with the richest information and lowest redundancy is selected for subsequent processing.

[0068] This example uses a forward selection algorithm called the continuous projection algorithm for hyperspectral band selection. The primary goal of the continuous projection algorithm is to minimize collinearity between bands. In hyperspectral data, adjacent bands are highly correlated, like two vectors pointing in nearly the same direction in vector space. The continuous projection algorithm uses a geometric projection method to continuously select bands pointing in new directions.

[0069] The spectral characteristics of cervical cancer are primarily concentrated in the hemoglobin absorption region (visible light) and the tissue scattering region (near-infrared). This medical prior knowledge can be used to guide the continuous projection algorithm, rather than forcing it to "blindly" search across the entire band.

[0070] The specific implementation can be:

[0071] Step 1: Regional division based on prior medical knowledge.

[0072] Based on the biological mechanisms of cervical cancer, the entire spectral range, such as 400-1000 nm, is divided into several key "regions of interest" (ROIs), including:

[0073] ROI 1 - Hemoglobin absorption region: approximately 500-600 nm. This region is critical for diagnosis and contains the characteristic peaks of oxyhemoglobin and deoxyhemoglobin. These wavelengths reflect tumor angiogenesis and hypoxia.

[0074] ROI 2 - Scattering Slope: Approximately 650-950 nm. The slope of the spectral curve in this region is primarily determined by microstructures such as nuclear size and cell density, and is of great diagnostic value for early precancerous lesions.

[0075] ROI 3 - Water absorption shoulder: approximately 950-1000 nm. This absorption peak near water, at 970 nm, can reflect the hydration status of tissues and serve as auxiliary information.

[0076] ROI 4 - Blue-green region: approximately 400-500 nm. This region may contain some fluorescence information and scattering information from superficial tissues.

[0077] Step 2: Weighted or forced choice within the region.

[0078] This embodiment allocates the target, for example, 64 bands, to these regions and guides the continuous projection algorithm:

[0079] This embodiment no longer randomly selects the first band, but forcibly selects the band that best represents the characteristics of the region from each ROI as the initial seed point.

[0080] Positioning method based on medical prior knowledge: For ROIs with clear biological diagnostic significance, directly locate the medically recognized characteristic wavelength as the seed point.

[0081] Application Example: In "ROI 1 - Hemoglobin Absorption Region" (approximately 500-600 nm), it is known that vascular proliferation in cervical lesions leads to changes in hemoglobin concentration, with distinct absorption peaks at approximately 542 nm and 577 nm. Therefore, the wavelength band closest to these two values ​​is forcibly selected as the initial seed point for this region. This ensures that subsequent wavelength selection consistently focuses on the most critical biomarker information for diagnosing cervical cancer.

[0082] Screening method based on the principle of maximizing information: For ROIs with relatively diffuse spectral features and no single recognized characteristic peak, a data-driven approach is used to determine the seed points.

[0083] Application Example: In "ROI 2 - Scattering Slope Region" (approximately 650-950 nm), the slope of the spectral curve primarily reflects changes in tissue microstructure, such as cell density and nuclear size. To identify the wavelength band that best reflects these changes, the information entropy or variance of each wavelength band within this ROI is calculated for a batch of training sample images. The wavelength band with the largest variance is selected as the seed point for this region, as the largest variance generally indicates that this wavelength band has the best discrimination between normal and diseased tissues and carries the richest classification information.

[0084] For example, in ROI 1, select the wavelengths around 542 nm and 77 nm.

[0085] In ROI 2, the wavelength bands around 700 nm and 900 nm were selected to define the scattering slope.

[0086] These "anchor" bands ensure that the most important biological information is not missed.

[0087] In each iteration of the continuous projection algorithm, different weights are assigned to the bands in different ROIs. For example, when calculating the length of the band vector, a weight coefficient W is multiplied.

[0088] Effective length = W * ||band vector||

[0089] Assign higher weights, such as W=1.5, to ROI 1 and ROI 2, and lower weights, such as W=0.8, to other regions. In this way, the continuous projection algorithm will prioritize finding the most informative bands in these key regions.

[0090] In each iteration of the continuous projection algorithm, different selection weights are assigned to bands in different ROIs through either static allocation driven by expert knowledge or dynamic learning driven by data. These weights are used to adjust the "effective length" of each candidate band vector during the projection step, thereby guiding the algorithm to prioritize bands in spectral regions with higher diagnostic value.

[0091] As an optional implementation, a static weighting method driven by expert knowledge involves presetting fixed weighting coefficients by experts based on the recognized medical importance of different spectral regions for cervical cancer diagnosis. Based on the medical prior knowledge provided in this embodiment, hemoglobin absorption and tissue scattering are key to distinguishing lesions.

[0092] For example, a higher weight coefficient, such as W = 1.5, is assigned to "ROI 1 (hemoglobin region)" and "ROI 2 (scattering region)." A lower weight, such as W = 0.8, is assigned to supporting regions like "ROI 3 (water absorption shoulder region)" and "ROI 4 (blue-green region)." When the continuous projection algorithm calculates the projection vector norm, this weight coefficient, W, is multiplied into the formula: effective length = W × ||band vector||. This directly increases the probability of selecting bands in highly important regions.

[0093] As an optional implementation, a data-driven adaptive weighting method specifically utilizes a small training dataset to dynamically determine weights for more objective and accurate weight assignment. Specifically, for each ROI, the average contribution of its internal bands to distinguishing healthy from diseased samples can be calculated. This contribution can be quantified using statistical metrics such as the Fisher discriminant ratio and the p-value of a t-test.

[0094] For example, the Fisher discriminant ratios (FDRs) for the bands in ROIs 1, 2, 3, and 4 are first calculated for distinguishing between two sample types. These means are then normalized to obtain adaptive weights W1, W2, W3, and W4 for each ROI. For example, if the band in ROI 2 (the scattering region) contributes most to sample differentiation, its weight W2 will be the highest. This approach allows weight assignment to automatically adapt to the inherent characteristics of the current dataset, prioritizing the bands most effective for the current detection task.

[0095] Step 3: Perform continuous projection algorithm within the region and merge.

[0096] Allocate the target quota of 64 bands to each ROI according to its importance. For example:

[0097] ROI 1 (Hemoglobin region): 20 places allocated

[0098] ROI 2 (scatter zone): 30 places allocated

[0099] ROI 3&4 (Other): 14 quotas in total

[0100] Then, the continuous projection algorithm is run independently within each ROI until the allocated quota is filled. Finally, the bands selected from all regions are merged to obtain the final 64-band subset.

[0101] In the first extraction unit, this embodiment adopts a background reconstruction module for feature extraction. The background reconstruction module contains N main modules connected in series. The value of N in the embodiment can be determined according to actual needs. In this embodiment, the value is 4. Each main module includes a visual state space module, a graph convolutional network and a gating mechanism. The gating mechanism is used to fuse the outputs of the visual state space branch and the graph convolutional network branch to reconstruct the background. The core principle is that in medicine, abnormal areas that often appear as lesions or inflammation areas can be regarded as abnormal areas during the reconstruction process because their information is larger than the background information. It is more difficult to reconstruct through reconstruction. After reconstructing the original data, the abnormality is detected by subtracting the reconstructed image from the original image.

[0102] The input of the background reconstruction module is the data after band selection , the output is the reconstructed background information .

[0103] In such Figure 2 As shown in the figure, the visual state space module and graph convolution extract spectral information as two independent branches, and then the features are dynamically fused through the gating mechanism.

[0104] The visual state-space module employs state-space modeling to capture global spatial dependencies while maintaining linear computational complexity. Each layer incorporates LayerNorm, linear projection, depthwise separable convolution, and SS2D selective scanning operations, outputting high-order spatial features through residual connections. This module captures long-range spatial dependencies in hyperspectral images, addressing the inadequacy of conventional convolution in modeling complex lesion structures.

[0105] like Figure 3 As shown, the visual state space module is used to extract spatial and spectral information. The visual state space module is used to input the hyperspectral data after band selection. The visual state space module includes two branches, the first branch and the second branch respectively include the first layer normalization, the first linear layer, the first activation function, the first SS2D and the second layer normalization in sequence, and the outputs of the first branch and the second branch are fused and processed by the second linear layer to obtain the first feature; the first feature and the input feature are added element by element to obtain the second feature; the second feature is processed by the third layer normalization, the third linear layer, the second activation function, the second SS2D and the fourth layer normalization in sequence to obtain the third feature; the feature output by the third layer normalization is processed by the fourth linear layer and the third activation function in sequence to obtain the fourth feature, the fourth feature and the third feature are fused and processed by the fifth linear layer to obtain the fifth feature, and the fifth feature and the second feature are added element by element to obtain the output feature of the visual state space module, namely the high-order spatial spectral feature F1.

[0106] The first SS2D input processing process is: W×C spatial state modeling, that is, slicing by row h, scanning the wide reference axis w and the spectral reference axis c for each fixed row index h∈[1,H], and taking out the matrix , perform SS2D-WC lateral + spectral state propagation on this plane.

[0107]

[0108] Among them, A w , A c , B is a learnable matrix or convolution kernel. Represents the "spatial state vector" at the w-th width position and channel c on the h-th row height dimension slice; the initial state , the output in the spectral dimension wc direction is: , the superscript W is the width of the hyperspectral data, the superscript C is the channel dimension of the hyperspectral data, the superscript H represents the height of the hyperspectral data, and P is the projection matrix representing the linear transformation, usually a 1×1 convolutional layer or a fully connected layer.

[0109] The same process is applied to the Height–Channel plane. The second SS2D process for the input is: H×C spatial state modeling, that is, slicing by column w, scanning the h and c axes, fixing the column index w∈[1,W], and taking out the matrix Similarly, do SS2D-HC longitudinal + spectral state propagation:

[0110]

[0111] in, 、 is a learnable matrix or convolution kernel; Represents the "spatial state vector" at the h-th height position and channel c on the w-th width dimension slice.

[0112] Get the output of the corresponding hc dimension: .

[0113] After adding the two side information element by element, they are used as the "spectral enhancement features" that are actually fed into spatial modeling:

[0114]

[0115] After the fusion features are normalized at the layer, the features are reshaped back to the original input feature dimension and the output is connected with the original feature residual. , and then do a real "plane space" state scan, which corresponds to the third SS2D module, outputting the spatial features. The output spatial features are , and The high-order spatial spectral feature F1 is obtained by element-by-element addition.

[0116] Cervical lesion areas usually show texture changes, blurred edges, and abnormal color; spatial state modeling has the ability to model global dependencies, capture the contextual characteristics of the lesion structure, and improve the ability to model the integrity of the lesion structure; at the same time, it maintains linear computational complexity, which is conducive to lightweight deployment.

[0117] In the second extraction unit, the graph convolutional network branch is used to build a graph in the spectral dimension to capture cross-band relationships. The input data of the graph convolutional network branch is , the output data is .

[0118] First, construct the hybrid gated adjacency matrix A. , based on the channel dimension C, i.e. the spectral dimension, as the node.

[0119] Physical prior adjacency matrix formula:

[0120]

[0121] It means that the smaller the wavelength difference is, the greater the edge weight is. Indicates the central wavelength value corresponding to the i-th spectral band, represents the central wavelength value corresponding to the jth spectral band; σ controls the attenuation width and can be set to the average bandwidth or learned during training.

[0122] Constructing an adaptive adjacency matrix based on Pearson correlation coefficient :

[0123]

[0124] Where x :,:,i and x :,:,j Defined as the vector of all pixel values ​​of the i-th and j-th spectral bands; cov(x:,:,i,x:,:,j) represents the covariance between two variables; and represents the standard deviation of the two variables.

[0125] The Pearson correlation coefficient of the batch pixel vectors is calculated and truncated to a non-negative value to reflect the span correlation in the actual data.

[0126] Hybrid gated adjacency matrix formula: , where β∈[0,1].

[0127] GCN regards each retained spectrum band as a graph node and implements node message passing through multi-layer graph convolution:

[0128]

[0129] in, is the adjacency matrix with self-loops added, is the degree matrix, is the node feature, are learnable weights, is the activation function.

[0130] This embodiment introduces the prior knowledge that channels with similar wavelengths are more correlated, so that the model is not affected by specific image data and has universal applicability.

[0131] This embodiment proposes a hybrid adjacency matrix construction method that integrates "physical priors" and "data adaptation" information. On the one hand, it leverages the physical continuity of the spectrum, namely wavelength proximity, to establish a stable and universal graph structure foundation, preventing the model from being excessively affected by data noise. On the other hand, by mining statistical correlations from real medical data, it captures non-local cross-band dependencies that are crucial for disease diagnosis. For example, a specific lesion may cause regular changes in reflectivity at both 540nm (the hemoglobin absorption region) and 800nm ​​(the tissue scattering region). Traditional graph structures that rely solely on wavelength proximity ignore these long-range correlations, while statistically adaptive matrices can discover and enhance these correlations through data analysis, allowing the graph network to learn more valuable features for diagnosis. Therefore, the essence of "data adaptation" is to enable the graph structure, namely the relationships between wavelengths, to self-adjust and optimize based on the specific spectral characteristics of the input image, rather than relying on a fixed, universal set of rules. This synergy between physics and data ensures that the final graph structure has both a strong skeleton of scientific principles and intelligent perception for specific tasks, making it more accurate and robust than any single method. This greatly enhances the application potential of graph convolutional networks in complex spectral medical diagnosis.

[0132] The graph convolutional network branch uses multi-layer graph convolution to model high-order relationships across bands. In the embodiment, the number of layers is 2, and the output is the enhanced spectral feature F2.

[0133] Spectral mapping (GCN) enhances cross-band discrimination capabilities: The key advantage of hyperspectral data lies in continuous band information, while the spectral band response of the lesion area may span multiple adjacent bands; GCN mapping can model nonlinear connections between bands and enhance the perception of weak spectral variations; it can more effectively model the spectral heterogeneity of CIN grades I to III, which helps with grading judgment.

[0134] In the dynamic fusion unit, e.g. Figure 5As shown in Figure 2, the gating mechanism dynamically fuses the features of the high-order spatial spectral feature F1 and the spectral feature F2, and generates them through 1×1 convolution and Sigmoid activation. The weight map of high-order spatial spectral features and spectral characteristics Perform adaptive weighting: .

[0135] The gated fusion mechanism (SSGM) dynamically balances spatial-spectral contributions: the importance of spatial and spectral features for anomaly detection varies in different samples. SSGM uses adaptive weights to fuse spatial-spectral features, enabling the model to automatically adjust its strategy in different scenarios, improving robustness and reducing overfitting of specific feature channels.

[0136] like Figure 4 As shown, the anomaly score R is defined as:

[0137]

[0138] Among them, X is the original data after band selection, For the reconstructed background data, after the abnormality score calculation, the larger the R value is, the more likely the pixel is abnormal. The lesion mask can be obtained by threshold segmentation. Figure 4 The areas with medium R values ​​have a high degree of overlap with the areas marked by doctors, and their AUC values ​​are calculated to be greater than 0.95.

[0139] Example 2

[0140] The purpose of this embodiment is to provide a cervical cancer detection method based on hyperspectral imaging, including:

[0141] Acquire hyperspectral cervical images;

[0142] According to the acquired hyperspectral cervical image, the visual state space is used to capture the long-range dependence of the hyperspectral cervical image in the spatial dimension and obtain the high-order spatial spectral features;

[0143] Graph convolutional networks are used to map hyperspectral cervical images and capture cross-band relationships to obtain spectral features.

[0144] Dynamically fusing the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image;

[0145] Based on the reconstructed background of hyperspectral cervical images, auxiliary detection of cervical cancer is achieved.

[0146] The specific operation process of each step involved in this embodiment is consistent with that in Example 1 and will not be repeated here.

[0147] In further embodiments, there is also provided:

[0148] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed by the processor, wherein when the computer instructions are executed by the processor, the method described in Example 2 is performed. For the sake of brevity, no further details are given here.

[0149] It should be understood that in this embodiment, the processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), off-the-shelf field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0150] The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0151] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in embodiment 2 is performed.

[0152] The method in Example 2 can be directly implemented and executed by a hardware processor, or by a combination of hardware and software modules within the processor. The software module can be located in a storage medium well-established in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. The storage medium is located in the memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not given here.

[0153] Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0154] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A cervical cancer detection system based on hyperspectral imaging, characterized in that: include: an acquisition unit configured to acquire a hyperspectral cervical image; The first extraction unit is configured to capture the long-range dependency of the hyperspectral cervical image in the spatial dimension using the visual state space based on the acquired hyperspectral cervical image, and obtain high-order spatial spectral features, specifically: The first SS2D module and the second SS2D module perform row slice scanning and column slice scanning respectively, and the scanning results are added element by element to obtain the spectrum enhancement feature; The spectral enhancement feature is scanned in a plane spatial state through the third SS2D module to obtain a high-order spatial spectral feature; The second extraction unit is configured to use a graph convolutional network to map the hyperspectral cervical image and capture cross-band relationships to obtain spectral features, specifically: Each retained spectral band in the hyperspectral cervical image is regarded as a graph node; the central wavelength values ​​corresponding to different spectral bands are used to construct a physical prior adjacency matrix; the correlation between pixel values ​​of different spectral bands is calculated using the Pearson correlation coefficient to construct an adaptive adjacency matrix; Based on the physical prior adjacency matrix and the adaptive adjacency matrix, a hybrid gated adjacency matrix is ​​constructed; the graph nodes and the hybrid gated adjacency matrix are input into a graph convolutional network to obtain spectral features; Physical prior adjacency matrix formula: It means that the smaller the wavelength difference is, the greater the edge weight is. Indicates the central wavelength value corresponding to the i-th spectral band, represents the central wavelength value corresponding to the jth spectral band; σ controls the attenuation width; Constructing an adaptive adjacency matrix based on Pearson correlation coefficient : Where x :,:,i and x :,:,j Defined as the vector of all pixel values ​​of the i-th and j-th spectral bands; cov(x:,:,i,x:,:,j) represents the covariance between two variables; and represents the standard deviation of the two variables; Hybrid gated adjacency matrix formula: , where β∈[0,1]; a dynamic fusion unit configured to dynamically fuse the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image; The detection unit is configured to realize auxiliary detection of cervical cancer based on the reconstructed background of the hyperspectral cervical image.

2. The cervical cancer detection system based on hyperspectral imaging according to claim 1, characterized in that: The acquisition unit also includes: using a continuous projection algorithm to select the optimal band of the acquired hyperspectral cervical image, and in each iteration of the continuous projection algorithm, assigning different weights to the bands in different regions of interest based on expert knowledge or dynamic learning.

3. The cervical cancer detection system based on hyperspectral imaging according to claim 1, characterized in that: In the detection unit, based on the reconstructed background of the hyperspectral cervical image, auxiliary detection of cervical cancer is achieved, specifically: Subtracting the hyperspectral cervical image from the reconstructed background of the hyperspectral cervical image to determine an abnormality score; According to the anomaly score, the lesion mask is obtained by threshold segmentation.

4. The cervical cancer detection system based on hyperspectral imaging according to claim 2, characterized in that: In the acquisition unit, a continuous projection algorithm is used to select the optimal band of the acquired hyperspectral cervical image, specifically: Perform regional segmentation on hyperspectral cervical images based on prior knowledge; Based on prior knowledge or the principle of maximizing information, a representative band is selected from each region of interest as the initial seed point; The continuous projection algorithm is used to select the bands in each region of interest. In each iteration of the continuous projection algorithm, different weights are assigned to the bands in different regions of interest based on expert knowledge or dynamic learning. The bands selected from all regions of interest are merged to obtain the optimal band.

5. A method for detecting cervical cancer based on hyperspectral imaging, characterized in that: include: Acquire hyperspectral cervical images; According to the acquired hyperspectral cervical image, the long-range dependence of the hyperspectral cervical image in the spatial dimension is captured using the visual state space to obtain the high-order spatial spectral features, specifically: The first SS2D module and the second SS2D module perform row slice scanning and column slice scanning respectively, and the scanning results are added element by element to obtain the spectrum enhancement feature; The spectral enhancement feature is scanned in a plane spatial state through the third SS2D module to obtain a high-order spatial spectral feature; A graph convolutional network is used to construct a map of the hyperspectral cervical image, capturing cross-band relationships to obtain spectral features, specifically: Each retained spectral band in the hyperspectral cervical image is regarded as a graph node; the central wavelength values ​​corresponding to different spectral bands are used to construct a physical prior adjacency matrix; the correlation between pixel values ​​of different spectral bands is calculated using the Pearson correlation coefficient to construct an adaptive adjacency matrix; Based on the physical prior adjacency matrix and the adaptive adjacency matrix, a hybrid gated adjacency matrix is ​​constructed; the graph nodes and the hybrid gated adjacency matrix are input into a graph convolutional network to obtain spectral features; Physical prior adjacency matrix formula: It means that the smaller the wavelength difference is, the greater the edge weight is. Indicates the central wavelength value corresponding to the i-th spectral band, represents the central wavelength value corresponding to the jth spectral band; σ controls the attenuation width; Constructing an adaptive adjacency matrix based on Pearson correlation coefficient : Where x :,:,i and x :,:,j Defined as the vector of all pixel values ​​of the i-th and j-th spectral bands; cov(x:,:,i,x:,:,j) represents the covariance between two variables; and represents the standard deviation of the two variables; Hybrid gated adjacency matrix formula: , where β∈[0,1]; Dynamically fusing the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image; Based on the reconstructed background of hyperspectral cervical images, auxiliary detection of cervical cancer is achieved.

6. The method for detecting cervical cancer based on hyperspectral imaging according to claim 5, wherein: After obtaining the hyperspectral cervical image, the method further includes: using a continuous projection algorithm to select the optimal band of the obtained hyperspectral cervical image, and in each iteration of the continuous projection algorithm, assigning different weights to the bands in different regions of interest based on expert knowledge or dynamic learning.

7. An electronic device, characterized in that: The system comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein the computer instructions are executed by the processor to complete the following steps: Acquire hyperspectral cervical images; According to the acquired hyperspectral cervical image, the long-range dependence of the hyperspectral cervical image in the spatial dimension is captured using the visual state space to obtain the high-order spatial spectral features, specifically: The first SS2D module and the second SS2D module perform row slice scanning and column slice scanning respectively, and the scanning results are added element by element to obtain the spectrum enhancement feature; The spectral enhancement feature is scanned in a plane spatial state through the third SS2D module to obtain a high-order spatial spectral feature; A graph convolutional network is used to construct a map of the hyperspectral cervical image, capturing cross-band relationships to obtain spectral features, specifically: Each retained spectral band in the hyperspectral cervical image is regarded as a graph node; the central wavelength values ​​corresponding to different spectral bands are used to construct a physical prior adjacency matrix; the correlation between pixel values ​​of different spectral bands is calculated using the Pearson correlation coefficient to construct an adaptive adjacency matrix; Based on the physical prior adjacency matrix and the adaptive adjacency matrix, a hybrid gated adjacency matrix is ​​constructed; the graph nodes and the hybrid gated adjacency matrix are input into a graph convolutional network to obtain spectral features; Physical prior adjacency matrix formula: It means that the smaller the wavelength difference is, the greater the edge weight is. Indicates the central wavelength value corresponding to the i-th spectral band, represents the central wavelength value corresponding to the jth spectral band; σ controls the attenuation width; Constructing an adaptive adjacency matrix based on Pearson correlation coefficient : Where x :,:,i and x :,:,j Defined as the vector of all pixel values ​​of the i-th and j-th spectral bands; cov(x:,:,i,x:,:,j) represents the covariance between two variables; and represents the standard deviation of the two variables; Hybrid gated adjacency matrix formula: , where β∈[0,1]; Dynamically fusing the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image; Based on the reconstructed background of hyperspectral cervical images, auxiliary detection of cervical cancer is achieved.

8. A computer-readable storage medium, characterized in that Used to store computer instructions, which, when executed by a processor, complete the following steps: Acquire hyperspectral cervical images; According to the acquired hyperspectral cervical image, the long-range dependence of the hyperspectral cervical image in the spatial dimension is captured using the visual state space to obtain the high-order spatial spectral features, specifically: The first SS2D module and the second SS2D module perform row slice scanning and column slice scanning respectively, and the scanning results are added element by element to obtain the spectrum enhancement feature; The spectral enhancement feature is scanned in a plane spatial state through the third SS2D module to obtain a high-order spatial spectral feature; A graph convolutional network is used to construct a map of the hyperspectral cervical image, capturing cross-band relationships to obtain spectral features, specifically: Each retained spectral band in the hyperspectral cervical image is regarded as a graph node; the central wavelength values ​​corresponding to different spectral bands are used to construct a physical prior adjacency matrix; the correlation between pixel values ​​of different spectral bands is calculated using the Pearson correlation coefficient to construct an adaptive adjacency matrix; Based on the physical prior adjacency matrix and the adaptive adjacency matrix, a hybrid gated adjacency matrix is ​​constructed; the graph nodes and the hybrid gated adjacency matrix are input into a graph convolutional network to obtain spectral features; Physical prior adjacency matrix formula: It means that the smaller the wavelength difference is, the greater the edge weight is. Indicates the central wavelength value corresponding to the i-th spectral band, represents the central wavelength value corresponding to the jth spectral band; σ controls the attenuation width; Constructing an adaptive adjacency matrix based on Pearson correlation coefficient : Where x :,:,i and x :,:,j Defined as the vector of all pixel values ​​of the i-th and j-th spectral bands; cov(x:,:,i,x:,:,j) represents the covariance between two variables; and represents the standard deviation of the two variables; Hybrid gated adjacency matrix formula: , where β∈[0,1]; Dynamically fusing the high-order spatial spectral features and the spectral features through a gated fusion mechanism to obtain a reconstructed background of a hyperspectral cervical image; Based on the reconstructed background of hyperspectral cervical images, auxiliary detection of cervical cancer is achieved.

Images