Medical hyperspectral image detection method, device and equipment and storage medium

Abstract

The invention provides a medical hyperspectral image detection method, device and equipment and a storage medium, and relates to the technical field of medical detection.The medical hyperspectral image detection method comprises the steps that a to-be-detected medical hyperspectral image is obtained, and the medical hyperspectral image comprises real labels of at least part of pixels; acquiring spectral width characteristics of the medical hyperspectral image; according to the spectral width characteristics, adopting a preset spectral width learning model to obtain a prediction label matrix of the medical hyperspectral image; the prediction label matrix comprises real labels of partial pixels in the medical hyperspectral image and prediction labels of other pixels; and determining a target focus area from the medical hyperspectral image based on the prediction label matrix. According to the invention, accurate and efficient focus detection and positioning are realized.

Description

Technical Field

[0001] This application relates to the field of medical testing, and more specifically, to a medical hyperspectral image detection method, apparatus, device, and storage medium. Background Technology

[0002] In the field of high-precision surgery, accurate and real-time intraoperative identification of lesion tissue from normal brain tissue is crucial to the success of the operation. Hyperspectral imaging technology captures optical information of the target region in hundreds of consecutive narrow bands, generating a three-dimensional data cube containing two spatial dimensions and one spectral dimension, which can reveal differences in tissue microstructure and biochemical composition that are imperceptible to the human eye.

[0003] However, current medical hyperspectral imaging tumor detection technology relies heavily on a large amount of precisely labeled data. This reliance is primarily limited by factors such as patient privacy, high data acquisition costs, and a scarcity of professional annotation resources, resulting in insufficient training data for models and severely restricting the performance and generalization ability of supervised algorithms. Furthermore, while the vast array of spectral bands provides rich information, it also contains a significant amount of redundant noise unrelated to classification. If key spectral features that are sensitive to tumor responses and have strong discriminative power cannot be effectively extracted, the accuracy and robustness of detection will be greatly compromised. Therefore, constructing high-precision feature representations from a limited number of labeled medical images has become a major challenge for improving lesion detection. Summary of the Invention

[0004] This invention provides a medical hyperspectral image detection method, apparatus, device, and storage medium, which can achieve accurate and efficient lesion detection and localization.

[0005] In a first aspect, embodiments of the present invention provide a medical hyperspectral image detection method, the method comprising: Acquire a medical hyperspectral image to be detected, the medical hyperspectral image including: at least some pixels of true labels; Obtain the spectral width features of the medical hyperspectral image; Based on the spectral width features, a preset spectral width learning model is used to obtain the predicted label matrix of the medical hyperspectral image; the predicted label matrix includes: the true labels of some pixels in the medical hyperspectral image, and the predicted labels of other pixels; Based on the predicted label matrix, the target lesion region is determined from the medical hyperspectral image.

[0006] Optionally, acquiring the spectral width features of the medical hyperspectral image includes: Based on the partial pixels and the other pixels in the medical hyperspectral image, the feature matrix of the medical hyperspectral image is obtained; The spectral width feature is obtained based on the feature matrix.

[0007] Optionally, obtaining the spectral width feature based on the feature matrix includes: Based on the preset feature mapping dimension and the preset first mapping matrix, the feature matrix is ​​mapped to obtain the mapping features of the medical hyperspectral image; Based on the preset feature enhancement dimension and the preset second mapping matrix, the mapping features are enhanced to obtain the enhanced features of the medical hyperspectral image; The spectral width feature is generated based on the mapping feature and the enhancement feature.

[0008] Optionally, the step of mapping the feature matrix according to a preset feature mapping dimension and a preset first mapping matrix to obtain the mapping features of the medical hyperspectral image includes: Based on the preset number of mapping feature node groups, the feature nodes of each pixel in the feature matrix are grouped to obtain multiple groups of mapping feature nodes. Based on the preset first mapping matrix, feature mapping is performed on each group of mapping feature nodes to obtain the mapping features of each group of mapping feature nodes; The mapping features of the medical hyperspectral image are obtained based on the mapping features of the multiple sets of mapping feature nodes.

[0009] Optionally, the step of enhancing the mapped features according to a preset feature enhancement dimension and a preset second mapping matrix to obtain the enhanced features of the medical hyperspectral image includes: Based on the preset number of enhanced feature node groups, the feature nodes of each pixel in the mapped feature are grouped to obtain multiple groups of enhanced feature nodes. Based on the preset second mapping matrix, feature enhancement is performed on each group of enhanced feature nodes to obtain the enhanced features of each group of enhanced feature nodes; The enhancement features of the medical hyperspectral image are obtained based on the enhancement features of the multiple sets of enhancement feature nodes.

[0010] Optionally, the step of obtaining the predicted label matrix of the medical hyperspectral image based on the spectral width features using a preset spectral width learning model includes: A pre-established regression optimization model is obtained as the spectral width learning model. The regression optimization model includes: a constraint term on the predicted label matrix, a constraint term on the regression coefficient matrix corresponding to the medical hyperspectral image, a constraint term on the reconstructed residual matrix corresponding to the spectral width feature, a constraint term on the auxiliary variable matrix, a constraint term on the spectral width feature, a constraint term on the latent space feature corresponding to the medical hyperspectral image, and a constraint term on the feature extraction matrix. Based on the spectral width characteristics, the regression optimization model is iteratively solved to obtain the feature extraction matrix, the latent space features, the auxiliary variable matrix, the reconstructed residual matrix, and the regression coefficient matrix; The predicted label matrix is ​​obtained based on the feature extraction matrix, the latent space features, the auxiliary variable matrix, the reconstructed residual matrix, and the regression coefficient matrix.

[0011] Optionally, determining the target lesion region from the medical hyperspectral image based on the predicted label matrix includes: The target lesion region is determined from the medical hyperspectral image based on the label values ​​of each pixel in the predicted label matrix, the preset labels of normal biological tissue, and the preset labels of lesion tissue.

[0012] Secondly, embodiments of the present invention also provide a medical hyperspectral image detection device, the device comprising: An acquisition module is used to acquire a medical hyperspectral image to be detected, wherein the medical hyperspectral image includes: at least some pixels of real labels; The acquisition module is also used to acquire the spectral width features of the medical hyperspectral image; The calculation module is used to obtain the predicted label matrix of the medical hyperspectral image based on the spectral width features and using a preset spectral width learning model; the predicted label matrix includes: the true labels of some pixels in the medical hyperspectral image, and the predicted labels of other pixels; The output module is used to determine the target lesion region from the medical hyperspectral image based on the predicted label matrix.

[0013] Thirdly, embodiments of the present invention also provide a medical hyperspectral image detection device, comprising: a processor, a memory, and a bus, wherein the memory stores program instructions executable by the processor, and when the medical hyperspectral image detection device is running, the processor communicates with the memory via the bus, and the processor executes the program instructions to perform the steps of the medical hyperspectral image detection method as described in any of the first aspects.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the medical hyperspectral image detection method as described in any of the first aspects.

[0015] This invention provides a medical hyperspectral image detection method, apparatus, device, and storage medium. It obtains corresponding spectral width features from medical hyperspectral images with partial pixel labels, then uses a pre-defined spectral width learning model to obtain a predicted label matrix for the medical hyperspectral image. Finally, based on this predicted label matrix, the target lesion region is determined from the medical hyperspectral image. By employing this invention, a high-quality global pixel-level predicted label matrix can be output using a pre-defined spectral width learning model, based on a small number of pixels with real labels, by constructing a spectral width matrix with strong representational capabilities. This achieves information enhancement from sparse labels to dense predictions, significantly reducing data preparation costs, solving the problem of lesion region segmentation in scenarios where medical hyperspectral image annotation data is missing, and improving the accuracy and efficiency of lesion detection. Attached Figure Description To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic flowchart of a medical hyperspectral image detection method provided by the present invention; Figure 2 This invention provides a schematic diagram of a process for obtaining spectral width features in medical hyperspectral image detection. Figure 3 This is a schematic diagram of another process for obtaining spectral width features in medical hyperspectral image detection provided by the present invention; Figure 4 This is a schematic diagram of the process for obtaining mapping features in medical hyperspectral image detection provided by the present invention; Figure 5 This is a schematic diagram of the process for obtaining enhanced features in medical hyperspectral image detection provided by the present invention; Figure 6 This invention provides a schematic diagram of the process for obtaining a predicted label matrix in medical hyperspectral image detection. Figure 7 This is a schematic diagram of a medical hyperspectral image detection device provided by the present invention; Figure 8 This is a schematic diagram of a medical hyperspectral image detection device provided by the present invention.

[0017] Icons: 1000, Acquisition module; 2000, Calculation module; 3000, Output module; 10, Medical hyperspectral image detection equipment; 11, Processor; 12, Memory; 13, Bus. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] Before providing a detailed explanation of the present invention, its application scenarios will be introduced first.

[0022] In the field of medical testing, hyperspectral images can be analyzed to extract richer biological tissue information than traditional RGB (Red, Green, Blue) images, thereby improving detection accuracy. However, current technologies heavily rely on large amounts of precisely labeled pixel-level data, which is extremely costly, cumbersome, and highly dependent on the time and experience of experts in medical practice. Furthermore, existing technologies often directly use raw spectral data or simple feature extraction methods, limiting their ability to handle complex lesion areas and affecting the final lesion identification results.

[0023] Based on this, the present invention proposes a medical hyperspectral image detection method, apparatus, device, and storage medium. It can utilize a small number of pixels with real labels to construct a spectral width matrix with strong representational capabilities, thereby outputting a high-quality global pixel-level predicted label matrix using a preset spectral width learning model. This achieves information enhancement from sparse labels to dense predictions, solving the problem of lesion region segmentation in scenarios where medical hyperspectral image annotation data is missing. It significantly reduces reliance on manual annotation and improves the accuracy and efficiency of lesion detection. The medical hyperspectral image detection method provided in the following embodiments of the present invention can be executed by a medical hyperspectral image detection device, which can be a desktop computer, laptop computer, or a portable smart terminal; the present invention does not limit this.

[0024] The following explanation, in conjunction with the accompanying drawings, uses several embodiments to illustrate the concepts. Figure 1 This is a flowchart illustrating a medical hyperspectral image detection method provided by the present invention, as shown below. Figure 1 As shown, this medical hyperspectral image detection method includes: S101, acquire the medical hyperspectral image to be detected, the medical hyperspectral image includes: at least some of the real labels of the pixels.

[0025] In this application, medical hyperspectral images generally refer to medical images acquired through hyperspectral imaging technology. Specifically, a hyperspectral imaging system can be used to scan a target area (such as human tissue) to obtain its reflection or radiation characteristics in different narrow spectral bands, ultimately generating a three-dimensional data cube, where each pixel contains spatial information and spectral information for multiple spectral bands. Optionally, existing medical hyperspectral images can also be downloaded from a database; this application does not impose any restrictions on this.

[0026] True labels refer to the known category annotations determined for certain pixels in a medical hyperspectral image, such as normal category or lesion category. The normal category refers to the pixel category of healthy tissue, and the lesion category refers to the pixel category of diseased tissue. Optionally, true labels can be obtained through manual annotation by technicians. For example, in a brain hyperspectral image, only a few pixels are labeled as lesions, meaning some pixels have true labels, while the majority of other pixels are unlabeled. This brain hyperspectral image can be used as input in this embodiment.

[0027] In one possible implementation, after acquiring the medical hyperspectral image to be detected, the image can be preprocessed, such as denoising and spectral correction, to improve image quality.

[0028] S102, Obtain the spectral width characteristics of the medical hyperspectral image.

[0029] After acquiring a medical hyperspectral image, traditional hyperspectral images have a fixed pixel dimension (e.g., 200 spectral bands), providing only a limited spectral response from a specific viewpoint. Therefore, to enrich the information of each pixel, it is necessary to perform width learning on the medical hyperspectral image to obtain the corresponding spectral width features. Spectral width features are a comprehensive feature representation used to describe the spectral differences and structural relationships between pixels. Specifically, spectral width features generate a new set of feature vectors by applying mapping and enhancement mechanisms to each pixel in the medical hyperspectral image to reflect the complex relationships between pixels.

[0030] Specifically, in one possible implementation, a linear transformation is performed on the input medical hyperspectral image to generate preliminary mapping features, which completes the initial expansion of the original features. Then, feature enhancement is performed based on the preliminary mapping features, and enhanced features with stronger discriminative power are formed through nonlinear transformation. Finally, the preliminary mapping features and enhanced features are fused to obtain the spectral width features of the medical hyperspectral image.

[0031] In practice, the data corresponding to each pixel in the medical hyperspectral image is used as input. Through feature mapping and enhancement, a spectral width feature with higher information density is obtained, which amplifies the differences between different pixels, lays the foundation for the subsequent generation of the prediction label matrix, improves the ability of the subsequent model to distinguish potential lesion areas in the medical hyperspectral image, and enhances the robustness of classification.

[0032] S103. Based on the spectral width characteristics, a preset spectral width learning model is used to obtain the predicted label matrix of the medical hyperspectral image.

[0033] The predicted label matrix includes: the true labels of some pixels in the medical hyperspectral image, and the predicted labels of other pixels.

[0034] The predicted label matrix is ​​a matrix containing category information for all pixels. In the predicted label matrix, the labels of some pixels come from real data, while the labels of others are predicted by a spectral width learning model. In this embodiment, the predicted label matrix is ​​generated by utilizing spectral width features in conjunction with a preset spectral width learning model.

[0035] The pre-defined spectral width learning model is a mathematical prediction model that uses the vectors corresponding to pixels with real labels in the spectral width features as training features. Here, the vector represents the coordinate position of the pixel in the multi-dimensional feature space. The known real labels are used as training targets. By adjusting the internal parameters of the spectral width learning model, the final output label is made as consistent as possible with the known real labels. In other words, the spectral width learning model summarizes the relationship between spectral width features and labels from pixels with real labels, thereby learning from the global medical hyperspectral image and finally obtaining the predicted label matrix.

[0036] In practice, the predicted label matrix corresponds spatially to the medical hyperspectral image to be detected, assigning a predicted category label to each pixel, thereby outlining the contour and extent of the lesion.

[0037] S104, based on the predicted label matrix, identifies the target lesion region from medical hyperspectral images.

[0038] Since each row of the prediction label matrix corresponds to a predicted label for one pixel, if the predicted label of a row is a lesion category, then the pixel corresponding to that row is determined to be located in the lesion region. By traversing all rows of the prediction label matrix, the locations of all pixels predicted as lesion categories are determined, and the target lesion region is further located.

[0039] For example, suppose a medical hyperspectral image contains 5000 pixels, of which 100 pixels are labeled as tumors, and the remaining pixels are unlabeled. In the predicted label matrix generated in this embodiment, the category labels of the remaining pixels are calculated by a spectral width learning model. Finally, by identifying all pixels predicted as tumors, the boundaries of the lesion region can be accurately drawn.

[0040] In the actual implementation of the step of determining the target lesion region based on the predicted label matrix, image visualization tools are usually used for display. These tools support overlaying the target lesion region onto corresponding CT (Computed Tomography), MRI (Magnetic Resonance Imaging), or PET (Positron Emission Tomography) images. This facilitates physicians' comprehensive assessment of the lesion's characteristics under different imaging modes, confirming and adjusting the detection results. Furthermore, clinical knowledge can be incorporated to correct the predicted results, improving diagnostic accuracy. Optionally, a report conforming to clinical standards can be automatically generated, containing quantitative information such as the number, location, volume, and maximum diameter of the lesions, and a rendered image containing the lesions. The following describes this embodiment through specific scenario examples. For instance, in a brain tumor resection surgery, in order to remove the tumor to the greatest extent while protecting normal functional brain tissue, this embodiment can scan the surgical area to generate a medical hyperspectral image, thereby obtaining the corresponding spectral width features and generating a prediction label matrix. This matrix contains the category label of each pixel. The prediction results can be binarized and superimposed on the brain tissue image, while simultaneously outlining the boundary contour of the tumor area and displaying it in a flashing or highlighted form, providing doctors with clear cutting guidance lines. It also calculates and displays the pixel area occupied by the tumor area in the current field of view in real time, assisting doctors in assessing the resection range.

[0041] In this embodiment, based on a small number of pixels with real labels, a spectral width matrix with strong representation capabilities can be constructed. This allows the use of a preset spectral width learning model to output a high-quality global pixel-level predicted label matrix, achieving information enhancement from sparse labels to dense predictions. This significantly reduces data preparation costs, solves the problem of lesion region segmentation in scenarios where medical hyperspectral image annotation data is missing, and improves the accuracy and efficiency of lesion detection. In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining spectral width features, the present invention also provides a possible implementation of a method for obtaining spectral width features in medical hyperspectral image detection. Figure 2 This is a schematic diagram illustrating the process of obtaining spectral width features in medical hyperspectral image detection provided by the present invention. For example... Figure 2 As shown, in S102 above, acquiring the spectral width features of a medical hyperspectral image includes: S201, Based on some pixels and other pixels in the medical hyperspectral image, obtain the feature matrix of the medical hyperspectral image.

[0042] In this model, "partial pixels" refers to pixels with true labels, belonging to the known normal or lesion categories; "other pixels" refers to unlabeled pixels, i.e., pixels not explicitly labeled as normal or lesion categories. By integrating the information from partial pixels and other pixels, a feature matrix reflecting the overall image characteristics is constructed. The feature matrix is ​​structured data formed by arranging partial pixels and other pixels according to their position and spectral information. It not only contains spectral information but also incorporates spatial distribution characteristics, providing a foundation for subsequent extraction of spectral width features.

[0043] Specifically, the set of pixels with real labels in a medical hyperspectral image is represented as a matrix. ,in This indicates the number of pixels with actual labels. This represents the number of spectral bands. Each row in the matrix represents a pixel with a real label, and each column represents the intensity of a spectral band. The corresponding real label matrix is ​​represented as follows: Where 2 represents the two different categories of image normality and lesion. It represents the set of real numbers. The first in OK Indicates the first The actual label of the 1st pixel. Specifically, if the 1st... If the true label of each pixel is a normal category, then and If the first If the true label of each pixel is the lesion category, then... and .

[0044] The set of unlabeled pixels in a medical hyperspectral image is represented as a matrix. ,in This indicates the number of unlabeled pixels. This represents the number of spectral bands; its corresponding unlabeled matrix is ​​represented as... It can be initialized to full matrix.

[0045] The feature matrix of the entire medical hyperspectral image is represented as follows: ,in This represents the total number of pixels in the image. The label matrix of the entire medical hyperspectral image to be tested represents the number of spectral bands. .

[0046] In practical implementation, for example, assuming a medical hyperspectral image contains 100 partial pixels and 9900 other pixels, the feature matrix can be a (10000×D) matrix, where D is the number of spectral bands. Each row corresponds to a pixel, and each column corresponds to the spectral value of a specific band. By arranging and combining partial pixels and other pixels according to their location and spectral information to form the feature matrix, the feature matrix can comprehensively describe the spectral and spatial information of the medical hyperspectral image, providing a consistent data format for subsequent processing.

[0047] S202, obtain the spectral width feature based on the feature matrix.

[0048] Feature matrix The rows correspond to pixels, and the columns correspond to spectral frequency bands. Feature matrix While containing basic spectral information, it suffers from limitations such as low dimensionality and information redundancy. By employing a series of feature transformations to obtain richer and more discriminative spectral width features, these limitations are overcome, providing optimal feature input for subsequent prediction tasks.

[0049] Specifically, for the feature matrix Feature mapping and feature enhancement are performed. First, feature mapping nodes are constructed to perform initial feature transformation. Then, the feature matrix is ​​mapped using these nodes. The features are mapped into a feature subspace and decomposed and extracted from multiple perspectives using parameter configuration to obtain mapped features. Then, based on the mapped features, deep transformation of features is achieved by constructing enhanced feature nodes. Through the transformation mechanism, the input mapped features are recombined and synthesized to construct enhanced features.

[0050] Finally, all mapped features generated in the mapping stage are combined with all enhanced features generated in the enhancement stage to obtain the spectral width feature. This combination process is achieved by concatenating along the feature dimension, preserving and integrating feature information from different sources. The final output spectral width feature simultaneously contains both the initially transformed mapped features and the deeply processed enhanced features, forming a matrix that is both informative and highly discriminative.

[0051] In this embodiment, the corresponding feature matrix is ​​extracted from the medical hyperspectral image, and finally the spectral width feature that can be used for analysis is generated, thereby enhancing the perception and discrimination ability of the subsequent model and improving the accuracy and stability of the diagnostic results.

[0052] In the above Figure 2 Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining spectral width features, the present invention also provides another possible implementation of the method for obtaining spectral width features in medical hyperspectral image detection. Figure 3 This is a schematic diagram illustrating another process for obtaining spectral width features in medical hyperspectral image detection provided by the present invention. (See diagram below.) Figure 3 As shown, in S202 above, the spectral width feature is obtained based on the feature matrix, including: S301, based on the preset feature mapping dimension and the preset first mapping matrix, the feature matrix is ​​mapped to obtain the mapping features of the medical hyperspectral image.

[0053] Among them, the preset feature mapping dimension This refers to the pre-defined number of dimensions used to transform the original feature space into a new feature space, for example... The first mapping matrix is ​​preset to be a set of randomly initialized weight matrices to ensure that the transformed features have good distribution characteristics. In this embodiment, the input is a feature matrix. .

[0054] Specifically, the feature matrix With the first mapping matrix Performing matrix multiplication to achieve linear projection of features involves transforming the original matrix into a linear projection of the features. 3D spectral features are projected onto the linear combination of spectral features. In the new feature space; then, a preset activation function (e.g., the sigmoid function) is used to activate the projected features. In this application, the linear activation function used during mapping is set to be... Linear functions are characterized by preserving the linear relationship of features, with their output exhibiting a simple proportional relationship to the input. In medical hyperspectral image analysis, using linear activation functions helps to preserve the original physical meaning of spectral features.

[0055] After the above transformation, the mapping features are obtained. ,in Represents the total dimension of the mapped features, and the mapped features Each row corresponds precisely in spatial location to a pixel in the medical hyperspectral image to be detected, and each column represents a new feature dimension formed through linear mapping. (This is related to the feature matrix.) In comparison, mapping features The recombination of features was achieved through linear transformation, which obtained new feature representations while preserving the original physical meaning of the spectra.

[0056] S302, enhance the mapping features according to the preset feature enhancement dimension and the preset second mapping matrix to obtain the enhanced features of the medical hyperspectral image.

[0057] Among them, the preset feature enhancement dimension This refers to enhancing the spatial dimension of feature representation after the initial mapping is completed. The preset second mapping matrix is ​​another randomly initialized weight matrix. By using the preset second mapping matrix, a secondary nonlinear transformation is performed on the mapping features of the medical hyperspectral image to strengthen key features and remove noise interference. The preset second mapping matrix is ​​set. In this step, the input is the mapped features. .

[0058] Specifically, mapping features With the preset second mapping matrix Performing matrix multiplication, this transformation will The enhanced feature space is projected onto the dimensional mapping features, thereby achieving the mapping features... The features are then recombined; then a preset nonlinear activation function is used to activate the projected features. The nonlinear function can obtain the complex nonlinear spectral feature relationships between different tissues, so that the enhanced features of the output medical hyperspectral image can maintain the original discrimination ability while having stronger robustness and generalization ability.

[0059] After the above transformation, the enhanced features are obtained. ,in Represents the total dimension of the augmented features, augmented features Each row corresponds precisely to a pixel in the medical hyperspectral image to be detected in terms of spatial location. Each column represents an independent enhancement feature dimension, which is a feature representation derived from the mapping features through nonlinear transformation. Each enhancement feature dimension captures the discrimination patterns of different tissues in terms of spectral features. Compared to, enhanced features Deep abstraction of features is achieved through nonlinear transformation, and a higher-level feature representation is obtained based on inheriting the discriminative information of the mapping features.

[0060] S303 generates spectral width features based on mapping and enhancement features.

[0061] The mapping features obtained above and enhanced features Optionally, as input, the mapping features can be processed first. and enhanced features Standardization is performed to eliminate differences in units and distributions across different feature spaces. Generation methods can include concatenation, weighted summation, and attention mechanisms; the specific implementation depends on the performance requirements and data characteristics of the application, ultimately yielding the spectral width features. .

[0062] In practice, by fusing two different features, the advantages of these two features can be fully utilized. The complementary effect of different features improves the separability of lesions from normal tissues, thereby enhancing the subsequent model's ability to perceive the target lesion area and significantly improving the accuracy of lesion location detection.

[0063] In this embodiment, the feature matrix is ​​first mapped to a low-dimensional but information-rich mapping space using a preset first mapping matrix and feature mapping dimension, forming the mapping features of the medical hyperspectral image. Then, based on the mapping features, the feature expression is further enhanced by using a preset second mapping matrix and feature enhancement dimension, removing noise and strengthening key information, forming the enhanced features of the medical hyperspectral image. Finally, the mapping features and enhanced features are fused to generate spectral width features, which serve as the key input for subsequent prediction tasks. The entire process is progressive, gradually optimizing feature representation, thereby improving the accuracy and stability of lesion detection.

[0064] In the above Figure 3 Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining mapping features, the present invention also provides a possible implementation of the method for obtaining mapping features in medical hyperspectral image detection. Figure 4 This is a schematic diagram illustrating the process of obtaining mapped features in medical hyperspectral image detection provided by the present invention. For example... Figure 4 As shown, in S301 above, the feature matrix is ​​mapped according to the preset feature mapping dimension and the preset first mapping matrix to obtain the mapped features of the medical hyperspectral image, including: S3011, according to the preset number of mapping feature node groups, the feature nodes of each pixel in the feature matrix are grouped to obtain multiple groups of mapping feature nodes.

[0065] Preset number of mapping feature node groups This refers to the number of groups into which each feature node in the feature matrix is ​​divided during the feature mapping process. The number of groups determines the parallel computing capability and the diversity of feature representation in subsequent mapping processes. For example, if set to... This means dividing the entire feature space into ten independent mapping units, each group processing feature nodes from a different subset. Based on previous embodiments, .

[0066] Multiple sets of mapped feature nodes are multiple feature sets formed after grouping, with each set corresponding to a group of mapped feature nodes. The grouping method of multiple sets of mapped feature nodes helps to reduce the complexity of a single mapping and can preserve the structural information of the original features.

[0067] S3012, according to the preset first mapping matrix, perform feature mapping on each group of mapping feature nodes to obtain the mapping features of each group of mapping feature nodes.

[0068] For each set of mapped feature nodes Using the preset first mapping matrix Obtained through linear projection And add the corresponding offset vector for that group. offset vector For each group of nodes The configured preset randomly initialized offset vector, Then, through the corresponding activation function of this group. Perform a linear transformation to obtain the mapping characteristics of this group. .

[0069] S3013, Based on the mapping features of multiple sets of mapping feature nodes, obtain the mapping features of the medical hyperspectral image.

[0070] Output of all mapped feature node groups By concatenating along the feature dimensions, a complete mapping feature is formed. .

[0071] This step integrates the mapping results of the previous multiple sets of mapping feature nodes to finally generate the mapping feature representation of the entire medical hyperspectral image.

[0072] In this embodiment, the feature matrix is ​​first grouped so that feature processing can be performed in parallel in multiple mapping units, which improves the overall computational efficiency. Then, the feature nodes of each group are mapped by a preset first mapping matrix, which enhances the discriminative ability of the features. Finally, the mapping features of each group are integrated to generate global mapping features for medical hyperspectral images, thereby providing reliable data support for subsequent model prediction.

[0073] In the above Figure 3 Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining enhanced features, the present invention also provides a possible implementation of a method for obtaining enhanced features in medical hyperspectral image detection. Figure 5 This is a schematic diagram illustrating the process of obtaining enhanced features in medical hyperspectral image detection provided by the present invention. Figure 5 As shown, in S302 above, the mapped features are enhanced according to the preset feature enhancement dimension and the preset second mapping matrix to obtain the enhanced features of the medical hyperspectral image, including: S3021, Based on the preset number of enhanced feature node groups, the feature nodes of each pixel in the mapped feature are grouped to obtain multiple groups of enhanced feature nodes.

[0074] Preset number of enhanced feature node groups This refers to the number of groups into which feature nodes in the mapped features are divided during the feature enhancement process. The number of groups determines how many subspaces the mapping space is divided into, used to independently extract and enhance specific types of feature information. For example, if the preset is... This means that all mapping features will be divided into 5 groups for parallel processing. Based on previous embodiments, .

[0075] By grouping the feature nodes in the mapping features, each group of nodes becomes independent of the others, thus avoiding interference between feature nodes and improving the accuracy and stability of feature extraction.

[0076] S3022, according to the preset second mapping matrix, perform feature enhancement on each group of enhanced feature nodes to obtain the enhanced features of each group of enhanced feature nodes.

[0077] For each set of mapped feature nodes Using a preset second mapping matrix Obtained through linear projection And add the corresponding offset vector for that group. offset vector For each group of augmented nodes The configured preset randomly initialized offset vector, Then, through the corresponding activation function of this group. By performing a nonlinear transformation, the enhanced features of this group are obtained. .

[0078] S3023, based on the enhancement features of multiple sets of enhancement feature nodes, the enhancement features of the medical hyperspectral image are obtained.

[0079] Output of all enhanced feature node groups The features are concatenated along the feature dimension to form a complete enhanced feature. .

[0080] This step integrates the enhancement results of multiple sets of enhanced feature nodes to ultimately generate an enhanced feature representation of the entire medical hyperspectral image.

[0081] In one possible implementation, by mapping features and enhanced features By concatenating the feature dimensions, the final spectral width feature is obtained. .

[0082] In this embodiment, by grouping the mapped features according to the preset number of enhanced feature node groups, and using the preset second mapping matrix to enhance the features of each group of enhanced feature nodes, and then generating enhanced features of the medical hyperspectral image based on the enhanced features of multiple groups of enhanced feature nodes, the efficient extraction and optimization of image features can be achieved, which helps to improve the recognition accuracy of the target lesion area of ​​the subsequent model, and ultimately provides more reliable technical support for medical diagnosis.

[0083] In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining the predicted label matrix, the present invention also provides a possible implementation of the method for obtaining the predicted label matrix in medical hyperspectral image detection. Figure 6 This is a schematic diagram illustrating the process of obtaining a predicted label matrix in medical hyperspectral image detection provided by the present invention. Figure 6 As shown, in S103 above, based on the spectral width characteristics, a preset spectral width learning model is used to obtain the predicted label matrix of the medical hyperspectral image, including: S401, Obtain the pre-established regression optimization model as the spectral width learning model.

[0084] The regression optimization model includes: constraints on the predicted label matrix, constraints on the regression coefficient matrix corresponding to the medical hyperspectral image, constraints on the reconstructed residual matrix corresponding to the spectral width feature, constraints on the auxiliary variable matrix, constraints on the spectral width feature, constraints on the latent space feature corresponding to the medical hyperspectral image, and constraints on the feature extraction matrix.

[0085] First, the pre-established regression optimization model is as follows:

[0086]

[0087] The meanings of each variable and parameter are as follows: It is the label matrix of the entire medical hyperspectral image to be tested; These are latent space features corresponding to medical hyperspectral images; It is the regression coefficient matrix corresponding to medical hyperspectral images; It is the reconstructed residual matrix corresponding to the spectral width feature; It is an auxiliary variable matrix; It is the feature extraction matrix; It is the spectral width matrix obtained previously.

[0088] Spectral width characteristics Constrained Lagrange multiplier matrix; Features of latent space Constrained Lagrange multiplier matrix; It is a constrained latent space feature The Laplace matrix; It is used to constrain the prediction of the label matrix. The Laplace matrix; It is the dimension of the hidden space; In order to control the regression coefficient matrix The preset parameter for the range of values ​​that an element can take; It is to control the constraint terms The preset parameters for the magnitude of the effect; It is to control the reconstruction residual matrix The parameters preset for sparsity; It is a penalty parameter.

[0089] In regression optimization models, Through latent space features and regression coefficient matrix Calculate the label matrix of the entire medical hyperspectral image to be detected; The regression coefficient matrix corresponding to medical hyperspectral images The constraint terms control the regression coefficient matrix The size of the range of values; The reconstructed residual matrix corresponding to the spectral width feature The constraint terms control the reconstruction of the residual matrix. sparsity; For auxiliary variable matrix The constraint terms are used to constrain latent space features. Laplace matrix Constraint auxiliary variable matrix Enhance Local similarity; For the predicted label matrix The constraint terms are obtained by using the Laplacian matrix to constrain the pixel labels. Constrained pixel label matrix Enhance Local similarity; Spectral width characteristics The constraint terms make The value is close to ; Latent space features corresponding to medical hyperspectral images The constraint terms make The value of is close to that of the auxiliary variable matrix. ; These are constraints on the feature extraction matrix, constraining the feature extraction matrix. It has orthogonality; It is the similarity matrix between latent space features. The associated Laplace matrix; It is the similarity matrix between the predicted label matrix and the target label matrix. The associated Laplace matrix; in, It is a similarity matrix corresponding to pixels with real labels, and its th... Line number Column position element The value is set to ; It is the similarity matrix corresponding to unlabeled pixels, and its th... Line number Column position element The value is based on two unlabeled pixels. Spatial distance between Calculated, specifically set as follows ,in The parameter used when calculating similarity between unlabeled pixels can be set to...

[0090] It is a similarity matrix between unlabeled pixels and pixels with real labels, and its th... Line number Column position element The value is based on two unlabeled pixels. Spatial distance between Calculated, specifically set as follows ,in To use parameters when calculating the similarity between unlabeled pixels and pixels with true labels, it can be set to .

[0091] Regression optimization models are mathematical models that optimize an objective function by introducing various constraints. In regression optimization models, different types of matrices are controlled by different constraint terms to ensure that the model can more accurately fit the data and improve its generalization ability during training. By using a pre-established regression optimization model as a spectral width learning model, the accuracy and reliability of target lesion region detection in medical hyperspectral images can be effectively improved.

[0092] S402, based on the spectral width characteristics, iteratively solve the regression optimization model to obtain the feature extraction matrix, latent space features, auxiliary variable matrix, reconstructed residual matrix, and regression coefficient matrix.

[0093] First, we introduce the iteration number. ,use , , , , , , , and Indicates the first The values ​​of each variable obtained after each iteration.

[0094] Then initialize. Preset the dimension of the hidden space. ,parameter , and The value; set the iteration number. Randomly initialize matrix ,set up , , , , and Create a matrix of all zeros; set the current penalty parameter. With the maximum penalty parameter Set the target function value Set stop threshold ; Next, the iteration number is incremented by one, that is... ; Solving the feature extraction matrix , No. After the next iteration, we obtain Represented as ,in and Through the Obtained by performing singular value decomposition, i.e. in, It is a left singular vector matrix. It is a right singular vector matrix. It is a singular value matrix; Solving for latent space features . No. After the next iteration, we obtain Represented as ;in and The calculation is as follows:

[0095] Solving the auxiliary variable matrix . No. After the next iteration, we obtain Represented as:

[0096] Solve for the reconstructed residual matrix . No. After the next iteration, we obtain Represented as:

[0097] in Is with The relevant diagonal matrix, its first... diagonal elements ; Solving the regression coefficient matrix . No. After the next iteration, we obtain Represented as:

[0098] S403. Based on the feature extraction matrix, latent space features, auxiliary variable matrix, reconstructed residual matrix, and regression coefficient matrix, the predicted label matrix is ​​obtained.

[0099] Solve for the label matrix of the entire medical hyperspectral image to be detected. . No. After the next iteration, we obtain Represented as .

[0100] It should be noted that, in this embodiment, the label matrix of the entire medical hyperspectral image to be detected... That is, the predicted label matrix described in the above embodiments, because for In this context, the prediction results include all pixels (including pixels with known true labels and pixels with unknown labels). The regression optimization model not only generates predictions for unlabeled pixels but also optimizes and adjusts the labels of pixels with true labels using constraint terms. This ensures that the predictions remain consistent with the true labels while maintaining smoothness and continuity with the predictions of spatially neighboring pixels. This global optimization mechanism effectively utilizes the spatial correlation between all pixels, improving the consistency of the overall prediction.

[0101] Finally, based on the established regression optimization model, the objective function value is calculated. .like , calculate at this time The predicted label matrix, as the final output, will back Rows as label matrices of unlabeled pixels Subsequently based on Identify the target lesion region from medical hyperspectral images; otherwise, solve the matrix using the following formula. , With parameters And then re-execute the above calculation of the feature extraction matrix in the order described above. To calculate The steps.

[0102]

[0103]

[0104]

[0105] In this embodiment, due to the first Latent space features after the second iteration and regression coefficient matrix Feature extraction matrices are also needed in the calculation. Auxiliary variable matrix Reconstructing the residual matrix Therefore, based on the feature extraction matrix, latent space features, auxiliary variable matrix, reconstructed residual matrix, and regression coefficient matrix, the predicted label matrix is ​​obtained.

[0106] In this embodiment, by constructing a regression optimization model containing multiple constraints and iteratively solving it using spectral width features, the accuracy of detecting target lesion regions in medical hyperspectral images can be effectively improved. This enhances the ability to model local structures and global features, thereby reducing false positives and false negatives, and providing more reliable support for clinical diagnosis.

[0107] In the above Figure 1 Based on the provided embodiments, in S104 above, determining the target lesion region from the medical hyperspectral image based on the predicted label matrix includes: S501, based on the label values ​​of each pixel in the predicted label matrix, the preset labels of normal biological tissues, and the preset labels of lesion tissues, determine the target lesion region from the medical hyperspectral image.

[0108] In the predicted label matrix, each row corresponds to the predicted label of one pixel.

[0109] Specifically, predict the label matrix The OK Indicates the first The predicted label value for each pixel. Optionally, if Then the first Each pixel label is classified as normal, meaning the default label for normal biological tissue is... (in );like Then the first Each pixel is determined as a lesion category, i.e., the preset label for the lesion tissue is... (in Traverse the predicted label matrix. For each row, the pixel locations of all predicted lesions are obtained, thus determining the target lesion region from the medical hyperspectral image. In practice, there are no restrictions on the actual representation of the preset labels; they can be freely selected according to the application.

[0110] Based on the comparison mechanism between the predicted label matrix and the preset labels, the lesion area can be accurately extracted from medical hyperspectral images, thereby improving the accuracy and efficiency of diagnosis, avoiding the tediousness of traditional manual annotation, and reducing the impact of human error.

[0111] In this embodiment, by comparing the predicted label matrix with the preset labels, efficient and automatic identification of lesion regions in medical hyperspectral images is achieved, and pixel-by-pixel classification is performed. This not only improves the automation level of lesion identification, but also enhances the consistency and repeatability of diagnostic results.

[0112] The following describes a surface detection device and electronic equipment for mining areas based on UAV aerial photography provided in this application. The specific implementation process and technical effects are described above and will not be repeated below.

[0113] Figure 7 This is a schematic diagram of a medical hyperspectral image detection device provided by the present invention, as shown below. Figure 7 As shown, the medical hyperspectral image detection device includes: The acquisition module 1000 is used to acquire the medical hyperspectral image to be detected, which includes: at least some of the true labels of the pixels; The acquisition module 1000 is also used to acquire the spectral width features of medical hyperspectral images; The calculation module 2000 is used to obtain the predicted label matrix of the medical hyperspectral image based on the spectral width features and a preset spectral width learning model. The predicted label matrix includes: the true labels of some pixels in the medical hyperspectral image, and the predicted labels of other pixels. Output module 3000 is used to determine the target lesion region from medical hyperspectral images based on the predicted label matrix.

[0114] Optionally, the acquisition module 1000 is also used to obtain the feature matrix of the medical hyperspectral image based on some pixels and other pixels in the medical hyperspectral image; and to obtain the spectral width feature based on the feature matrix.

[0115] Optionally, the acquisition module 1000 is further configured to map the feature matrix according to the preset feature mapping dimension and the preset first mapping matrix to obtain the mapping features of the medical hyperspectral image; enhance the mapping features according to the preset feature enhancement dimension and the preset second mapping matrix to obtain the enhanced features of the medical hyperspectral image; and generate spectral width features according to the mapping features and the enhanced features.

[0116] Optionally, the acquisition module 1000 is further configured to group the feature nodes of each pixel in the feature matrix according to the preset number of mapping feature node groups to obtain multiple groups of mapping feature nodes; perform feature mapping on each group of mapping feature nodes according to the preset first mapping matrix to obtain the mapping features of each group of mapping feature nodes; and obtain the mapping features of the medical hyperspectral image according to the mapping features of multiple groups of mapping feature nodes.

[0117] Optionally, the acquisition module 1000 is further configured to group the feature nodes of each pixel in the mapping feature according to the preset number of enhanced feature node groups to obtain multiple groups of enhanced feature nodes; perform feature enhancement on each group of enhanced feature nodes according to the preset second mapping matrix to obtain the enhanced features of each group of enhanced feature nodes; and obtain the enhanced features of the medical hyperspectral image according to the enhanced features of multiple groups of enhanced feature nodes.

[0118] Optionally, the acquisition module 1000 is also used to acquire a pre-established regression optimization model as a spectral width learning model. The regression optimization model includes: constraint terms on the predicted label matrix, constraint terms on the regression coefficient matrix corresponding to the medical hyperspectral image, constraint terms on the reconstructed residual matrix corresponding to the spectral width feature, constraint terms on the auxiliary variable matrix, constraint terms on the spectral width feature, constraint terms on the latent space feature corresponding to the medical hyperspectral image, and constraint terms on the feature extraction matrix.

[0119] Optionally, the calculation module 2000 is also used to iteratively solve the regression optimization model based on the spectral width characteristics to obtain the feature extraction matrix, latent space features, auxiliary variable matrix, reconstructed residual matrix, and regression coefficient matrix; and to obtain the prediction label matrix based on the feature extraction matrix, latent space features, auxiliary variable matrix, reconstructed residual matrix, and regression coefficient matrix.

[0120] Optionally, the output module 3000 is also used to determine the target lesion region from the medical hyperspectral image based on the label values ​​of each pixel in the predicted label matrix, the preset labels of normal biological tissues, and the preset labels of lesion tissues.

[0121] These modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more digital signal processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). Alternatively, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a system-on-a-chip (SOC).

[0122] Figure 8 This is a schematic diagram of a medical hyperspectral image detection device provided by the present invention. The device may be a computing device or server with computing processing capabilities.

[0123] The medical hyperspectral image detection device 10 includes a processor 11, a storage medium 12, and a bus 13. The storage medium 12 stores program instructions executable by the processor 11. When the medical hyperspectral image detection device 10 is executed, the processor 11 communicates with the storage medium 12 via the bus 13, and the processor 11 executes the program instructions to perform the above-described method embodiment. The specific implementation and technical effects are similar and will not be described in detail here.

[0124] Optionally, the present invention also provides a program product, such as a computer-readable storage medium, including a program that, when executed by a processor, is used to perform the above-described method embodiments.

[0125] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0126] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0127] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0128] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0129] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A medical hyperspectral image detection method, characterized in that, The method includes: Acquire a medical hyperspectral image to be detected, the medical hyperspectral image including: at least some pixels of true labels; Obtain the spectral width features of the medical hyperspectral image; Based on the spectral width features, a preset spectral width learning model is used to obtain the predicted label matrix of the medical hyperspectral image; the predicted label matrix includes: the true labels of some pixels in the medical hyperspectral image, and the predicted labels of other pixels; Based on the predicted label matrix, the target lesion region is determined from the medical hyperspectral image.

2. The method according to claim 1, characterized in that, The acquisition of the spectral width features of the medical hyperspectral image includes: Based on the partial pixels and the other pixels in the medical hyperspectral image, the feature matrix of the medical hyperspectral image is obtained; The spectral width feature is obtained based on the feature matrix.

3. The method according to claim 2, characterized in that, The step of obtaining the spectral width feature based on the feature matrix includes: Based on the preset feature mapping dimension and the preset first mapping matrix, the feature matrix is ​​mapped to obtain the mapping features of the medical hyperspectral image; Based on the preset feature enhancement dimension and the preset second mapping matrix, the mapping features are enhanced to obtain the enhanced features of the medical hyperspectral image; The spectral width feature is generated based on the mapping feature and the enhancement feature.

4. The method according to claim 3, characterized in that, The step of mapping the feature matrix according to a preset feature mapping dimension and a preset first mapping matrix to obtain the mapping features of the medical hyperspectral image includes: Based on the preset number of mapping feature node groups, the feature nodes of each pixel in the feature matrix are grouped to obtain multiple groups of mapping feature nodes. Based on the preset first mapping matrix, feature mapping is performed on each group of mapping feature nodes to obtain the mapping features of each group of mapping feature nodes; The mapping features of the medical hyperspectral image are obtained based on the mapping features of the multiple sets of mapping feature nodes.

5. The method according to claim 3, characterized in that, The step of enhancing the mapping features according to a preset feature enhancement dimension and a preset second mapping matrix to obtain the enhanced features of the medical hyperspectral image includes: Based on the preset number of enhanced feature node groups, the feature nodes of each pixel in the mapped feature are grouped to obtain multiple groups of enhanced feature nodes. Based on the preset second mapping matrix, feature enhancement is performed on each group of enhanced feature nodes to obtain the enhanced features of each group of enhanced feature nodes; The enhancement features of the medical hyperspectral image are obtained based on the enhancement features of the multiple sets of enhancement feature nodes.

6. The method according to claim 1, characterized in that, The step of obtaining the predicted label matrix of the medical hyperspectral image based on the spectral width features and using a preset spectral width learning model includes: A pre-established regression optimization model is obtained as the spectral width learning model. The regression optimization model includes: a constraint term on the predicted label matrix, a constraint term on the regression coefficient matrix corresponding to the medical hyperspectral image, a constraint term on the reconstructed residual matrix corresponding to the spectral width feature, a constraint term on the auxiliary variable matrix, a constraint term on the spectral width feature, a constraint term on the latent space feature corresponding to the medical hyperspectral image, and a constraint term on the feature extraction matrix. Based on the spectral width characteristics, the regression optimization model is iteratively solved to obtain the feature extraction matrix, the latent space features, the auxiliary variable matrix, the reconstructed residual matrix, and the regression coefficient matrix; The predicted label matrix is ​​obtained based on the feature extraction matrix, the latent space features, the auxiliary variable matrix, the reconstructed residual matrix, and the regression coefficient matrix.

7. The method according to claim 1, characterized in that, The step of determining the target lesion region from the medical hyperspectral image based on the predicted label matrix includes: The target lesion region is determined from the medical hyperspectral image based on the label values ​​of each pixel in the predicted label matrix, the preset labels of normal biological tissue, and the preset labels of lesion tissue.

8. A medical hyperspectral image detection device, characterized in that, The device includes: An acquisition module is used to acquire a medical hyperspectral image to be detected, wherein the medical hyperspectral image includes: at least some pixels of real labels; The acquisition module is also used to acquire the spectral width features of the medical hyperspectral image; The calculation module is used to obtain the predicted label matrix of the medical hyperspectral image based on the spectral width features and using a preset spectral width learning model; the predicted label matrix includes: the true labels of some pixels in the medical hyperspectral image, and the predicted labels of other pixels; The output module is used to determine the target lesion region from the medical hyperspectral image based on the predicted label matrix.

9. A medical hyperspectral imaging detection device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores program instructions executable by the processor. When the medical hyperspectral image detection device is running, the processor communicates with the memory via the bus, and the processor executes the program instructions to perform the steps of the medical hyperspectral image detection method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, performs the medical hyperspectral image detection method as described in any one of claims 1-7.

Images