Intelligent Screening System and Method for Perianal Lesions Combining Polarization Imaging and Hyperspectral Imaging

Abstract

This invention relates to the fields of optical imaging and artificial intelligence algorithm technology, and discloses an intelligent screening system and method for perianal lesions that integrates polarization imaging and hyperspectral imaging. The method includes the following steps: acquiring polarization hyperspectral image data of the area to be screened, wherein the polarization hyperspectral image data contains polarization information and spectral information; preprocessing the polarization hyperspectral image data to generate standardized polarization hyperspectral data; and simultaneously acquiring polarization hyperspectral image data, extracting polarization feature parameters and spectral feature parameters, and performing feature-level fusion to generate a fused feature vector. This method can simultaneously obtain comprehensive information on tissue microstructure and biochemical components, overcome the false positive interference caused by the inability of single-modal imaging to distinguish between inflammation, edema, and scar tissue, reduce the misdiagnosis rate, and improve the reliability of screening results.

Description

Technical Field

[0001] This invention relates to the fields of optical imaging and artificial intelligence algorithm technology, specifically to an intelligent screening system and method for perianal lesions that integrates polarization imaging and hyperspectral imaging. Background Technology

[0002] Polarization imaging utilizes the polarization properties of light waves to detect physical properties that traditional intensity imaging cannot capture. Hyperspectral imaging acquires spectral information across continuous narrow bands, providing rich data on material composition and biochemical characteristics. Intelligent screening uses deep learning artificial intelligence algorithms to fuse and analyze multimodal image data, automatically identifying lesion features.

[0003] Currently, due to the diverse clinical manifestations of perianal lesions and the insidious nature of early lesions, single-modality imaging techniques often fail to fully reflect the pathological state of tissues when screening for perianal lesions. When relying solely on spectral information, it is impossible to effectively distinguish false positive interference caused by inflammation, edema, and scar tissue.

[0004] Therefore, we propose an intelligent screening system and method for perianal lesions that integrates polarization imaging and hyperspectral imaging to address the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent screening system and method for perianal lesions that integrates polarization imaging and hyperspectral imaging, solving the problem mentioned in the background art of the inability to effectively distinguish false positive interference caused by inflammation, edema and scar tissue.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent screening system and method for perianal lesions integrating polarization imaging and hyperspectral imaging, wherein the method includes the following steps: S1. Collect polarization hyperspectral image data of the area to be screened, wherein the polarization hyperspectral image data includes polarization information and spectral information; S2. Preprocess the polarization hyperspectral image data to generate standardized polarization hyperspectral data; S3. Based on the standardized polarization hyperspectral data, extract polarization feature parameters and spectral feature parameters, and perform feature-level fusion to generate a fused feature vector; S4. Input the fused feature vector into a pre-trained deep learning classification model to identify and classify lesion areas and generate preliminary diagnostic results; S5. Based on the preliminary diagnostic results and the standardized polarization hyperspectral data, conduct a risk assessment of the malignancy of the lesion and generate risk level data; S6. Generate a comprehensive screening report based on the risk level data and the preliminary diagnosis results.

[0007] Preferably, the step S1 of acquiring polarization hyperspectral image data of the area to be screened includes the following steps: S11. Using a polarization hyperspectral imaging device, the area to be screened is scanned within a preset band range, and hyperspectral images under different polarization states are acquired simultaneously. S12. Register and fuse the acquired hyperspectral images of different polarization states to generate a polarization hyperspectral image data cube containing Stokes parameters.

[0008] Preferably, the generation of normalized polarization hyperspectral data in step S2 includes the following steps: S21. Perform dark current correction, radiometric calibration, and atmospheric correction on the polarization hyperspectral image data cube to generate a reflectance data cube. S22. The reflectance data cube is denoised and spectrally smoothed, and a normalization method is used to eliminate the influence of uneven illumination to generate the standardized polarization hyperspectral data.

[0009] Preferably, generating the fused feature vector in step S3 includes the following steps: S31. Extract polarization feature parameters characterizing the microstructure of the tissue surface from the standardized polarization hyperspectral data. The polarization feature parameters include linear polarization degree, polarization angle, and Mueller matrix elements. S32. Extract spectral characteristic parameters characterizing tissue biochemical components from the standardized polarization hyperspectral data, wherein the spectral characteristic parameters include reflectance, absorption depth, and spectral derivative in a specific band; S33. Using an attention-based feature fusion network, the polarization feature parameters and the spectral feature parameters are adaptively weighted and fused to generate the fused feature vector.

[0010] Preferably, generating the preliminary diagnostic result in step S4 includes the following steps: S41. Obtain polarization hyperspectral data of historical cases and their corresponding pathological labels to construct a training dataset; S42. The deep learning model with a hybrid architecture of convolutional neural network and Transformer is trained using the training dataset to obtain the deep learning classification model. S43. Input the fused feature vector into the deep learning classification model, output the probability values ​​of different lesion types, and determine the preliminary diagnosis result according to the preset threshold.

[0011] Preferably, generating risk level data in step S5 includes the following steps: S51. Based on the preliminary diagnostic results, locate the suspected lesion area and extract the target spectral curve of the suspected lesion area from the standardized polarization hyperspectral data; S52. The target spectral curve is compared with a pre-stored standard spectral database of different severity levels, and the risk level data is calculated and generated by combining the polarization degree change rate in the polarization characteristic parameter through a risk assessment algorithm.

[0012] Preferably, generating a comprehensive screening report in step S6 includes the following steps: S61. The preliminary diagnosis results, the risk level data, and the location information of the suspected lesion area in the polarization hyperspectral image data are correlated and integrated. S62. Call the preset report template to generate the comprehensive screening report containing text descriptions, image tags and quantitative indicators.

[0013] Preferably, the specific steps of the adaptive weighted fusion in S33 are as follows: S331. Map the polarization feature parameter and the spectral feature parameter to the same feature space respectively to generate a first feature map and a second feature map; S332. Calculate the channel attention weights and spatial attention weights of the first feature map and the second feature map, and generate an attention weight matrix; S333. Based on the attention weight matrix, perform element-wise weighted summation on the first feature map and the second feature map to generate the fused feature vector.

[0014] Preferably, the specific operation steps of the risk assessment algorithm in S52 are as follows: S521. Perform spectral angle matching and correlation coefficient calculation between the target spectral curve and the standard curves of each level in the standard spectral database to generate a spectral similarity score. S522. Extract the polarization degree change rate of the suspected lesion area and normalize it as the polarization anomaly coefficient. S523. Input the spectral similarity score and the polarization anomaly coefficient into the logistic regression model, calculate and output the risk level data.

[0015] Preferably, the system includes: The data acquisition module uses the polarization hyperspectral imaging unit to acquire the raw image data of the area to be screened, and outputs standardized polarization hyperspectral data through the data preprocessing unit; The feature extraction and fusion module receives the standardized polarization hyperspectral data, extracts feature parameters through the polarization feature extraction unit and the spectral feature extraction unit respectively, and outputs a fused feature vector by the feature fusion unit. The intelligent diagnosis module receives the fused feature vector, performs lesion classification through the deep learning inference unit, and outputs risk level data through the risk assessment unit. The report generation module receives the risk level data and preliminary diagnostic results, and generates a comprehensive screening report through the information integration unit and the report output unit.

[0016] Compared with existing technologies, this invention provides an intelligent screening system and method for perianal lesions that integrates polarization imaging and hyperspectral imaging, and has the following beneficial effects: 1. In this invention, by synchronously acquiring polarization hyperspectral image data, extracting polarization feature parameters and spectral feature parameters and performing feature-level fusion to generate a fused feature vector, comprehensive information on tissue microstructure and biochemical components can be obtained simultaneously. This overcomes the false positive interference caused by the inability of single-modal imaging to distinguish between inflammation, edema and scar tissue, reduces the misdiagnosis rate, and improves the reliability of screening results.

[0017] 2. In this invention, by preprocessing the polarization hyperspectral image data, standardized polarization hyperspectral data is generated, eliminating the interference of factors such as uneven illumination, surface curvature, and secretions on polarization measurement. At the same time, by locating suspected lesion areas and extracting target spectral curves, and combining the rate of change of polarization degree for risk assessment, the system can identify local abnormal polarization signals in real time, reduce the deviation in polarization feature extraction caused by changes in surface morphology, and ensure the accuracy of lesion boundary determination.

[0018] 3. In this invention, lesion identification and classification are performed by inputting the fused feature vector into a pre-trained deep learning classification model. Then, based on the preliminary diagnosis results and standardized polarization hyperspectral data, the malignancy risk of the lesion is assessed to generate risk level data. Finally, a comprehensive screening report is generated based on the risk level data and the preliminary diagnosis results. This achieves full automation and objectivity from data collection to report output, reduces the tedious process of doctors manually comparing spectral curves and morphological features, lowers the risk of missed diagnoses, and enables dynamic risk stratification based on the tissue characteristics of different individuals, thereby improving the accuracy and efficiency of personalized screening. Attached Figure Description

[0019] Figure 1 This is a flowchart of the intelligent screening method for perianal lesions that integrates polarization imaging and hyperspectral imaging according to the present invention; Figure 2 This is a schematic diagram of the intelligent screening system for perianal lesions that integrates polarization imaging and hyperspectral imaging according to the present invention. Detailed Implementation

[0020] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-2 This intelligent screening system and method for perianal lesions, which integrates polarization imaging and hyperspectral imaging, includes the following steps: S1. Collect polarization hyperspectral image data of the area to be screened. The polarization hyperspectral image data contains polarization information and spectral information. S2. Preprocess the polarization hyperspectral image data to generate standardized polarization hyperspectral data; S3. Based on standardized polarization hyperspectral data, extract polarization feature parameters and spectral feature parameters, and perform feature-level fusion to generate a fused feature vector; S4. Input the fused feature vector into a pre-trained deep learning classification model to identify and classify lesion areas and generate preliminary diagnostic results. S5. Based on the preliminary diagnostic results and standardized polarization hyperspectral data, conduct a risk assessment of the malignancy of the lesion and generate risk level data; S6. Generate a comprehensive screening report based on risk level data and preliminary diagnostic results.

[0022] The steps involved in acquiring polarization hyperspectral image data of the area to be screened in S1 are as follows: S11. Using a polarization hyperspectral imaging device, the area to be screened is scanned within a preset band range, and hyperspectral images under different polarization states are acquired simultaneously. S12. Register and fuse the acquired hyperspectral images of different polarization states to generate a polarization hyperspectral image data cube containing Stokes parameters.

[0023] Generating normalized polarization hyperspectral data in S2 includes the following steps: S21. Perform dark current correction, radiometric calibration, and atmospheric correction on the polarization hyperspectral image data cube to generate a reflectance data cube. The specific implementation method is as follows: First, dark current correction is performed to eliminate the inherent thermal noise of the detector itself. A completely black image is acquired under unexposed and unilluminated conditions to serve as the dark field image. For the actual acquired raw hyperspectral image, its spatial location... and band The dark current correction formula is as follows: ; in, Represents the original image in pixel coordinates Place, No. grayscale value of the band This represents the dark field reference value for the corresponding location and band. The corrected image data; This step ensures the purity of the image substrate and eliminates systematic errors caused by temperature drift; Secondly, radiometric calibration is performed to establish a quantitative relationship between image grayscale values ​​and absolute radiance. A reference image is acquired using a standard radiation source under the same imaging conditions. The known reflectance of the white board is assumed to be... The average grayscale value of the collected whiteboard image is Then correct the gain and offset The radiometric calibration formula for any pixel can be obtained through linear fitting: ; in, This represents the radiance value received by the sensor. This represents the gain coefficient related to the band. This represents the bias coefficient related to the band. The corrected image data; This formula converts dimensionless gray values ​​into physical radiation quantities with physical dimensions, ensuring the accuracy of energy measurements across different wavebands. Finally, atmospheric correction is performed to eliminate the absorption and scattering effects of water vapor and aerosols in the air on the light signal, extracting the true reflectance characteristics of the Earth's surface. A near-infrared strong reflectance method based on a radiative transfer model is used to calculate the solar irradiance at the top of the atmosphere. The formula for generating the reflectance numerical cube is defined as follows: ; in, This represents the actual reflectance of the generated reflectance numerical cube at the corresponding location and wavelength. Pi This represents the radiance value received by the sensor. Solar spectral irradiance, The solar zenith angle; This step successfully removed environmental interference and yielded a pure cube of the object's surface reflectance. S22. Denoising and spectral smoothing are performed on the reflectance data cube, and a normalization method is used to eliminate the influence of uneven illumination, generating standardized polarization hyperspectral data. The specific implementation method is as follows: First, denoising is performed using a nonlocal mean filtering algorithm in the spatial domain. Redundant information from similar structures in the image is utilized to smooth noise while preserving edge details. The discrete calculation formula is as follows: ; in, Indicates the position after noise reduction and band The output reflectivity value at that location, Represents the search window Other locations inside Reflectance values ​​in the same wavelength band Indicates position and The weighting coefficients between them Normalization factor; This formula effectively suppresses salt-and-pepper noise and high-frequency white noise, resulting in a smoother image noise floor. Secondly, spectral smoothing is performed. To eliminate random spikes and fluctuations on individual spectral curves, the Savitzky-Golay convolution smoothing method is used. This method replaces the original data points by performing local polynomial least-squares fitting on the spectral data within a sliding window. For the ... The spectral vector of each pixel, its smoothed reflectance The calculation formula is: ; in, The reflectance after smoothing. Indicates the first The center wavelength of each band The radius of the smooth window is defined. The offset on both sides of the center band is The reflectivity of adjacent bands, The expected Savitzky–Golay convolution coefficients are... For indexing, For indexing; This operation reduces spectral noise variance while preserving key morphological features such as spectral peaks and absorption valleys; Finally, a normalization method is used to eliminate localized uneven illumination caused by poor probe fit and tissue curvature. The whitening concept from standard normal variable transformation is introduced, and the smoothed spectral curve is normalized pixel-by-pixel maximum-minimum value to stretch the spectral dynamic range and unify the dimensions. The calculation formula for standardized polarization hyperspectral data is as follows: ; in, This is the final generated standard polarization hyperspectral data. For the current pixel at wavelength The smooth reflectance value at that location, and These represent the minimum and maximum values ​​obtained from the entire spectral curve of that pixel, respectively. This formula allows for the independent scaling of the spectral curve at each spatial location. Within a unified standard range, the influence of local light intensity differences on subsequent spectral feature extraction is eliminated, ensuring data scale consistency.

[0024] Generating the fused feature vector in S3 includes the following steps: S31. Extract polarization feature parameters characterizing the microstructure of tissue surfaces from standardized polarization hyperspectral data. The polarization feature parameters include linear polarization degree, polarization angle, and Mueller matrix elements. The specific implementation method is as follows: First, based on the generated standardized polarization hyperspectral data, in a specific band Below, four basic Stokes vector components are obtained through the polarization modulation unit. , , , Based on this, a complete Mueller matrix is ​​constructed. To describe the polarization transformation characteristics of light propagating in tissue, for each spatial position on the two-dimensional imaging plane. The formula for calculating its Mueller matrix is: ; in, For Muller matrix, Represents total light intensity. Represents a mixture of linear and circular polarization components. These represent the intensities of horizontal, vertical, and 45-degree linearly polarized light, as well as right-handed and left-handed circularly polarized light, respectively. Each element in the matrix Each corresponds to a specific polarization scattering mechanism, completely preserving the scattering information of the tissue's microstructure; Secondly, based on the Mueller matrix, the degree of linear polarization and the polarization angle are calculated. The degree of linear polarization reflects the strength of the depolarization effect on the tissue surface, while the polarization angle indicates the dominant direction of tissue fiber and cell arrangement. The calculation formulas are as follows: ; ; In the above formula, For linear polarization degree, It is the polarization angle. Total light intensity and These are the linear polarization components in the horizontal and vertical directions, respectively; Due to the disordered cell arrangement, diseased tissue often leads to a decrease in linear polarization degree and a disordered distribution of polarization angle. The extracted Mueller matrix elements, linear polarization degree, and polarization angle together constitute a rich set of polarization characteristic parameters. S32. Extract spectral characteristic parameters characterizing tissue biochemical components from standardized polarization hyperspectral data. These spectral characteristic parameters include reflectance, absorption depth, and spectral derivative in specific wavelength bands. The specific implementation method is as follows: First, the reflectance of specific wavelength bands is extracted, and characteristic wavelength bands closely related to common pathological changes are selected. Then, the selected characteristic wavelength bands are analyzed. Its reflectivity Directly taken from standardized polarization hyperspectral data The calculation formula is: ; in, Characteristic bands Reflectivity at that location To standardize the values ​​of polarization hyperspectral data in the corresponding bands, For indexing; The level of reflectivity directly reflects the absorption strength of specific biochemical molecules in this wavelength band. Secondly, in order to quantify the intensity of the absorption peak, the absorption depth is calculated. Assuming in the characteristic band There is a distinct absorption peak at this point. Using the reflectance of the nearby baseline bands with no or weak absorption as a reference, the formula for calculating the absorption depth is: ; in, For absorption depth, The reflectance at the center of the characteristic absorption peak, The baseline reflectance fitted to this absorption peak; The greater the absorption depth, the higher the concentration of a specific biochemical component at that location, which usually corresponds to the hypoxic state of the lesion area; Finally, to eliminate the effects of instrument noise and some baseline drift, and to highlight the subtle trends in the spectral curves, the spectral derivative was calculated. In particular, for calculating the first-order spectral derivative The calculation formula is as follows: ; in, Indicates in band The first spectral derivative value at that point, and These represent the reflectance values ​​of adjacent high and low bands, respectively. The wavelength interval between two adjacent bands; The spectral derivative can effectively amplify the small spectral shifts caused by specific molecular vibrations, improve the anti-interference ability of feature extraction, and the extracted specific band reflectance, absorption depth and first-order spectral derivative together construct the spectral characteristic parameters characterizing tissue biochemical components. S33. An attention-based feature fusion network is used to adaptively weight and fuse polarization feature parameters and spectral feature parameters to generate a fused feature vector.

[0025] Generating preliminary diagnostic results in S4 includes the following steps: S41. Obtain the polarization hyperspectral data of historical cases and their corresponding pathological labels, construct a training dataset, and finally complete the training dataset. Defined as a binary tuple consisting of a set of feature vectors and a set of labels: ; In the formula, For the training dataset, Represents the total number of historical case samples. It is the first The feature vector of each sample It is the first The true pathological label of each sample For indexing; S42. Train a deep learning model with a hybrid architecture of convolutional neural network and Transformer using the training dataset to obtain a deep learning classification model. During the model training process, introduce the cross-entropy loss function. To quantify the difference between the model's predicted results and the actual pathological labels, the calculation formula is defined as follows: ; In the formula, For loss function, Represents the total number of categories in pathological classification. The real label is in the first The values ​​that a class can take. It is the probability value output by the model after Softmax normalization. For indexing; During training, the pre-built training dataset will be used. The network is fed in batches, and the weight parameters are iteratively updated using the backpropagation algorithm and optimizer to minimize the loss function. When the loss value converges to a stable state, the trained deep learning classification model is obtained. S43. Input the fused feature vector into the deep learning classification model, output the probability values ​​of different lesion types, and determine the preliminary diagnosis result according to the preset threshold. Input the fused feature vector into the established deep learning classification model. The fully connected layer of the model will calculate the unnormalized classification score through linear transformation and nonlinear activation. Then, the Softmax function is used to map the score to a probability distribution. The calculation formula is: ; In the formula, To use the Softmax function to map scores to a probability distribution, The representative model calculates the first The original score of the lesion-like disease, The total number of categories, For indexing, The representative model calculates the first The original score of the lesion type; After this step, the system outputs a vector containing the probability values ​​of each category. ; Finally, a preset confidence threshold is set. System traversal Of all probability values, the category with the highest probability is selected as the prediction result. when If the condition is met, the area is determined to be the type of lesion in the preliminary diagnosis; otherwise, it is marked as "to be reviewed". This process achieves an automated closed loop from raw data to clinical diagnostic conclusions.

[0026] Generating risk level data in S5 includes the following steps: S51. Based on the preliminary diagnostic results, locate the suspected lesion area and extract the target spectral curve of the suspected lesion area from the standardized polarization hyperspectral data. S52. Compare the target spectral curve with a pre-stored standard spectral database of different severity levels, and combine it with the polarization degree change rate in the polarization characteristic parameter to calculate and generate risk level data through a risk assessment algorithm.

[0027] Generating a comprehensive screening report in S6 includes the following steps: S61. Link and integrate the preliminary diagnostic results, risk level data, and location information of suspected lesion areas in polarized hyperspectral image data; S62. Call the preset report template to generate a comprehensive screening report that includes text descriptions, image tags, and quantitative indicators.

[0028] The specific steps for adaptive weighted fusion in S33 are as follows: S331. Map the polarization feature parameters and spectral feature parameters to the same feature space respectively to generate a first feature map and a second feature map. In specific implementation, let the polarization feature vector extracted through the aforementioned steps be... The spectral eigenvector is By using fully connected layers and reshaping operations, these features are mapped back to the spatial dimension. The calculation formulas for the generated first and second feature maps are as follows: ; ; in, and They represent spatial locations respectively The mapping values ​​of polarization characteristics and spectral characteristics at the location, and This is the weight matrix. and For bias terms, It is a transpose operator; S332. Calculate the channel attention weights and spatial attention weights of the first feature map and the second feature map to generate an attention weight matrix. The specific implementation method is as follows: First, channel attention weights are calculated by compressing spatial information through global average pooling, and then a multilayer perceptron is used to generate importance weights for each channel. The formula for calculating the channel attention weight matrix is ​​as follows: ; in, Here is the channel attention weight matrix. Indicates the feature map Perform global average pooling and output channel statistics. This represents a multilayer perceptron. This represents the Sigmoid activation function; Secondly, spatial attention weights are calculated by concatenating the average pooling and max pooling results along the channel dimension to generate a spatial feature map. Then, spatial weights are generated through a convolutional layer. The formula for calculating the spatial attention weight matrix is ​​as follows: ; in, Here is the spatial attention weight matrix. and These represent global average pooling and global max pooling at the channel level, respectively. Indicates use Feature fusion is performed using convolutional kernels of varying sizes. Use the Sigmoid activation function; The final attention weight matrix Composed of channel weights and spatial weights, typically implemented through element-wise multiplication: ; in, This indicates element-wise multiplication. This is the final attention weight matrix. For channel attention weights, Spatial attention weights; S333. Based on the attention weight matrix, the first feature map and the second feature map are summed element-wise with weights to generate a fused feature vector. The specific implementation method is as follows: First, the attention weight matrix Applied to the first feature map respectively Second feature map Feature enhancement: ; ; in, This indicates element-wise multiplication. and These represent the polarization feature map and spectral feature map after attention weighting, respectively. , The original polarization feature map and spectral feature map are shown. Subsequently, to generate the final fused feature vector, the weighted two-dimensional feature map needs to be flattened into a one-dimensional vector and then concatenated to form the fused feature vector. The calculation formula is: ; in, This means expanding the two-dimensional feature map into a one-dimensional vector by rows and columns. This indicates a vector concatenation operation.

[0029] The final generated It contains comprehensive features that have been filtered by attention and integrate polarization and spectral information, which can be directly input into subsequent classifiers for lesion identification.

[0030] The specific steps for the risk assessment algorithm in S52 are as follows: S521. Match the target spectral curve with standard curves of various levels in the standard spectral database using spectral angles and calculate correlation coefficients to generate a spectral similarity score. The specific implementation method is as follows: First, spectral angle matching calculation is performed. The spectral angle is measured by calculating the angle between two spectral curves in a specific multidimensional space. It is insensitive to the overall scaling of illumination intensity. Let the normalized spectral vector of the target pixel be... The first in the standard database The standard reference spectral vector of the class is spectral angle The formula for calculating the similarity score is: ; in, For spectral similarity scores, This represents the dot product operation of two vectors. Let represent the Euclidean norms of the target spectral vector and the standard spectral vector, respectively. This represents the total number of spectral bands. and Representing the target spectrum and standard spectrum respectively in the 1st... Reflectance values ​​at each wavelength band For indexing, For indexing; Secondly, the correlation coefficient is calculated to measure the linear correlation of the spectral curves. Correlation coefficient between standard spectrum and target spectrum The calculation formula is: ; in, The average reflectance of the target spectrum, The average reflectance of the standard spectrum; Finally, to comprehensively evaluate the similarity, the spectral angles and correlation coefficients are normalized and then weighted and summed to generate the final spectral similarity score: ; in, The spectral angles calculated above, and Let be the weighting coefficient, satisfying , That is, the first Similarity score between the standard spectrum and the target spectrum. For the first The correlation coefficient between the standard spectrum and the target spectrum; S522. Extract the polarization degree change rate of the suspected lesion area and normalize it as the polarization anomaly coefficient. The specific implementation method is as follows: First, select a local neighborhood window within the suspected lesion area and calculate the degree of linear polarization at the center of that neighborhood. And the linear polarization degree of all other image points in the neighborhood. Then, the spatial variation coefficient of polarization is calculated and defined as the rate of change of polarization. : ; in, Intra-neighborhood polarization degree standard deviation This represents the average degree of polarization within the neighborhood. and The calculation formulas are as follows: ; ; In the formula, Represents the defined local neighborhood space range. This represents the total number of image points contained within the neighborhood. Finally, in order to convert the calculation results into dimensionless coefficients that can be used for subsequent successive sweeps, the rate of change of polarization degree was... Min-Max normalization is performed to generate polarization anomaly coefficients. : in, This is the currently calculated rate of change of polarization. and These are the minimum and maximum values ​​of the rate of change of polarization degree calculated for all regions of interest in the entire detected image, respectively. The range of values ​​is normalized to between, The larger the value, the greater the deviation of the polarization characteristics of the region from normal tissue, and the higher the suspicion of anomaly; S523. Input the spectral similarity score and polarization anomaly coefficient into the logistic regression model, calculate and output the risk level data. Let the feature vector of the input logistic regression model be... This includes spectral similarity scores. and polarization anomaly coefficient : ; in, transpose Logistic regression models calculate the probability of belonging to a specific risk level by linearly combining features and applying a sigmoid activation function. The calculation formula is as follows: ; in, For linear combination terms, the calculation formula is: ; In the above formula, For the bias term of the model, and Spectral similarity scores and polarization anomaly coefficient The corresponding weighting coefficients, It is a natural constant; Finally, based on the preset classification threshold , the probability value Risk level labels converted to binary classification : ; in, The representative was deemed high-risk. If the risk level is determined to be low, the system will output a quantitative risk assessment result after this step.

[0031] The system includes: The data acquisition module uses the polarization hyperspectral imaging unit to acquire the raw image data of the area to be screened, and outputs standardized polarization hyperspectral data through the data preprocessing unit; The feature extraction and fusion module receives standardized polarization hyperspectral data, extracts feature parameters through the polarization feature extraction unit and the spectral feature extraction unit respectively, and outputs a fused feature vector by the feature fusion unit. The intelligent diagnosis module receives fused feature vectors, classifies lesions through a deep learning inference unit, and outputs risk level data through a risk assessment unit. The report generation module receives risk level data and preliminary diagnostic results, and generates a comprehensive screening report through the information integration unit and the report output unit.

[0032] The operation steps of the intelligent screening system and method for perianal lesions integrating polarization imaging and hyperspectral imaging are as follows: Step 1: Acquisition of polarization hyperspectral image data: Using a polarization hyperspectral imaging device, the area to be screened is scanned within a preset band range, and hyperspectral images under different polarization states are acquired simultaneously. The acquired hyperspectral images under different polarization states are registered and fused to generate a polarization hyperspectral image data cube containing Stokes parameters. This data contains both polarization and spectral information, providing the original data foundation for subsequent multimodal analysis.

[0033] Step 2: Data preprocessing to generate standardized polarization hyperspectral data: Dark current correction, radiometric calibration, and atmospheric correction are sequentially performed on the polarization hyperspectral image data cube to remove detector noise, establish a quantitative relationship between gray values ​​and radiance, and eliminate atmospheric scattering and absorption interference, generating a reflectance data cube. Subsequently, the reflectance data cube is subjected to denoising and spectral smoothing. Nonlocal mean filtering is used to suppress spatial noise, Savitzky-Golay convolution smoothing is used to eliminate random peak fluctuations in the spectral curve, and finally, the maximum-minimum normalization method is used to eliminate the influence of uneven illumination. The spectral curve of each spatial location is independently scaled to a unified standard range of 0 to 1 to generate standardized polarization hyperspectral data.

[0034] Step 3: Extraction and feature-level fusion of polarization and spectral feature parameters: Two types of feature parameters are extracted from standardized polarization hyperspectral data. The first type is polarization feature parameters characterizing the microstructure of tissue surface, including linear polarization degree, polarization angle, and Mueller matrix elements. The orderliness and depolarization characteristics of tissue cells are analyzed by calculating the Stokes vector and Mueller matrix. The second type is spectral feature parameters characterizing tissue biochemical components, including reflectance, absorption depth, and spectral derivative in specific bands. The concentrations of hemoglobin and water biochemical molecules are reflected by analyzing the absorption intensity and spectral curve trends in characteristic bands. An attention-based feature fusion network is used to adaptively weight and fuse polarization feature parameters with spectral feature parameters to generate a fused feature vector, thereby achieving complementary enhancement of multimodal information.

[0035] Step 4: Deep learning classification model for lesion identification and classification: The polarization hyperspectral data of historical cases and their corresponding pathological labels are obtained to construct a training dataset. This dataset is used to train a deep learning model with a hybrid architecture of convolutional neural network and Transformer to obtain a deep learning classification model. The fused feature vector is input into the model, and the probability values ​​of different lesion types are output through fully connected layers and the Softmax function. The preliminary diagnosis result is determined according to the preset confidence threshold, so as to realize the automatic identification and classification of lesion areas.

[0036] Step 5: Risk assessment of lesion malignancy: Based on the preliminary diagnostic results, suspected lesion areas are located, and target spectral curves of the suspected lesion areas are extracted from standardized polarization hyperspectral data. The target spectral curves are compared with pre-stored standard spectral databases of different malignancy levels. Spectral similarity scores are generated by spectral angle matching and correlation coefficient calculation. At the same time, the polarization degree change rate of the suspected lesion areas is extracted, normalized, and used as the polarization anomaly coefficient. The spectral similarity scores and polarization anomaly coefficients are input into a logistic regression model to calculate and output risk level data.

[0037] Step 6: Generation of comprehensive screening report: By linking and integrating preliminary diagnostic results, risk level data, and location information of suspected lesion areas in polarization hyperspectral image data, and calling preset report templates, a comprehensive screening report containing text descriptions, image markers, and quantitative indicators is generated, completing the entire process of automated screening from data acquisition to diagnostic output.

[0038] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart screening system and method for perianal lesions integrating polarization imaging and hyperspectral imaging, characterized in that, The method includes the following steps: S1. Collect polarization hyperspectral image data of the area to be screened, wherein the polarization hyperspectral image data includes polarization information and spectral information; S2. Preprocess the polarization hyperspectral image data to generate standardized polarization hyperspectral data; S3. Based on the standardized polarization hyperspectral data, extract polarization feature parameters and spectral feature parameters, and perform feature-level fusion to generate a fused feature vector; S4. Input the fused feature vector into a pre-trained deep learning classification model to identify and classify lesion areas and generate preliminary diagnostic results; S5. Based on the preliminary diagnostic results and the standardized polarization hyperspectral data, conduct a risk assessment of the malignancy of the lesion and generate risk level data; S6. Generate a comprehensive screening report based on the risk level data and the preliminary diagnosis results.

2. The intelligent screening method for perianal lesions fusion of polarization imaging and hyperspectral imaging according to claim 1, characterized in that, The steps involved in acquiring polarization hyperspectral image data of the area to be screened in S1 are as follows: S11. Using a polarization hyperspectral imaging device, the area to be screened is scanned within a preset band range, and hyperspectral images under different polarization states are acquired simultaneously. S12. Register and fuse the acquired hyperspectral images of different polarization states to generate a polarization hyperspectral image data cube containing Stokes parameters.

3. The intelligent screening method for perianal lesions fusion with polarization imaging and hyperspectral imaging according to claim 2, characterized in that, The generation of normalized polarization hyperspectral data in step S2 includes the following steps: S21. Perform dark current correction, radiometric calibration, and atmospheric correction on the polarization hyperspectral image data cube to generate a reflectance data cube. S22. The reflectance data cube is denoised and spectrally smoothed, and a normalization method is used to eliminate the influence of uneven illumination to generate the standardized polarization hyperspectral data.

4. The intelligent screening method for perianal lesions fusion of polarization imaging and hyperspectral imaging according to claim 3, characterized in that, The generation of the fused feature vector in S3 includes the following steps: S31. Extract polarization feature parameters characterizing the microstructure of the tissue surface from the standardized polarization hyperspectral data. The polarization feature parameters include linear polarization degree, polarization angle, and Mueller matrix elements. S32. Extract spectral characteristic parameters characterizing tissue biochemical components from the standardized polarization hyperspectral data, wherein the spectral characteristic parameters include reflectance, absorption depth, and spectral derivative in a specific band; S33. Using an attention-based feature fusion network, the polarization feature parameters and the spectral feature parameters are adaptively weighted and fused to generate the fused feature vector.

5. The intelligent screening method for perianal lesions fusion with polarization imaging and hyperspectral imaging according to claim 4, characterized in that, The process of generating a preliminary diagnostic result in S4 includes the following steps: S41. Obtain polarization hyperspectral data of historical cases and their corresponding pathological labels, and construct a training dataset; S42. The deep learning model with a hybrid architecture of convolutional neural network and Transformer is trained using the training dataset to obtain the deep learning classification model. S43. Input the fused feature vector into the deep learning classification model, output the probability values ​​of different lesion types, and determine the preliminary diagnosis result according to the preset threshold.

6. The intelligent screening method for perianal lesions fusion polarization imaging and hyperspectral imaging according to claim 5, characterized in that, The process of generating risk level data in S5 includes the following steps: S51. Based on the preliminary diagnostic results, locate the suspected lesion area and extract the target spectral curve of the suspected lesion area from the standardized polarization hyperspectral data; S52. The target spectral curve is compared with a pre-stored standard spectral database of different severity levels, and the risk level data is calculated and generated by combining the polarization degree change rate in the polarization characteristic parameter through a risk assessment algorithm.

7. The intelligent screening method for perianal lesions fusion polarization imaging and hyperspectral imaging according to claim 6, characterized in that, The process of generating a comprehensive screening report in S6 includes the following steps: S61. The preliminary diagnosis results, the risk level data, and the location information of the suspected lesion area in the polarization hyperspectral image data are correlated and integrated. S62. Call the preset report template to generate the comprehensive screening report containing text descriptions, image tags and quantitative indicators.

8. The intelligent screening method for perianal lesions by fusing polarization imaging and hyperspectral imaging according to claim 4, characterized in that, The specific steps of adaptive weighted fusion in S33 are as follows: S331. Map the polarization feature parameter and the spectral feature parameter to the same feature space respectively to generate a first feature map and a second feature map; S332. Calculate the channel attention weights and spatial attention weights of the first feature map and the second feature map, and generate an attention weight matrix; S333. Based on the attention weight matrix, perform element-wise weighted summation on the first feature map and the second feature map to generate the fused feature vector.

9. The intelligent screening method for perianal lesions fusion polarization imaging and hyperspectral imaging according to claim 6, characterized in that, The specific operation steps of the risk assessment algorithm in S52 are as follows: S521. Perform spectral angle matching and correlation coefficient calculation between the target spectral curve and the standard curves of each level in the standard spectral database to generate a spectral similarity score. S522. Extract the polarization degree change rate of the suspected lesion area and normalize it as the polarization anomaly coefficient. S523. Input the spectral similarity score and the polarization anomaly coefficient into the logistic regression model, calculate and output the risk level data.

10. A smart screening system for perianal lesions integrating polarization imaging and hyperspectral imaging, used to implement the smart screening method for perianal lesions integrating polarization imaging and hyperspectral imaging as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition module uses the polarization hyperspectral imaging unit to acquire the raw image data of the area to be screened, and outputs standardized polarization hyperspectral data through the data preprocessing unit; The feature extraction and fusion module receives the standardized polarization hyperspectral data, extracts feature parameters through the polarization feature extraction unit and the spectral feature extraction unit respectively, and outputs a fused feature vector by the feature fusion unit. The intelligent diagnosis module receives the fused feature vector, performs lesion classification through the deep learning inference unit, and outputs risk level data through the risk assessment unit. The report generation module receives the risk level data and preliminary diagnostic results, and generates a comprehensive screening report through the information integration unit and the report output unit.

Images