A method and system for lipid plaque detection combining OCT polarization and spectral information

Abstract

This invention discloses a method and system for detecting lipid plaques that combines OCT polarization and spectral information. OCT scans the vascular lumen, acquiring interference spectral signals from two orthogonal polarization channels and PS-OCT images. Feature extraction is performed on the PS-OCT images and interference spectral signals to obtain texture feature matrices and spectral feature matrices. The spectral feature matrices are then classified or clustered to obtain an initial 3D lipid plaque probability distribution. The feature matrices are input into a preset classification model, outputting a lipid plaque projection image. Disconnected noise in the lipid plaque projection image is then cleaned to generate a 2D lipid plaque projection mask. The final 3D lipid plaque probability distribution is obtained by combining the initial 3D lipid plaque probability distribution and the 2D lipid plaque projection mask. This invention improves the performance of OCT tissue structure chromatography and the accuracy, sensitivity, and specificity of NIRS in lipid plaque detection.

Description

Technical Field

[0001] This invention relates to a lipid plaque detection method and system in the field of biomedical imaging, specifically to a lipid plaque detection method and system that combines OCT polarization and spectral information. Background Technology

[0002] OCT (Optical Computational Transmission) is a promising imaging modality with advantages such as high resolution, high scan rate, label-free operation, and non-invasive biopsy. By combining an endoscopic probe with OCT, three-dimensional microscopic images of internal organs and tissues, including the airway, digestive tract, urinary tract, and cardiovascular system, can be provided. Of particular interest is coronary artery disease, one of the most common and fatal diseases worldwide. OCT can provide microscopic structural information of the coronary artery wall, aiding in the diagnosis of coronary artery disease. Furthermore, in vivo examinations using cardiovascular OCT can provide information such as fibrous cap thickness and macrophage accumulation, helping to detect vulnerable plaques that can lead to myocardial infarction.

[0003] However, analyzing only the microstructure of coronary arteries is insufficient; combining chemical and molecular information is also necessary to improve the diagnostic accuracy of atherosclerosis. In recent years, NIRS imaging has been successfully applied in clinical testing. Based on the absorption of near-infrared light by organic molecules, the probability of lipid cores deep within the arterial wall can be analyzed using spectral information. An imaging technique combining OCT and NIRS would better meet clinical needs. Therefore, how to organically combine OCT and NIRS to enhance the accuracy of microscopic tissue structure detection and improve the precision of lipid plaque identification is an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a lipid plaque detection method and system that combines OCT polarization and spectral information. This method can provide depth-resolved tissue birefringence information, enhance the accuracy of microscopic tissue structure detection, and improve the precision of lipid plaque identification.

[0005] The technical solution adopted in this invention is as follows, including the following steps:

[0006] Step S1: Use PS-OCT to scan the blood vessel lumen, acquire the interference spectrum signals of the two orthogonal polarization channels, and obtain the PS-OCT images of the two orthogonal polarization channels based on the interference spectrum signals of the two orthogonal polarization channels respectively;

[0007] In practice, the interference spectral signals of the two orthogonal polarization channels are processed sequentially by background suppression, spectral shaping and fast Fourier transform to obtain the PS-OCT images of the two orthogonal polarization channels respectively.

[0008] Step S2: Extract texture features from the PS-OCT image to obtain a texture feature matrix;

[0009] Step S3: Generate initial 3D lipid plaque probability distribution

[0010] First, spectral features are extracted from the interference spectrum signal to obtain a spectral feature matrix; then, the spectral feature matrix is ​​classified or clustered to obtain an initial 3D lipid plaque probability distribution.

[0011] Step S4: Input all the texture feature matrices and spectral feature matrices obtained in steps S2 to S3 into the preset classification model and output the lipid plaque projection image;

[0012] Step S5: Clean the non-connected noise in the lipid plaque projection image to obtain a 2D lipid plaque projection mask;

[0013] Step S6: Obtain the final 3D lipid plaque probability distribution based on the initial 3D lipid plaque probability distribution and the 2D lipid plaque projection mask, and determine the location of the lipid plaques based on the final 3D lipid plaque probability distribution.

[0014] The extraction methods for the texture feature matrix in step S2 include:

[0015] The PS-OCT images of two orthogonal polarization channels are averaged, and then the texture features are extracted from the averaged PS-OCT image to obtain the averaged texture feature matrix.

[0016] Alternatively, texture features can be extracted from the PS-OCT images of the two orthogonal polarization channels separately to obtain the texture feature matrices of each of the two orthogonal polarization channels.

[0017] The extraction methods for the spectral feature matrix in step S3 include:

[0018] The interference spectral signals of the two orthogonal polarization channels are averaged, and then the spectral features of the averaged interference spectral signals are extracted to obtain the averaged spectral feature matrix. The averaged spectral feature matrix is ​​then divided into a short-wavelength spectral feature matrix and a long-wavelength spectral feature matrix.

[0019] Alternatively, the spectral features of the interference spectral signals of the two orthogonal polarization channels can be extracted separately to obtain the spectral feature matrices of the two orthogonal polarization channels, and the spectral feature matrices of the two orthogonal polarization channels can be divided into short-wavelength spectral feature matrices and long-wavelength spectral feature matrices.

[0020] The methods for classifying / clustering the spectral feature matrix in step S3 include:

[0021] The average spectral feature difference is obtained based on the spectral feature matrix. When the average spectral feature difference is greater than a preset threshold, it is a lipid plaque. The initial 3D lipid plaque probability distribution is obtained based on the location of the lipid plaque.

[0022] Alternatively, eigenvalues ​​can be obtained from the spectral feature matrix, and all the obtained eigenvalues ​​can form the principal component eigenvalue space. Finally, the principal component eigenvalue space is clustered to obtain the initial 3D lipid plaque probability distribution.

[0023] Alternatively, a short- and long-wavelength spectral feature space can be constructed based on the spectral feature matrix, and the short- and long-wavelength spectral feature space can be clustered to obtain the initial 3D lipid plaque probability distribution.

[0024] Alternatively, a multiple linear regression model can be used to classify the spectral feature space of short and long wavelength bands to obtain a classification curve, and the initial 3D lipid plaque probability distribution can be obtained based on the classification curve.

[0025] The methods for obtaining the average spectral feature difference include:

[0026] The average spectral feature difference is obtained by directly subtracting the averaged short-band spectral feature matrix from the long-band spectral feature matrix.

[0027] Alternatively, the spectral feature difference can be calculated by directly subtracting the short-band spectral feature matrix and the long-band spectral feature matrix of each of the two orthogonal polarization channels, and then the spectral feature difference of the two orthogonal polarization channels can be averaged and used as the average spectral feature difference.

[0028] The methods for obtaining the feature values ​​include:

[0029] The averaged short-band spectral feature matrix and long-band spectral feature matrix are processed by averaging method to obtain the short-band attenuation coefficient curve and long-band attenuation coefficient curve at different wavelengths. Then, principal component analysis is performed on the short-band attenuation coefficient curve and long-band attenuation coefficient curve to obtain their respective eigenvalues.

[0030] Alternatively, the averaging method can be applied to the short-wavelength spectral feature matrices and long-wavelength spectral feature matrices of the two orthogonal polarization channels respectively to obtain the short-wavelength attenuation coefficient curves and long-wavelength attenuation coefficient curves of the two orthogonal polarization channels at different wavelengths. Then, principal component analysis can be performed on the short-wavelength attenuation coefficient curves and long-wavelength attenuation coefficient curves of the two orthogonal polarization channels to obtain their respective eigenvalues.

[0031] The construction methods for the short and long wavelength spectral feature spaces include:

[0032] A short- and long-wavelength spectral feature space is constructed based on the averaged short-wavelength spectral feature matrix and long-wavelength spectral feature matrix. The short- and long-wavelength spectral feature space obtained by this method can be used for clustering and classification.

[0033] Alternatively, short and long band spectral feature spaces can be constructed for each of the two orthogonal polarization channels based on the short-band spectral feature matrices and long-band spectral feature matrices of the two orthogonal polarization channels. The short and long band spectral feature spaces of the two orthogonal polarization channels can be averaged to obtain the short and long band spectral feature spaces used for clustering.

[0034] Alternatively, based on the short-band and long-band spectral feature matrices of the two orthogonal polarization channels, short-band and long-band spectral feature spaces for classification can be constructed for each of the two orthogonal polarization channels.

[0035] In step S3, the specific steps for extracting spectral features from the interference spectral signal to obtain the spectral feature matrix are as follows:

[0036] First, the interference spectrum signal is divided into N spectral band signals, where N is an integer greater than 1. For each spectral band signal, a wavenumber-resolved OCT intensity signal is obtained through a short-time Fourier transform algorithm.

[0037] The average system noise is removed from the OCT intensity signal along the depth direction to obtain the denoised OCT intensity signal;

[0038] Linear fitting is performed on the depth direction of the denoised OCT intensity signal to obtain an intensity-depth curve. The slope of each position on the intensity-depth curve is used as the spectral feature of the current position to obtain the spectral features of the target tissue region and obtain an N-frame spectral feature matrix of the vascular lumen tissue.

[0039] Alternatively, the spectral characteristics of the target tissue region can be calculated using the following formula:

[0040]

[0041] Where μ(k,x,y,z) represents the spectral characteristics at the current position (x,y,z) of the current band k, x represents the coordinate value of the fast scan direction in the OCT detection scan, y represents the coordinate value of the slow scan direction in the OCT detection scan, z represents the coordinate value of the depth direction, the depth direction is the direction perpendicular to the plane formed by the fast scan direction and the slow scan direction, i.e., the optical axis direction, A′(k,x,y,z) represents the intensity of the denoised OCT signal at the current position (x,y,z) of the current band k, σ represents the physical size in air corresponding to each pixel in the depth direction, and n represents the refractive index of the target tissue region.

[0042] Step S2, which involves extracting texture features from the PS-OCT image, specifically includes:

[0043] Extract one or more of the following from PS-OCT images: mean, variance, standard deviation, homogeneity, contrast, dissimilarity, entropy, second moment of angle, and correlation matrix.

[0044] The classification model in step S4 includes one or more classification models based on convolutional neural networks, fully convolutional neural networks, U-net networks, and generative adversarial networks.

[0045] A lipid plaque detection system combining OCT polarization and spectral information, characterized in that:

[0046] Includes PS-OCT imaging equipment, scanning equipment, and signal processing equipment;

[0047] Or it may include a PS-OCT imaging device, an NIRS imaging device, a scanning device, and a signal processing device;

[0048] The PS-OCT imaging device is used for intracavitary tissue imaging and spectral detection, the NIRS imaging device is used for intracavitary tissue spectral detection, the scanning device is used for intracavitary tissue imaging, and the signal processing device is used for lipid plaque analysis of the acquired PS-OCT and NIRS signals.

[0049] The lipid plaque detection system employs one of the following methods:

[0050] The PS-OCT imaging device and the NIRS imaging device share the same PS-OCT light source, and lipid plaques are identified by analyzing the spectrum of the PS-OCT light source.

[0051] Alternatively, the PS-OCT imaging device and the NIRS imaging device can be connected to the PS-OCT light source and the NIRS light source respectively, and lipid plaques can be identified by analyzing the spectrum of the PS-OCT light source.

[0052] Alternatively, the PS-OCT imaging device and the NIRS imaging device can be connected to the PS-OCT light source and the NIRS light source respectively, and lipid plaques can be identified by analyzing the spectrum of the NIRS light source.

[0053] The PS-OCT imaging device employs alternating A-line encoding, frequency reuse, depth encoding, cross-sampling, or parallel detection. The center wavelength of the PS-OCT imaging device's operating band is between 1200-1350 nanometers, and the bandwidth is greater than 100 nanometers.

[0054] This invention utilizes endoscopic PS-OCT for intravascular imaging, employing a dual photodetector structure to record interference spectral signals of two orthogonal polarization states in three-dimensional space from the intima to the adventitia. A dual-channel OCT structural image is generated based on the orthogonal polarization states, and speckle suppression is achieved through image averaging. A subspectral depth analysis model is used to calculate spectral features, where the spectrum is divided into N bands (N is an integer greater than 1). The dual polarization channels are combined with multispectral features, along with texture features extracted from the OCT data, to perform feature recognition and detection of lipid plaques, acquiring NIRS images. This invention improves the performance of OCT tissue structure chromatography and enhances the accuracy, sensitivity, and specificity of NIRS in lipid plaque detection.

[0055] Compared with the prior art, the present invention has the following beneficial effects and advantages:

[0056] 1. The lipid plaque analysis method, device and probe of the present invention, which combines PS-OCT imaging and NIRS imaging, can enhance the detection accuracy of microscopic tissue structure, increase the depth of spectral penetration into tissue, and improve the accuracy of lipid plaque identification, and has outstanding technical effects.

[0057] 2. The present invention improves the accuracy of lipid plaque identification by using a lipid plaque analysis method and device based on PS-OCT imaging and NIRS imaging.

[0058] 3. This invention is based on PS-OCT imaging and NIRS imaging, which can effectively reduce OCT speckle noise.

[0059] 4. This invention is based on PS-OCT imaging and NIRS imaging, which can provide tissue birefringence information and improve the ability of OCT to distinguish fibrous layers. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the imaging system of the present invention, wherein (a) is a schematic diagram of an imaging system in which the PS-OCT device and the NIRS device are connected to the same light source, and (b) is a schematic diagram of an imaging system in which the PS-OCT device and the NIRS device are connected to different light sources.

[0061] Figure 2 This is a schematic flowchart of the lipid plaque analysis method according to an embodiment of the present invention;

[0062] Figure 3 A schematic diagram illustrating the combination of dual polarization channels, texture features, and spectral features in a lipid plaque analysis method according to an exemplary embodiment of the present invention;

[0063] Figure 4The figures shown are experimental results of a phantom body according to an exemplary embodiment of the present invention. (a) is a tomographic diagram of the phantom body in a rectangular coordinate system; (b) is a DR attenuation coefficient diagram in a rectangular coordinate system. The lipid attenuation coefficient is significantly higher than the collagen attenuation coefficient; (c) is the attenuation spectrum of lipids and collagen within the bandwidth of the catheter-OCT system; (d) is the GMM clustering result of the DR attenuation coefficient in the subspectral system; and (e) is the lipid distribution map of IVOCT-NIRS in polar coordinates.

[0064] Among them: 1001-PS-OCT light source; 1002-NIRS light source; 1003-wavelength division multiplexer; 1004-double cladding coupler; 1005-low noise photodetector; 102-90:10 fiber coupler; 103-first polarization controller; 104-second polarization controller; 105-single-mode fiber; 106-polarization-maintaining fiber; 107-first circulator; 108-second circulator; 109-50:50 fiber coupler; 110 - Reference arm collimating lens; 111- Reference arm focusing lens; 112- Plane mirror; 113- Fiber optic rotary connector; 114- Linear motor; 115- DC brushless motor; 116- Probe; 117- Sample; 118- Third polarization controller; 119- Fourth polarization controller; 120- First polarization beam splitter; 121- Second polarization beam splitter; 122- First photoelectric balance detector; 123- Second photoelectric balance detector; 124- Signal processor. Detailed Implementation

[0065] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this document. It should be noted that these descriptions and examples are merely illustrative and should not be construed as limiting the scope of the present invention. The scope of protection of the present invention is defined by the appended claims, and any modifications based on the claims of the present invention are within the scope of protection of the present invention.

[0066] This invention provides a lipid plaque detection method and system that combines OCT polarization and spectral information, which can enhance the detection accuracy of microstructure, increase the depth of spectral penetration into the tissue, and improve the accuracy of lipid plaque identification.

[0067] The method of the present invention includes the following steps:

[0068] Step S1: Use PS-OCT to scan the blood vessel lumen and acquire the interference spectral signals of the two orthogonal polarization channels. The interference spectral signals are specifically the interference spectral signals in the two-dimensional / three-dimensional space below the surface of the blood vessel lumen. The interference spectral signals of the two orthogonal polarization channels are processed by background suppression, spectral shaping and fast Fourier transform respectively to obtain the PS-OCT images of the two orthogonal polarization channels.

[0069] Step S2: Extract texture features from the PS-OCT image to obtain a texture feature matrix;

[0070] Step S3: Generate an initial 3D lipid plaque probability distribution, including:

[0071] S301. Extract spectral features from the interference spectral signal to obtain the spectral feature matrix;

[0072] S302. Classify or cluster the spectral feature matrix to obtain the initial 3D lipid plaque probability distribution;

[0073] Step S4: Input all the texture feature matrices and spectral feature matrices obtained in steps S2 to S3 into the preset classification model for training, and output the lipid plaque projection image after training.

[0074] Step S5: Clean the non-connected noise of the lipid plaque projection image to obtain a 2D lipid plaque projection mask. Specifically, this step involves creating an A-line mask for the texture feature rectangle using a deep learning classification model.

[0075] Step S6: Obtain the final 3D lipid plaque probability distribution based on the initial 3D lipid plaque probability distribution and the 2D lipid plaque projection mask, and determine the location of the lipid plaques based on the final 3D lipid plaque probability distribution.

[0076] The methods for extracting the texture feature matrix in step S2 include:

[0077] The PS-OCT images of two orthogonal polarization channels are averaged, and then the texture features are extracted from the averaged PS-OCT image to obtain the averaged texture feature matrix.

[0078] Alternatively, texture features can be extracted from the PS-OCT images of the two orthogonal polarization channels separately to obtain the texture feature matrices of each of the two orthogonal polarization channels.

[0079] The extraction methods for the spectral feature matrix in step S301 include:

[0080] The interference spectral signals of the two orthogonal polarization channels are averaged, and then the spectral features of the averaged interference spectral signals are extracted to obtain the averaged spectral feature matrix.

[0081] Alternatively, the spectral features of the interference spectral signals of the two orthogonal polarization channels can be extracted separately to obtain the spectral feature matrices of each of the two orthogonal polarization channels.

[0082] The methods for classifying / clustering the spectral feature matrix in step S302 include:

[0083] The averaged spectral feature matrix is ​​divided into a short-wavelength spectral feature matrix and a long-wavelength spectral feature matrix. Then, the spectral feature difference is calculated by directly subtracting the short-wavelength spectral feature matrix and the long-wavelength spectral feature matrix. When the spectral feature difference is greater than a preset threshold, it is a lipid plaque. Based on the location of the lipid plaque, the initial 3D lipid plaque probability distribution is obtained.

[0084] Alternatively, the spectral feature matrices of the two orthogonal polarization channels can be divided into short-wavelength spectral feature matrices and long-wavelength spectral feature matrices, respectively. Then, the spectral feature difference is calculated by directly subtracting the short-wavelength spectral feature matrices and long-wavelength spectral feature matrices of the two orthogonal polarization channels. The spectral feature difference of the two orthogonal polarization channels is then averaged and used as the average value of the spectral feature difference. When the average value of the spectral feature difference is greater than a preset threshold, it is a lipid plaque. Based on the location of the lipid plaque, an initial 3D lipid plaque probability distribution is obtained.

[0085] Alternatively, the averaged spectral feature matrix can be divided into a short-wavelength spectral feature matrix and a long-wavelength spectral feature matrix. Then, the short-wavelength spectral feature matrix and the long-wavelength spectral feature matrix are processed by averaging to obtain the short-wavelength attenuation coefficient curve and the long-wavelength attenuation coefficient curve at different wavelengths. Then, principal component analysis is performed on the short-wavelength attenuation coefficient curve and the long-wavelength attenuation coefficient curve to obtain their respective eigenvalues. All the obtained eigenvalues ​​form the principal component eigenvalue space. Finally, the principal component eigenvalue space is clustered to obtain the initial 3D lipid plaque probability distribution.

[0086] Alternatively, the spectral feature matrices of the two orthogonal polarization channels can be divided into short-wavelength spectral feature matrices and long-wavelength spectral feature matrices, respectively. Then, the averaging method is applied to the short-wavelength spectral feature matrices and long-wavelength spectral feature matrices of the two orthogonal polarization channels to obtain the short-wavelength attenuation coefficient curves and long-wavelength attenuation coefficient curves of the two orthogonal polarization channels at different wavelengths. After principal component analysis is performed on the short-wavelength attenuation coefficient curves and long-wavelength attenuation coefficient curves of the two orthogonal polarization channels, their corresponding eigenvalues ​​are obtained. All the obtained eigenvalues ​​form the principal component eigenvalue space. Finally, the principal component eigenvalue space is clustered to obtain the initial 3D lipid plaque probability distribution.

[0087] Alternatively, the averaged spectral feature matrix can be divided into a short-wavelength spectral feature matrix and a long-wavelength spectral feature matrix; a short-wavelength and long-wavelength spectral feature space can be constructed based on the short-wavelength and long-wavelength spectral feature matrices, and the short-wavelength and long-wavelength spectral feature space can be clustered to obtain the initial 3D lipid plaque probability distribution.

[0088] Alternatively, the spectral feature matrices of the two orthogonal polarization channels can be divided into short-wavelength spectral feature matrices and long-wavelength spectral feature matrices, respectively. Based on the short-wavelength spectral feature matrices and long-wavelength spectral feature matrices of the two orthogonal polarization channels, short-wavelength and long-wavelength spectral feature spaces of the two orthogonal polarization channels can be constructed respectively. The short-wavelength and long-wavelength spectral feature spaces of the two orthogonal polarization channels can be averaged, and then the averaged short-wavelength and long-wavelength spectral feature spaces can be clustered to obtain the initial 3D lipid plaque probability distribution.

[0089] Alternatively, the averaged spectral feature matrix can be divided into a short-wavelength spectral feature matrix and a long-wavelength spectral feature matrix; a short-wavelength and long-wavelength spectral feature space can be constructed based on the short-wavelength and long-wavelength spectral feature matrices, and a classification curve can be obtained by classifying the short-wavelength and long-wavelength spectral feature space using a multiple linear regression model. Based on the classification curve, the initial 3D lipid plaque probability distribution can be obtained.

[0090] Alternatively, the spectral feature matrices of the two orthogonal polarization channels can be divided into short-wavelength spectral feature matrices and long-wavelength spectral feature matrices, respectively. Based on the short-wavelength and long-wavelength spectral feature matrices of the two orthogonal polarization channels, short-wavelength and long-wavelength spectral feature spaces of the two orthogonal polarization channels can be constructed. The short-wavelength and long-wavelength spectral feature spaces of the two orthogonal polarization channels can be classified using a multiple linear regression model to obtain classification curves. Based on the classification curves, the initial 3D lipid plaque probability distribution can be obtained.

[0091] In practice, the N-frame spectral feature matrix of vascular lumen tissue is divided into short-band spectral feature matrix and long-band spectral feature matrix according to the center wavelength of the OCT working band. The short-band spectral feature matrix and long-band spectral feature matrix are then classified / clustered to achieve the detection of lipid plaques.

[0092] In step S301, the specific steps for extracting spectral features from the averaged / two orthogonal polarization channels to obtain the spectral feature matrices of each of the averaged / two orthogonal polarization channels are as follows:

[0093] First, the interference spectrum signal is divided into N spectral band signals, where N is an integer greater than 1. For each spectral band signal, a wavenumber-resolved OCT intensity signal is obtained through a short-time Fourier transform algorithm.

[0094] The average system noise is removed from the OCT intensity signal along the depth direction to obtain the denoised OCT intensity signal;

[0095] Linear fitting is performed on the depth direction of the denoised OCT intensity signal to obtain an intensity-depth curve. The slope of each position on the intensity-depth curve is used as the spectral feature of the current position to obtain the spectral features of the target tissue region and obtain an N-frame spectral feature matrix of the vascular lumen tissue.

[0096] Alternatively, the spectral characteristics of the target tissue region can be calculated using the following formula:

[0097]

[0098] Where μ(k,x,y,z) represents the spectral characteristics at the current position (x,y,z) of the current band k, x represents the coordinate value of the fast scan direction in the OCT detection scan, y represents the coordinate value of the slow scan direction in the OCT detection scan, z represents the coordinate value of the depth direction, the depth direction is the direction perpendicular to the plane formed by the fast scan direction and the slow scan direction, i.e., the optical axis direction, A′(k,x,y,z) represents the intensity of the denoised OCT signal at the current position (x,y,z) of the current band k, σ represents the physical size in air corresponding to each pixel in the depth direction, and n represents the refractive index of the target tissue region.

[0099] Step S2 involves extracting texture features from the PS-OCT image, specifically including:

[0100] One or more of the following parameters are extracted from PS-OCT images: mean, variance, standard deviation, homogeneity, contrast, dissimilarity, entropy, second moment of angle, and correlation matrix, for lipid plaque classification and identification.

[0101] The classification model in step S4 includes one or more classification models based on convolutional neural networks, fully convolutional neural networks, U-net networks, and generative adversarial networks.

[0102] This invention provides a lipid plaque detection system that combines OCT polarization and spectral information, comprising two connection methods:

[0103] The first form:

[0104] It includes a PS-OCT imaging device for intracavitary tissue imaging and spectral detection. In this case, the PS-OCT imaging device can be used for both PS-OCT imaging and NIRS imaging. That is, in the first form, the PS-OCT imaging device can also be an NIRS imaging device.

[0105] Includes a scanning device, including an endoscopic scanner for imaging intracavitary tissues;

[0106] It includes one or more signal processing devices for lipid plaque analysis of acquired PS-OCT and NIRS signals.

[0107] The second form:

[0108] Alternatively, it may include a PS-OCT imaging device for intracavitary tissue imaging and spectral detection, for PS-OCT imaging;

[0109] Includes an NIRS imaging device for intracavitary tissue spectral detection, used for NIRS imaging;

[0110] Includes a scanning device, including an endoscopic scanner for imaging intracavitary tissues;

[0111] It includes one or more signal processing devices for lipid plaque analysis of acquired PS-OCT and NIRS signals.

[0112] The scanning device includes an optical fiber rotary connector 113, a linear motor 114, a DC brushless motor 115, and a probe 116. The PS-OCT imaging device mainly consists of a PS-OCT light source 1001, a 90:10 optical fiber coupler 102, a first polarization controller 103, a second polarization controller 104, a single-mode fiber 105, a polarization-maintaining fiber 106, a first circulator 107, a second circulator 108, a 50:50 optical fiber coupler 109, a reference arm collimating lens 110, a reference arm focusing lens 111, a plane mirror 112, a third polarization controller 118, a fourth polarization controller 119, a first polarization beamsplitter 120, a second polarization beamsplitter 121, a first photoelectric balance detector 122, and a second photoelectric balance detector 123. In the second form, the NIRS imaging device mainly consists of an NIRS light source 1002, a wavelength division multiplexer 1003, a double-clad coupler 1004, and a low-noise photodetector 1005.

[0113] The PS-OCT imaging device employs one of the following:

[0114] The PS-OCT imaging device employs alternating A-line encoding, with a center wavelength of 1200-1350 nm and a bandwidth greater than 100 nm in the operating band.

[0115] Alternatively, a frequency-reused PS-OCT imaging device with a center wavelength of 1200-1400 nanometers and a bandwidth greater than 100 nanometers in the operating band.

[0116] Alternatively, a depth-coded PS-OCT imaging device with a center wavelength of 1200-1400 nanometers and a bandwidth greater than 100 nanometers.

[0117] Alternatively, a cross-sampling PS-OCT imaging device can be used, with a center wavelength of 1200-1400 nanometers and a bandwidth greater than 100 nanometers.

[0118] Alternatively, a parallel detection PS-OCT imaging device can be used, with a center wavelength of 1200-1400 nanometers and a bandwidth greater than 100 nanometers.

[0119] The lipid plaque detection system of this invention employs one of the following:

[0120] The PS-OCT imaging device and the NIRS imaging device are connected to the same PS-OCT light source. Lipid plaques are identified by analyzing the spectrum of the PS-OCT light source. This is the first form, where the PS-OCT imaging device and the NIRS imaging device are the same device.

[0121] Alternatively, the PS-OCT imaging device and the NIRS imaging device can be connected to the PS-OCT light source and the NIRS light source respectively, and lipid plaques can be identified by analyzing the spectrum of the PS-OCT light source, which is the second form;

[0122] Alternatively, the PS-OCT imaging device and the NIRS imaging device can be connected to the PS-OCT light source and the NIRS light source respectively, and lipid plaques can be identified by analyzing the spectrum of the NIRS light source, which is the second form.

[0123] The probe is installed in a lipid plaque detection system. Figure 1 (a) is a schematic diagram of an imaging system according to the present invention. In this case, the PS-OCT imaging device and the NIRS imaging device are connected to the same PS-OCT light source. The system includes a PS-OCT light source 1001, a 90:10 fiber coupler 102, a first polarization controller 103, a second polarization controller 104, a single-mode fiber 105, a polarization-maintaining fiber 106, a first circulator 107, a second circulator 108, a 50:50 fiber coupler 109, a reference arm collimating lens 110, a reference arm focusing lens 111, a plane mirror 112, a fiber optic rotary connector 113, a linear motor 114, a DC brushless motor 115, a probe 116, a sample 117, a third polarization controller 118, a fourth polarization controller 119, a first polarization beam splitter 120, a second polarization beam splitter 121, a first photoelectric balance detector 122, a second photoelectric balance detector 123, and a signal processor 124.

[0124] The initial beam output from the PS-OCT light source 1001 is incident on a 90:10 fiber coupler 102, and distributed into the sample arm and reference arm at a 90:10 ratio. The beam entering the reference arm has its polarization state stabilized by a first polarization controller 103, and is then incident on a single-mode fiber 105. The outgoing light enters port 1 of a first circulator 107, exits from port 2, and sequentially enters the reference arm collimating lens 110, the reference arm focusing lens 111, and the plane mirror 112. The light reflected by the plane mirror 112 returns to the first circulator 107 and exits through port 3. The beam entering the sample arm has its polarization state stabilized by a second polarization controller 104, and is then incident on a polarization-maintaining fiber 106. The outgoing light enters port 1 of a second circulator 108, exits from port 2, passes through a fiber optic rotary connector 113, and then enters the probe 116. The scanning of probe 116 is controlled by driving linear motor 114 and brushless DC motor 115. Probe 116 is used to scan sample 117. The light reflected from the sample returns from probe 116 to the second circulator 108 and exits through port 3 of the second circulator 108. The reflected light from the sample arm and the reference arm interferes in 50:50 fiber couplers 109 and enters the third polarization controller 118 and the fourth polarization controller 119 in a 50:50 ratio. The outgoing light enters the first polarization beamsplitter 120 and the second polarization beamsplitter 121, respectively, and the four beams generated enter the first photoelectric balance detector 122 and the second photoelectric balance detector 123. The electrical signals of photoelectric balance detectors 122 and 123 are received by a high-speed digital acquisition card and transmitted to signal processor 124 for processing.

[0125] Figure 1 (b) is a schematic diagram of another imaging system of the present invention, in which the PS-OCT imaging device and the NIRS imaging device each use two light sources. The system includes a PS-OCT light source 1001, a NIRS light source 1002, a wavelength division multiplexer 1003, a double cladding coupler 1004, a low-noise photodetector 1005, a 90:10 fiber coupler 102, a first polarization controller 103, a second polarization controller 104, a single-mode fiber 105, a polarization-maintaining fiber 106, and a first circulator 1. 07. Second circulator 108. 50:50 fiber coupler 109. Reference arm collimating lens 110. Reference arm focusing lens 111. Plane mirror 112. Fiber optic rotary connector 113. Linear motor 114. DC brushless motor 115. Probe 116. Sample 117. Third polarization controller 118. Fourth polarization controller 119. First polarization beam splitter 120. Second polarization beam splitter 121. First photoelectric balance detector 122. Second photoelectric balance detector 123. And signal processor 124.

[0126] The beam output from the NIRS light source 1002 is coupled into the PS-OCT optical path via a wavelength division multiplexer 1003, and then enters the probe 116 via a double-clad coupler 1004. The probe 116 in this system is made of double-clad fiber, where the inner cladding transmits the PS-OCT beam and the outer cladding transmits the NIRS beam, which is then converged onto the sample by the imaging probe. The NIRS beam returning from the sample is received by a low-noise photodetector 1005. The PS-OCT beam returning from the sample is received by a first photoelectric balance detector 122 and a second photoelectric balance detector 123.

[0127] For example, the PS-OCT light source selects a high-speed swept frequency light source to achieve the high-speed imaging performance of the system, uses polarization-maintaining fiber to generate the delay of orthogonal polarization state, and performs polarization state diversity output through fiber-type polarization beam splitter to realize polarization OCT imaging.

[0128] Figure 2 The diagram shown is a schematic flowchart of a lipid plaque detection method according to an embodiment of the present invention. The lipid plaque analysis method of this embodiment includes the following steps:

[0129] Step S1: Acquire intracavitary PS-OCT images.

[0130] For example, an OCT device is used for image acquisition. The circumferential scanning speed of the rotary motor is set to 50 rpm, and the retraction speed of the linear motor is set to 13 mm / s. Each set of data contains 200 frames, and each frame contains 2000 A-lines. The 200 frames of images are arranged sequentially according to the scanning order.

[0131] Step S2: Extract the texture feature matrix based on the intracavitary PS-OCT image;

[0132] Step S3: Generate initial 3D lipid plaque probability distribution

[0133] Step S301: Extract the spectral feature matrix based on the interference spectral signal;

[0134] For example, processing PS-OCT images in Cartesian coordinates involves first segmenting the lumen boundary and guidewire using existing techniques. Then, by manually selecting the guidewire shadow area on the maximum intensity projection map, the A-lines value of the guidewire artifact is set to zero. Finally, a feature matrix is ​​extracted, including texture and spectral features from both polarization channels. Specifically, this can be achieved by performing sub-spectral calculations on the spectral features of polarization channel one and polarization channel two, extracting texture features such as contrast, correlation, energy, and homogeneity. The spectrum is divided into short-wavelength and long-wavelength spectra. Furthermore, the spectral features of polarization channel one and polarization channel two can be combined before sub-spectral calculations, or vice versa. This step involves various methods of combining polarization dual channels and dual spectra. Averaging the structural images generated by the dual polarization channels effectively suppresses speckle noise and improves the imaging effect of OCT on microscopic tissue structures.

[0135] Step S302: Classify or cluster the spectral feature matrix to obtain the initial 3D lipid plaque probability distribution;

[0136] For example, the A-line optical features of the first 150 pixels (~0.75 mm) below the vascular lumen boundary of the dual polarization channel are simultaneously input into the Gaussian Mixture Model (GMM) for cluster analysis among lipid pixels.

[0137] The spectral attenuation coefficients μt (1250-1300nm) and μt (1300-1350nm) of the PS-OCT 3D dataset were calculated. Then, the optical features of the first 150 pixels (approximately 0.75mm) below the A-line vessel lumen boundary were input into the GMM clustering model to obtain the probability of lipid plaques. The GMM clustering parameters were set as follows: number of classes M = 2 (representing two clusters, lipid and non-lipid), convergence error ε = 10. -8 The maximum number of iterations is n = 100.

[0138] Step S4: Input the feature matrix into the preset classification model to obtain the lipid plaque projection image.

[0139] For example, the A-line pixels are first moved radially (z) to flatten the PS-OCT feature image to a 1-pixel position (z direction), and speckle noise is reduced using a (5,5)(zx) kernel Gaussian filter and averaged dual polarization channel data. Then, the A-line features of the first 200 pixels (~1 mm) below the vessel lumen boundary are input into a CNN for lipid classification between A-lines.

[0140] Specifically, considering computational cost and accuracy, the A-line optical and texture features of the first 200 pixels (~1 mm) below the vascular lumen boundary are input into a deep CNN, outputting lipid probability values ​​between A-lines. The CNN structure consists of 9 layers, including 1 input layer, 3 convolutional layers, 2 max-pooling layers, and 3 fully connected layers. Different filter sizes (32, 64, and 96) and kernel sizes (11, 9, and 7) are applied in convolutional layers 2, 4, and 6, generating relevant feature maps with a stride of 2 pixels. Batch normalization and linear rectified function layers are then used to accelerate model training convergence. The max-pooling layer has a pool size of 2 pixels to reduce dimensionality, prevent overfitting, and make the model unaffected by slight transformations, distortions, and translations. The network ends with three fully connected layers, the first two of which include 100 output units with linear rectified functions and culling layers, while the last layer includes two output units with softmax activation functions. Finally, a binary-based projection mask f(x,y) is obtained, where 1 represents lipids and 0 represents other values.

[0141] Step S5: Clean the non-connected noise in the lipid plaque projection image to determine the 2D lipid plaque projection mask.

[0142] For example, a fully connected conditional random field (CRF) is used to remove noise from a two-dimensional projection surface (xy) mask.

[0143] Since A-line classification may ignore the spatial similarity between adjacent A-lines, a fully connected CRF is used to cleanse the noisy projected (xy) lipid mask. Pairwise edge potentials for each probability classification are defined by a linear combination of Gaussian kernels to improve classification performance. There are two types of Gaussian kernels (smooth kernel and appearance kernel):

[0144]

[0145] Among them, CRF(f i +f j ) represents the binary projection mask matrix, ω1 and ω2 are weighting factors, exp() represents the exponential function with base e, and p i p represents the spatial position of row i. j Indicates the spatial location of column j, I i I represents the intensity at position i in the matrix. j θ represents the intensity at position j. α θ is a configurable parameter representing the size and shape of the neighborhood. β θ is the proximity parameter. γ This is the similarity parameter. The first term is the smoothing kernel. Remove small, isolated regions from the CNN prediction results. Second term: appearance kernel. Optimization is performed based on the assumption that adjacent pixels with similar intensity may belong to the same category. Since the pixel intensity in the projected image is affected by the probe distance, we use spectral feature values ​​instead of probe distance for optimization.

[0146] Step S6: Obtain the final 3D lipid plaque probability distribution based on the initial 3D lipid plaque probability distribution and the 2D lipid plaque projection mask, and determine the location of the lipid plaques based on the final 3D lipid plaque probability distribution. The specific method for obtaining the final 3D lipid plaque probability distribution in Step S6 is as follows: combine the initial 3D lipid plaque probability distribution and the 2D lipid plaque projection mask, and remove spectral feature artifacts from the normal blood vessel wall and calcification boundaries to obtain the final 3D lipid plaque probability distribution.

[0147] For example, the identified lipid A-line mask is applied to the GMM 3D result to avoid the attenuation artifacts caused by the light and dark boundaries in the final result, and a 3D lipid probability distribution is obtained.

[0148] Figure 3 The diagram illustrates the combination of dual polarization channels, texture features, and spectral features in the lipid plaque detection method of an exemplary embodiment of the present invention, including the extraction and combination of dual-channel spectral features and texture features.

[0149] Specifically, this explanation focuses on single-channel data (the feature extraction steps are the same for dual-channel data), performing a short-time Fourier transform on the axial scanning interferometric spectrum. The axial scanning slices are subjected to Fourier transform after applying a Hanning window, with the number of spectrum segments preferably set to 2. Wavelength-related spectral features are calculated using a deep analytical model.

[0150] First, a short-time Fourier transform (STFT) is performed on the spectral signal to obtain the axial OCT intensity signal A(k,z) that is resolvable along the depth z wavenumber k:

[0151]

[0152] Among them, STFT[S int [(k)] represents the short-time Fourier transform of the spectrum, S int (k′) is the interference spectrum signal with wavenumber k′, where k′ represents the wavenumber, and w(k′-k;Δk) is the analysis window function. The window function is fixed and the Fourier transform is performed by sliding the spectral interference signal. Here we choose the Hanning window, but other windows can also be used, such as the Kaiser window.

[0153] Then, based on the wavenumber-resolved OCT intensity signal characteristics in the depth direction, the spectral characteristics of different bands are calculated, including:

[0154] The average system noise is removed from the OCT intensity signal along the depth direction to obtain the denoised OCT intensity signal, which is set using the following formula:

[0155] A′(k,z)=A0(k,z)-B(k,z)

[0156] Where A′(k,z) represents the denoised OCT intensity signal corresponding to the coordinate value z of the current depth and band, A0(k,z) is the OCT intensity signal corresponding to the current depth coordinate z and the current band k, and B(k,z) is the average system noise corresponding to the coordinate value z in the current depth direction. The average system noise is specifically calculated as follows: without placing a sample, a blank scan is performed to obtain the system noise. The system noise is then averaged along the depth direction for each XY plane, i.e., the system noise values ​​in the X and Y directions are averaged sequentially to obtain a one-dimensional average system noise distribution along the depth direction.

[0157] The spectral characteristics of the target tissue region are calculated based on the depth-direction characteristics of the denoised OCT signal, specifically:

[0158] The depth-direction features of the denoised OCT signal are linearly fitted in the depth direction to obtain an intensity-depth curve. The slope of each position on the intensity-depth curve is used as the spectral feature of the current position to obtain the spectral features of the target tissue region.

[0159] Alternatively, the spectral characteristics of the target tissue region can be set using the following formula:

[0160]

[0161] Where μ(k,x,y,z) represents the spectral characteristics at the current position (x,y,z) of the current band k, x represents the coordinate value of the fast scan direction in the OCT detection scan, y represents the coordinate value of the slow scan direction in the OCT detection scan, z represents the coordinate value of the depth direction, the depth direction is the direction perpendicular to the plane formed by the fast scan direction and the slow scan direction, i.e., the optical axis direction, A′(k,x,y,z) represents the intensity of the denoised OCT signal at the current position (x,y,z) of the current band k, σ represents the physical size in air corresponding to each pixel in the depth direction, and n represents the refractive index of the target tissue region.

[0162] Texture feature extraction is achieved using the gray-level co-occurrence matrix (GLCM). The GLCM is an L×L feature matrix, defined as I(i,j) in the image, which calculates the spatial proximity frequency between pixels with gray value i and pixels with gray value j. Offsets (Dx, Dz) are used to describe the spatial relationships.

[0163] D x =D·cos(θ),D z =D·sin(θ)

[0164] Among them, D x and D z θ represents the offsets in the x and z directions, respectively, D is the step size between pixels, and θ is the spatial direction. The gray-level co-occurrence matrix is ​​calculated in Cartesian coordinates. At that time, the grayscale value of the PS-OCT image is G = 64, θ = 90°, and D = 2. Therefore, four [presumably referring to calculations] are performed for each single pixel in the n×n neighborhood (here, n = 11). Extract contrast Correlation energy and homogeneity Four texture features:

[0165]

[0166]

[0167]

[0168]

[0169]

[0170]

[0171] Where i and j are the row and column positions of the matrix, μ i and μ j All are the mean, σ i and σ j Both are variances. For co-occurrence matrix, This indicates that the row and column positions of the matrix are multiplied by the gray-level co-occurrence matrix.

[0172] There are multiple ways to combine dual-channel, texture features and spectral features. For example, the spectra of the dual polarization channels are first averaged, and then short-band and long-band spectral feature images are obtained. The difference in spectral features is calculated by directly subtracting the short-band and long-band spectral feature images. When the difference in spectral features is greater than a preset threshold, it is considered to be a lipid plaque. Then, an A-line mask is made for the texture features by constructing a deep learning model to remove spectral feature artifacts of normal blood vessel walls and calcification boundaries, and finally the distribution of lipid plaques is obtained.

[0173] Figure 4The results of the phantom experiment obtained using this embodiment are shown. Lipid plaque information was obtained from the acquired PS-OCT images using the lipid plaque analysis method described in this invention. For example... Figure 4 As shown in (a), the left side of the phantom is composed of lipids, and the right side is composed of collagen. The lipid component cannot penetrate near-infrared light due to its high attenuation properties, while the collagen component, due to its low attenuation properties, allows for deep penetration of near-infrared light. Figure 4 This characteristic can also be verified in the depth resolution attenuation coefficient plot of (b). The lipid component has a high attenuation coefficient, while the collagen component has a low attenuation coefficient. The lipid attenuation coefficient is significantly higher than that of collagen. The attenuation spectra of lipids and collagen within the bandwidth of the OCT system are as follows: Figure 4 As shown in (c), the decay spectrum of collagen components is generally flat, while the decay spectrum of lipid components shows differences in both long and short wavelengths. Figure 4 (d) Clustering results of attenuation coefficients analyzed by spectral depth. Points represent classified lipid pixels (lipid probability > 0.6), and crosses represent classified collagen pixels. Figure 4 (e) shows the OCT-NIRS lipid distribution map in polar coordinates. In actual processing, different colors can be used to represent different components; specifically, yellow can represent lipids, and red can represent non-lipids. Figure 4 (e) uses color to represent the probability of lipids, where white represents lipids and black represents non-lipids. The closer to white (the real color is yellow), the higher the probability that the component is a lipid.

[0174] The above experimental results fully demonstrate that the lipid plaque analysis method, device and probe combining PS-OCT imaging and NIRS imaging involved in this invention can enhance the detection accuracy of microscopic tissue structure, increase the depth of spectral penetration into the tissue, and improve the accuracy of lipid plaque identification, thus exhibiting outstanding and significant technical effects.

[0175] The results of lipid detection in the phantom using PS-OCT-NIRS were consistent with the known components, demonstrating the effectiveness of the proposed method.

Claims

1. A method of lipid plaque detection combining OCT polarization and spectral information, characterized by, Includes the following steps: Step S1: Use PS-OCT to scan the blood vessel lumen, acquire the interference spectrum signals of the two orthogonal polarization channels, and obtain the PS-OCT images of the two orthogonal polarization channels based on the interference spectrum signals of the two orthogonal polarization channels respectively; Step S2: Extract texture features from the PS-OCT image to obtain a texture feature matrix; Step S3: Extract spectral features from the interference spectrum signal to obtain a spectral feature matrix; classify / cluster the spectral feature matrix to obtain the initial 3D lipid plaque probability distribution; Step S4: Input all texture feature matrices and spectral feature matrices into the preset classification model and output lipid plaque projection images; Step S5: Clean the non-connected noise in the lipid plaque projection image to obtain a 2D lipid plaque projection mask; Step S6: Obtain the final 3D lipid plaque probability distribution based on the initial 3D lipid plaque probability distribution and the 2D lipid plaque projection mask, and determine the location of the lipid plaques based on the final 3D lipid plaque probability distribution. The methods for classifying / clustering the spectral feature matrix in step S3 include: The average spectral feature difference is obtained based on the spectral feature matrix. When the average spectral feature difference is greater than a preset threshold, it is a lipid plaque. The initial 3D lipid plaque probability distribution is obtained based on the location of the lipid plaque. Alternatively, eigenvalues ​​can be obtained from the spectral feature matrix, and all the obtained eigenvalues ​​can form the principal component eigenvalue space. Finally, the principal component eigenvalue space is clustered to obtain the initial 3D lipid plaque probability distribution. Alternatively, a short- and long-wavelength spectral feature space can be constructed based on the spectral feature matrix, and the short- and long-wavelength spectral feature space can be clustered to obtain the initial 3D lipid plaque probability distribution. Alternatively, a multiple linear regression model can be used to classify the spectral feature space of short and long bands to obtain a classification curve, and the initial 3D lipid plaque probability distribution can be obtained based on the classification curve. The methods for obtaining the average spectral feature difference include: The average spectral feature difference can be calculated by directly subtracting the averaged short-band spectral feature matrix from the long-band spectral feature matrix; or by directly subtracting the short-band spectral feature matrix from the long-band spectral feature matrix of each of the two orthogonal polarization channels to calculate the spectral feature difference, and then averaging the spectral feature differences of the two orthogonal polarization channels as the average spectral feature difference. The methods for obtaining the feature values ​​include: The averaged short-wavelength spectral feature matrix and the long-wavelength spectral feature matrix are processed by averaging method to obtain the short-wavelength attenuation coefficient curve and the long-wavelength attenuation coefficient curve at different wavelengths. Then, principal component analysis is performed on the short-wavelength attenuation coefficient curve and the long-wavelength attenuation coefficient curve to obtain their respective eigenvalues. Alternatively, the averaging method can be applied to the short-wavelength spectral feature matrices and long-wavelength spectral feature matrices of the two orthogonal polarization channels respectively to obtain the short-wavelength attenuation coefficient curves and long-wavelength attenuation coefficient curves of the two orthogonal polarization channels at different wavelengths. Then, principal component analysis can be performed on the short-wavelength attenuation coefficient curves and long-wavelength attenuation coefficient curves of the two orthogonal polarization channels to obtain their respective eigenvalues. The construction methods for the short and long wavelength spectral feature spaces include: Construct short- and long-band spectral feature spaces based on the averaged short-band and long-band spectral feature matrices; Alternatively, short and long band spectral feature spaces can be constructed for each of the two orthogonal polarization channels based on the short-band spectral feature matrices and long-band spectral feature matrices of the two orthogonal polarization channels. The short and long band spectral feature spaces of the two orthogonal polarization channels can be averaged to obtain the short and long band spectral feature spaces used for clustering. Alternatively, based on the short-band and long-band spectral feature matrices of the two orthogonal polarization channels, short-band and long-band spectral feature spaces for classification can be constructed for each of the two orthogonal polarization channels. In step S3, the specific steps for extracting spectral features from the interference spectral signal to obtain the spectral feature matrix are as follows: First, the interference spectrum signal is divided into N spectral band signals, where N is an integer greater than 1. For each spectral band signal, a wavenumber-resolved OCT intensity signal is obtained through a short-time Fourier transform algorithm. The average system noise is removed from the OCT intensity signal along the depth direction to obtain the denoised OCT intensity signal; Linear fitting is performed on the depth direction of the denoised OCT intensity signal to obtain an intensity-depth curve. The slope of each position on the intensity-depth curve is used as the spectral feature of the current position to obtain the spectral features of the target tissue region and obtain an N-frame spectral feature matrix of the vascular lumen tissue. Alternatively, the spectral characteristics of the target tissue region can be calculated using the following formula: ； in, Current position of current band k (x,y,z) Spectral characteristics at that location x These are the coordinate values ​​for the fast scan direction in OCT detection scanning. y These are the coordinates of the slow scan direction during OCT detection scanning. z These are the coordinate values ​​in the depth direction, which is perpendicular to the plane formed by the fast and slow scan directions, i.e., the optical axis direction. Current position of current band k (x,y,z) The intensity of the denoised OCT signal. The physical size in air corresponding to each pixel in the depth direction. n The refractive index of the target tissue region.

2. The lipid plaque detection method combining OCT polarization and spectral information according to claim 1, characterized in that: The extraction method of the texture feature matrix in step S2 includes: The PS-OCT images of two orthogonal polarization channels are averaged, and then the texture features are extracted from the averaged PS-OCT image to obtain the averaged texture feature matrix. Alternatively, texture features can be extracted from the PS-OCT images of the two orthogonal polarization channels separately to obtain the texture feature matrices of each of the two orthogonal polarization channels.

3. The lipid plaque detection method combining OCT polarization and spectral information according to claim 1, characterized in that: The extraction methods for the spectral feature matrix in step S3 include: The interference spectral signals of the two orthogonal polarization channels are averaged, and then the spectral features of the averaged interference spectral signals are extracted to obtain the averaged spectral feature matrix. The averaged spectral feature matrix is ​​then divided into a short-wavelength spectral feature matrix and a long-wavelength spectral feature matrix. Alternatively, the spectral features of the interference spectral signals of the two orthogonal polarization channels can be extracted separately to obtain the spectral feature matrices of the two orthogonal polarization channels. The spectral feature matrices of the two orthogonal polarization channels can then be divided into short-wavelength spectral feature matrices and long-wavelength spectral feature matrices.

4. The lipid plaque detection method combining OCT polarization and spectral information according to claim 1, characterized in that: Step S2, which involves extracting texture features from the PS-OCT image, specifically includes: Extract one or more of the following from PS-OCT images: mean, variance, standard deviation, homogeneity, contrast, dissimilarity, entropy, second moment of angle, and correlation matrix; The classification model in step S4 includes one or more classification models based on convolutional neural networks, fully convolutional neural networks, U-net networks, and generative adversarial networks.

5. A lipid plaque detection system combining OCT polarization and spectral information for implementing the method of any one of claims 1 to 4, characterized in that: Includes PS-OCT imaging equipment, scanning equipment, and signal processing equipment; Or it may include a PS-OCT imaging device, an NIRS imaging device, a scanning device, and a signal processing device; The PS-OCT imaging device is used for intracavitary tissue imaging and spectral detection, the NIRS imaging device is used for intracavitary tissue spectral detection, the scanning device is used for intracavitary tissue imaging, and the signal processing device is used for lipid plaque analysis of the acquired PS-OCT and NIRS signals.

6. The lipid plaque detection system combining OCT polarization and spectral information according to claim 5, characterized in that: The lipid plaque detection system employs one of the following methods: The PS-OCT imaging device and the NIRS imaging device are connected to the same PS-OCT light source, and lipid plaques are identified by analyzing the spectrum of the PS-OCT light source. Alternatively, the PS-OCT imaging device and the NIRS imaging device can be connected to the PS-OCT light source and the NIRS light source respectively, and lipid plaques can be identified by analyzing the spectrum of the PS-OCT light source or the NIRS light source. The PS-OCT imaging device employs alternating A-line encoding, frequency reuse, depth encoding, cross-sampling, or parallel detection. The center wavelength of the PS-OCT imaging device's operating band is 1200-1350 nanometers, and the bandwidth is greater than 100 nanometers.

Images