Cardiovascular oct image plaque automatic recognition method and system based on deep learning

Abstract

The application provides a cardiovascular OCT image plaque automatic identification method and system based on deep learning, and relates to the technical field of medical image processing. The method comprises the following steps: locating a predetermined radial interval corresponding to the optical distance from the imaging prism to the outer wall of the guide catheter along the radial depth dimension of the original OCT image; detecting a ring-shaped high-brightness reflection band continuously distributed along the angle direction in the radial interval to obtain a catheter reflection artifact mask area; and constructing a radial illumination attenuation field based on the eccentric displacement vector of the imaging prism relative to the lumen center axis in the current frame and the attenuation coefficient of the probe light beam in the blood and the blood vessel tissue. The application improves the accuracy of deep learning in automatic identification of cardiovascular OCT plaques.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for automatic identification of cardiovascular OCT image plaques based on deep learning. Background Technology

[0002] In the clinical application of cardiovascular optical coherence tomography (OCT), the use of deep learning methods for automatic pixel-level segmentation of plaque components, such as fibrous caps, lipid pools, and calcified nodules, has become a research hotspot. However, when directly processing raw polar coordinate OCT images, the segmentation reliability of existing deep learning-based automatic recognition methods is easily affected by inherent physical artifacts and illumination inhomogeneities in the images.

[0003] Specifically, when the guiding catheter is positioned at the opening of the left coronary artery, a ring-shaped, bright, strongly reflective band corresponding to the outer wall of the catheter will appear in the original OCT image along the radial depth direction. This is the Fresnel reflection artifact of the catheter itself. This artifact is very similar to a thin fibrous cap or superficial calcified plaque in terms of brightness intensity and local morphology. At the same time, due to the off-center position of the imaging catheter within the lumen and the attenuation of the probe beam in the blood and vascular tissue, the image will show significant illumination attenuation in the circumferential and radial directions, forming an uneven distribution with one side brighter and the other side extremely dark. This results in a significant weakening of the signal of the real vessel wall structure far from the catheter. In the above situation, if directly... When uncorrected OCT images are fed into convolutional neural networks for segmentation, the networks often struggle to distinguish ductal reflection structures from real plaques based on physical mechanisms. In an imaging frame near the opening of a left main coronary artery, a bright reflection ring region adjacent to the duct is labeled as a fibrous cap with high confidence by the convolutional neural network, resulting in a false positive. Meanwhile, the real thin fibrous cap located in the dark area of ​​attenuated illumination in the same frame is missed due to insufficient signal contrast, and the boundary of the adjacent lipid pool is also prone to displacement or breakage. Such misidentifications caused by ductal artifacts and off-center dark areas limit the accuracy and robustness of existing deep learning-based automatic plaque analysis methods. Summary of the Invention

[0004] This invention provides a method and system for automatic identification of cardiovascular OCT image plaques based on deep learning, which improves the accuracy of automatic identification of cardiovascular OCT plaques by deep learning.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] In a first aspect, a method for automatic identification of cardiovascular OCT image plaques based on deep learning, the method comprising:

[0007] Step 1: Acquire raw OCT images in polar coordinate format for each frame, acquired in real time at the ostium of the left coronary artery via the guiding catheter;

[0008] Step 2: Along the radial depth dimension of the original OCT image, locate the predetermined radial interval corresponding to the optical distance from the imaging prism to the outer wall of the guiding catheter, detect the annular high-brightness strong reflection band continuously distributed along the angular direction within the radial interval, and obtain the catheter reflection artifact mask area; based on the eccentric displacement vector of the imaging prism relative to the central axis of the lumen in the current frame and the attenuation coefficient of the probe beam in blood and vascular tissue, construct the radial illumination attenuation field.

[0009] Step 3: Based on the catheter reflection artifact masking area, the artifact area in the original OCT image is shielded, and the illuminance compensation that is inversely mapped to the radial illuminance attenuation field is applied to the vessel wall area outside the masking area to obtain the standardized OCT image.

[0010] Step 4: Input the standardized OCT image into the pre-trained convolutional neural network to perform pixel-level semantic segmentation and obtain a pixel-by-pixel patch category probability map, including the probability distribution channels of calcified nodules, fibrous caps and lipid pools.

[0011] Step 5: Perform Boolean masking on the plaque category probability map using the catheter reflection artifact masking area to filter out the Fresnel reflection ring artifact area of ​​the catheter itself, extract the connected domain boundaries of each plaque type in the masked probability map and determine the plaque category, and obtain the automatic plaque identification result that excludes the interference of catheter artifacts and eccentric illumination dark areas.

[0012] Secondly, a deep learning-based automatic plaque recognition system for cardiovascular OCT images includes:

[0013] The acquisition module is used to acquire each frame of raw OCT image in polar coordinate format in real time at the opening of the left coronary artery by the guiding catheter;

[0014] The module is used to locate a predetermined radial interval corresponding to the optical distance from the imaging prism to the outer wall of the guiding catheter along the radial depth dimension of the original OCT image, detect the annular high-brightness strong reflection band continuously distributed along the angular direction within the radial interval, and obtain the catheter reflection artifact mask area; based on the eccentric displacement vector of the imaging prism relative to the central axis of the lumen in the current frame and the attenuation coefficient of the probe beam in blood and vascular tissue, a radial illumination attenuation field is constructed.

[0015] The standardization module is used to shield artifact regions in the original OCT image based on the catheter reflection artifact mask region, and apply illuminance compensation to the vessel wall region outside the mask region, which has an inverse mapping relationship with the radial illuminance attenuation field, to obtain a standardized OCT image.

[0016] The segmentation module is used to input standardized OCT images into a pre-trained convolutional neural network to perform pixel-level semantic segmentation and obtain pixel-by-pixel patch category probability maps, including probability distribution channels for calcified nodules, fibrous caps and lipid pools.

[0017] The identification module is used to perform Boolean masking on the plaque category probability map using the catheter reflection artifact masking area, filter out the Fresnel reflection ring artifact area of ​​the catheter itself, extract the connected domain boundary of each plaque type in the masked probability map and determine the plaque category, and obtain the automatic plaque identification result that excludes the interference of catheter artifacts and eccentric illumination dark areas.

[0018] Thirdly, a computing device includes:

[0019] One or more processors;

[0020] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0021] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0022] The above-described solution of the present invention has at least the following beneficial effects:

[0023] By locating and generating a catheter reflection artifact mask region, artifact interference is shielded during the image standardization stage, and illuminance compensation is performed on the vessel wall region through inverse mapping with the radial illuminance attenuation field. This effectively corrects the non-uniform signal attenuation caused by the eccentricity of the imaging prism. Simultaneously, after deep learning semantic segmentation, the same catheter reflection artifact mask is used again to perform Boolean masking on the plaque category probability map, eliminating false positive plaque regions caused by Fresnel reflection rings. This collaboratively solves the two types of problems—artifact false detection and dark area missed detection—within the same framework, improving the pixel-level segmentation accuracy and recognition reliability of plaque types such as calcified nodules, fibrous caps, and lipid pools. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of an automatic identification method for cardiovascular OCT images based on deep learning, provided by an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of an automatic cardiovascular OCT image plaque recognition system based on deep learning, provided in an embodiment of the present invention. Detailed Implementation

[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0027] like Figure 1 As shown, embodiments of the present invention propose an automatic plaque identification method for cardiovascular OCT images based on deep learning, the method comprising the following steps:

[0028] Step 1: Acquire raw OCT images in polar coordinate format for each frame, acquired in real time at the ostium of the left coronary artery via the guiding catheter;

[0029] Step 2: Along the radial depth dimension of the original OCT image, locate the predetermined radial interval corresponding to the optical distance from the imaging prism to the outer wall of the guiding catheter, detect the annular high-brightness strong reflection band continuously distributed along the angular direction within the radial interval, and obtain the catheter reflection artifact mask area; based on the eccentric displacement vector of the imaging prism relative to the central axis of the lumen in the current frame and the attenuation coefficient of the probe beam in blood and vascular tissue, construct the radial illumination attenuation field.

[0030] Step 3: Based on the catheter reflection artifact masking area, the artifact area in the original OCT image is shielded, and the illuminance compensation that is inversely mapped to the radial illuminance attenuation field is applied to the vessel wall area outside the masking area to obtain the standardized OCT image.

[0031] Step 4: Input the standardized OCT image into the pre-trained convolutional neural network to perform pixel-level semantic segmentation and obtain a pixel-by-pixel patch category probability map, including the probability distribution channels of calcified nodules, fibrous caps and lipid pools.

[0032] Step 5: Perform Boolean masking on the plaque category probability map using the catheter reflection artifact masking area to filter out the Fresnel reflection ring artifact area of ​​the catheter itself, extract the connected domain boundaries of each plaque type in the masked probability map and determine the plaque category, and obtain the automatic plaque identification result that excludes the interference of catheter artifacts and eccentric illumination dark areas.

[0033] In this embodiment of the invention, by locating and generating a catheter reflection artifact mask region, artifact interference is shielded during the image standardization stage, and illuminance compensation is performed on the vessel wall region by inverse mapping with the radial illuminance attenuation field. This effectively corrects the non-uniform signal attenuation caused by the eccentricity of the imaging prism. Simultaneously, after deep learning semantic segmentation, the same catheter reflection artifact mask is used again to perform Boolean masking on the plaque category probability map, eliminating false positive plaque regions caused by Fresnel reflection rings. This collaboratively solves the two types of problems of artifact false detection and dark area missed detection within the same framework, improving the pixel-level segmentation accuracy and recognition reliability of plaque types such as calcified nodules, fibrous caps, and lipid pools.

[0034] In a preferred embodiment of the present invention, step 1 above may include:

[0035] In this embodiment of the invention, during the clinical examination of optical coherence tomography (OCT), the imaging guide catheter is first inserted through the radial or femoral artery and pushed retrogradely along the arterial path to the opening of the left main coronary artery. The imaging window of the catheter is positioned precisely in the lumen of the left main coronary artery near the aortic sinus. After confirming that the guide catheter is in place, a detection beam is emitted at high speed using a single-wavelength near-infrared light source (center wavelength of about 1300nm) through the imaging fiber optic probe built into the catheter. The beam is reflected by a micro prism at the tip of the catheter and emitted along the direction perpendicular to the long axis of the catheter, performing a 360-degree circumferential scan of the vessel wall. At the same time, driven by the external retraction device of the catheter, the imaging probe is pulled back distally along the left main coronary artery at a constant rate, such as 20mm / s or 40mm / s, to achieve dynamic tomographic imaging of the entire vessel.

[0036] During each circumferential scanning cycle, the backscattered light interference signal acquired by the interferometer is Fourier transformed to obtain a one-dimensional depth-resolved reflectivity profile along the probe light propagation path, i.e., a single A-line signal. By stitching together hundreds of A-line signals arranged at equal angular intervals in the scanning order, a two-dimensional polar coordinate format original OCT image is formed with the imaging probe as the center, radial depth as the vertical axis, and scanning angle as the horizontal axis. In this image, each column represents an A-line depth scan line in an angular direction, and each row represents the circumferential pixel intensity distribution at a certain fixed radial depth. The acquisition system outputs these polar coordinate original images at a real-time frame rate, and each frame corresponds to the imaging result of the complete cavity cross-section of the probe at a certain axial position.

[0037] In a preferred embodiment of the present invention, step 2 above may include:

[0038] In this embodiment of the invention, step 220 involves defining a starting depth and an ending depth along the radial depth dimension of the original OCT image based on the known optical distance between the outer wall of the guiding catheter and the imaging prism, thus obtaining a predetermined radial interval. Specifically, after acquiring the original OCT image in polar coordinate format for the current frame, the radial depth interval corresponding to the reflection from the outer wall of the catheter needs to be determined from the image. The determination of this interval depends on the optical distance from the exit surface of the imaging prism to the end of the outer wall of the guiding catheter. This optical distance is not preset without basis, but rather a reliable parameter established based on the geometric optical calibration process before the catheter leaves the factory. The imaging guiding catheter must undergo several calibrations before leaving the factory or when it is first connected to the OCT system. For optical parameter calibration, a miniature prism is encapsulated at the end of the imaging fiber optic probe embedded in the catheter. A fixed physical optical path, composed of an optically transparent medium within the catheter cavity (such as air, saline solution, or a low-refractive-index filler), exists between the prism's exit face and the outer surface of the catheter's outermost transparent sheath. During calibration, the catheter tip is placed in a standard medium with a known refractive index, and the system acquires a still scan image. Due to the abrupt change in the medium's refractive index, a very strong Fresnel reflection occurs at the interface between the catheter's outer wall and the external medium. This reflection peak appears as a bright spike on the original A-line signal. The system automatically detects the optical delay of this reflection peak relative to the prism surface (reference zero-depth position) and uses the formula... Perform the conversion, among which For the measured optical distance, Group refractive index of the medium filling the catheter (provided by the catheter material handbook). For the physical distance to be determined, the system will calculate the physical distance. Refractive index of the filling medium It is also stored in the system configuration file corresponding to the unique identifier of the catheter, as the geometric optical parameters of the catheter model.

[0039] During step 220, based on the currently used catheter model, two parameters corresponding to that catheter are read from the configuration file: the physical distance from the prism exit surface to the outer wall of the catheter. and the group refractive index of the catheter lumen medium. Combined with the current OCT system's sampling resolution in the depth direction (The unit is micrometers per pixel, determined by the center wavelength of the light source and the sampling bandwidth of the spectrometer, and is a system constant.) The physical optical distance is converted into the pixel span in the image using the following formula: ;

[0040] In the formula, This indicates the number of pixel rows that need to be traversed from the prism surface to the reflecting surface of the outer wall of the conduit. This is the total optical distance. This represents the optical depth increment corresponding to a single pixel. In the original OCT image, the pixel row indices in the radial depth dimension are arranged from lightest to darkest, with row 0 fixed at the zero-depth reference position (i.e., the starting depth) of the imaging prism exit surface. Therefore, the starting depth of the predetermined radial interval is set to... =0, then the termination depth needs to be increased by a tolerance offset based on the theoretical pixel span. To fully cover the actual width of the reflective zone on the outer wall of the catheter. The values ​​are determined based on the following: the reflection peak on the outer wall of the duct is not an ideal single pixel, but rather exhibits a certain width due to the influence of the system's axial point spread function. Its full width at half maximum (FWHM) typically occupies a certain area in the depth direction. 1 pixel, The number of pixels at half maximum width and full width at half maximum (FWHM) of the reflection peak can be obtained during the system calibration phase; simultaneously, the optical distance during the calibration process... The measurements exhibit slight fluctuations, and their standard deviation... Pixel fluctuation after conversion To ensure that the reflection band is completely included under various permissible deviations, take This value is typically in the 3-5 pixel range; therefore, the termination depth is set to... The radial depth row index range thus defined [ , This constitutes a predetermined radial interval, which corresponds to the depth region in which the outer wall of the guide catheter produces Fresnel reflections in the image.

[0041] Step 221: Extract pixel intensity values ​​in all angular directions within a predetermined radial interval to obtain an intensity projection curve along the angular direction; specifically, this includes: receiving the predetermined radial interval generated in step 220, whose depth row index upper and lower bounds are respectively... and For the original OCT polar coordinate image ,in Indicates the pixel row index along the radial depth direction. Representing the pixel column index along the circumferential angular direction, the following angle-wise energy accumulation operation is performed to obtain a one-dimensional intensity projection curve; specifically, for each fixed angle column... = Extract from depth line in this angle direction = arrive = The intensity values ​​of all pixels within the range are counted, and these intensity values ​​are summed one by one to generate a single total intensity value as the projection response for that angle. This summation process iterates through all pixels. After listing the angles, a one-dimensional discrete intensity projection curve is constructed with the angle direction as the independent variable and the cumulative total intensity as the dependent variable. Its mathematical expression is =1, 2, ..., In the formula: Radial depth indexing of the original polar coordinate OCT image Angle Index The pixel grayscale value at that location; The starting depth row index for the predetermined radial interval corresponds to the zero-depth reference position of the imaging prism exit surface; The index for the termination depth of the predetermined radial interval corresponds to the rear boundary of the outer wall reflection zone of the catheter. Angle Index The intensity projection value at that point reflects the overall reflected energy within the catheter outer wall region along that angle. This summation operation is equivalent to performing energy integration within a fixed interval in the radial direction. Since the Fresnel reflection generated by the catheter outer wall has a continuous circumferential distribution and its intensity is significantly higher than the blood scattering background, the cumulative value at a single angle... It can effectively enhance the strong reflection signal of the outer wall of the catheter, while smoothing and canceling the randomly distributed blood scattering noise during the radial accumulation process, thereby improving the contrast between the target signal and the background noise.

[0042] Step 222: Apply an adaptive threshold to the intensity projection curve, and mark continuous angle segments with intensity values ​​consistently higher than the adaptive threshold as candidate strong reflection zones; specifically, this includes: using the intensity projection curve output in step 221... To process objects, =1, 2, ..., ,in To determine the total number of columns of circumferential angles in the image, first calculate the arithmetic mean of the one-dimensional curve along the angular direction. and standard deviation , The DC background level, which characterizes the projection intensity, is mainly contributed by blood scattering and tissue background signals; This reflects the degree of fluctuation and dispersion in projected intensity caused by strong reflection from the outer wall of the duct, and an adaptive intensity threshold is constructed accordingly. Its form is ,in For threshold sensitivity factor, The value is determined by the statistical distribution characteristics of the background component of the projection curve. When the non-reflective angular region on the outer wall of the duct dominates, the corresponding... They can be approximated as independent random variables following a Gaussian distribution, with their mean and standard deviation being respectively... and Based on the probability characteristics of the Gaussian distribution, if we want to control the probability of background noise being misjudged as strong reflection to an extremely low level (e.g., less than 0.13%), we can take... =3 achieves approximately 99.87% confidence boundary; to improve detection sensitivity for weaker reflection segments, the value can be appropriately reduced. Value; For the typical strong reflection scenario of Fresnel reflection on the outer wall of the OCT catheter, its reflection intensity is usually much higher than the background by several standard deviations. Therefore, while ensuring low false alarms, it is necessary to consider detection robustness. The standard setting is... =2.5, in clinical applications, The preset range is generally 2.0-3.0, and is optimized and fixed to 2.5 at the factory according to the catheter model and imaging environment.

[0043] Next, the projection curve is scanned point by point along the angular direction, and each angle is indexed. Binary labeling ,in =1 indicates that the angle point is a threshold point, corresponding to a potential location of strong reflection from the outer wall of the catheter; =0 indicates a point that does not exceed the threshold; after obtaining the binary labeled sequence, a one-dimensional connected component scanning algorithm is used to extract continuous strong reflection segments. The detailed algorithm process is as follows: initialize an empty interval list, set a traversal pointer... =1, when ≤ At that time, if =0 Increment and continue scanning; if =1, record the starting index of the current segment. = Then it continues to increase. until > or =0, at this point the record terminates the index. = -1, in this interval Add to list, and from the current Continue scanning; after the scan is complete, a set of continuous angle segments is obtained. ,in This is an index for consecutive segments, with values ​​ranging from 1 to... ; This represents the total number of effective continuous strong reflection segments that are ultimately retained after minimum width filtering.

[0044] To eliminate spurious short segments caused by localized scattering enhancement of blood or random noise, the angular span width of each candidate segment is calculated. And based on the preset minimum width threshold Filter and retain only ≥ The section, The setting has a clear physical basis. The outer wall of the imaging guide catheter is physically continuous throughout the entire circumference. Even if it is obstructed by the guidewire or tissue, its Fresnel reflection band should occupy at least a certain continuous angular range in the OCT image. Considering the lateral resolution of the OCT system (typically about 20-30 μm), the outer diameter of the catheter (usually 0.5-1.0 mm), and the arc length corresponding to each column of angle sampling, the minimum continuous strong reflection angle span will not be less than the number of pixel columns corresponding to about 10°. Combining the typical imaging angle resolution (for example, 900 columns correspond to 360°, then each column is 0.4°), 10° corresponds to about 25 columns. Taking into account the minimum allowable continuity under the worst-case scenario, we take... =5 pixel columns (corresponding to approximately 2°-4°, depending on the actual sampling density) can effectively filter out random noise segments with a width of 1-4 pixels, while completely preserving the true outer wall reflection zone of the duct. After this filtering, a final set of continuous angle segments is obtained as the circumferential coverage range of the candidate strong reflection zone.

[0045] Step 223: In polar coordinates, set all pixel positions corresponding to the candidate strong reflection band within a predetermined radial interval as foreground and the remaining pixels as background to form the duct reflection artifact mask region; specifically including: receiving the predetermined radial interval determined in step 220 [ , ], and the set of continuous angle segments corresponding to a set of candidate strong reflection zones output in step 222. Each angle segment is indexed by a starting angle. and termination angle index Define the binary mask image to be generated in polar coordinate image space with the exact same size as the original OCT image. This is used to accurately mark the area where the reflection artifacts on the outer wall of the catheter are located.

[0046] Let the original OCT polar coordinate image be... The radial depth of the rows is The number of circumferential angle columns is Its row index Corresponding radial position, column index To correspond to the angular coordinates, first create a two-dimensional matrix. Its dimensions are And initialize all elements to the background identifier 0, i.e. This matrix is ​​the binary mask image to be generated. Initially, the entire image region is marked as the background. Subsequently, each consecutive strong reflection segment in the candidate angle segment set is processed one by one. For the ... Each angle segment Take its starting column index and terminating column index In the mask image Perform a rectangular area fill operation on the top, filling all areas that meet the row coordinates. exist Within the range, and column coordinates exist The pixels within the range are set as foreground identifier 1. This operation can be expressed as follows: ,when and This rectangular region corresponds to a radial depth from […] in the polar coordinate image. Extending to , along the circumferential angle from Crossing to A continuous band of blocks, covering the reflection range of the duct's outer wall within that angle segment, traversing all... For each angle segment, the above filling operation is performed sequentially. After processing all angle segments, a complete catheter reflection artifact mask image is obtained. In this binary mask, foreground pixels with a value of 1 precisely correspond to the location of the bright ring artifact caused by Fresnel reflection from the outer wall of the guiding catheter in the original OCT image; background pixels with a value of 0 correspond to the vascular wall tissue, blood, and other non-catheter reflection areas.

[0047] Step 224: Obtain the eccentric displacement vector of the imaging prism relative to the central axis of the cavity in the current frame, and determine the actual optical distance of the imaging prism to the cavity wall surface in each angular direction; specifically including: based on the polar coordinate OCT image of the current frame... ,in =1, ..., Radial depth index (from shallow to deep). The sequence is a circumferential angle sequence; the catheter reflection artifact mask generated in step 223. Where 1 represents an artifact pixel and 0 represents a valid organized pixel; the coordinates of the imaging prism rotation center. The data is read from the system geometry calibration file; first, the lumen wall contour is extracted, and the artifact regions in each angle column are shielded using a catheter reflection artifact mask. Only valid pixels with a mask value of 0 are operated on. For each angle column, the original intensity is first smoothed radially by a moving average with a window width of 5 pixels to suppress speckle noise. The absolute value of the intensity difference between adjacent radial positions is calculated to obtain the gradient sequence of the entire column. The gradient sequence is scanned from the shallow end to the deep end to find the first radial index that meets the peak condition. The gradient at this point is a local maximum, and the amplitude is more than 3 times the median value of the gradient in this column. This index is used as a candidate point of the lumen wall. If a certain angle... If no peak value meeting the criteria is found in the degree column, the candidate point at that location is marked as invalid. After all angle columns have been processed, the resulting candidate wall radial indices are arranged by angle to form a one-dimensional sequence. Median filtering is applied to this sequence along the angle dimension. A sliding window with a width of 11 angle points is taken with the current angle as the center. The median of the valid values ​​within the window is used to replace the center value. If the difference between the original center value and the median exceeds a preset threshold of 10 pixels, replacement is performed. Invalid values ​​in the window are not included in the median calculation. After filtering, linear interpolation is performed on the still missing angle positions to finally obtain a circumferentially continuous and smooth radial index sequence of the lumen wall surface. The validity of each angle is also indicated.

[0048] Next, the lumen center is fitted. First, the above lumen wall contour point set is transformed from polar coordinates to Cartesian coordinates. According to the system calibration, the radial index is mapped to the physical distance. Combined with the corresponding angle, the point set is obtained through the standard polar coordinate-Cartesian mapping algorithm. ,in =1, ..., Furthermore, points corresponding to invalid angles are removed. Based on this, a robust circle fitting method based on algebraic distance is used to iteratively solve for the center of the lumen. The specific steps are as follows: Let the algebraic equation of the circle be... Then the coordinates of the center of the circle are ,radius For each contour point involved in the fitting It can list information about unknown parameters. linear equations All Solving the system of equations at each effective point, we obtain an overdetermined linear system of equations. ,in = , for A ×3 matrix, the first Behavior , for dimensional column vector, the first Each component is To minimize the sum of squared algebraic distances, we need to find the least squares solution. In actual calculations, normal equations can be directly constructed. The system of equations has a size of 3×3, and the parameters can be obtained by solving it. The initial estimate; if the matrix If the expression is close to singular, then use singular value decomposition. Perform direct decomposition to obtain a more stable least-squares solution. After obtaining the parameters, immediately calculate the initial circle center. and radius .

[0049] Calculate the algebraic residual for each point involved in the fitting process. Find the median of all the absolute values ​​of the residuals, and calculate the median absolute deviation (MAD) = 1.4826 × median ( -median(f)|), points with residual absolute values ​​greater than 3×MAD are identified as outliers and removed. The circle fitting process is then re-executed using the remaining inliers to update the parameters. ) and center, radius, repeat this process until the change in center coordinates between two consecutive iterations is less than The system outputs stable cavity center coordinates after reaching millimeters or the preset maximum number of iterations. and equivalent radius If the total number of valid points is too small, causing fitting failure, motion compensation interpolation is performed using the fitting results of adjacent frames; after obtaining the lumen center, the eccentric displacement vector can be calculated, and the coordinates of the imaging prism rotation center can be read from the calibration data. The center of this cavity and the center of the aforementioned lumen are located in the same Cartesian coordinate system, so the difference between the two can be calculated directly. This vector fully describes the offset and direction of the prism center relative to the geometric center of the tube cross-section.

[0050] Determine the actual optical distance at each angle, for each circumferential angle. From the center of the prism in the Cartesian plane Unit direction vector along that angle Construct a ray and calculate the geometric position of the intersection point between the ray and the circle of the tube. Take the intersection point along the positive direction of the ray as a point on the surface of the tube wall, and directly calculate the distance from the prism center to the intersection point using the Euclidean distance between the two points. This distance is the angle at which the probe beam is angled. The actual optical path length traversed by the upper part of the blood vessel to reach the lumen wall is used to obtain the actual optical distance vector at each angle. (Unit: mm), radial index vector of the tube wall surface and eccentric displacement vector .

[0051] Step 225: Decompose the actual optical distance into the blood-filling path length and the vascular tissue penetration path length, and couple these with the attenuation coefficients of the probe beam in the blood and vascular tissue, respectively, to obtain the cumulative path attenuation at each angle. Specifically, this includes: based on the actual optical distance vectors at each angle output in step 224... The unit is millimeters; the system's pre-calibrated blood attenuation coefficient. The unit is This coefficient, determined through in vitro measurements of blood optical properties, describes the exponential attenuation intensity per unit length of probe light propagating in blood. In typical intravascular optical coherence tomography (ICP-CT) wavelengths (approximately 1.3 μm), the common range for the whole blood attenuation coefficient is approximately 0.3 mm. -1 Up to 3.0mm -1 The specific value is affected by factors such as the wavelength of the light source, hematocrit, blood flow velocity, and red blood cell aggregation state.

[0052] First, path medium decomposition is performed. At the surface of the lumen wall, the probe beam just reaches the interface from the blood to the vessel wall and has not yet penetrated into the vessel tissue. Therefore, the entire optical path is completely filled by the blood medium, and the penetration length within the vessel tissue is zero. This clear medium boundary provides a unified initial condition for recursively calculating the attenuation factor along the radial depth. This means that at the lumen wall surface, there is no need to consider the influence of tissue composition differences on light attenuation, and the local attenuation at all angles is uniquely determined by the optical path length of the blood segment. The path cumulative attenuation factor at the lumen wall surface is calculated according to the Lambert-Beer law for light propagation in a homogeneous medium. When a monochromatic collimated beam passes through a thickness of... When the medium absorbs and scatters light, the intensity of the emitted light is... With incident light intensity The ratio shows an exponentially decreasing relationship. ,in The attenuation coefficient of this medium (including absorption and scattering contributions, i.e., the extinction coefficient) is used to apply this law to the current scene, for any angle. The optical distance the probe beam travels in the blood after exiting the prism is determined in step 224. The attenuation along this path is contributed solely by the blood medium, and the blood attenuation coefficient falls within the aforementioned calibration range. Therefore, the ratio of the residual illuminance at the lumen wall surface to the original illuminance is the path cumulative attenuation factor. , can be represented as To ensure numerical stability, calculations exceeding the normal range are performed. (For example, abnormal optical distance values ​​caused by tube obstruction or incomplete contour) are first truncated to avoid overflow in exponential calculations.

[0053] Perform recursive initialization, using the above calculations. Record the angles one by one as the starting values ​​for the two-dimensional attenuation recursion along the radial direction into the tissue depth in step 226. At the same time, record the corresponding radial index of the lumen wall surface. This information is also preserved and passed to subsequent steps to locate the physical starting point of the lumen surface on each A-line, enabling the recursive calculation to accurately extend from the blood-tissue interface into deeper tissues, obtaining the path cumulative attenuation vector at each angle of the lumen wall surface. and the corresponding radial index vector of the lumen wall surface. .

[0054] Step 226: Using each angular direction of the polar coordinate image as a reference, the cumulative attenuation is calculated along the radial depth from shallow to deep to obtain a two-dimensional radial illuminance attenuation field of the same size as the original OCT image. The value of each pixel represents the proportion of illuminance lost by the probe beam due to propagation attenuation at the corresponding location. Specifically, this includes: based on the cumulative attenuation vector of the cavity wall surface path at each angle output in step 225... The radial index vector of the lumen wall surface output in step 225 Image size ,in This represents the number of sampling points in the radial (depth) direction. Number of sampling points in the circumferential angle; tissue attenuation coefficient (Unit: mm) -1 The attenuation coefficient (AFC) describes the exponential attenuation intensity per unit length of probe light propagating within the blood vessel wall tissue. It includes contributions from tissue absorption and scattering. In intravascular optical coherence tomography (OCT), the near-infrared band (center wavelength approximately 1.3 μm) is commonly used, and the AFC typically ranges from approximately 2.0 mm. -1 Up to 6.0mm -1 Blood attenuation coefficient (Unit: mm) -1 ), consistent with the blood attenuation coefficient used in step 225; pixel radial physical spacing (Unit: mm) represents the actual physical distance between adjacent radial sampling points, which is determined by the radial resolution calibrated by the system.

[0055] First, create a decay matrix, and initialize a two-dimensional matrix with the same size as the polar coordinate image. row index =1, 2, ..., Corresponding to radial depths from shallow to deep (i.e., extending from the conduit towards the tube wall), the column indexes correspond to the circumferential angles. ( =1, 2, ..., All elements are initialized to 0. The elements in the text represent spatial locations. At this location, the ratio of the remaining illuminance to the original emitted illuminance after the probe light has attenuated through blood and / or tissue is defined as follows: For radial index Calculate in sequence In the formula This is the radial index of the current recursion; The attenuation value is the value of the previous radial position (shallower layer) in the same angular column; The tissue attenuation coefficient; The product is the radial physical interval of the pixels. This represents the total attenuation of the light beam as it passes through an area of ​​one radial pixel thickness. This is the transmittance, which reflects the multiple of light intensity retained after passing through a pixel. The recursive process is executed sequentially from shallow to deep along the radial direction, ensuring that the attenuation value of the deeper layers is accumulated and multiplied on the basis of the shallow layers.

[0056] Pushing into the lumen (blood zone, superficial direction) For radial index Reverse restoration of attenuation in the blood pathway In the formula This represents the attenuation value of a deeper layer (closer to the surface) within the same angle column; This refers to the blood attenuation coefficient. The meaning is the same as above. Since the light beam actually propagates from shallow to deep, if it travels back from the surface towards the light source, the light intensity should gradually increase, and the attenuation factor monotonically increases and approaches 1. This recursive formula can be obtained by multiplying by... To achieve a reasonable recovery of the shallow illuminance ratio and ensure a lower attenuation factor near the conduit in the lumen, which conforms to physical laws; if a certain angle is... Radial indexing of the corresponding lumen wall surface is caused by mask occlusion, signal loss, or contour detection failure. If invalid, then the recursive starting point of the column is generated using an angle-based neighbor interpolation method. The specific interpolation process is as follows: in the angle sequence, the current angle index is used... Search outwards from the center to both sides to find the nearest valid angle columns on the left and right, respectively denoted as... and And record the corresponding angle value. and Effective surface radial index and and surface attenuation value and If only one valid column is found (e.g., at the start or end boundary of the image), the valid radial index and surface attenuation value of that side are directly copied as the estimated value of the current column; if valid columns are found on both sides, linear interpolation is performed based on the angular position, and the interpolation weights are calculated first. ,in From the current perspective In the interval The relative position ratio on (0≤ ≤1), then estimate the surface radial index and surface attenuation value of the current column. , Interpolation Round to the nearest integer radial index as the equivalent lumen wall surface position for that column; and use... Using this position as the initial recursive value, the attenuation factor is calculated for both the deep and shallow regions according to the recursive rules of the effective columns mentioned above. After completing the recursion for all angle columns, a complete two-dimensional radial illuminance attenuation field is obtained. .

[0057] Improve the segmentation consistency and quantitative measurement accuracy of structures such as lumen boundaries, plaque composition, and support beams in OCT images.

[0058] In a preferred embodiment of the present invention, step 3 above may include:

[0059] In this embodiment of the invention, step 330 involves performing pixel-by-pixel position matching between the catheter reflection artifact mask area and the original OCT image, and setting the grayscale values ​​of pixels falling within the mask area to invalid values ​​to obtain an artifact shielding image that preserves the vessel wall area; specifically, the input data includes two two-dimensional images of identical size. The first image is a catheter reflection artifact binary mask image generated in step 223. This mask image is a two-dimensional matrix with a height of rows, corresponding to the radial depth direction in polar coordinate space, and a width of columns, corresponding to the circumferential angular direction. Each element in the matrix takes a value of 0 or 1: a pixel with a value of 1 marks the ring-shaped bright artifact area formed by Fresnel reflection from the outer wall of the catheter; a pixel with a value of 0 corresponds to the effective imaging area outside the artifact area, such as the blood vessel wall tissue, blood, and lumen background. The second image is the original polar coordinate format OCT image acquired in step 1, which is also a two-dimensional matrix of rows and columns. The gray value of each element reflects the quantized measurement value of the backscattered light intensity at the corresponding spatial location. The gray value range depends on the quantization bit depth of the imaging system. Based on this, an artifact masking image is created. A two-dimensional floating-point matrix with the same size as the original OCT image is allocated. The matrix is ​​initialized and assigned values. All elements of the matrix are traversed row by row and column by column. The value of each element is uniformly set to the floating-point non-numeric (NaN) defined by the IEEE 754 standard. Specifically, the assignment operation is performed for each combination of row indices from 0 to -1 and column indices from 0 to -1. After initialization, all pixels in the matrix are in the NaN state, which means that no effective image information has been assigned. NaN is selected as the unified marker of invalid values.

[0060] After matrix initialization, all pixel positions are traversed in row-major order. For each coordinate determined by the radial depth index and circumferential angle index, the binary label at that position is read from the duct reflection artifact mask image. If it is equal to 1, it indicates that the current coordinate is within the duct reflection artifact region, and the element at the corresponding position in the matrix retains its initial NaN value and is not modified. If it is equal to 0, it indicates that the current coordinate belongs to the effective imaging region, and the original gray value at that position is read from the original OCT image and directly assigned to the element at the same coordinate in the matrix, i.e., the assignment is performed. After the above traversal and conditional assignment operations are completed for all pixel positions, the matrix is ​​the desired artifact masking image.

[0061] Step 331: Extract the position indices of each effective pixel located outside the mask region in the artifact masking image, and retrieve the corresponding attenuation value from the radial illumination attenuation field based on the position index; specifically, this includes: based on the artifact masking image output in step 330... The effective pixel locations store the backscattering intensity of the blood vessel wall, while invalid pixels (such as catheter reflection artifact areas) have been marked as NaN; the two-dimensional radial illuminance attenuation field output in step 226 ,and Strictly of the same size, each element represents the proportion of remaining illuminance at the corresponding coordinate position in the lighting path. At the beginning of processing, an empty dynamic list is created and named... The list is initialized as a linear sequence container that supports on-demand expansion (e.g., a vector in contiguous memory or a linked list in linked storage), containing no records and having a length of zero. Each record to be stored is logically designed as a tuple structure that can hold a two-dimensional position coordinate and two scalar values, i.e. ( This is for subsequent unified storage and traversal; it also masks artifact images. Perform a full row-first traversal, using a double nested loop, with the outer loop using a radial index. Take 1, 2, ... in sequence. The inner layer for each fixed Let the angle index Take 1, 2, ... in sequence. During this process, each location is accessed pixel by pixel. And perform the following judgment and extraction operations,

[0062] Validity determination and location, grayscale recording, reading If the value at the specified location is equal to NaN (i.e., a non-numeric identifier as defined by the IEEE 754 standard), it means that the pixel has been identified as a catheter reflection artifact or masking area in step 330 and does not contain valid blood vessel wall tissue information. Therefore, it is ignored directly, no recording is made, and the process continues to the next location. If the value is not equal to NaN, the current location is determined to be a valid pixel, and its row index is immediately set. and column indexes It is saved as a spatial location identifier, and the original grayscale value of the pixel is recorded as... , here This represents the measured backscatter intensity of the tissue point without considering illuminance attenuation.

[0063] Extract the corresponding attenuation value from the attenuation field to obtain the position index of the effective pixel. Then, using the exact same coordinates As an access key, from the two-dimensional radial illuminance attenuation field Extract the corresponding local attenuation value Each element of the attenuation field matrix A quantitatively describes the remaining illuminance fraction caused by factors such as blood absorption, scattering, and tissue optical attenuation throughout the entire two-way optical path of the probe light, from the duct lens, through the blood and tissue to the spatial point, and then back along the same path. The physical meaning of this value is that, in the ideal case without any attenuation, the normalized illuminance at this location should be 1.0; in reality, some photons are lost, so the effective illuminance reaching this point and capable of exciting backscattering only accounts for a portion of the original emitted illuminance. This proportion is... In mathematics, Strictly limited to the interval (0,1], and always greater than 0; in computer implementation, It is stored as a double-precision floating-point number; to ensure the stability of subsequent division operations, the theoretical lower bound of 0 is often replaced with a very small positive number (such as...). or ),thus The actual typical values ​​are mostly distributed between 0.01 and 0.98. The area immediately adjacent to the outer surface of the catheter sheath has an extremely short optical path. It can approach 1.0; while in deeper tissues far from the ducts or in areas obscured by dense red blood cell shadows, It may be as low as 0.05 or even smaller, but it will never be zero.

[0064] Values ​​can be retrieved directly using simple row and column indices. Then, the local illumination attenuation factor specific to that effective pixel is extracted. It corresponds strictly one-to-one with the pixel coordinates, ensuring accurate spatial registration. At this point, all the necessary information about the current effective pixel is complete, including its row position. Column position Original grayscale value and local attenuation value These four data items are packaged into a single logical record. The storage format can be a structure containing four floating-point numbers, a tuple, or a sequentially concatenated pair. In the current implementation, it is uniformly stored in the form of (...). The quadruple representation of ) then calls the list. The append method at the end inserts the record entry to the end of the list; as the traversal progresses, each time a valid pixel is encountered, The length of the loop dynamically increases by 1, and its internal entries are strictly arranged according to the row-major traversal spatial order. This ordering characteristic facilitates subsequent fast backtracking based on position indexes or spatial coherence analysis. When the double loop traversal is complete... After tracing all pixel positions, the extraction of effective information from the entire artifact masking image is complete. At this point, the dynamic list... It has fully collected the three-dimensional information (position, gray level, attenuation) of all effective tissue pixels in the lumen region of the blood vessel wall, and each item faithfully corresponds to an unmasked tissue observation point in the image.

[0065] Step 332 involves taking the attenuation value in the radial illuminance attenuation field as input, obtaining the corresponding illuminance compensation coefficient according to a preset inverse mapping relationship, and then multiplying the illuminance compensation coefficient with the original grayscale value of the corresponding effective pixel for correction to obtain the uniform grayscale value after illuminance compensation; specifically, this includes: based on the effective pixel record set output in step 331... Each of these items is in the form of The quadruple records the row index, column index, and original grayscale value of the pixel, respectively. and the local attenuation value taken from the two-dimensional radial illuminance attenuation field. To stabilize the ideal inverse mapping, the compensation coefficient is defined as follows: In the formula, The preset values ​​need to take into account both the numerical lower bound of the attenuation field estimation and the system's background noise level. The attenuation field is usually constructed with a very small lower bound value. (For example This is to avoid situations where the denominator is zero later, but this lower limit is only used to ensure the safety of numerical calculations and does not represent the reliable measurement lower limit of actual physical illuminance; what truly determines the reasonable upper limit of the compensation coefficient is the noise equivalent gray value of the imaging system, when... When the value is too small, the corresponding original signal The noise has been severely attenuated, and its amplitude may be close to or below the noise floor. At this point, strong compensation will only amplify the noise to a level comparable to the tissue signal, creating false structures. Therefore, The value should be slightly greater than or equal to the attenuation threshold at which noise begins to dominate, in order to limit the maximum compensation gain. In a typical OCT / OFDI intravascular imaging system, the noise floor is approximately 0.1% to 1% of the maximum unsaturated signal intensity, corresponding to an acceptable minimum compensable illuminance ratio of approximately Therefore, select = Based on clear engineering principles, it ensures that the compensation coefficient in the region near the attenuation limit is approximately 500, while the compensation coefficient for pixels much smaller than this threshold is also blocked to within approximately 1000, avoiding excessive gain from invalid pixels. For systems with lower or higher noise levels, adjustments can be made accordingly. to or However, the overall level remained at Around the same magnitude, to balance the degree of illuminance restoration and the requirements for noise suppression, the preset... , is a fixed positive number.

[0066] Under this preset value, the illuminance compensation coefficient The value of is strictly limited to a finite interval, when the attenuation is weakest, i.e. When = 1, the compensation coefficient is at its minimum. With almost no additional amplification applied, it indicates that the original illuminance was already close to uniform. When the attenuation is extremely strong, When it approaches its physically valid minimum (e.g.) Even smaller, but constrained by ), The contribution to the denominator is far less than The compensation coefficient tends to its maximum value. This upper limit ensures that even under the most pessimistic optical path, the grayscale magnification will not exceed 1000 times, thus protecting the overall image quality; for the vast majority of effective pixels located within the blood vessel wall lumen, the attenuation value... The compensation coefficient is typically distributed between 0.01 and 0.98. The main activity ranges from approximately 1.02 to 100, within which illuminance recovery is effective and noise amplification is controllable. By this definition, the inverse mapping transforms the unbounded ideal reciprocal compensation into a bounded stable mapping, maintaining a reasonable balance of signal-to-noise ratio while restoring grayscale consistency.

[0067] Based on the aforementioned pre-defined reverse mapping relationship and For the set of valid pixel records Perform a line-by-line traversal and update, processing each entry in the set as an object, and sequentially execute the following sequential operations to read the attenuation value of that pixel from the entry. This value comes directly from the decay field. The elements at the same position represent the remaining illuminance share, according to the formula The illuminance compensation coefficient is calculated using double-precision floating-point numbers to ensure numerical accuracy. Here, the regularization parameter is directly involved in the calculation as a constant denominator increment, without needing to be reset for each pixel; the original grayscale value is read directly. That is, the measured intensity at that location in the artifact masking image, and perform multiplicative compensation. This product stretches the original grayscale by the required magnification factor, resulting in a nearly uniform grayscale representation of tissue points from different optical path attenuations, and the newly generated compensation coefficient... and homogenized gray values Append to the current record to form an extended tuple Replace or update entries in the list in place, maintaining the original row-first traversal order. When the traversal is complete, the set is... All valid pixels have undergone illuminance compensation correction, and each record carries complete compensation process parameters.

[0068] Step 333: Keep the uniform grayscale values ​​of all valid pixels unchanged, and keep the invalid values ​​of invalid pixels unchanged, to reconstruct a standardized OCT image of the same size as the original OCT image; specifically, this includes: based on the artifact masking image output in step 330. Size and the defined invalid value identifier (For example, NaN in the IEEE 754 standard, or a sentinel value of -1.0 that exceeds a specific grayscale range); the effective pixel record set output by step 332 Each record contains a row index. Column index and the homogenized gray value after illuminance compensation (All are finite floating-point numbers), create a blank image matrix and initialize it with invalid values, and construct a two-dimensional array based on the input size information. This makes it have the exact same shape as the artifact masking image. To fully preserve the numerical precision and dynamic range of the compensated grayscale, the data type of this matrix is ​​generally selected as double-precision floating-point (float64) or single-precision floating-point (float32). Subsequently, for each element of the entire matrix, from row 0, column 0 to... Line number Unify and assign invalid values This filling operation must logically be strictly consistent with the invalid values ​​used in the mask image of step 330. The specific implementation depends on the selected numerical computation environment and the data characteristics of the invalid values. If the invalid value used in step 330 is NaN (usually a special floating-point value defined by IEEE 754), then the corresponding NaN constant (such as numpy.nan, float('nan')) is used to fill all positions of the matrix. In practice, batch initialization functions provided by numerical libraries (such as numpy.full(( , The function `np.nan` generates a matrix in one step where all elements are NaN, avoiding pixel-by-pixel assignment. Importantly, the system-defined NaN value must be used here, rather than a custom floating-point number, to ensure that it is completely consistent with the binary representation of invalid values ​​in the upstream mask image, so that the subsequent invalid value detection function can work correctly.

[0069] If step 330 uses a numerical sentinel (such as -1.0, and this value cannot conflict with any valid grayscale value), then the same value is used for filling. During initialization, a matrix with all elements set to zero can be generated first, and then the entire matrix can be assigned the value -1.0. The key to this approach is that the selected value must be literally identical to the invalid value set in step 330 (consistent type and literal value) to avoid logical errors caused by implicit conversions between integers and floating-point numbers (e.g., -1 and -1.0 are judged to be unequal in some weakly typed scenarios). Through the above initialization process, the image matrix... Every cell in physical memory carries an unambiguous invalidation mark, and there is no undefined state that has not been assigned a value. At this point, the image is visually equivalent to a completely masked blank image.

[0070] Fill the compensated grayscale values ​​according to the effective pixel records, and iterate through the effective pixel record set generated in step 332. For each entry in the set, extract its position index. and the compensated grayscale value ,by For row number, As column index, directly in the matrix Positioning to the first Line number The column's cells, and the invalid value that the cell originally held. Overwrite as ,because Only valid tissue pixels identified as non-artifacts in step 330 are included. Therefore, all write operations will not fall on the catheter imaging area, blood residue artifacts, or other masked locations. For these invalid locations, the invalid values ​​assigned from the initialization stage will remain unchanged throughout the process and will not be contaminated by any valid grayscale values. The filling process can be implemented by traversing the records and assigning values ​​point by point. In terms of code implementation, it is usually a single loop that writes directly according to coordinates. If the record set itself is already arranged in order according to pixel position, batch indexing operations can also be used to speed up the writing, but the logical effect is the same: the compensated grayscale of each valid pixel is accurately placed in its corresponding spatial position, while invalid pixels retain the invalid mark in place.

[0071] when After all records in the matrix have been processed, The effective pixel areas are filled with uniform grayscale values ​​after illumination compensation, while the invalid areas retain invalid value identifiers that are completely consistent with the input masking image. The image as a whole has no undefined areas other than holes or missing values, achieving a unified approach to grayscale standardization of effective information and mask protection of invalid information; Standardized OCT Image , for one The image contains a floating-point matrix where all valid pixel locations identified as the vessel wall lumen store the uniformized grayscale values ​​after illumination compensation. The values ​​have been stripped of the effects of optical attenuation and only reflect the relative backscattering characteristics of the tissue itself; all locations marked as artifacts or invalid areas in step 330 retain the same invalid values ​​as the artifact-masking image. .

[0072] Enhance the inherent contrast between different tissue layers in the image to avoid false bright spots or pseudo-structures caused by noise spikes in the reconstructed image.

[0073] In a preferred embodiment of the present invention, step 4 above may include:

[0074] In this embodiment of the invention, step 440 involves inputting a standardized OCT image into the encoding branch of a pre-trained convolutional neural network to extract multi-scale deep feature maps characterizing the microstructure of the blood vessel wall and the optical features of plaques. Specifically, this includes: the pre-trained deep convolutional neural network backbone typically requires a three-channel RGB image as input, and the weights are trained according to the statistical characteristics of the three-channel distribution. To fully utilize the low-level feature extraction capability of the pre-trained weights, single-channel... Copy and expand along the channel dimension to a three-channel image The specific operation involves filling the same grayscale value into the red, green, and blue channels respectively, so that... In any pixel The three channels at the location have equal intensity and are all original grayscale values. This processing does not change the spatial structure of the image, but only matches the input shape of the pre-trained network.

[0075] A pre-trained encoding branch architecture is constructed. The backbone network of the encoding branch is selected from the Residual Network (ResNet) architecture pre-trained on large-scale natural image datasets such as ImageNet, preferably ResNet-34 or ResNet-50. The global average pooling layer and fully connected classification layer at the end of the original network are removed, and only all convolutional layers, batch normalization layers and activation layers from input to the fourth stage output are retained. This pre-trained network has learned rich edge, texture and shape representations from massive natural images, which can be transferred to OCT images to identify optical feature changes in fiber cap thickness, lipid pool boundary and calcification surface. All convolutional layer weights are initialized with pre-trained values, and the mean and variance estimates of the batch normalization layer are also directly adopted from the pre-trained statistics or adjusted according to the fine-tuning strategy.

[0076] Define the forward computation flow of the encoding branch, which consists of four downsampling stages connected in series. The spatial resolution of the output of each stage is halved successively, while the number of channels increases several times, forming a multi-scale feature pyramid. The first stage (shallow feature extraction) is performed on the three-channel input. A 2D convolution with a kernel size of 7×7 and a stride of 2 is applied, with 64 output channels. After this convolution, batch normalization is performed, and a ReLU non-linear activation function is applied to each element. The convolutional feature map is then further downsampled using a max pooling operation with a kernel size of 3×3 and a stride of 2. This final output feature map is obtained from this stage. The space dimensions are With 64 channels, pooling and large-step convolution preserve a significant amount of spatial detail. The first stage captured subtle textural variations and micro-boundary cues within the pipe wall; the second stage, with The input is processed through a set of residual blocks. In ResNet-34, the second stage contains three residual blocks. Each residual block consists of two convolutional layers with a kernel size of 3×3, followed by batch normalization and ReLU activation. The output of a residual block equals its input plus the output of the second convolutional layer. If the number of input and output channels is inconsistent, a 1×1 convolution is used in the skip connections for channel alignment. In the first residual block of this stage, the stride of the first convolutional layer is set to 2 to achieve 2x downsampling, while the stride of the convolutions in the remaining residual blocks is 1. After this stage, the spatial resolution is further halved, the number of channels increases to 128, and the output feature map is... The size is At this point, the features begin to encode more complex tissue region attributes, such as continuous thickness variations of the fibrous cap; in the third stage, The input is fed into the third stage, which contains 5 residual blocks in ResNet-34. The structure is similar to the second stage. The first convolutional stride of the first residual block is 2 to achieve downsampling, and the remaining strides are 1. The output feature map is then processed. The size is The number of channels is 256; the receptive field at this scale covers a larger area of ​​the vessel wall, and can perceive the regional aggregation and dispersion of lipid pools and the multi-point distribution pattern of calcified nodules; the fourth stage (deep semantic encoding). The input to the fourth stage contains three residual blocks. The first residual block is downsampled, and the output is a feature map. Size is With 512 channels, this feature map has the largest receptive field and the strongest semantic abstraction ability. It directly responds to the high brightness and strong attenuation features of calcified nodules, the uniform and moderate signal of fibrous caps, and the low scattering region of lipid pools, providing global contextual information for plaque category identification.

[0077] Step 441: The multi-scale deep feature map is passed to the decoding branch for spatial resolution restoration and cross-layer feature fusion to reconstruct a semantic feature tensor including boundary information. Specifically, the decoding branch adopts a symmetrical U-shaped structure, and through step-by-step transposed convolutional upsampling and cross-layer skip connections, it gradually parses the highly abstract semantics in the deepest feature map and fuses them with shallow details, ultimately restoring a semantic feature tensor with the same spatial resolution as the original input and containing clear organizational boundaries. The first layer of decoding (by...) Fusion generate ), at the deepest level As initial features, a 2x upsampling is first applied through a transposed convolutional layer. This transposed convolution has a kernel size of 2×2, a stride of 2, 512 input channels, and 256 output channels. The weights of the transposed convolution are randomly initialized and updated during training. It uses learnable interpolation parameters to upsample the spatial size from... Zoom in At the same time, the number of channels is compressed by half to 256 to obtain the upsampled feature map. .

[0078] Will Feature map at the same resolution as the coding branch (size The feature map is concatenated along the channel dimension (256 channels), resulting in a feature map with 256 + 256 = 512 channels. This concatenation operation seamlessly connects deep semantics with mid-scale region features, providing context for boundary localization. Subsequently, two consecutive convolutional layers are applied to the concatenated feature map for deep fusion. Each convolutional layer has a kernel size of 3×3 and a stride of 1, with boundary padding to maintain the spatial dimensions. The first convolution has 512 input channels and 256 output channels; the second convolution has 256 input and output channels. Each convolutional layer is followed by batch normalization and a ReLU activation function, ultimately outputting a fused feature map. Its size is Number of channels: 256.

[0079] Second layer decoding (by...) Fusion generate ),Will As input, upsampling is performed using a transposed convolution with a kernel size of 2×2 and a stride of 2. The number of input channels is 256, and the number of output channels is set to 128. The spatial dimensions are determined by... Double to The upsampled feature map is obtained. ;Will With coding branches (size The fusion feature map is formed by concatenating 128 channels along the channel dimension, resulting in a fusion feature map with 128 + 128 = 256 channels. Two consecutive 3×3 convolutions are then performed on this fusion map. The first convolution reduces the number of channels from 256 to 128, while the second convolution maintains 128 channels. Each convolution is followed by batch normalization and ReLU, and the output feature map is then generated. ,size Number of channels: 128.

[0080] The third layer of decoding (by...) Fusion generate ),right Perform transposed convolution upsampling with a kernel size of 2×2 and a stride of 2, with 128 input channels and 64 output channels. The spatial dimensions are determined by... Expand to ,produce ;Will With shallow detail features (size The fusion feature map, with 64 channels, is concatenated along the channel dimension to obtain 128 channels. It is then processed by two 3×3 convolutions: the first reduces the number of channels from 128 to 64, while the second maintains 64. Each layer undergoes post-batch normalization and ReLU activation, resulting in the final output. The size is With 64 channels, the feature map has now incorporated rich boundary and texture information.

[0081] right Perform another transposed convolution upsampling, with a kernel size of 2×2, a stride of 2, 64 input channels, and a preset number of output channels. (For example, 128 or 256 can be used), directly restoring the space size to the original input size. The upsampled feature map is obtained. Immediately afterwards, to Apply a core with a size of 3×3, a step size of 1, and a maximum number of input / output channels. The convolutional layer is followed by batch normalization and ReLU. This convolution is used to smooth the pixel block effect caused by upsampling and integrate neighborhood information, outputting the final semantic feature tensor. .

[0082] Step 442 involves performing pixel-by-pixel activation calculations on the semantic feature tensor using multi-channel classification convolutional kernels to obtain independent channel probability distributions corresponding to calcified nodules, fibrous caps, and lipid pools. Specifically, this includes constructing pixel-level classification convolutional layers to achieve the mapping from high-dimensional semantic features to plaque category scores. A classification mapping head consisting of 1×1 convolutions is added on top. This convolutional layer does not change the spatial size of the feature map, but only performs independent linear transformations and implicit nonlinear combinations on the channel vectors of each pixel. The specific settings are as follows: convolution kernel size 1×1, indicating that each output channel only performs point-by-point cross-channel weighted summation of the input features, without spatial neighborhood information mixing; input channel number Csem; output channel number K=3, corresponding to the three types of plaque components to be identified: calcific nodule, fibrous cap, and lipid pool; stride 1; no spatial downsampling is used; padding with 0 (for 1×1 convolution, padding with 0 does not affect any value); bias is enabled with learnable bias vectors. = Weight initialization: Convolutional weights are randomly initialized (e.g., He or Glorot initialization), and biases are initialized to zero. All parameters are optimized through backpropagation during subsequent training. The weights of this convolutional layer can be organized into a four-dimensional tensor. Reshape it into a two-dimensional matrix For spatial location eigenvectors at location Convolution is equivalent to matrix multiplication. ,in This is the unnormalized score (logits) for the three categories corresponding to this pixel. Corresponding calcified nodules, Corresponding fiber cap, Corresponding to the lipid pool, this linear mapping will The 3D features are compressed into 3D discriminative information, while retaining the linear combination ability of the features of each channel.

[0083] logits The numerical range of is unrestricted and cannot be directly used as a probability interpretation. To obtain the normalized probability of each pixel belonging to each patch category, the Softmax activation function is applied along the channel dimension (i.e., between the three categories). This function has differentiable and exponentially normalized properties, and its calculation form is as follows: In the formula For spatial location The logit value of the k-th class (from the aforementioned convolution output); The natural exponential function maps any real number to the positive real number field, guaranteeing a non-negative probability; the denominator... To exponentiate and sum all three types of logits, and use the sum as a normalization factor, we ensure that the sum of the three probabilities is always 1; For the final output of the k-th class probability value, satisfying and The Softmax operation is performed on a pixel-by-pixel basis, meaning it is applied to each pixel individually. The above calculations are performed on the three-dimensional vectors in coordinates, without cross-pixel interaction. The calculations can be efficiently parallelized in a deep learning framework using a channel-dimensional softmax layer, resulting in a size that is still [value missing]. The probability tensor P is calculated and then split into three independent two-dimensional probability graphs along the channel dimension. Take the 0th channel of P, which is the first type probability value for all pixel locations, corresponding to the confidence level of the calcified nodule. Take the first channel of P, which corresponds to the confidence level of the fiber cap. Take the second channel of P, which corresponds to the confidence level of the lipid pool. Each probability map is... The floating-point matrix has values ​​between 0 and 1. For invalid pixel regions in the original image (such as the outside of the lumen or the guide wire artifact region), the feature vector at the corresponding position can be set to zero or ignored by the preceding mask.

[0084] Step 443 involves spatially stacking the probability distributions of independent channels according to channel dimensions to obtain a plaque category probability map whose dimensions are aligned with the standardized OCT image. Specifically, this includes determining the stacking order and data structure, and, based on clinical reading habits and anatomical importance, defining the channel arrangement order of the three plaque components as follows: Channel 0 contains the probability of calcified nodules. The probability of the fiber cap in channel 1 The probability of the second channel lipid pool This order is fixed for all subsequent analysis, visualization, and quantitative algorithms, ensuring consistency in channel semantics. A spatial stacking operation is performed, using tensor stacking functions (such as torch.stack, tf.stack, or np.stack) along the channel axis (the last dimension, axis=-1) to combine the three probability maps into a single three-dimensional tensor within deep learning frameworks (such as PyTorch, TensorFlow) or numerical computing environments. The specific operation is equivalent to initializing a tensor with shape... All-zero tensors Then assign values ​​layer by layer:

[0085] For all and Performing the above assignment is a pure data recombination process; it does not perform any numerical transformations, activations, or normalizations. Therefore, it does not change the already labeled probability values, nor does it introduce new parameters. Since the classification mapping head in step 442 uses 1×1 convolution and pixel-wise softmax, the spatial coordinates of its output probability map are consistent with the input semantic features. Strict point-to-point correspondence; and The image size is restored to its original polar coordinates after upsampling in the decoding branch. Furthermore, without any cropping or shifting, the final probability map spatial coordinates are identical to the standardized OCT image output in step 333. Perfect alignment, specifically represented by a probability map. The Line number The column pixels correspond to the depth coordinate index in the polar image. The angle column index is The location; the probability vector at that location. The tendency of the tissue spot to be quantified on the three plaque components was directly quantified.

[0086] Achieve high-fidelity reconstruction of tissue boundaries, and improve the objectivity, repeatability and full automation of quantitative analysis of plaques in OCT images.

[0087] In a preferred embodiment of the present invention, step 5 above may include:

[0088] In this embodiment of the invention, step 550 involves using the duct reflection artifact mask area as a Boolean filtering template to perform a logical zeroing operation on the pixel probability values ​​at the corresponding coordinate positions in the patch category probability map, eliminating the interference area of ​​the duct Fresnel reflection ring, and obtaining a clean patch probability map. Specifically, this includes: using the duct artifact mask to perform spatially selective filtering on the probability map, uniformly setting the probability values ​​of the three channels of all pixels marked as artifacts to zero to prevent artifact areas from being misjudged as patch components by subsequent threshold segmentation, mask alignment and shape verification, and verification... shape and Are the first two dimensions consistent (both are)? If the previous steps ensured that the two were generated synchronously, this alignment will naturally hold, requiring no additional coordinate transformation; every spatial coordinate of the two will be identical. All of them point to the same tissue location in the polar coordinate system, so they can be directly mapped using simple logical indexing, thus revealing the two-dimensional artifact mask. Expanded into a suppression factor map with the same dimensions as the probability map. This is to enable element-wise multiplication, specifically by performing multiplication for each spatial location. Read mask value Define the inhibitory factor at this position that affects all channels. =1 (artifact), then the factors of all three channels at that position are set to 0; if If the value is 0 (not an artifact), then all factors are set to 1. ,in In practice, this can be achieved by inverting the two-dimensional mask and adding a third dimension of length 1 (e.g., ...). It automatically expands into a three-dimensional shape using a broadcast mechanism. No need to explicitly allocate memory for a three-dimensional array.

[0089] Pure probability diagram Through the original probability diagram With inhibitory factor plot Element-wise multiplication yields In the formula This represents the channel index, where 0 corresponds to calcified nodules, 1 corresponds to fibrous caps, and 2 corresponds to lipid pools. For the filtered first Class probability value; This is the inhibitory factor for this category at this position, taking either 0 or 1. This multiplication applies to all... Synchronous execution on each element can be efficiently accomplished using operators from deep learning frameworks or broadcast multiplication in NumPy, for any pixel marked as an artifact. ,because At this position, all channels are 0; the multiplication operation sets all three channel probabilities to zero. In subsequent processing, the probability sum of the three channels for these pixels is zero, so they are naturally regarded as unorganized areas (background) and will not be activated by the confidence threshold of any patch category. For effective organized pixels that are not artifacts, the factor is 1, the probability value is preserved as is, and the three-channel sum is still approximately 1 (affected by floating-point rounding), thus maintaining the integrity of the probability meaning. Maintain and Same data type (usually float32), shape ( The probability value of the artifact region is set to zero precisely, and no numerical perturbation is introduced into the other regions, nor is the relative probability ratio between effective pixels changed.

[0090] Step 551: Apply a preset segmentation confidence threshold to each probability distribution channel in the clean patch probability map to obtain a binarized region mask for each patch type; specifically, this includes the clean probability map output in step 550. ,size The three channels represent the probabilities of calcified nodules, fibrous caps, and lipid pools, respectively. The probability values ​​for artifact regions are all zero. The sum of the three channel probabilities for each pixel in the effective tissue region is 1. A preset segmentation confidence threshold is used. All are scalar values ​​between 0 and 1, acting on the probability channels of calcified nodules, fibrous caps, and lipid pools, respectively. The specific criteria and typical values ​​for the thresholds are as follows: Calcified nodule threshold. Calcification appears as a region of strong reflection and clear boundaries in OCT images. Deep neural networks typically exhibit highly concentrated responses to this feature, with probability values ​​often close to 0 or 1. To suppress discrete false positives caused by imaging noise or minor artifacts, a higher threshold should be selected. The recommended threshold is obtained by fine-tuning the threshold based on an independent validation set, using the criterion of maximizing the Youden index (Youden's index = Sensitivity + Specificity - 1). =0.75; Fiber cap threshold The fibrous cap is a band-like structure covering the surface of the lipid pool. It has clear boundaries but significant thickness variations. In the probability map, areas with moderate to high values ​​may exhibit tailing. A medium threshold is recommended to balance contour integrity and boundary sharpness. =0.50; Lipid pool threshold Because lipid pools have weak signal attenuation and blurred boundaries, the maximum probability of network output may be lower than that of calcification or fibrous caps. However, the interpretation of vulnerable plaques requires high sensitivity. Therefore, a relatively low threshold is used to ensure the detection rate. =0.35. All of the above thresholds can be adjusted by physicians according to diagnostic needs. The high specificity mode can be uniformly increased by 0.1 to 0.15, and the high sensitivity mode can be uniformly decreased by 0.1 to 0.15. The system defaults to using the above-mentioned verified recommended values.

[0091] Thresholding is performed independently on each category channel of the purity probability map, generating three binarized foreground masks to provide candidate regions for subsequent connected component analysis. Thresholding is performed channel by channel, traversing all spatial locations of the probability map. For each location, perform a comparison operation by channel; for calcified nodules (k=0), if... ,but ,otherwise ; Fiber cap channel (k=1), if ,but ,otherwise ; Lipid pool channels (k=2), if ,but ,otherwise Formal expression as ,in .

[0092] The multi-class characteristic of threshold determination means that, since the thresholds for each channel are applied independently, a pixel may simultaneously satisfy the conditions of multiple channels, meaning the same pixel may meet the conditions of multiple channels. , , The probability of each element being 1 allows the algorithm to explicitly preserve overlapping class regions, enabling subsequent steps (553) to arbitrate the region probability peaks and complete the final single-class assignment, avoiding boundary truncation caused by hard classification. Since the probability value of artifact regions is already zero, they will not meet any threshold conditions, and all channel masks will have a value of 0, automatically eliminating interference. Three binary masks are used. , , All based on size It is stored as a Boolean or integer matrix, where the foreground value is 1 and the background value is 0. All operations are element-wise comparisons and can be completed in one go based on vectorized computation.

[0093] Step 552: Perform connected component topology analysis on each binary region mask to extract the outer contour boundary coordinate sequence of each independent patch region and calculate the region's geometric topology attributes. Specifically, this includes: performing connected component labeling on each binary mask, extracting the outer contour coordinate sequence of each independent connected region, and calculating the region's geometric topology attributes, including area, perimeter, centroid, minimum bounding rectangle size, and eccentricity. Attribute calculation uses contour point sets or contour moments. For connected component labeling and contour extraction, 8-adjacent connected component analysis (such as OpenCV connectedComponentsWithStats or an equivalent algorithm) is used on each binary mask Bk. A unique label is assigned to each connected component. For each region with a label greater than 0, its outer contour boundary point sequence is extracted. The outer contour is extracted using a boundary tracking algorithm (such as the Suzuki-Abe algorithm) and stored as a closed point set. The point set is arranged in clockwise or counterclockwise order, representing the region's outer boundary in polar coordinates. Location. Geometric topological property calculation: For each connected component, calculate the following properties based on its contour point set or internal pixel set (all calculations are based on polar coordinate pixel units, without conversion to Cartesian coordinates to ensure dimensional consistency). The area is the total number of pixels contained in the connected component; if physical area is required, it can be converted in subsequent steps based on pixel size. The perimeter is the pixelated length of the calculated outer contour, using the sum of Euclidean distances between contour points (adjacent points are counted as unit distance or diagonal distance). The pixel length of the contour is obtained, and the centroid coordinates are calculated using the ratio of the first and zero moments of the contour to the centroid of the region. The algorithm calculates the arithmetic mean of the coordinates of all pixels within the contour, and then uses the rotating caliper method to find the minimum area bounding rectangle of the contour point set, recording its width. ,high (All in pixel coordinates) and rotation angle; eccentricity is calculated by constructing the covariance matrix using the second-order central moments of the contour and then calculating its eigenvalues. ( ), eccentricity is defined as This value is a dimensionless number between 0 and 1, describing the degree to which the shape of the region deviates from a circle (0 represents a circle, and the larger the value, the flatter and longer it is). The calculation step only uses the covariance matrix eigenvalue algorithm and does not expand the matrix expression. If the same pixel is contained by multiple category masks, this stage retains independence and will be arbitrated according to the probability peak in step 553.

[0094] Step 553: Based on the confidence peak value of pixels within the region enclosed by each outer contour boundary sequence in the corresponding probability distribution channel, determine the patch category to which each connected component belongs; specifically, this includes: the list of connected component records output in step 552, where each record contains the following fields. Connected Components The set of coordinates of all pixels inside is represented as ,in The set of row and column indices in polar coordinates is obtained in step 552 by contour filling or connected component labeling after extracting the outer contour, ensuring that all pixels within the covered area are included; area represents the pixel area of ​​the connected component. ,Right now The number of elements; the probability map of clean patches output in step 550. ,size The three channels are arranged along the last dimension. Channel index 0 corresponds to the probability of calcified nodules, channel 1 corresponds to the probability of fibrous caps, and channel 2 corresponds to the probability of lipid pools. The probability values ​​of each spatial location are non-negative. The sum of the three channels in the effective region is close to 1, while the probability of the three channels in the artifact region is 0. For each connected region generated in step 552, the maximum confidence level in the region is independently calculated on the three channels of the purity probability map. By comparing the peak values ​​of the three channels, the single patch category to which the connected region finally belongs is determined, which resolves the cross-category overlap ambiguity that may be caused by threshold segmentation, and removes regions that do not meet the minimum confidence standard or have too small an area.

[0095] Pixel coordinate extraction from connected components: read from the currently processed connected component records. The field retrieves the region. The set of all pixel coordinates In the formula For the first The row index (depth direction) of a pixel in the polar coordinate image, with a value range of 0 < < , For the first The column index of each pixel (angular direction), with a value range of 0 < < ; Let be the total number of pixels in the connected region, i.e., the area of ​​the region, and ≥1.

[0096] For each plaque category channel (In order, calcification, fibrous cap, lipid pool), from the probability plot Extracting sets from The probability values ​​of the corresponding channels for all pixels covered are calculated, and the maximum value is taken as the confidence peak of that category in this region. ,in For channel In the region The maximum probability value within reflects whether the region belongs to the first... Highest confidence level of plaque-like components, numerical range For traversing regions The coordinates of each pixel in the image. The purity probability plot in coordinates First The probability values ​​of each channel; if the connected component falls entirely within the artifact removal region, then the probability of each of the three channels for all pixels is zero, and the peak values ​​of the three categories are all zero. ;

[0097] Compare the peak confidence levels of the three channels, and select the channel corresponding to the highest confidence level as the final patch component label for that connected component. In the formula The index for the winning channel is 0, representing calcified nodules, 1, representing fibrous caps, and 2, representing lipid pools. The maximum value is unique. The results are given directly. When two or three channels simultaneously achieve the same maximum value, the classification is determined according to a preset clinical priority order: lipid pool (channel 2) > fibrous cap (channel 1) > calcified nodules (channel 0). This priority is based on the fact that the lipid pool is the core feature of vulnerable plaques and has the highest risk of clinical missed detection; when the confidence levels are the same, lipid components should be identified first. The fibrous cap covers the surface of the lipid pool, and when it coexists with the lipid pool, the lipid pool should also be detected first. Calcified nodules usually exist in isolation, and their probability map response is relatively sharp, rarely being exactly equal to other categories in terms of peak value. At the same time, the peak value of the winning channel is extracted as the overall classification confidence level for that region. This value records the highest support strength for the assigned category within the region. After assigning a category label, two validity checks are performed, discarding unreliable or extremely weak regions and excluding them from the final patch identification results. Low-confidence filtering is also applied. If no clear plaque composition is found in the area, it will be excluded. To limit the minimum confidence level, a typical value of 0.10 is used, which is far below the segmentation threshold. This is only used to exclude scattered connected components caused by pure noise or artifact residues. Minimum area filtering is applied if the region area... If the area is deemed clinically insignificant and likely represents imaging noise, it will be removed. This is the minimum clinically significant area threshold, expressed in pixels, with a typical value of 10 pixels. This value can be converted to the actual area based on the OCT image resolution, for example, by converting it to square millimeters if necessary, but here we operate directly in the pixel domain; only when... and Only when this condition is met is the connected region retained as a valid patch area.

[0098] For the valid connected components that have passed the filtering, add two new fields to the records in step 552. String tags, based on The mapping is as Calcific Nodule, Fibrous Cap, or LipidPool. For floating-point values, equal to Discarded connected components are directly removed from the record list and no longer participate in subsequent steps. The original structure of the connected component list is maintained, and each valid connected component is supplemented with the final assigned patch category label and category confidence peak.

[0099] Step 554 involves spatially binding the assigned plaque category, outer contour boundary coordinate sequence, and geometric topological attributes to obtain the automatic plaque identification result. Specifically, this includes constructing a plaque object for each identified valid plaque region. This object contains the following key information: the category field uses explicit string identifiers, such as calcified nodules recorded as "Calcific Nodule", fibrous caps as "Fibrous Cap", and lipid pools as "Lipid". The "Pool" label comes directly from the attribution decision in step 553. The contour field stores the coordinate sequence of the outer boundary of the region, storing multiple polar coordinate points in list form. Each point consists of radial and angular indices, which can be used to directly draw closed boundaries on the polar coordinate image. Alternatively, it can be mapped back to Cartesian coordinates through coordinate system transformation parameters for overlay display. The area field records the total number of pixels occupied by the patch, and the perimeter field gives the pixel length obtained by accumulating the distances between adjacent points of the contour. Both reflect the absolute size and boundary complexity of the patch. The centroid field gives the center position of the region in the polar coordinate image, including radial and angular coordinates, which facilitates spatial positioning. The circumscribed rectangle information describes the smallest rotation rectangle that encloses the region, including the width, height, and rotation angle of the rectangle, used to grasp the macroscopic extension direction of the patch. The eccentricity field quantitatively describes the degree of deviation of the patch shape from a circle. The closer the value is to zero, the closer it is to a circle, and the closer it is to one, the narrower and longer it is, which helps in the auxiliary judgment of patch morphology. The confidence field directly reuses the peak value of the category confidence determined in step 553. It is a value between zero and one. The highest probability support for the region belonging to the determined category is represented, providing a quantitative reference for the uncertainty of the result. In addition, a pixel-level Boolean mask can be generated as needed, marking the location of the plaque as true and the rest as false. Next, all plaque objects are aggregated into a global list, and each plaque is sorted according to its area size or radial position of the centroid, so that the entities in the list have a consistent arrangement logic. This list can be serialized into a common exchange format such as JSON to ensure compatibility with downstream modules. Complete coordinate system metadata and process parameters are embedded at the top level of the output data packet. The coordinate system metadata includes the pixel size of the polar coordinate image, i.e., the number of radial sampling points and the number of angular scan lines. It also explains the physical mapping relationship from the center of the lumen to the depth of the outer wall of the blood vessel corresponding to the radial range, the arc coverage of the circumference or a specific sector corresponding to the angular range, and the polar coordinate semantics used for the contour coordinates. The process parameters record information such as the segmentation thresholds, minimum confidence limits, minimum area thresholds and software versions used in this analysis, forming a traceability chain for the entire process. Finally, a complete automatic plaque identification result data packet is output.

[0100] It provides objective, spatially continuous, and directly superimposed pixel-level confidence for artifact removal, threshold segmentation, connected component analysis, and patch quantization, improving the consistency and efficiency of automatic identification of vulnerable patches.

[0101] like Figure 2 As shown, embodiments of the present invention also provide an automatic plaque recognition system for cardiovascular OCT images based on deep learning, comprising:

[0102] The acquisition module is used to acquire each frame of raw OCT image in polar coordinate format in real time at the opening of the left coronary artery by the guiding catheter;

[0103] The module is used to locate a predetermined radial interval corresponding to the optical distance from the imaging prism to the outer wall of the guiding catheter along the radial depth dimension of the original OCT image, detect the annular high-brightness strong reflection band continuously distributed along the angular direction within the radial interval, and obtain the catheter reflection artifact mask area; based on the eccentric displacement vector of the imaging prism relative to the central axis of the lumen in the current frame and the attenuation coefficient of the probe beam in blood and vascular tissue, a radial illumination attenuation field is constructed.

[0104] The standardization module is used to shield artifact regions in the original OCT image based on the catheter reflection artifact mask region, and apply illuminance compensation to the vessel wall region outside the mask region, which has an inverse mapping relationship with the radial illuminance attenuation field, to obtain a standardized OCT image.

[0105] The segmentation module is used to input standardized OCT images into a pre-trained convolutional neural network to perform pixel-level semantic segmentation and obtain pixel-by-pixel patch category probability maps, including probability distribution channels for calcified nodules, fibrous caps and lipid pools.

[0106] The identification module is used to perform Boolean masking on the plaque category probability map using the catheter reflection artifact masking area, filter out the Fresnel reflection ring artifact area of ​​the catheter itself, extract the connected domain boundary of each plaque type in the masked probability map and determine the plaque category, and obtain the automatic plaque identification result that excludes the interference of catheter artifacts and eccentric illumination dark areas.

[0107] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0108] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0109] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0110] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatic plaque identification in cardiovascular OCT images based on deep learning, characterized in that, The method includes: Step 1: Acquire raw OCT images in polar coordinate format for each frame, acquired in real time at the ostium of the left coronary artery via the guiding catheter; Step 2: Along the radial depth dimension of the original OCT image, locate the predetermined radial interval corresponding to the optical distance from the imaging prism to the outer wall of the guiding catheter, detect the annular high-brightness strong reflection band continuously distributed along the angular direction within the radial interval, and obtain the catheter reflection artifact mask area; based on the eccentric displacement vector of the imaging prism relative to the central axis of the lumen in the current frame and the attenuation coefficient of the probe beam in blood and vascular tissue, construct the radial illumination attenuation field. Step 3: Based on the catheter reflection artifact masking area, the artifact area in the original OCT image is shielded, and the illuminance compensation that is inversely mapped to the radial illuminance attenuation field is applied to the vessel wall area outside the masking area to obtain the standardized OCT image. Step 4: Input the standardized OCT image into the pre-trained convolutional neural network to perform pixel-level semantic segmentation and obtain a pixel-by-pixel patch category probability map, including the probability distribution channels of calcified nodules, fibrous caps and lipid pools. Step 5: Perform Boolean masking on the plaque category probability map using the catheter reflection artifact masking area to filter out the Fresnel reflection ring artifact area of ​​the catheter itself, extract the connected domain boundaries of each plaque type in the masked probability map and determine the plaque category, and obtain the automatic plaque identification result that excludes the interference of catheter artifacts and eccentric illumination dark areas.

2. The method for automatic identification of cardiovascular OCT image plaques based on deep learning according to claim 1, characterized in that, Along the radial depth dimension of the original OCT image, a predetermined radial interval corresponding to the optical distance from the imaging prism to the outer wall of the guiding catheter is located. A continuously distributed annular high-brightness, strong-reflection band along the angular direction within this radial interval is detected, yielding the catheter reflection artifact mask region, including: Based on the known optical distance between the outer wall of the guiding catheter and the imaging prism, the starting and ending depths are defined along the radial depth dimension of the original OCT image to obtain the predetermined radial interval. Extract pixel intensity values ​​in all angular directions within a predetermined radial interval to obtain intensity projection curves along the angular directions; An adaptive threshold is applied to the intensity projection curve, and continuous angle segments with intensity values ​​consistently higher than the adaptive threshold are marked as candidate strong reflection zones; In polar coordinates, all pixel positions corresponding to the candidate strong reflection band within a predetermined radial interval are set as foreground, and the remaining pixels are set as background, forming the duct reflection artifact mask region.

3. The method for automatic identification of cardiovascular OCT image plaques based on deep learning according to claim 2, characterized in that, Based on the eccentric displacement vector of the imaging prism relative to the central axis of the lumen in the current frame and the attenuation coefficient of the probe beam in blood and vascular tissue, a radial illumination attenuation field is constructed, including: Obtain the eccentric displacement vector of the imaging prism relative to the central axis of the cavity in the current frame, and determine the actual optical distance of the imaging prism to the cavity wall surface in each angular direction; The actual optical distance is decomposed into the blood-filling path length and the vascular tissue-penetrating path length, and coupled with the attenuation coefficient of the probe beam in the blood and the attenuation coefficient in the vascular tissue, respectively, to obtain the cumulative path attenuation in each angular direction. Using each angular direction of the polar coordinate image as a reference, the attenuation is accumulated along the radial depth from shallow to deep to obtain a two-dimensional radial illuminance attenuation field of the same size as the original OCT image. The value of each pixel represents the proportion of illuminance lost by the probe beam due to propagation attenuation at the corresponding position.

4. The method for automatic identification of cardiovascular OCT image plaques based on deep learning according to claim 3, characterized in that, The actual optical distance is decomposed into the blood-filling path length and the vascular tissue penetration path length. These are then coupled with the attenuation coefficients of the probe beam in blood and vascular tissue, respectively, to obtain the cumulative path attenuation in each angular direction, including: Based on the actual optical distance from the imaging prism to the lumen wall surface in the current angular direction, and the interface position determined by the difference in optical properties between blood and vascular tissue along the radial scan line in the current angular direction, the actual optical distance is decomposed into the blood filling path length from the imaging prism to the interface and the vascular tissue penetration path length from the interface to the lumen wall surface. The blood path attenuation is calculated based on the blood filling path length and the attenuation coefficient of the probe beam in the blood; the tissue path attenuation is calculated based on the vascular tissue penetration path length and the attenuation coefficient of the probe beam in the vascular tissue. The blood path attenuation and tissue path attenuation are added together to obtain the cumulative path attenuation in the current angle direction. The cumulative path attenuation is obtained by traversing all angle directions in the polar coordinate image.

5. The method for automatic identification of cardiovascular OCT image plaques based on deep learning according to claim 4, characterized in that, Based on the catheter reflection artifact masking region to shield artifact regions in the original OCT image, illuminance compensation, which has an inverse mapping relationship with the radial illuminance attenuation field, is applied to the vessel wall region outside the masking region to obtain a standardized OCT image, including: The catheter reflection artifact mask area is matched pixel by pixel with the original OCT image, and the gray value of the pixels falling within the mask area is set to invalid value to obtain an artifact mask image that preserves the vessel wall area. Extract the position index of each effective pixel located outside the mask region in the artifact masking image, and retrieve the corresponding attenuation value from the radial illumination attenuation field based on the position index; Using the attenuation value in the radial illuminance attenuation field as input, the corresponding illuminance compensation coefficient is obtained according to the preset inverse mapping relationship. The illuminance compensation coefficient is then multiplied and corrected with the original gray value of the corresponding effective pixel to obtain the uniform gray value after illuminance compensation. By keeping the uniform grayscale values ​​of all valid pixels unchanged and invalid pixels retaining their invalid values, a normalized OCT image of the same size as the original OCT image is reconstructed.

6. The method for automatic identification of cardiovascular OCT image plaques based on deep learning according to claim 5, characterized in that, Standardized OCT images are input into a pre-trained convolutional neural network to perform pixel-level semantic segmentation, resulting in pixel-by-pixel plaque category probability maps, including probability distribution channels for calcified nodules, fibrous caps, and lipid pools, including: Standardized OCT images are input into the encoding branch of a pre-trained convolutional neural network to extract multi-scale deep feature maps that characterize the microstructure of the blood vessel wall and the optical features of plaques. Multi-scale deep feature maps are passed to the decoding branch for spatial resolution restoration and cross-level feature fusion to reconstruct a semantic feature tensor including boundary information. The semantic feature tensor is activated pixel by pixel by multi-channel classification convolution kernels, and independent channel probability distributions corresponding to calcified nodules, fibrous caps and lipid pools are obtained respectively. The probability distributions of independent channels are spatially stacked according to the channel dimension to obtain a patch category probability map whose dimension is aligned with the standardized OCT image.

7. The method for automatic identification of cardiovascular OCT image plaques based on deep learning according to claim 6, characterized in that, Boolean masking is performed on the plaque category probability map using the catheter reflection artifact masking region to filter out the Fresnel reflection ring artifact region of the catheter itself. The connected component boundaries of each plaque type in the masked probability map are extracted and the plaque category is determined. The automatic plaque identification results, excluding interference from catheter artifacts and eccentric illumination dark areas, are obtained, including: Using the duct reflection artifact mask area as a Boolean filter template, a logical zeroing operation is performed on the pixel probability value at the corresponding coordinate position in the patch category probability map to remove the interference area of ​​the duct Fresnel reflection ring and obtain a clean patch probability map. A preset segmentation confidence threshold is applied to each probability distribution channel in the pure patch probability map to obtain a binarized region mask for each patch type. Perform connected component topology analysis on each binarized region mask to extract the outer contour boundary coordinate sequence of each independent patch region and calculate the region's geometric topology properties; Based on the confidence peak of pixels within the region enclosed by each outer contour boundary sequence in the corresponding probability distribution channel, the patch category to which each connected region belongs is determined; Spatial data binding is performed on the patch category, outer contour boundary coordinate sequence, and geometric topological attributes to obtain the automatic patch identification results.

8. A deep learning-based automatic plaque recognition system for cardiovascular OCT images, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to acquire each frame of raw OCT image in polar coordinate format in real time at the opening of the left coronary artery by the guiding catheter; The module is used to locate a predetermined radial interval corresponding to the optical distance from the imaging prism to the outer wall of the guiding catheter along the radial depth dimension of the original OCT image, detect the annular high-brightness strong reflection band continuously distributed along the angular direction within the radial interval, and obtain the catheter reflection artifact mask area; based on the eccentric displacement vector of the imaging prism relative to the central axis of the lumen in the current frame and the attenuation coefficient of the probe beam in blood and vascular tissue, a radial illumination attenuation field is constructed. The standardization module is used to shield artifact regions in the original OCT image based on the catheter reflection artifact mask region, and apply illuminance compensation to the vessel wall region outside the mask region, which has an inverse mapping relationship with the radial illuminance attenuation field, to obtain a standardized OCT image. The segmentation module is used to input standardized OCT images into a pre-trained convolutional neural network to perform pixel-level semantic segmentation and obtain pixel-by-pixel patch category probability maps, including probability distribution channels for calcified nodules, fibrous caps and lipid pools. The identification module is used to perform Boolean masking on the plaque category probability map using the catheter reflection artifact masking area, filter out the Fresnel reflection ring artifact area of ​​the catheter itself, extract the connected domain boundary of each plaque type in the masked probability map and determine the plaque category, and obtain the automatic plaque identification result that excludes the interference of catheter artifacts and eccentric illumination dark areas.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images