A single input ps-oct high robust imaging method

Abstract

This invention discloses a Bayesian optimization method driven by spectral domain polarization uniformity (SD-DOPU) for single-input polarization-sensitive optical coherence tomography (PS-OCT) imaging, aiming to solve the problems of polarization trajectory distortion, optical axis distortion, and severe overestimation of phase delay in low signal-to-noise ratio (SNR) regions of PS-OCT. The method uses SD-DOPU as a pixel-level SNR weight to perform spatial domain adaptive filtering on the interference signal; then, using the Poincaré sphere cross-section circle parameter as the optimization objective, a probabilistic model is constructed by combining prior constraints of optical axis spatial continuity and the physical range of phase delay; finally, the optimal parameters are solved using Bayesian maximum a posteriori probability estimation to calculate the local intrinsic optical axis and the true phase delay. Experimental results demonstrate that this invention effectively overcomes the robustness defects of algorithms under low SNR conditions, significantly improves the analytical accuracy of birefringence characteristics, and provides a reliable imaging method for the early diagnosis of diseases in complex retinal tissues and other birefringent biological tissues.

Description

Technical Field

[0001] This invention belongs to the field of polarization-sensitive optical coherence tomography (PS-OCT) technology, specifically involving a Bayesian optimization method driven by spectral domain polarization uniformity (SD-DOPU) for single-input PS-OCT imaging. This method primarily addresses the challenge of robust resolution of birefringence characteristics in traditional single-input PS-OCT in low signal-to-noise ratio regions such as deep tissues or tissues with strong light absorption characteristics. It provides high-precision imaging technology support for the early disease progression of birefringent biological tissues in clinical practice. Background Technology

[0002] Optical coherence tomography (OCT) is a medical diagnostic technique that utilizes the interference principle of low-coherence light to perform non-invasive and high-resolution three-dimensional tomographic imaging of complex biological tissues. Polarization-sensitive optical coherence tomography (PSOCT), as a core functional extension of this technology, can non-invasively and quantitatively image birefringent tissues such as collagen fibers in the human body, detecting optical axes and delays, and providing specific functional imaging contrast that surpasses traditional morphological features. The human body contains various birefringent tissues, and their birefringence properties change with the progression of certain diseases, making this imaging technology show significant clinical translational potential in multiple medical subfields. In ophthalmological clinical diagnosis, it can be used to assess the degree of optic nerve damage in glaucoma, predict the development trend of pathological myopia, and monitor the structural integrity of the retinal nerve fiber layer. In cardiovascular disease assessment, it can characterize the physical thickness and optical attenuation characteristics of the fibrous cap on the surface of atherosclerotic plaques to assess their acute rupture risk. In respiratory medicine, it can be used with endoscopic probes to achieve in vivo in situ detection and quantification of the severity of pulmonary fibrosis lesions. Furthermore, in dermatological clinical practice, this technique is often used for non-invasive assessment of the actual tissue penetration depth of thermal burns and the maturity of postoperative scar tissue. These extensive clinical applications indicate that polarization-sensitive optical coherence tomography (OCT) holds promise as an effective functional optical biopsy tool for the systematic diagnosis of fibrotic diseases, early determination of tumor invasion boundaries, and complex fundus tissue diseases.

[0003] Currently, mainstream polarization-sensitive optical coherence tomography systems in academia and industry primarily rely on multiple polarization input modulation mechanisms, including time-division multiplexing, frequency-division multiplexing, and space-division multiplexing schemes. The core logic of these multi-input schemes is to construct a complete Jones or Mueller matrix through multiple independent measurements or channel multiplexing, thereby comprehensively resolving the polarization information of the sample. However, these complex schemes, while increasing system hardware complexity, also incur specific physical performance costs. Specifically, time-division multiplexing systems sacrifice scanning imaging speed and are more sensitive to the physiological movements of biological organisms, easily producing motion artifacts. Frequency-division multiplexing and space-division multiplexing directly halve the usable imaging depth of the system, and signal aliasing is prevalent during frequency domain reconstruction. A more critical technical bottleneck lies in the reduced polarization sensitivity of multi-input systems due to wavelength dependence. Within the extremely wide spectral range of modern ultra-wideband swept-frequency light sources, the overall polarization response characteristics of the system change drastically with the angular frequency corresponding to the wavelength, making it physically impossible to strictly maintain a stable orthogonal relationship between two polarized beams across the entire wide spectral range. This wavelength dispersion problem introduces polarization state distortion and crosstalk between channels, thereby reducing the polarization extraction sensitivity and data reliability of multi-input systems in deep tissues, which restricts their practical application in high-resolution medical imaging scenarios.

[0004] In contrast, single-input polarization-sensitive optical coherence tomography systems employ a single circular polarization state for full-field illumination, physically avoiding the polarization sensitivity degradation problem caused by wavelength dependence. This represents a potential technological path for developing next-generation high-speed, high-resolution polarization imaging. Early single-input systems, lacking algorithmic compensation and hardware dynamic correction mechanisms, could not decouple sample birefringence from background birefringence caused by the system's single-mode fiber. Furthermore, subsequent static standard calibration methods were prone to failure due to fiber bending or probe rotation.

[0005] In recent years, existing polarization state tracking algorithms based on discrete differential geometry have made progress. These algorithms analyze the spatial evolution trajectory of a single incident polarization state on the surface of a Poincaré sphere in three-dimensional mathematical space, achieving local analytical polarization state resolution with a minimal hardware system, and to some extent solving the problem of birefringence decoupling in the system. However, traditional polarization state tracking algorithms based on discrete differential geometry are still limited by the deep computational challenges posed by low signal-to-noise ratio environments when dealing with complex biological tissues. In fact, low signal-to-noise ratio regions are not only found in deep tissues due to optical attenuation (such as the posterior sclera), but are widely distributed in various physiological and pathological structures. For example, the retinal pigment epithelium (RPE), with its strong light absorption properties, causes an exponential decay in optical intensity, leading to a sharp drop in signal strength within and beneath it. Transitional layers such as the inner retinal layer (INL) and outer retinal layer (ONL), lacking significant birefringence evolution dynamics, often exhibit typical low signal-to-noise ratio (SNR) and weak birefringence physiological layers. Furthermore, in specific pathological conditions such as severe lens opacity (e.g., cataracts) or abnormal thickening of the retinal wall, effective energy transmission is severely hindered, similarly creating extremely low SNR interference sources within or beneath the tissue. In these widespread low SNR regions, the deep superposition of speckle noise from strong tissue absorption and scattering with detector shot noise severely distorts the smooth evolution trajectory of the polarization vector due to high-frequency random noise, resulting in physically unreasonable spatial geometric calculation parameters. During geometric derivation, when a local polarization vector point is dominated by random noise, the spatial tangent vector formed by adjacent data points undergoes a drastic spatial flip, directly causing random jumps and directional disturbances in the subnormal vector calculated through the outer product, i.e., the local optical axis. Meanwhile, the scattered and disordered polarization vector points severely reduce the diameter of the evolutionary cross-section circle obtained by the final spatial fitting. Under the intervention of the nonlinear geometric projection amplification effect, this leads to a significant increase in the absolute phase delay value calculated by the inverse trigonometric function. This pseudo-phase delay overestimation and nonlinear phase crosstalk phenomenon caused by random noise fluctuations mask the weak polarization signal characteristics of real birefringent tissues, limiting the clinical translational potential of single-input polarization-sensitive optical coherence tomography for the non-destructive detection of lesions in complex structures and deep tissues. Summary of the Invention

[0006] To address the core technical problems of existing single-input polarization-sensitive optical coherence tomography systems in the low signal-to-noise ratio region, such as severe distortion of polarization evolution trajectory, drastic optical axis disorder, and severe overestimation of local phase delay, this invention provides a Bayesian optimization single-input imaging analytical method driven by spectral domain polarization consistency.

[0007] The technical solution of the present invention includes the following core implementation steps.

[0008] Step 1 acquires the original real-valued interferometric signal from a single-input system with dual channels. A complex-valued analytic signal containing amplitude and phase information is reconstructed using Hilbert transform or fast Fourier transform. Then, the Stokes parameters of the full spectrum are extracted based on the amplitude and phase difference between the two orthogonal channels. To overcome polarization distortion caused by low signal-to-noise ratio without sacrificing spatial resolution, the complete spectral interferometric data is divided into K independent spectral sub-bands in the wavenumber domain using a Gaussian window function. The local Stokes parameters of each sub-band are calculated, and the spectral polarization uniformity is calculated by taking the average amplitude of the normalized Stokes vector between sub-bands, constructing a global pixel-level signal-to-noise ratio weighted resonant image. The physical mechanism is that traditional methods rely on spatial averaging to calculate polarization uniformity, which severely blurs tissue boundaries. This method utilizes the characteristic that different wavelength components of a swept-frequency light source are generated independently in time. The light signal reflected from real birefringent tissue exhibits highly stable and uniform polarization uniformity across different spectral bands, while system noise manifests as random phase perturbations in different spectral bands, leading to mutual cancellation of vectors between sub-bands and extremely low uniformity. Based on these physical properties, this step achieves precise mathematical discrimination between physiological signals and random background noise while preserving the details of high-frequency spatial structure.

[0009] Step 2, based on the signal-to-noise ratio weighted recalculation image constructed using the spectral domain polarization consistency described in the previous steps, performs adaptive filtering in the two-dimensional spatial domain. This step adaptively adjusts the spatial scale boundary of the two-dimensional filtering window by evaluating the confidence level of the center pixel, and uses the spectral domain polarization consistency value as the relative weight coefficient for the weighted average. This achieves smoothing of isolated noise points while obtaining edge-preserving full-spectrum Stokes parameters after filtering.

[0010] Step 3 uses the spatial parameters of the cross-sectional circle formed by fitting the polarization evolution trajectory on the surface of a three-dimensional Poincaré sphere as the core optimization objective, and constructs a weighted Gaussian likelihood function based on the polarization consistency weight in the spectral domain. Simultaneously, combining the physical aggregation characteristics of natural birefringence in biological tissues, a spherical Gaussian distribution prior constraint characterizing the spatial continuity of the optical axis and a half-normal distribution prior constraint characterizing the physical range of phase delay are introduced to construct a complete global prior probabilistic physical model. The phase delay per micrometer in real birefringent tissue is distributed within a reasonable biophysical range, and due to the traction continuity of collagen fibers, the change of the tissue optical axis in three-dimensional space is usually stable and continuous.

[0011] Step 4 uses a Bayesian maximum a posteriori probability estimation framework to solve for the optimal set of parameters for the fitted cross-sectional circle that maximizes the overall posterior probability. This allows for the robust acquisition of the optimized cumulative optical axis and cumulative phase delay under low signal-to-noise ratio conditions from a noisy background. Finally, by constructing a three-dimensional matrix layer by layer along the depth direction and performing a reverse rotation operation, the polarization effects accumulated by the birefringence of the shallower layers above are gradually removed, and the depth-resolved local optical axis and true phase delay of the sample at the current depth are calculated.

[0012] Compared with existing conventional analysis techniques, the present invention has the following core beneficial effects.

[0013] Firstly, the system's anti-interference and robustness in low signal-to-noise ratio (SNR) environments are improved. This invention proposes using spectral polarization consistency as a criterion for physically distinguishing real tissue signals from noise. Since the different spectral components of the swept-frequency light source are generated sequentially and independently in physical time, the system can construct an accurate pixel-level SNR weight map by reconstructing the full-spectrum data in the frequency domain and calculating the average amplitude of the multi-subband normalized Stokes vector. Through further adaptive spatial filtering and depth-weighted likelihood function construction, the algorithm effectively suppresses trajectory distortion and singularity flipping of polarization states on the surface of the Poincaré sphere, ensuring that the system can still extract accurate polarization structure information in extremely deep tissues and low SNR regions.

[0014] Secondly, it reduces the high estimation error of phase delay in deep inference. Traditional discrete differential geometry methods, when faced with low signal-to-noise ratio limits, measure local phase delays that deviate from physical reality. The fundamental reason is that noise causes the actual evolution trajectory to deviate from the ideal cross-section, resulting in an excessively small spatial dimension of the fitted cross-section circle. This inevitably inflates the absolute value of the calculated local angle when solving the inverse trigonometric function. This invention introduces a prior physical range constraint based on a semi-normal distribution, combined with Bayesian maximum a posteriori probability statistical inference, to effectively eliminate extremely high-probability extreme noise distortion events during the energy function minimization process. This mechanism accurately and smoothly restores the phase delay to a reasonable tissue physical range. For example, in experiments on complex birefringent artificial samples, it successfully brought the abnormal mean of up to 10.272 degrees per micrometer back to the true level of 4.567 degrees per micrometer, improving the accuracy of tissue quantitative analysis and the distribution concentration of measurement results.

[0015] Furthermore, it effectively corrects the visual phenomenon of optical axis disorder in deep tissues. Biological tissues in nature, such as collagen fibers within the skin or sclera, possess structural physical continuity; their polarization optical axes should be smoothly and continuously distributed in three-dimensional anatomical space. This invention introduces a priori constraint model for the spatial continuity of optical axes based on a spherical Gaussian distribution. This mathematical construction aligns with the traction physical laws of biological collagen tissues. This ensures that the estimated local intrinsic optical axes remain smooth and continuous even against noisy backgrounds, significantly reducing the extreme value mutations in optical axes caused by Gaussian random noise. The mean absolute error and global standard deviation of its quantitative assessment both show significant performance improvements.

[0016] Finally, it demonstrates excellent clinical translational value. This optimized method is based on a revolutionary underlying algorithm architecture, requiring no hardware modifications to existing optical core components and seamlessly compatible with mainstream swept-frequency single-input imaging systems in existing medical institutions. In future clinical applications, this algorithm is expected to overcome the interference of low signal-to-noise ratio regions caused by shallow high-scattering tissues and deep signal attenuation on imaging accuracy. The system of this invention is expected to present the true birefringence characteristics of tissues with high resolution and without motion artifacts, enabling non-invasive exploration of birefringent biological tissues in the early stages of disease and dynamic tracking of subsequent disease progression, providing a medical image analysis tool with extremely high clinical translational value. Attached Figure Description

[0017] Figure 1 This is a three-dimensional schematic diagram of the Poincaré sphere showing the polarization evolution trajectory distortion phenomenon under increasingly low signal-to-noise ratio conditions.

[0018] Figure 2 This is the overall module architecture and step-by-step algorithm flowchart of the Bayesian global optimization method driven by the physical characteristics of polarization consistency in the spectral domain, as described in this invention.

[0019] Figure 3 This is a systematic comparison of the construction of multi-layer noise models, the spatial evolution of polarization trajectories, and the estimation accuracy of local optical axis and absolute phase delay by different algorithms in the numerical simulation experiments of this invention.

[0020] Figure 4 This is a comparison of the local intrinsic optical axis and depth phase delay analysis results of imaging a 3D-printed artificial sample with a continuously gradient customized birefringent microstructure in space, based on the present invention. Detailed Implementation

[0021] To make the technical content of this invention clearer and easier to understand, the implementation details and derivation of the technical solution proposed in this invention will be further explained in detail below with reference to the accompanying drawings (including algorithm logic diagrams) and specific implementation verification cases. It should be noted that this invention can be embodied in many different forms of embodiments, and the scope of protection of this invention is not limited to the preferred embodiments mentioned below.

[0022] First, a detailed implementation description of the system's underlying hardware platform construction. This embodiment employs a broadband swept-frequency microelectromechanical system (MEMS) vertical-cavity surface-emitting laser source with an operating frequency of 400 kHz, a center wavelength set at 1060 nm, and a spectral dynamic tuning range of 100 nm. The absolute intensity point spread function axial half-maximum width measured in free-space air medium is 5 μm. The collimated probe beam size entering the test sample is 0.67 mm, corresponding to a theoretical optical lateral resolution of 44 μm for the sample. The system's effective depth of focus is designed to be 2.9 mm, and the instrument-measured 3 dB sensitivity roll-off limit ranging depth is 3.5 mm, sufficient to cover complete deep tissue structures. During single-input polarization imaging detection, the original light wave output from the broadband light source is first split into two beams by a fiber coupler with a splitting ratio of 75:25, injected into the probe sample arm and the delay reference arm respectively. The beam inside the sample arm passes through a first-stage polarization controller, an optical circulator responsible for suppressing back echoes, and a second-stage polarization controller before being smoothly emitted from the terminal collimator. The physical function of the two-stage polarization controller here is to adjust and ensure that the intrinsic polarization state of the incident sample light is purified into a stable linear polarization state before reaching the subsequent open spatial optical path. Subsequently, a quarter-wave plate with a slow axis at a 45-degree spatial angle relative to the input linearly polarized state is installed at the intersection of the spatial optical paths in the sample arm. This waveplate utilizes the birefringence phase delay effect of the crystal material to convert the linearly polarized light into purely circularly polarized light. Finally, this probe circularly polarized light beam is then scanned by a grating system using a two-dimensional high-speed scanning mirror and a 4f lens with a specific focal length and relayed onto the biological sample under test. Simultaneously, the system's reference arm includes a high-precision optical delay line for matching the interference optical path and an independent polarization controller for balancing polarization dispersion and initial polarization adjustment. The reference light wave and the returning sample light wave carrying information about the depth scattering of the sample's deep microstructure ultimately converge in a broadband beamsplitter, resulting in optical interference. This mixed interference signal is then directly fed into a polarization beamsplitter, where it is separated into two orthogonal polarization electric field components, horizontal and vertical, which are simultaneously received by two broadband dual-frequency balanced photodetectors. The balanced detection architecture eliminates the inherent common-mode intensity noise of the light source, highlighting the characteristics of the interference fringes. Finally, the detector converts the high-frequency interference light signal into an analog electrical signal, which is then transmitted to a high-speed data acquisition card to complete dual-channel data acquisition and digital mapping.

[0023] In low signal-to-noise ratio physical environments, such as when the system's imaging depth continuously extends into the sclera, causing the backscattered elastic light signal to decay exponentially, various random broadband noises, especially intratissue speckle noise and detector background thermal noise, will inevitably gradually take over and dominate the signal. For example... Figure 1 As shown in (a), in the shallow, high signal-to-noise ratio region, the extracted polarization state vector can smoothly undergo geometrically regular rotational evolution around the true intrinsic optical axis of the biological tissue on a three-dimensional Poincaré sphere. However, as... Figure 1 As intuitively revealed in (b), with the intensifying effect of background noise accumulation due to imaging depth, the evolution trajectory of the vector coordinates representing the polarization states of each depth layer on the Poincaré sphere deviates from the physically accurate perfect cross-sectional circular orbit. This uncontrollable three-dimensional trajectory distortion causes the traditional discrete differential geometry algorithm, which relies on local differentiation, to produce an exponentially amplified geometric calculation bias when calculating the binormal vector by taking the three-dimensional outer product based on the evolving cut vectors formed by adjacent vectors. This leads to the disordered flipping of the fitted tangent plane in three-dimensional space. This flipping ultimately results in disordered absolute optical axes in the calculated output, while the accumulated phase delay value increases unreasonably.

[0024] The core algorithm implementation process of the present invention includes three main modules: spectral domain polarization consistency calculation and spatial mapping, nonlinear spatial domain adaptive dynamic filtering, and Bayesian maximum a posteriori global optimization combined with underlying physical prior constraints.

[0025] Step 1 primarily involves rigorous extraction and spatial mapping of polarization consistency features from the original physical interference signal to the spectral domain. For example... Figure 2 As shown in the initial "raw interference data" and subsequent symmetrical branches, the system first acquires the raw real-valued interference fringe signals of the horizontal and vertical polarization channels output by a dual-channel balanced photodetector. To simultaneously extract the hidden amplitude and phase features from these real-valued interference fringes, the system introduces a discrete Hilbert transform or fast Fourier transform operation after performing nonlinear wavenumber domain resampling on the signal. This operation eliminates negative frequency components, rigorously reconstructing the real-valued signal into an analytic signal in the complex domain, thereby accurately separating the instantaneous amplitude and phase information of the two orthogonal polarization channels as a function of depth. Based on the acquired horizontal channel amplitude, vertical channel amplitude, and the instantaneous phase difference between the two channels, the system calculates the full-spectrum broadband Stokes parameter matrix representing the current optical polarization state. The specific mathematical extraction formula is expressed as follows:

[0026]

[0027] Where parameter I represents the total reflected light intensity scalar, parameters Q and U represent the two linearly polarized energy components in orthogonal Cartesian coordinates, parameter V represents the circularly polarized energy component, and A is the analytical amplitude of the corresponding channel. This represents the phase difference between channels. Subsequently, to overcome the problem of a sharp drop in signal-to-noise ratio caused by light wave attenuation in deep tissues, this invention utilizes the spectral output characteristics of a swept-frequency light source. That is, the spectral components of different center wavelengths of the light source are generated sequentially in emission time and their phases are independent of each other, to perform purely physical-level discrimination between high-confidence signals and random noise. At the specific computational level, the system introduces digital frequency division multiplexing technology, using a Gaussian window function of a specific bandwidth to uniformly or partially overlapwise divide the acquired complete broadband spectral data into K independent spectral segments in the wavenumber domain before the fast Fourier transform. In this specific embodiment, these are set to 9 segments, each undergoing an independent inverse Fourier transform to calculate and extract the normalized core Stokes parameters of the k-th local spectral segment. For example... Figure 2 As shown in the “Calculate SD-DOPU” node, spectral domain polarization uniformity is defined as the magnitude of the average normalized Stokes 3D vector within these K independent subbands:

[0028]

[0029] The theoretical range of the calculated result is strictly mapped and defined as a closed interval [0, 1]. The closer the value is to 1, the higher the probability that the independent pixel belongs to the real and effective reflected signal of the tissue; conversely, the closer the value is to 0, the more likely the pixel is currently being interfered with by strong broadband speckle or shot noise. The system completely maps and extracts all the calculated values ​​of the entire two-dimensional global tomographic image, thus constructing an accurate pixel-level signal-to-noise ratio weighted benchmark map in one go. This calculation method based on frequency domain segmentation rather than spatial domain smoothing fundamentally avoids the tissue edge blurring effect caused by traditional polarization uniformity algorithms.

[0030] Step 2 mainly focuses on adaptive filtering within the spatial domain. For example... Figure 2The "Spatial Domain Weighted Filtering" module and the dashed control flow line introduced from the right illustrate how, to effectively suppress high-frequency speckle noise in extremely deep tissue layers while preserving the physical edge details and high-frequency optical resolution of real microscopic birefringent tissues to the maximum extent, the system utilizes the spectral consistency map constructed in the previous step to perform adaptive spatial filtering on the fluctuating original Stokes parameters, dynamically adjusting both the spatial scale of the filtering window and the weighting coefficients. Regarding the dynamic adjustment of the spatial scale, the system algorithm dynamically changes the coverage size of the two-dimensional spatial filtering kernel window in real time based on the polarization consistency value strictly corresponding to the currently traversed center pixel. For high-confidence pure signal regions with values ​​close to 1, the algorithm shrinks to use a very small window, such as a 3x3 pixel array, for slight filtering to avoid erroneously blurring key tissue edges; for extremely low-confidence deep noise regions with values ​​close to 0, the two-dimensional window scale is gradually expanded to a very large smoothing scale, such as dynamically adjustable to a 21x21 pixel array, to perform large-scale mean smoothing and denoising processing on this region. Regarding the dynamic adjustment of the weighting coefficients, within the currently dynamically determined physical filtering window, the algorithm performs non-uniform weighted averaging based on spectral consistency confidence. By assigning large computational fusion weights to neighboring pixels with high confidence values, they can dominate the multi-dimensional smoothing process of the window, thereby suppressing the numerical interference of surrounding low signal-to-noise ratio pixels and refining the evolutionary characteristic trajectory of the true polarization signal.

[0031] Step 3 mainly involves prior probability constraints and mathematical modeling of the probability likelihood function. (See reference...) Figure 2 The fifth symmetric branch structure transforms the original algebraic geometric problem of directly analyzing birefringence information on a three-dimensional Poincaré sphere into a statistical problem of solving a stable maximum a posteriori probability estimation of the best-fit cross-sectional circular parameter set on the sphere. This assumes a specified optical detection depth... Within a finite depth moving window, the normally observed set of Stokes-Poincaré vectors is: The set of parameter matrices for the desired cross-section circle that needs to be approximated and optimized is set as follows: It contains three key components, including the orientation vector. The cumulative optical axis direction three-dimensional vector after rigorous normalization constraints satisfies its... The fundamental physical law that the norm equals 1. Parameters Let be the absolute coordinate center vector of the fitted cross-section circle in the three-dimensional Poincaré sphere space, a real scalar. This refers to the precise local single-layer physical phase delay value that the system ultimately needs to calculate. Figure 2 In the likelihood function modeling stage of the right branch, for the M independent polarization state measurement discrete points contained within the current computational depth window, affected by the single birefringence physical effect of the tissue, its purely theoretical evolution trajectory should be attached to the circumferential curve of the circle at that specific cross-section. Let the first... The geometric Euclidean distance from each measured 3D point to the edge of the theoretical cross section is: Meanwhile, considering the Gaussian statistical noise characteristics widely present in current high-frequency optical measurement systems, the spectral polarization consistency value extracted in the previous step is used as the core weight to control the contribution of each discrete data point to the overall energy model. The larger the weight value, the more severe the probability penalty for a small deviation error at that measured point on the overall optimization model. Based on this mechanism, a weighted Gaussian likelihood penalty function that resists outlier interference is constructed:

[0032]

[0033] Where the variance parameter This is the tolerance parameter for the comprehensive background noise standard deviation, determined by the overall system environment assessment. Figure 2 In the physical prior probability modeling stage of the left branch, to prevent the system from falling into the local probability extremum trap caused by singular noise during the nonlinear mathematical fitting process, the algorithm introduces two prior distribution mathematical constraints that fit the underlying physical characteristics of biological tissues. The first is the physical constraint of the spatial smoothness and continuity of the tissue's optical axis. Because the optical axes of collagen fibers within biological tissues, such as the posterior sclera of the eyeball or the nerve fiber layer of the retina, typically exhibit gradual structural changes and continuous orientation in three-dimensional Poincaré space, with very few random abrupt changes occurring at non-lesion interfaces, the statistical model assumes that the local optical axis direction probability between adjacent depth layers follows a specific spherical Gaussian distribution physical constraint, i.e., the Von Mises-Fisher spatial orientation distribution.

[0034]

[0035] Reference direction This is the deterministic reference optical axis direction, for which stable optimization estimation has been completed in the previous depth layer. Core parameters. To regulate the physical concentration penalty parameter of this spherical distribution, a larger value indicates a more stringent rigid physical constraint on the continuous smoothness of the biological collagen structure by the system algorithm. (Constant) This is the normalization constant that ensures the total probability of the probability density function converges to 1 over the entire integration domain. Secondly, there is the constraint of the biophysical limit range of single-layer phase delay. For real birefringent medical tissues, the relative phase delay of a single micrometer-thick layer has a clear physical and biological limit boundary. The wildly overestimated phase delay values ​​often output by traditional unconstrained algorithms are, in the physical world, low-probability events caused by the extreme shrinkage of the fitted cross-section under noise pressure. Therefore, the algorithm stipulates that the locally calculated phase delay probability must follow a strict half-normal one-sided distribution, thereby strongly suppressing and completely eliminating the probability of abnormally large values ​​through probability penalty.

[0036]

[0037] Among them, core control parameters This is the theoretical variance limit of the limiting phase delay, predetermined by the empirical attributes of the target biological tissue before algorithm initialization. The total prior distribution model probability of the final system comprehensive evaluation is composed of the product of the above two independent prior probability physical constraints, and its adjustment balance equation includes a custom fine-tuning weight system for precisely adjusting the intensity of the optical axis absolute smoothness regularization penalty and the intensity of the phase absolute amplitude regularization.

[0038] Step 4 focuses on solving the global optimization problem for the Bayesian maximum a posteriori probability. For example... Figure 2 As shown at the converging endpoint, according to the classical Bayesian probability theorem, the absolute posterior probability of the multidimensional parameter set is mathematically proportional to the product of the previously constructed measurement likelihood function and the physical prior probability—two core indices. The ultimate computational goal of this invention is to find the optimal parameter matrix set that maximizes the global posterior probability of the system. In computer engineering implementation, since directly maximizing the target probability function containing a large number of exponential terms is equivalent to finding and minimizing its overall negative logarithm, this probability estimation process can be transformed into a fast optimization problem of minimizing the global system energy function:

[0039]

[0040] Subsequently, the minimum value of the aforementioned composite target energy function is obtained through a mathematical optimization solution strategy. Under extremely low signal-to-noise ratio and deep speckle noise interference, the system can accurately pinpoint the optimal set of physical parameter matrices for the fitted cross-sectional circle at the current depth, and then robustly calculate the cumulative optical axis direction and cumulative phase delay values ​​of the current tissue layer with extremely high noise resistance. Finally, the system utilizes the calculated robust and noise-free cumulative parameter combination to execute a depth-analytical three-dimensional rotational reverse derivation process. Specifically, by sequentially performing three-dimensional orthogonal reverse rotation operations, the cumulative polarization interference effects of the shallow birefringence effect covering the deep target tissue are gradually stripped away, accurately decoupling and calculating the local true optical axis hidden at the current specific depth in the deep, complex, multi-layered biomedical tissue.

[0041] To comprehensively and quantitatively demonstrate the absolute performance improvement of the Bayesian global optimization system method driven by the physical properties of polarization consistency in the spectral domain compared with the traditional method based on simple discrete differential geometry, the implementation scheme of this invention was verified by theoretical numerical quantitative simulation analysis, comparative imaging experiments of 3D printed birefringent gradient artificial samples, and in vivo human eye fundus imaging experiments.

[0042] In the quantitative verification phase of basic numerical simulation, a single-layer complex simulation environment optical physics model was constructed, in which the standard deviation of background noise increases nonlinearly and exponentially with the imaging depth of the simulated optical detector. This model pre-sets a standard physical baseline where the theoretical true phase delay of the internal homogeneous anisotropic birefringent medium is a constant 3.5 degrees per micrometer. For example... Figure 3 The visualization of (a1-a2) includes a noise standard deviation degradation curve that increases non-linearly with tissue depth and a spatial distribution intensity map of pure noise on B-scan tomography. (Comparison) Figure 3 The two-dimensional evolution maps of the core optical axis and the physical maps of absolute phase delay output in (c1-c3) and (d1-d3) show experimental data indicating that traditional differential geometry methods relying on local derivatives are extremely sensitive to high-frequency shot noise. When the statistical standard deviation of the environmentally injected noise slightly exceeds the lower threshold of 0.1, the optical axis direction of the algorithm framework, which forcibly fits the Poincaré sphere model, undergoes random and drastic flipping and deviation. Simultaneously, the phase delay output by the algorithm exhibits a geometrically amplified artifact-like overestimation. However, when the method of this invention, which includes a low signal-to-noise ratio spectral discrimination mechanism and a posterior optimization core idea, is applied to this same set of low signal-to-noise ratio, harsh simulation data, the system utilizes a dual computational constraint system of data-fidelity likelihood assessment and continuous prior physical constraints to effectively converge and suppress the extremely discrete and uncontrolled free drift problem of the polarization evolution trajectory on the Poincaré sphere. Regarding the overall quantitative mathematical statistical error index analysis, such as... Figure 3As shown in the comparative histograms (f1-f2) and (g1-g4), the innovative method proposed in this paper reduces the average absolute calculation error and the fluctuation of the system's global standard deviation in determining the spatial direction of the deep core optical axis to an extremely low optimal tolerance level. This result curbs the problem of infinite divergence of mathematical errors that inevitably occurs in deep extrapolation, ensuring that even in a deep environment with strong noise and heavy pollution, the output physiological parameter estimation results still maintain a high degree of consistency with the pre-set physical reality standard answer benchmark, proving the feasibility of our method.

[0043] In the high-precision imaging verification of continuously gradient customized real birefringent physical artificial samples, in order to rigorously test the robustness and error correction capability of the overall algorithm system in a complex optical axis distribution physical model with fully known microstructure and varied spatial arrangement, the project team adopted precision fused deposition modeling fully automated 3D modeling and manufacturing technology to customize and prepare a complex solid ring-shaped birefringent polymer artificial sample with an extremely continuous and smooth spiral gradient of 360-degree internal optical axis spatial arrangement. For example... Figure 4 (ad) The photographic reference of the object, the basic image of the optical tomography structure, and the original hybrid Stokes physical polarization intensity map and the signal-to-noise ratio weighted map together confirm that the effective polarization signal inevitably suffers a sharp attenuation and physical degradation of the signal-to-noise ratio when penetrating deep regions of thick solids. Figure 4 As visually presented by (eh), in the extremely deep signal-to-noise ratio collapse attenuation region corresponding to the thick structure of the artificial sample, traditional differential geometry methods, whether generating cross-sectional tomographic maps or frontal polar coordinate projection maps, inevitably produce a large number of discrete, bright yellow overestimation errors in their output absolute phase delay quantitative spectra. Correspondingly, such as Figure 4 The deep analysis results in (il) show that the optimization mechanism of this invention eliminates the influence of the low signal-to-noise ratio region, and intuitively demonstrates the performance advantages of the new architecture in deeply suppressing physical diffuse noise interference and correcting the absolute overestimation error of phase at deep depths.

[0044] In summary, this invention proposes a polarization characteristic analysis method based on the deep fusion of spectral domain polarization consistency (SD-DOPU) driving and Bayesian maximum a posteriori probability estimation. This framework constructs a systematic computational system from three core dimensions: accurate distinction between signal and noise, spatial domain adaptive filtering suppression, and prior constraint optimization incorporating the physical properties of biological tissues. This method fundamentally overcomes the inherent robustness defects of traditional single-input polarization-sensitive interferometric imaging systems in low signal-to-noise ratio environments, effectively solving problems such as drastic abrupt changes in optical axis direction and severe overestimation of phase delay caused by distortion of polarization evolution trajectory. While fully retaining the advantages of the minimalist hardware architecture of single-input systems, this technical solution not only provides a high-precision quantitative tool for the early functional diagnosis of birefringent tissue diseases of the fundus such as glaucoma and pathological myopia, but also offers a novel, robust, and non-invasive method for quantitative medical imaging analysis of deep tissues, including in vivo quantitative monitoring of cardiovascular plaques, early assessment of respiratory fibrosis, and non-invasive detection of skin burn depth.

Claims

1. A Bayesian optimization method driven by spectral domain polarization uniformity (SD-DOPU) for single-input polarization-sensitive optical coherence tomography (PS-OCT) imaging, characterized in that, To reduce polarization trajectory distortion in the low signal-to-noise ratio region of the single-input PS-OCT method, the method includes the following steps: Step 1: Acquire the interference signal from the dual-channel acquisition of the single-input PS-OCT system, and calculate the Stokes parameters of the full spectrum; divide the complete spectral data into... For each spectral band, the Stokes parameter is calculated, and then the SD-DOPU is calculated to construct a pixel-level signal-to-noise ratio weight map; Step 2: Based on the signal-to-noise ratio weight map constructed by the SD-DOPU, perform adaptive filtering in the spatial domain. By dynamically adjusting the spatial scale of the filtering window and the weighting coefficients of the weighted average, obtain the filtered full-spectrum Stokes parameters. Step 3: Using the cross-sectional circle parameters formed by the polarization evolution trajectory on the Poincaré sphere as the optimization target, construct a likelihood function based on the SD-DOPU, and combine the physical properties of tissue birefringence to introduce prior constraints on optical axis continuity and phase delay range to construct a prior probability model; Step 4: Solve for the cross-sectional circle parameters that maximize the posterior probability using Bayesian maximum a posteriori probability estimation (MAP) to obtain the optimized cumulative optical axis and phase delay under low signal-to-noise ratio. Finally, calculate the local optical axis of the sample by three-dimensional rotation.

2. The method according to claim 1, characterized in that, In step 1, the complete spectral data is divided into After sub-bands, the formula for calculating the spectral domain polarization uniformity SD-DOPU is as follows: in, , and The first Normalized Stokes vector components of each spectral subband.

3. The method according to claim 1, characterized in that, The adaptive filtering in step 2 specifically includes: Spatial scale adjustment: The size of the filtering window is dynamically adjusted according to the SD-DOPU value of the center pixel. Small window filtering is used in high confidence signal areas where SD-DOPU is close to 1, and large window filtering is used in low confidence noise areas where SD-DOPU is close to 0. Weight coefficient adjustment: Within a defined filtering window, perform a weighted average based on SD-DOPU, giving higher weights to pixels with high SD-DOPU and suppressing the contribution of pixels with low SD-DOPU.

4. The method according to claim 1, characterized in that, In step 3, the potential cross-sectional circle parameter is set as follows: ,in For the optical axis, Center of the cross section Let the local phase delay be denoted as ; let the th phase delay be denoted as . Measurement data points The distance to the cross-sectional circle is Using pixel weights SD-DOPU as weights The constructed Gaussian distribution likelihood function is: in, The standard deviation of noise. This represents the total number of observation points within a finite depth window.

5. The method according to claim 4, characterized in that, The prior probability model in step 3 consists of optical axis continuity constraints and phase delay range constraints. The total prior probability formula is: The continuity of the optical axis is assumed to have a probability distribution that follows a spherical Gaussian distribution (Von Mises-Fisher, VMF). in For the estimated optical axis of the previous depth, For concentration parameters, The normalization constant is used; the phase delay range follows a half-normal distribution. in It is the variance of the phase delay; and These are the weighting coefficients.

6. The method according to claim 1 or 5, characterized in that, The goal of the Bayesian maximum a posteriori probability estimation in step 4 is to find the cross-sectional circular parameter that maximizes the a posteriori probability. : It is equivalent to minimizing the energy function. The formula for the energy function is: By solving for the minimum value of the energy function, the optimal optical axis and phase estimation under low signal-to-noise ratio conditions can be obtained.

7. A single-input PS-OCT system implementing the method as described in any one of claims 1 to 6, characterized in that, include: Frequency-sweeping laser source; Couplers for splitting light to the sample arm and reference arm; an optical path structure disposed in the sample arm including a polarization controller, a circulator, and a quarter-wave plate that converts linearly polarized light into circularly polarized light; an optical delay line disposed in the reference arm; and a balanced detector for receiving the horizontal and vertical components of the interference signal. The system also includes a data processing unit for performing dual-channel acquisition of interference signals, spectral division and SD-DOPU calculation, adaptive filtering, and executing a Bayesian optimization algorithm driven by spectral domain polarization consistency to analyze the polarization characteristics of the sample.

Images