1.IntroductionOptical coherence tomography (OCT)1 can acquire high resolution cross-sectional images of biological samples by measuring their backscattered signals and has become a widely used tool for microstructure analysis,2 disease diagnosis, and biomedical imaging.3,4 However, OCT image contrast comes from the small range of in-sample refractive index variations, around 1.3 to 1.5,5 and this makes accurate real-time assessment of biological tissue challenging. To obtain accurate tissue anatomical structure information, there has been a growing interest in OCT attenuation coefficient imaging,5,6 which measures the light decay associated with the absorption and scattering. Following Lambert–Beer’s law, the irradiance exhibits an exponential decay along the penetrating depth , which can be written as where is the initial irradiance value, is the attenuation coefficient. With a low , a large or , or a combination of these factors, the decayed irradiance could be low, significantly limiting the visibility in deeper layers.7,8 Many OCT attenuation coefficient models have been developed to correct this attenuation loss and increase potential contrast for different tissue characterization, such as assessing the skin wound healing,9 atherosclerotic plaques,10,11 and filtering blebs after trabeculectomy.12 To quantify the attenuation of tissue, one straightforward approach is to assume a homogeneous sample and fit the OCT A-scan signals with exponential decay functions.7,13 Due to the noise reduction and pre-averaging step, the attenuation coefficient is not depth-resolved, resulting in a loss of details of the inner structures of samples.Inspired by ultrasonic imaging techniques,14 a depth-dependent model was proposed to measure local attenuation properties.15,16 It bypasses the fitting step and noticeably outperforms in turbid samples with strong axial variations in light scattering. Later, several methods have been designed to improve upon this method by point spread function (PSF)17,18 as a practical light source and taking the noise floor into account.19 Another type of improved algorithms have been developed to mitigate the distal end errors,20,21 which is caused by the Vermeer’s model hypothesis (). However, all these computational approaches assumed a constant backscattering of the attenuated light, denoted as the backscattered fraction . In theory, depends strongly on the particle size, relative refractive index,22 and the numerical aperture (NA) of the OCT system.23 Previous studies proved that there is a significant difference in the backscattered fraction measured from intravascular OCT images.24 Therefore, the assumption of fixed can result in considerable under- or overestimation of tissue attenuation. More recently, Cannon et al.22 proposed an improved model that can measure the depth-resolved attenuation coefficients and the layer-resolved backscattering fractions simultaneously, and thus able to reduce measurement errors for heterogeneous samples. However, this method requires pre-segmentation of each layer, and an inaccurate segmentation could seriously affect the subsequent attenuation estimation. Furthermore, it is still challenging to precisely measure for a complex tissue containing fuzzy boundaries and noise. In this study, we make a systematical analysis of over- and underestimation of attenuation coefficients due to the assumption of a constant backscattering fraction and present an alternative and practical depth-dependent attenuation characterization method without requiring a pre-segmentation step. As shown in Eq. (1), the decay rate of irradiance relies on , whereas the acquired OCT intensity is determined by the product of all three variables, , , and . To decouple their contributions, an underdetermined equation set with initial conditions was deduced. Approximated and were then solved by iterations and updated simultaneously. Several constraint conditions were introduced to control the iteration process in practice. The stability and convergency of the equation set is discussed in the later section. Based on the estimated optical properties of samples, a model for correcting the light attenuation loss is also demonstrated. Our method is first verified using numerically simulated OCT A-scan signals and Monte Carlo simulated numerical phantom models. Further validations are performed using multi-layer silicone-dye- phantoms and bovine retina imaged with our in-house built SS-OCT (sweep-source OCT). Compared with raw OCT images of the retinal tissue, our attenuation-corrected OCT profiles are more accurate assessing the retinal microvasculature. We believe that such an accurate tissue characterization of and and the attenuation compensation model can be a highly effective diagnostic tool for analyzing complex and heterogeneous biological samples. 2.Method2.1.Theory2.1.1.Iteration-based computation of μ and RAs shown in Eq. (1), the irradiance value () decreases when the light propagates through a medium with the loss. In our SS-OCT system, the acquired digital signal intensity detected at the depth is a portion of the irradiance value, given as15 where is a converting factor related to the detection quantum efficiency and digitization process.23 Note that the axial PSF calibration, sensitivity roll-off, and noise floor related to in our SS-OCT system are discussed later in Sec. 2.3. The factor 2 is due to the round trip between the detector and the sample. According to the scattering theory, the attenuation coefficient is determined by the scattering coefficient and absorption coefficient . is negligible compared to in the near-infrared region.25 depends on the geometric size of the scattering particle, their volume density, the incident wavelength, and refractive indices.26 On the contrary, is not affected by the volume density and is only determined by the other factors above. For an ideal sphere case, it can be written as where is the angle between the incoming wave and backscattered wave (), and is the volume scattering function, indicating the differential scattering cross section of particles per unit volume.23,27 The backscattering fraction is the ratio of the integral of over the acceptance angles determined by NA to the integral of over full solid angle. Another relevant parameter is the scattering anisotropy, defined as the average cosine of the scattering angle In the following section, we do not discuss the calculation of . It is only used for analyzing the scattering behaviors of samples and building a numerical phantom using the Monte Carlo simulation.Here, we assume the sample surface is at depth 0 and the light decays to zero at infinity. The penetration depth is also unlimited. Similar to Eq. (2), the initial value of the signal intensity can be written as Next, we continue the work by Vermeer’s method15 and introduce an integral term of from Eq. (2): Equation (6) is deduced by the integration by parts. Following Ref. 15, we divide Eq. (2) by Eq. (6) and substitute by Eq. (1), we obtain To better separate from other variables, we move them to left-hand side (LHS), shown as If the backscattering fraction is assumed constant, the derivatives of will be zero and Eq. (9) will degenerate to the basic attenuation coefficient estimation model,15 given as For multi-layer samples with varying backscattering fractions, using the simplified model in Eq. (10) can lead to both under- and overestimation of over the entire imaging range. Both , , and the derivative of can affect this type of measurement errors. To calculate the backscattering fraction , we unitize the model introduced by Liu et al.24 It subdivides a heterogeneous sample into portions small enough to be homogeneous at any depth, e.g., , where nd . Under this condition, Eq. (2) can be modified as where at depth . is treated as 1 since its thickness is small. We take the logarithm of Eq. (11) and plug Eq. (1) into it, and obtain where the logarithm of has a linear relationship both with the integral of and the logarithm of divided by . This assumption simplifies calculations and ensures an accurate estimation of in a discrete form.Equations (9) and (12) form a nonlinear integral equation set, which describes a strong coupling between and and their contributions to the acquired OCT signal intensity, given as They are transcendental equations including exponents and logarithms and thus there are no explicit general solutions. Moreover, the whole equation set is underdetermined since they are both derived from Eq. (2). To address this, we adopted an extended form of a stationary iterative method to obtain an approximated solution to this nonlinear system, where only a single unknown variable is on the LHS in Eq. (13). It is defined as Richardson iteration,28 where is the number of iterations, is the iteration matrix, and is a constant vector. Its key idea is the successive approximation from to , where only depends on but not the previous values . The initial values and are calculated by Eqs. (10) and (12), respectively. The digitized irradiance value is acquired at the beginning of experiments, detailed in Sec. 2.3. Owing to the strong coupling in Eq. (13), several constraint conditions are added to better approximate the ground truth . Without them, the variables might remain unchanged after initialization, resulting in infinite but similar loops (detailed in Sec. 2.1.2). As stated in Eq. (13), the recursion from to is determined by the derivative of the backscattering fraction , and any level of changes in at depth can influence the iteration procedure ahead of it. As a result, is fuzzily processed and only large changes in are retained. The rate of changes of was measured by . It is reasonable to ignore the small signal fluctuation, because it may come from measuring errors, such as the system speckle noise, whereas the large variation is caused by the sample structures. Through multiple tests, we confirmed that four to five times the average derivative was normally an appropriate threshold for all the following simulations and experiments to reduce the noise and preserve the signal edge. Each region with small variation of was substituted by its average value. Due to the strong coupling between and , omitting this fuzzy processing step or using a very small threshold can result in duplicated outputs in each loop. On the contrary, if changes of are not considered, our estimation equation will degenerate to the basic attenuation estimation model in Eq. (10).2.1.2.Convergence and constraint conditionsThe convergence of our equation set is determined by its iteration matrix. In a linear system, for all initial , the convergent solution is proved to exist when the induced norm of iteration matrix satisfies that .28 In the Supplemental Material, we demonstrated that the induced norm can be larger than 1 in certain cases. Therefore, there is no guarantee that a global convergent solution will be obtained. However, local optimal values can be approximated given that the initial value by Vermeer’s model15 is usually close to ground truth value , which has the same order of magnitude. Here, we added practical constraints on the iteration process to prevent unnatural distortions of and , given as are the scaling factors that can be adjusted according to an actual situation. Equation (15a) is based on the range of , which is a unitless ratio between 0 and 1. Equation (15b) describes the fluctuation of depth-dependent around its average value computed by Eq. (2). Equation (15c) prevents overcompensation of the attenuation . As stated in Eqs. (15d) and (15e), the program will stop executing once it achieves a local convergence or the loop is completed after a predetermined iteration value.2.1.3.Attenuation compensation modelAs shown in Eq. (2), the acquired signal intensity is badly affected by the decayed irradiance , leading to the deeper region of thick specimens undetectable. On the contrary, the ideal signal intensity without the influence of the attenuated irradiance of the layers above can be assumed as where the initial irradiance takes place of the attenuated irradiance . With more accurate and obtained by our model, we could achieve a better intensity compensation for the attenuation loss in OCT images. Previous studies9,29,30 took the attenuation map as a representation of the compensated intensity map , where is proportional to with being a constant. This is not applicable to a multi-layer tissue with complex scattering behaviors. Our model works better on heterogeneous samples by taking the backscattering fraction into intensity compensation, which allows a more precise reconstruction of OCT signal intensity.2.2.Manufacturing Process of SamplesTo validate the performance of our method, two-layer silicone elastomer-based phantoms were fabricated using a well-known protocol.31 First, viscous suspensions including particles, a black dye (Higgins®), and a base elastomer (Sylgard® 284 Silicone Elastomer MicroLubrol) along with a curing agent were well blended. Different backscattering properties were obtained by changing the concentration of particles and the dyes, which has a different dependency on the light scattering and absorption.31 The mass concentration of and a black dye are set to 0.1 w%, 0.2 w%, 0.25 w%, 0.3 w%, and 0.5 w%, 0.75 w%, 1.5 w%, 2 w%, respectively. The thickness of each cured mixture was around 300 to . After curing, they were imaged and stacked into different two-layer phantoms. The tissue study was conducted using ex vivo healthy bovine retinas, which were dissected to obtain an open-sky view for OCT imaging. 2.3.OCT System Setup and System CalibrationAn in-house built SS-OCT system that operated at a 100 kHz sweep rate with a sweeping range of 100 nm and a center wavelength of 1060 nm was used for all experiments. The axial resolution was in air and in biological samples, whereas the lateral resolution was . The NA of our OCT system is 0.05. Taking axial PSF and the sensitivity roll-off into account, the OCT signal intensity defined in Eq. (2) is modified as where and denote the sensitivity roll-off and the axial PSF, respectively. is the depth of focus related to the sample surface. Benefited from the long coherence length of our system, , the effect of the sensitivity roll-off is negligible within the sample imaging depth of . Moreover, is expressed as32 Here, denotes the Rayleigh length of the incident Gaussian beam. A highly reflective sample was imaged twice at different depths,33 whose signal intensities were denoted as image1 and image2. The difference of the surface depth in these two images was measured as . According to Eq. (17), the normalized signal intensity of image1 should be equal to the normalized signal intensity of image2 at the depth . Therefore, the values of and can be estimated by exhaustive search to minimize the following equation: The norm of all 1024 A-scans in the B-scan was computed to reduce errors and exclude outliers. The background noise is removed by subtracting the background signals obtained without the sample. The term is acquired by the global integration of OCT signals (unit is A.U.) from a strong reflector, calculated as 20,068 A.U. × mm.2.4.Signal ProcessingOur approach is based on the iterative calculation and all the real-time OCT data were saved for offline processing using Matlab. The first step is to remove the axial PSF from the acquired signal intensity . The averaging is used to remove the speckle noise and increase SNR with a temporal window size of 5 B-mode images. Then, the signal intensity of the noise floor is identified and replaced by a fitting signal, for example, as shown in Fig. 1(a). The intensity of the noise floor at depth is defined as 50 dB. The average attenuation coefficient near the distal end is obtained by fitting the signal intensity at the depth , which is given as where the intensity at the starting point is defined as 58 dB. The intensity of the noise floor deeper than is extrapolated to infinity from Eq. (20). The curve fitting and extrapolation can reduce the effect of the noise floor; otherwise it can cause the underestimation of 19 when using Eq. (10), as shown in blue in Fig. 1(b). Moreover, the basic assumption of Eq. (10) is satisfied since the predictive signal intensity travels deep enough that the irradiance light decays away completely at that penetration depth.15 Because of neglecting the varying , the measurement of based on Eq. (10) deviates from the ideal value. Next, initial values are computed using Eqs. (10) and (12), whereas the average backscattering fraction of the sample is computed (detailed in Sec. 2.1.2). Once the program is executed, the stationary iteration method is utilized to resolve a transcendental equation set Eq. (13), by a successive approximation of . To decouple the equation set, the backscattering fraction is blurred in each loop. Only larger changes of at a specified threshold level are taken into account to rectify the measurement error of . The program stops automatically if one of the criteria is not met. Finally, the optimized tissue profiles is used for calculating OCT images with attenuation compensation based on Eq. (16). The detailed sequence processing is shown in Algorithm 1.Algorithm 1Tissue characterization by stationary iterative method
2.5.Numerical Simulation2.5.1.A-mode numerical imagingA numerical simulation of the OCT depth profile was performed using a single scattering model by defining and . The calculation was based on Eq. (2), with a constant value . Each digital phantom contained four layers and their bottom layers were set to be infinitely long, where all the light decayed to zero. All phantoms were placed at a depth of , with an axial resolution of . The A-line simulation depicted an ideal situation that avoided the estimation bias of , which neglected many interference factors, including the multiple scattering effects, a limited imaging depth of the system, the speckle noise, and the noise floor. The ground truth values of the region were computed by fitting with an exponential function in Eq. (20); the ground truth values was computed by Cannon’s layer-based model as Ground truth values are also computed in the same way in the subsequent experiments.2.5.2.Monte Carlo simulationThe basic theory of Monte Carlo modeling for light transport (MCML) is well-described by Jacques and Wang,34 and it is a powerful tool for analyzing laser–matter interactions, especially for multi-layer materials. Here, we chose the Monte Carlo approach to simulate the light propagation through a complex layered retinal model and applied our method for tissue characterization. The retinal model has a complex layered structure,35 and it is a suitable model to evaluate and compare different attenuation characterization models. The overall geometrical retinal model is shown in Table 1. The 13-layer geometry was based on the common four-layer retinal model containing retina, retinal pigment epithelium (RPE), choroid with 70% blood, and sclera.36,37 To subdivide the neural retina into 10 layers, 20 OCT images from five ex-vivo bovine eyes were collected and segmented manually. To preserve edges and reduce speckle noise, each image was then divided into 10 regions of interest (RoIs), with each RoI containing 100 A-scans. These A-scans were then averaged individually. In our 13-layer model, the thickness of each sublayer was determined according to the layer segmentation result and the common four-layer model.36,37 The absorption coefficient , scattering coefficient , and anisotropy of each sublayer were determined by using an exhaustive search within plus or minus average values of the neural retina from the original four-layer model36,37 to minimize the mean squared error between simulated A-lines and the average A-lines of each RoI. The simulated result was on a logarithmic scale with gamma correction. The simulation was conducted in the same open-sky view in order it to be consistent with our bovine retinal study, in which the eye was dissected and the vitreous was removed. The MCML algorithm was modified from Wang’s MCML programs.34 The weight of photon packages is initialized as 1 and it decreases at each interaction step, due to effects such as photon absorption and scattering. is the total weight of phantoms that reflect from a reference arm, and is the total weight of phantoms that backscatter from the sample and fit the detecting condition (i.e., NA, detector position, and size). The detected OCT signal intensity is convolved by the following equation:37 where is a constant value determined by the light source, and are the total weights of photons with optical path difference , and is the coherence length of the light source. The related coefficients NA of 0.05, the coherence length of , and the transversal scanning step of were used. Four hundred consecutive A-lines were simulated with 5,000,000 weighted photons. No sensitivity roll-off in depth and axial PSF were considered here.Table 1Retina’s layers, thickness, and optical properties.
3.Result3.1.Numerical Simulation3.1.1.Numerical simulation: A-modeTo initially test the feasibility of our method, two heterogeneous digital phantoms, each containing four layers, were simulated, with a thickness of 0.5 mm for the top three layers. The attenuation coefficient and backscattering fraction of each layer were set to be distinct. No additional constraints between and were established. As shown in Figs. 2(a) and 2(b), the logarithmic OCT A-line intensity (square of magnitude) decreases linearly, which agrees with Eqs. (1) and (2). The simulated backscattering fraction of the first phantom is plotted in Fig. 2(c). The attenuation profiles measured by the conventional model using Eq. (10),15 and our algorithm is plotted in Figs. 2(d) and 2(e). Confounded by the layer-dependent of samples, a large estimation bias can be seen using the conventional method without iteration (). The amount of measurement bias varies across the depth, and it is more pronounced near the boundary of each layer, which is marked in gray. Note that this type of measuring error reduces to zero in the bottom layer, which indicates that the error at a certain depth is only related to the derivative of below that point, as we previously analyzed in Eq. (9). On the contrary, we obtained a more accurate estimation result if we utilize the stationary iteration method for and substitute for . As shown in Fig. 2(c), the initial value appears to be constant within the whole penetrating depth, , and deviates from the true value as shown by the dashed line. When the iteration starts, both the updated coefficients and closely match their theoretical values, shown in Figs. 2(c)–2(e). As shown in Figs. 2(c) and 2(d), if the threshold of is not appropriately selected, for example, when it falls within the range of and , the inter-layer variations at will be neglected by our algorithm. Moreover, it leads to a significant underestimation of the attenuation coefficient within the range of 0 to and across the entire imaging depth. However, even under ideal conditions without Gaussian noise, our method cannot achieve zero estimation errors between and true values as shown in Figs. 2(c) and 2(e). These discrepancies stem from the absence of global convergence of our algorithm (see the Supplemental Material for details). Figure 2(f) shows the local averaging step of . It keeps larger variations near boundaries while ignoring small ones caused by the local structures of tissues and noise, which improves the stability of our algorithm. After the optical properties were corrected, we applied the attenuation compensation method given by Eq. (16) to correct the OCT signal intensity. As shown in Figs. 2(a) and 2(b), the intensity distribution with attenuation compensation is consistent with the distribution of the samples’ optical properties with distinct and . Compared to the initial OCT A-line intensity, it provides more accurate morphological information over the entire imaging depth without the attenuation loss. 3.1.2.Monte Carlo simulation: B-modeBased on the Monte Carlo modeling mentioned above, we recorded the photon trajectory and the corresponding OCT intensity of the retinal tissue model. The simulated results have a distinct layered structure that matches real samples as shown in Figs. 3(a) and 3(b). Compared to frequently used four-layer retinal model, our modified version has a complex layered structure and matches our ex-vivo samples better, as shown in Figs. 3(a) and 3(b). The regions above the internal limiting membrane (ILM) were air both in Figs. 3(a) and 3(b). A representative OCT A-scan image from dashed red RoI in Fig. 3(b) is shown as a blue line [Fig. 3(c)]. There is no significant exponential decrease within each layer because of its small thickness, absorption, and scattering coefficients, which makes it difficult to compute the ground truth by a fitting curve. Therefore, we only fit the intensity distribution and computed the ground truth of five thicker layers, i.e., the ganglion cell layer (GCL), the inner plexiform layer (IPL), the outer nuclear layer (ONL), the photoreceptor outer segment layer, and the choroid. Due to a lower NA (NA = 0.05) and larger scattering anisotropy , the signal intensity decays under the impact of multiple scatter light, leading to a smaller , namely .23 Figures 3(d)–3(h) demonstrate the backscattering fraction and attenuation maps and corresponding comparisons with theoretical values. To have a better comparison, we applied gamma correction38 to enhance the image contrast in attenuation map [Figs. 3(d) and 3(e)]. Governed by the scattering coefficients and anisotropy , the ideal backscattering fraction first rises and then declines. It leads to various degrees of estimation biases of as shown in Fig. 2(f). Simulation results proved that our iteration method could calculate the samples’ depth-dependent backscattering fraction and leverage it to rectify the over- or under-estimation of attenuation. We also found that iterations of our model stop with significantly fewer epochs when compared with the refractive-index-matched single-scattering A-line simulation in Fig. 2. This situation is consistent with our subsequent experimental results. A possible reason is that the multiple scattering, the axial PSF, and the scattering anisotropy39 work together to affect the OCT signal distribution, leading to a larger interlayer change of . The adaptive iterative feedback of our algorithm could monitor the iteration procedure and stop it in time, which mitigates the effect of over-compensation. However, there still exists a slight error between and the true values in Figs. 3(f) and 3(g). They are caused by a lack of a global convergence of our algorithm (see the Supplemental Material for details). Suitable constraint conditions can improve the accuracy of our iterated approximating method. 3.2.Phantom ExperimentEight distinct single-layer phantoms were constructed and imaged to evaluate our method. Note the specular reflection signals were removed. The calculations of the ideal values of and were discussed in Sec. 2.5.1, as shown in Figs. 4(a) and 4(b). The experiment results agree with the theoretical analysis that has a linear dependence on the scatterer concentration without the influence of the absorber (Pearson correlation coefficient ).15,31 More importantly, it demonstrates that the ratio of the absorber concentration to the scatter concentration has an adverse effect on the backscattering faction . To validate our method, these single-layer phantoms were stack together; examples of the A-line signal intensity is shown in Fig. 4(c). The signal intensity is rescaled between 0 and 255. The first layered phantom contained 0.1 w%, 0.3 w% of , 0.5 w%, 0 w% of the black dye, whereas the second layered phantom contained 0.2 w%, 0.1 w% of , 0 w%, 1.5 w% of the black dye, from the top to bottom. Figures 5(a), 5(b), and 5(e) show OCT intensity and attenuation maps for the first phantom. A representative A-line shown by the red region is plotted in Fig. 5(c). Due to the increasing backscattering ratio [Fig. 5(g)], the attenuation profile in the top layer is underestimated severely. Simultaneously, we could notice a significant over-estimation bias in the same layer of the second phantom, making almost homogeneous [Fig. 5(i)]. On the contrary, the attenuation profiles using our method accurately matches the ground truth throughout the whole imaging depth as shown in Figs. 5(f) and 5(j). Our method is more robust against the backscattering variation and contains less estimation bias. Compared with other recent methods, it does not need a preliminary inter-layer segmentation, which is often difficult due to the fuzzy boundaries of OCT intensity profiles and the noise induced by the abnormal distributions of scatterers. Furthermore, it measures depth-dependent backscattering profiles as shown in Fig. 5(d). The map of shows significant inter-layer variations and very little intra-layer variations. The loss of the intra-layer resolution is caused by the averaging of , where the local changes of are neglected during the iteration [Fig. 2(f)]. Overall, the respective mean backscattering profiles of all A-scans agree well with the ideal value , as shown in Fig. 5(g). To mitigate the influence of the noise floor, the OCT intensity was extrapolated to infinity as explained earlier. Therefore, the measured and might not match. To avoid this scenario, below sample bottoms was set to zero. 3.3.Bovine Retinal ExperimentTo further test our method, an ex vivo bovine retinal tissue was imaged by our OCT system, as shown in Fig. 6(a). Figures 6(b), 6(d), and 6(e) show the corresponding attenuation and the backscattering fraction maps. Representative average A-lines of , , and of the red RoI are shown in Figs. 6(f)–6(h). Certain retinal layers can be quite thin, 17 to 40 thick with various types of cells, making it hard to precisely annotate the boundaries. The OCT signal intensity was divided into three parts and the ground truth values of in each part were computed by Eq. (20), respectively. The first layer consists of ILM through to outer plexiform layer (OPL), the second layer consists of ONL through to ELM, and the third layer is the outer retinal layer. No layer segmentation was conducted in our iterative method. As shown in Figs. 6(b) and 6(h), using the conventional model leads to over- and under-estimation errors of when neglecting inter-layer variations of . Moreover, the measured attenuation profile of the outer retinal layer is severely polluted due to the blood vessel, which casts shadows over the bottom layers, highlighted by the white RoIs. Our model could track the large variation of adaptively and automatically, as shown in Fig. 6(e), to rectify the value given by Eq. (13). The mean values of precisely matched the ground truth values, as shown in Fig. 6(f). Figure 6(c) shows the processed OCT image with attenuation compensation by leveraging both and , given by Eq. (16). Due to the light attenuation, the reflected signal intensity from the external limiting membrane (ELM) and sclera in Fig. 6(a) is faint and easily overshadowed by the speckle noise. On the contrary, our intensity compensation method [Fig. 6(c)] enhances their visibility, especially in the vascular area of the underlying choroid tissue, highlighted by the blue RoIs. It also eliminated shadow artifacts, and no overamplified background noise is observed when using our algorithm. 4.DiscussionWe proposed and presented a depth-resolved attenuation estimation algorithm that calculates the depth-dependent backscattering fraction profile for SS-OCT imaging. The proposed method does not assume the backscattering fraction as a constant in modeling the attenuation profiles and eliminates the under- and over-estimation problems it causes. Our proposed method provides a more precise characterization of tissue without explicit interlayer segmentation. The method is based on an iterative model using the nonlinear relationship between the irradiance, OCT intensity, and unknown optical properties of tissue. To decouple equations, our approach neglects the small variations in the backscattering fraction and applies appropriate constraint conditions for the solution convergence. It does not assume or to be constant at any depth nor requires a layer segmentation preprocessing step. Furthermore, we utilized the corrected values of and to compensate for the OCT signal intensity loss caused by the decreased irradiance, thus reducing the stripe noise and improving the global visibility of the OCT image. The attenuation coefficient and the backscattering fraction are two important characteristics of samples that could be derived from OCT images. Governed by the different combinations of scatterers’ characteristics, such as the geometric size, the volume density, and the nucleus, they represent the tissue characteristics from different perspectives.24 On the other hand, they tend to be highly coupled to interacting light and highly correlated with the governing equations. Therefore, their analytical solutions cannot be acquired directly without knowledge of other characteristics. To address this issue, the recent studies often measured them individually while assuming the other variables to be constant.15,22 These methods typically rely on accurate layer segmentation, which may fail to perform the task when there are gradual changes in tissue structure or extremely thin tissue layers, e.g., the retinal tissue. Inspired by these issues, we blurred the backscattering term and utilized it to rectify the estimation bias of by iterative procedures. The fuzzy process maintains the significant variation of while replacing the minor ones with a local average. The related thresholds can be adjusted automatically during each iteration, which applies to different ranges of . This process is similar to a layer segmentation, except that it is adaptive with dynamic changes and only determined by . Therefore, it is more sensible than the conventional intensity-based segmentation methods that are easily disturbed by both and . It is often hard to segment the OCT intensity profile successfully when the effects of and make boundaries blurry [shown in red in Fig. 4(c)]. During the iteration, the attenuation profile over the whole depth can be corrected automatically without the layer-by-layer analysis. Moreover, we believe that our method can be used in the visible spectrum when light absorption cannot be neglected. In this scenario, attenuation is measured as , and is measured as a ratio of the backscattered light to the attenuated light.15 Although the light absorption can reduce the magnitude of , we can still rectify the measurement error of by computing . We acknowledge that our existing fuzzy processing of needs to be fine-tuned for tissue with less distinguishable layers, to extract the structure information from the noise. The metric of variation of also needs to be improved. If these prerequisites are met, the attenuation profile over the entire depth can be corrected automatically by the derivative of in Eq. (13) without the layer-by-layer analysis. It is crucial to design appropriate constraint conditions to control the loops when it has no global convergence (see the Supplemental Material). The empirical results indicate that two to four loops are typically enough to yield a reasonable solution for both representative phantoms and real tissues. In Fig. 7, we summarized the mean percentage errors between estimated values and true values obtained from all simulated and experimental results in Figs. 3, 5, and 6. Based on Fig. 7, the first loop usually has relatively little impact on the accuracy. However, the second to fourth loops prove to be sufficient in reducing errors to below 10% for both tissues and phantoms used. Based on this information and the low complexity of each loop, our model has a strong potential for OCT-based real-time tissue assessment. One possible solution to enhance its performance is to convert the existing Matlab codes into a more efficient programming language. In addition, employing parallel processing with GPUs could further optimize the computation time, reducing it to hundreds of milliseconds per B-scan (1024 A-scans). Our proposed constraint conditions are also based on reasonable ranges of tissue’s optical properties, degree of compensation, and iterative times. For a wider application, we acknowledge that they might need further refinement to suit a particular tissue. The large drop in estimation errors near each interface can be another way to further refine constraint conditions, indicated by the black arrow in Fig. 2(e). The presence of multiple scattering is also an important impact on the quantitative attenuation characterization, as we discussed in Figs. 3 and 6 of retina images. Under this circumstance, the high scattering anisotropy and the small scattering angle jointly lead to a larger contribution of highly forward directed scattering behavior, with the rise of the backscattering fraction.23 It emphasizes the importance of our iterative computation model, which can reduce additive errors caused by the pixelwise variation of . Our results suggest that multiple scattering may have more influence near each interface, and the effect on estimating could be reduced by the better control of the iterations. It is worth noting that backscattering fractions calculated by our method can highlight the layer-to-layer boundaries, which is helpful in various clinical applications needing tissue layer segmentation. The further work might include analyzing the analytic solutions of optical properties under the multiple scattering modeling. Our proposed method is evaluated through Monte Carlo simulation and OCT imaging. Both the iterative attenuation estimation model and the intensity-based compensation model demonstrate promising results over a wider depth, which can be advantageous for imaging heterogeneous tissues with a complex and subtle layered structure, such as intravascular tissue,10,13 retina,41 and brain.42 5.ConclusionIn this study, we presented a novel depth-resolved attenuation and backscattering analysis method that could remedy the estimation bias induced by depth-wise variations in backscattering fraction . It also enables a depth-resolved calculation of the backscattering fraction, and an intensity-based compensation model is derived utilizing the corrected tissues’ optical properties. The methodology was validated through a simulated OCT A-line image, the Monte Carlo model, and experimental OCT imaging. The results demonstrated that this algorithm is capable of precisely estimating attenuation coefficients and the depth-dependent backscattering fractions in complex scattering samples, without any preliminary steps, such as layer segmentation. Therefore, it has great potential for the quantitative characterization of heterogeneous structures and for enlarging the related detectable regions of tissue for OCT imaging applications. Code, Data, and Materials AvailabilityThe data that support the findings of this study are available from the corresponding author upon reasonable request. ReferencesD. Huang et al.,
“Optical coherence tomography,”
Science, 254
(5035), 1178
–1181 https://doi.org/10.1126/science.1957169 SCIEAS 0036-8075
(1991).
Google Scholar
J. U. Kang et al.,
“Endoscopic functional Fourier domain common path optical coherence tomography for microsurgery,”
IEEE J. Sel. Top. Quantum Electron., 16
(4), 781
–792 https://doi.org/10.1109/JSTQE.2009.2031597 IJSQEN 1077-260X
(2010).
Google Scholar
G. W. Cheon et al.,
“Accurate real-time depth control for CP-SSOCT distal sensor based handheld microsurgery tools,”
Biomed. Opt. Express, 6
(5), 1942
–1953 https://doi.org/10.1364/BOE.6.001942 BOEICL 2156-7085
(2015).
Google Scholar
R. K. Wang et al.,
“Three dimensional optical angiography,”
Opt. Express, 15
(7), 4083
–4097 https://doi.org/10.1364/OE.15.004083 OPEXFF 1094-4087
(2007).
Google Scholar
D. J. Faber et al.,
“Quantitative measurement of attenuation coefficients of weakly scattering media using optical coherence tomography,”
Opt. Express, 12
(19), 4353 https://doi.org/10.1364/OPEX.12.004353 OPEXFF 1094-4087
(2004).
Google Scholar
P. Gong et al.,
“Parametric imaging of attenuation by optical coherence tomography: review of models, methods, and clinical translation,”
J. Biomed. Opt., 25
(4), 040901 https://doi.org/10.1117/1.JBO.25.4.040901 JBOPFO 1083-3668
(2020).
Google Scholar
S. Chang et al.,
“Attenuation compensation for optical coherence tomography imaging,”
Opt. Commun., 282
(23), 4503
–4507 https://doi.org/10.1016/j.optcom.2009.08.030 OPCOB8 0030-4018
(2009).
Google Scholar
X. Li, S. Liang and J. Zhang,
“Contrast improvement for swept source optical coherence tomography image of sub-surface tissue,”
Proc. SPIE, 10053 100532R https://doi.org/10.1117/12.2251650 PSISDG 0277-786X
(2017).
Google Scholar
B. Ghosh,
“Attenuation corrected-optical coherence tomography for quantitative assessment of skin wound healing and scar morphology,”
J. Biophotonics, 14 e202000357 https://doi.org/10.1002/jbio.202000357
(2020).
Google Scholar
C. Xu et al.,
“Characterization of atherosclerosis plaques by measuring both backscattering and attenuation coefficients in optical coherence tomography,”
J. Biomed. Opt., 13
(3), 034003 https://doi.org/10.1117/1.2927464 JBOPFO 1083-3668
(2008).
Google Scholar
G. van Soest et al.,
“Atherosclerotic tissue characterization in vivo by optical coherence tomography attenuation imaging,”
J. Biomed. Opt., 15
(1), 011105 https://doi.org/10.1117/1.3280271 JBOPFO 1083-3668
(2010).
Google Scholar
M. Kosugi et al.,
“Usefulness of polarization-sensitive optical coherence tomography- derived attenuation-coefficient images to visualize the internal structure of the filtering bleb,”
Curr. Eye Res., 46
(4), 606
–609 https://doi.org/10.1080/02713683.2020.1825749
(2022).
Google Scholar
Y. Gan et al.,
“Automated classification of optical coherence tomography images of human atrial tissue,”
J. Biomed. Opt., 21
(10), 101407 https://doi.org/10.1117/1.JBO.21.10.101407 JBOPFO 1083-3668
(2016).
Google Scholar
D. I. Hughes and F. A. Duck,
“Automatic attenuation compensation for ultrasonic imaging,”
Ultrasound Med. Biol., 23
(5), 651
–664 https://doi.org/10.1016/S0301-5629(97)00002-1 USMBA3 0301-5629
(1997).
Google Scholar
K. A. Vermeer et al.,
“Depth-resolved model-based reconstruction of attenuation coefficients in optical coherence tomography,”
Biomed. Opt. Express, 5
(1), 322 https://doi.org/10.1364/BOE.5.000322 BOEICL 2156-7085
(2014).
Google Scholar
M. J. A. Girard et al.,
“Shadow removal and contrast enhancement in optical coherence tomography images of the human optic nerve head,”
Investig. Ophthalmol. Vis. Sci., 52
(10), 7738
–7748 https://doi.org/10.1167/iovs.10-6925 IOVSDA 0146-0404
(2011).
Google Scholar
G. T. Smith et al.,
“Automated, depth-resolved estimation of the attenuation coefficient from optical coherence tomography data,”
IEEE Trans. Med. Imaging, 34
(12), 2592
–2602 https://doi.org/10.1109/TMI.2015.2450197 ITMID4 0278-0062
(2015).
Google Scholar
B. Ghafaryasl et al.,
“Analysis of attenuation coefficient estimation in Fourier-domain OCT of semi-infinite media,”
Biomed. Opt. Express, 11
(11), 6093 https://doi.org/10.1364/BOE.403283 BOEICL 2156-7085
(2020).
Google Scholar
K. Li et al.,
“Robust, accurate depth-resolved attenuation characterization in optical coherence tomography,”
Biomed. Opt. Express, 11
(2), 672 https://doi.org/10.1364/BOE.382493 BOEICL 2156-7085
(2020).
Google Scholar
M. M. Amaral et al.,
“General model for depth-resolved estimation of the optical attenuation coefficients in optical coherence tomography,”
J. Biophotonics, 12
(10), e201800402 https://doi.org/10.1002/jbio.201800402
(2019).
Google Scholar
J. Liu et al.,
“Optimized depth-resolved estimation to measure optical attenuation coefficients from optical coherence tomography and its application in cerebral damage determination,”
J. Biomed. Opt., 24
(3), 035002 https://doi.org/10.1117/1.jbo.24.3.035002 JBOPFO 1083-3668
(2019).
Google Scholar
T. M. Cannon, B. E. Bouma and N. Uribe-Patarroyo,
“Layer-based, depth-resolved computation of attenuation coefficients and backscattering fractions in tissue using optical coherence tomography,”
Biomed. Opt. Express, 12
(8), 5037 https://doi.org/10.1364/BOE.427833 BOEICL 2156-7085
(2021).
Google Scholar
M. Almasian et al.,
“Validation of quantitative attenuation and backscattering coefficient measurements by optical coherence tomography in the concentration-dependent and multiple scattering regime,”
J. Biomed. Opt., 20
(12), 121314 https://doi.org/10.1117/1.JBO.20.12.121314 JBOPFO 1083-3668
(2015).
Google Scholar
S. Liu,
“Tissue characterization with depth-resolved attenuation coefficient and backscatter term in intravascular optical coherence tomography images,”
J. Biomed. Opt., 22
(9), 096004 https://doi.org/10.1117/1.jbo.22.9.096004 JBOPFO 1083-3668
(2017).
Google Scholar
S. L. Jacques,
“Optical properties of biological tissues: a review,”
Phys. Med. Biol., 58
(11), R37
–R61 https://doi.org/10.1088/0031-9155/58/11/R37 PHMBA7 0031-9155
(2013).
Google Scholar
C. F. Bohren and D. R. Huffman,
“Absorption and scattering by a sphere,”
Absorption and Scattering of Light by Small Particles, 82
–129 Wiley Online Books(
(1998). Google Scholar
H. Horvath et al.,
“Relationship between fraction of backscattered light and asymmetry parameter,”
J. Aerosol. Sci., 91 43
–53 https://doi.org/10.1016/j.jaerosci.2015.09.003 JALSB7 0021-8502
(2015).
Google Scholar
C. T. Kelley, Iterative Methods for Linear and Nonlinear Equations, Society for Industrial and Applied Mathematics(
(1995). Google Scholar
H. Zhou et al.,
“Attenuation correction assisted automatic segmentation for assessing choroidal thickness and vasculature with swept-source OCT,”
Biomed. Opt. Express, 9
(12), 6067 https://doi.org/10.1364/BOE.9.006067 BOEICL 2156-7085
(2018).
Google Scholar
A. Taghavikhalilbad et al.,
“Semi-automated localization of dermal epidermal junction in optical coherence tomography images of skin,”
Appl. Opt., 56
(11), 3116
–3121 https://doi.org/10.1364/AO.56.003116 APOPAI 0003-6935
(2017).
Google Scholar
D. M. de Bruin et al.,
“Optical phantoms of varying geometry based on thin building blocks with controlled optical properties,”
J. Biomed. Opt., 15
(2), 025001 https://doi.org/10.1117/1.3369003 JBOPFO 1083-3668
(2010).
Google Scholar
T. G. Van Leeuwen, D. J. Faber and M. C. Aalders,
“Measurement of the axial point spread function in scattering media using single-mode fiber-based optical coherence tomography,”
IEEE J. Sel. Top. Quantum Electron., 9
(2), 227
–233 https://doi.org/10.1109/JSTQE.2003.813299 IJSQEN 1077-260X
(2003).
Google Scholar
N. Dwork et al.,
“Automated estimation of OCT confocal function parameters from two B-scans,”
in Conf. Lasers and Electro-Opt.,
AW1O.4
(2016). Google Scholar
S. L. Jacques and L. Wang,
“Monte Carlo modeling of light transport in tissues,”
Opt. Response Laser-Irradiat. Tissue, 2607
(713), 73
–100 https://doi.org/10.1007/978-1-4757-6092-7_4
(1995).
Google Scholar
N. Mahabadi and Y. Al Khalili, Neuroanatomy, RetinaStatPearls [Internet], StatPearls Publishing, Treasure Island, Florida
(2022). Google Scholar
M. Hammer et al.,
“Optical properties of ocular fundus tissues: an in vitro study using the double-integrating-sphere technique and inverse Monte Carlo simulation,”
Phys. Med. Biol., 40
(6), 963
–978 https://doi.org/10.1088/0031-9155/40/6/001 PHMBA7 0031-9155
(1995).
Google Scholar
A. Gerakis et al.,
“Monte Carlo modeling of corneal and retinal optical coherence tomography imaging,”
in 8th IEEE Int. Conf. Bioinf. Bioeng. BIBE 2008,
(2008). https://doi.org/10.1109/BIBE.2008.4696840 Google Scholar
C. A. Poynton,
“SMPTE tutorial: “Gamma” and its disguises: the nonlinear mappings of intensity in perception, CRTs, film, and video,”
SMPTE J., 102
(12), 1099
–1108 https://doi.org/10.5594/J01651 SMPJDF 0036-1682
(1993).
Google Scholar
K. K. Bizheva, A. M. Siegel and D. A. Boas,
“Path-length-resolved dynamic light scattering in highly scattering random media: the transition to diffusing wave spectroscopy,”
Phys. Rev. E, 58
(6), 7664
–7667 https://doi.org/10.1103/PhysRevE.58.7664
(1998).
Google Scholar
Q. Wang et al.,
“Thickness of individual layers at the macula and associated factors: the Beijing Eye Study 2011,”
BMC Ophthalmol., 20
(1), 49 https://doi.org/10.1186/s12886-019-1296-6
(2020).
Google Scholar
G. Thepass, H. G. Lemij and K. A. Vermeer,
“Attenuation coefficients from SD-OCT data: structural information beyond morphology on RNFL integrity in glaucoma,”
J. Glaucoma, 26
(11), 1001
–1009 https://doi.org/10.1097/IJG.0000000000000764 JOGLES
(2017).
Google Scholar
J. Lefebvre et al.,
“Whole mouse brain imaging using optical coherence tomography: reconstruction, normalization, segmentation, and comparison with diffusion MRI,”
Neurophotonics, 4
(4), 041501 https://doi.org/10.1117/1.NPh.4.4.041501
(2017).
Google Scholar
BiographyYaning Wang is a PhD student in the Department of Electrical and Computer Engineering at the Johns Hopkins University. She received her BS degree in optical engineering from Huazhong University of Science and Technology in 2019. Her current research interests include optical coherence tomography, multimodal optical imaging systems, and learning-based medical image processing. |
Signal attenuation
Optical coherence tomography
Backscatter
Tissues
Signal intensity
Monte Carlo methods
Biological samples