Sensitivity factors for diffuse optics – derivation of equations for finite changes Authors: Aleh Sudakou, Stanislaw Wojtkiewicz, Roman Maniewski, Adam Liebert Published: 2026 ABSTRACT AbstractWe present analytical equations for time-resolved signals in diffuse optics and their sensitivity factors in homogeneous media, derived from perturbation-based equations valid for infinitesimal changes and then extended to finite changes in optical properties. Time-resolved optical measurements are becoming increasingly available due to technological advancements, expanding their use in various techniques such as near-infrared spectroscopy (NIRS). These measurements acquire distributions of times of flight (DTOF) of photons, and the measurands considered in this study are attenuation (A), mean time of flight (m1), and variance (V) of the DTOF. Established methods for recovering changes in the absorption coefficient (Δμa) from ΔA, Δm1, or ΔV use perturbation-based sensitivity factors, which are valid for infinitesimal changes. We derived sensitivity factor equations that relate ΔA, Δm1, and ΔV to finite (including large) changes in absorption (Δμa), reduced scattering coefficient (Δμs′) , and source-detector distance (Δr). The derivations rely on solutions of the diffusion equation (DE) for homogeneous infinite (IM) and semi-infinite (SM) media within the diffusion approximation (μs′≫μa) , without introducing additional assumptions. We also present analytical equations for the modified Beer–Lambert law (MBLL) that are valid for finite Δμa in IM and SM, which can be directly applied in continuous-wave NIRS data analysis. Sensitivity factors and the MBLL require knowledge of the baseline optical properties and the differential pathlength factor (DPF), and we assessed how errors in these parameters affect the recovered Δμa. We also assessed cross-talk, in which scattering changes Δμs′ lead to spuriously recovered absorption changes Δμa, and vice versa. The proposed framework can be used to improve the accuracy of methods for estimating changes in optical properties. Introduction Perturbation theory is a widely used mathematical approach for deriving equations that relate changes in a measurand to changes in a parameter by linearizing their relation, enabling recovery of the parameter change from the measured signal change. When the relation is nonlinear, such equations are valid only for infinitesimal changes. When approximating a nonlinear dependency by a linear function, the deviation on the left side of the baseline is opposite to that on the right side. Therefore, finite changes recovered using such expressions may be underestimated and overestimated for positive and negative changes, respectively, or opposite depending on the curvature of the function. One approach for a more accurate recovery involves evaluating two changes (from the baseline to the perturbed state, and from the perturbed state to the baseline), where one will be underestimated and the other overestimated, and combining the two results. In this study, we apply this approach to make a finite-change correction to equations used in diffuse optics for recovering changes in optical properties from changes in measurands, such as light intensity changes. Diffuse optics uses light to probe optically turbid media, which has many applications in biological tissues [1,2]. In particular, near-infrared spectroscopy (NIRS) is a non-invasive and safe technique for assessing tissue changes, e.g., in the brain, related to vital physiological parameters such as oxygenation and oxygen metabolism [3–6]. The clinical usefulness of NIRS has been demonstrated by a wide range of clinical applications and trials [7–9]. The Beer–Lambert law (BLL) establishes a linear relation between measured absorbance (A) and the concentration of chromophores (C) in the medium, making it widely used across many scientific fields [10–12]: A=εCd , where d is the distance light travelled and ε is the molar extinction coefficient of a chromophore. The BLL is the fundamental relation used for recovering concentrations C, but it is valid only in non-scattering media as it does not account for light losses due to scattering and assumes that d is independent of C [10–12]. In NIRS, light can penetrate deep into tissues (a few cm) and reemerge at the surface [3], which is enabled by tissue’s high scattering and low absorption in the NIR range [13–15]. This introduces light losses due to scattering and the mean optical pathlength d for the mean distance that light travels. The modified Beer-Lambert law (MBLL) is commonly used to recover changes (ΔC) from measured ΔA , under the assumption that the scattering properties remain constant in time and d is independent of ΔC [16]. The second assumption is not valid in scattering media since d decreases for an increase in C, as illustrated in [17]. Therefore, ΔA is non-linearly proportional to ΔC , and the MBLL is valid only for infinitesimal ΔC . Among several measurement approaches available in diffuse optics (continuous-wave [6], frequency-domain [18], and multi-distance [19]), the time-resolved measurements provide the most information about light propagation, enabling data analysis methods based on other measurands in addition to A [20,21]. A laser emits pulses of light (up to ∼80 MHz repetition rate), and a single-photon detector coupled to a time-correlated single-photon counting card records photons arrival times (up to ∼10 ps resolution) [22], allowing to compute the distribution of times of flight of photons (DTOF) [21]. An analogue of the BLL in time domain was discussed in [23], where the shape of the temporal profile of a propagating light pulse was characterized. The common measurands used to recover changes in optical properties include changes in statistical moments of DTOFs [24,25], integrated intensity over time windows defined on the time axis of the DTOF, and ratios of such time windows [26]. These methods typically rely on perturbation theory and assume that changes in measurands are linearly proportional to changes in optical properties, which is valid only for infinitesimal changes. Although time-resolved measurements can be used to estimate absolute optical properties [21,27], and thereby their changes, this requires adequate integration time to acquire sufficient photon counts, which is challenging in monitoring dynamic changes. In this study, we use perturbation theory to derive equations for sensitivity factors that relate ΔA and changes in statistical moments of DTOFs (mean time of flight Δm1 and variance ΔV ) to infinitesimal changes in absorption (δμa) , reduced scattering (δμs′) , and source-detector distance (δr) . This method for recovering changes in μa was developed by Liebert et al. [25] and has become well-established [28]. Next, we extend the perturbation-based equations to finite (including large) parameter changes (Δμa, Δμs′ , and Δr). We consider homogeneous infinite medium (IM) and semi-infinite medium (SM) geometries [29]. The IM applies to measurements in which the source and detector are positioned inside a medium and far from boundaries compared to r. The SM applies to measurements on the surface of a medium, such as those obtained with optodes placed on tissues or other optically turbid medium. The manuscript is organized as follows. In the Methods, we outline the derivations, which can be reproduced using the expressions provided in the Appendix, and present the derived equations for the sensitivity factors, along with equations for the MBLL that are valid for finite Δμa . In the Results, we verify the equations and show the improvement in parameter recovery, assess the impact of errors in the baseline optical properties or differential pathlength factor (DPF) , and evaluate the cross-talk due to changes in a second optical property. We discuss the implications and limitations of the presented equations and the derivation approach. Lastly, in the Appendix, we present closed-form solutions for Δμa for some of the derived equations. Table 1 summarizes the main equations in this study, grouped according to their purpose in the derivation and recovery of optical property changes. Table 1.Main equations grouped according to their purpose.DescriptionEquationsAnalytical solutions for the DTOF for IM and SM geometriesEqs. (1) and (2)DTOF moments in terms of optical propertiesEqs. (4)–(8)Formula used to extend sensitivity factor equations to finite changes Eq. (10) Sensitivity factors forinfinitesimal δμa and finite ΔμaEqs. (11)–(22)infinitesimal δμs′ and finite Δμs′ Eqs. (23)–(34)finite ΔrEqs. (35)–(40) Modified Beer-Lambert law (MBLL) Eq. (41) Differential pathlength factor (DPF) in terms of optical properties Eq. (42) MBLL for finite Δμafor IM, and a closed-form solution for Δμa Eqs. (44)–(45) for SM Eq. (46) Expressions substituted into Eq. (10) to derive sensitivity factors for finite Δμa (Eqs. (17)–(22))Eqs. (48)–(53)Formulas for recovering finite Δμa (Eqs. (17)–(22)), but without requiring nonlinear solversEqs. (54)–(57) Main equations grouped according to their purpose. Methods Statistical moments of the distribution of times of flight of photons We model the light source as a pulsed laser represented by a Dirac-delta light impulse. The source and detector are represented by two points in the medium. The photon fluence rate (φ) per incident photon (mm−2s−1) that reaches a detector located at a distance r from a light source is given by the following solution of the diffusion equation (DE) for an infinite medium (IM) [27]: where D is the diffusion coefficient, c is the speed of light in the medium, and t is the arrival time of photons. We assume that D depends only on scattering, D=1/(3μs′) , because it was experimentally found that D is independent of μa in the diffusion regime [30,31]. Boundary conditions are not considered for IM. The time-resolved reflectance R(r,t) , which represents the detected fraction of launched photons per unit area per unit time (mm−2 s−1) exiting the surface at a distance r from the light source, is given by the following solution of the DE for a semi-infinite medium (SM) [27] with the zero-boundary condition [32,33]. We made the approximation r2+z02≈r , since typical measurements are carried out at relatively large source-detector separations r≫z0 , where z0=1/μs′ is the effective depth at which the source is positioned. R(r,t) models the DTOF measured in time-resolved diffuse optics experiments. The k-th order normalized non-central statistical moments (mk) of the DTOF are defined as: The zeroth-order moment (m0) is the integral of φ(r,t) or R(r,t) over time, which corresponds to the total number of detected photons (Ntot) as a fraction of the number of incident photons. It can be used to calculate changes in attenuation ΔA=ln(Ntot,0/Ntot)=ln(I0/I) , where Ntot,0 and Ntot,0 correspond to the reference state and after a parameter change, respectively, analogous to the reference (I0) and after a parameter change (I) light intensities. The first-order moment is the mean time of flight (m1). The second-order central moment is the variance (V) of the DTOF, which is related to non-central moments: V=⟨(t−⟨t⟩)2⟩=m2−m12 . The third-order central moment (m3,C) is given by: m3,C=m3−3m1m2+2m13 . Equations (4)–(6) express the moments of the DTOF as functions of the medium’s optical properties. Noise increases rapidly with increasing moment order, and we previously found that moments beyond V are not useful in practice because of their high noise [34,35]. Therefore, we did not include m3,C in further analysis. The equations for IM and SM (Eqs. (4)–(6)) differ only by dimensionless factors containing a ratio. Since this ratio is always less than 1, the values of moments in SM are smaller than those in IM for the same optical properties, and the difference decreases with increasing r and μs′ , and with decreasing μa . Rearranging these equations allows expressing the optical properties as a function of the measured moments (Eqs. (7) and (8)), which is an established method for estimating the optical properties of a homogeneous medium [24]. Derivation of sensitivity factors Here, we outline the derivation of analytical equations for sensitivity factors (SF) that are valid for infinitesimal changes (denoted SFδ ) and finite changes (denoted SFΔ ). The measured values of A, m1, and V are influenced not only by the medium’s optical properties and the measurement geometry, but also by the laser, detector, and other instrumental factors. For the first four moments, these instrumental effects can be accounted for by measuring the instrument response function (IRF) and subtracting the moments of the IRF [36]. If the IRF remains constant during a measurement, the values of ΔA,Δm1,ΔV , and Δm3,C depend solely on changes in optical properties of the medium, which forms the basis of a well-established method for recovering Δμa from changes in moments [25]. In this study, we additionally consider Δμs′ and Δr . The equations for SFδ valid for infinitesimal δμa,δμs′ , or δr correspond to: where M represents the considered measurand (A, m1, or V). SF are referred to as the mean partial pathlength (MPP) for ΔA , the mean time of flight sensitivity factor (MTSF) for Δm1 , and the variance sensitivity factor (VSF) for ΔV . Throughout this study, the subscripts denote absorption (a), scattering (s), and source-detector distance (r). To derive equations for MTSFδ and VSFδ based on perturbation theory, we differentiated the expressions for m1 and V, respectively, given in the previous section, with respect to μa,μs′ , or r. To derive an equation for MPPδ , we differentiated A=ln(1/m0) , as defined in the previous section, with respect to μa,μs′ , or r. Using Δμa as an example, the method based on sensitivity factors for recovery of Δμa uses the relation Δμa=ΔM/SFa . To extend the validity of SFaδ to finite changes, i.e., derive SFaΔ , we implement the following approach. As mentioned in the Introduction, we recover two changes (relative to baseline, and relative to perturbed state), where one is overestimated and the other is underestimated, and combine the two results. Accurate Δμa can be recovered using the following formula: where μa∗=μa+Δμa , and SFa(μa) and SFa(μa∗) are the sensitivity factor expressions for the state before and after Δμa , respectively. Δμa,1 is the change recovered using SFa(μa) , and Δμa,2 is the change recovered using SFa(μa∗) . As shown in Eq. (10), we can derive an expression for SFaΔ such that Δμa can be recovered using the relation Δμa=ΔM/SFaΔ . All expressions for SFa(μa) and SFa(μa∗) , which we substituted into Eq. (10) to derive the SFaΔ equations, are provided in the Appendix, and they were obtained using the following approach. To derive MPPaΔ for IM and SM, we substituted the corresponding MPPaδ(μa) and MPPaδ(μa=μa∗) for SFa(μa) and SFa(μa∗) , respectively, in Eq. (10). The same approach was used to derive all SFsΔ expressions, i.e., MPPsΔ,MTSFsΔ , and VSFsΔ , for IM and SM, by substituting the corresponding SFsδ(μs′) and SFsδ(μs′=μs′∗) into Eq. (10). In contrast, to derive MTSFaΔ and VSFaΔ for IM, the expressions for SFa(μa) and SFa(μa∗) contain both μa and μa∗ . This makes the magnitudes of overestimation and underestimation in Δμa,1 and Δμa,2 equal, so that they cancel in Eq. (10). Next, to derive MTSFaΔ and VSFaΔ for SM, we use their equations for IM and the relation between the equations for IM and SM for moments (Eqs. (4)–(6)). The expressions in the Appendix provide the information needed to reproduce the derivations. m1 and V in IM are proportional to r (Eqs. (4) and (5)), and hence MTSFrδ and VSFrδ in IM are valid for any magnitude Δr . In SM, m1 and V are nearly proportional to r. Also, A in IM and SM is nearly proportional to r, which results from A=ln(1/m0) , where m0 is the integral of Eqs. (1) and (2) for IM and SM, respectively. Therefore, we derived the SFrΔ equations by obtaining SFrδ (differentiating the moments with respect to r) and replacing r with rav=r+Δr/2 , which is the mean of the initial and changed r. Sensitivity factors for Δμa Equations (11)–(13) express SFaδ for IM in terms of optical properties or moments, whereas Eqs. (14)–(16) express them for SM, in terms of moments. The values of SFaδ decrease as μa increases, hence the magnitude of the measured changes in moments grows less with increasing Δμa , leading to an underestimation in the recovered Δμa . The equations for IM and SM (Eqs. (11)–(13)) differ by the same dimensionless factors as in the equations for moments (Eqs. (4)–(6)). This similarity is used to derive the MTSFaΔ and VSFaΔ equations for SM, as explained in the previous section. Equations (17)–(19) express SFaΔ in terms of optical properties, while Eqs. (20)–(22) express them in SM, where μa∗=μa+Δμa , and μe=3μaμs′ , and μe∗=3μa∗μs′ , and m1∗=m1+Δm1 . These equations for finite changes (Eqs. (17)–(22)) simplify to those for infinitesimal changes (Eqs. (11)–(16)) as Δμa approaches 0. Sensitivity factors for Δμs′ Equations (23)–(25) express SFsδ in terms of optical properties, while Eqs. (26)–(28) express them in terms of moments, where μs′∗=μs′+Δμs′ , and μe∗=3μaμs′∗ , and V∗=V+ΔV . SFsδ are proportional to μs′−1/2 , similar as MPPaδ is proportional to μa−1/2 , except MPPsδ has an additional term containing 1/μs′ . Therefore, the finite-change correction will have a similar improvement for Δμs′ recovery as for Δμa recovery using MPPa , as shown in Fig. 1. Analysis of the validity of the sensitivity factor equations for recovery of (a) Δμa using Eqs. (11)–(22), (b) Δμs′ using Eqs. (23)–(34), and (c) Δr using Eqs. (35)–(40). Results are shown for SFδ (valid for infinitesimal changes; red, blue, and green curves) and SFΔ (valid for finite changes; magenta, cyan, and purple curves). Equations (29)–(31) express SFsΔ for IM in terms of optical properties or moments, whereas Eqs. (32)–(34) express SFsΔ for SM. These equations for finite changes (Eqs. (29)–(34)) simplify to those for infinitesimal changes (Eqs. (23)–(28)) as Δμs′ approaches 0. Sensitivity factors for Δr Equations (35)–(37) express SFrΔ in terms of optical properties, while Eqs. (38)–(40) express them in terms of moments. The change in interoptode distance Δr is a known parameter during measurements and hence it is not meaningful to recover it from the measured data. However, we present SFrΔ equations because they can be used to recover absolute optical properties from multi-distance measurements [37]. In such data analyses, Δr is a known parameter, SFrΔ is unknown, which is a function of μa and μs′ , and the changes in moments are the measurands. In contrast, the sensitivity factor approach is typically used to recover the changed parameter rather than the baseline optical properties. The variable rav is the average of the initial and changed values of r, given by rav=r+Δr/2 . MTSFrΔ and VSFrΔ in IM (Eqs. (36) and (37)) do not depend on r, consistent with the equations for m1 and V (Eqs. (4) and (5)). Modified Beer–Lambert law for finite Δμa The equation for MPPaδ (Eq. (11)), which relates ΔA to an infinitesimal δμa , is analogous to the modified Beer–Lambert law (MBLL), but expressed in terms of optical properties rather than the differential pathlength factor (DPF) . Here, we derive analytical equations for the MBLL that are valid for finite Δμa by expressing MPPaΔ in terms of DPF. The MBLL, which is valid for infinitesimal δμa , is [16]: where DPF (dimensionless) is a scaling factor that relates r to the mean pathlength travelled by photons [38]. G is an unknown geometry-dependent factor that accounts for light losses due to scattering and measurement geometry. Assuming G remains constant during a measurement, Eq. (41) allows recovering changes in absorption using the relation: ΔA=rDPFΔμa . By definition, DPF=cm1/r=MPPaδ/r . Using the equation for MPPaδ (Eq. (11)), we express DPF in terms of optical properties for IM and SM: As expected, DPF increases with μs′ and decreases with μa , whereas the MBLL assumes that DPF remains constant after Δμa . DPF is independent of r for the IM, but there is a small dependence on r for the SM. To derive analytical equations for the MBLL that are valid for finite Δμa in IM and SM, we express MPPaΔ (Eq. (17)) in terms of MPPaδ (Eq. (11)). For the IM geometry, the relation is: We substitute the following two definitions: MPPaΔ=ΔA/Δμa and MPPaδ=rDPF . The resulting equation for the MBLL that is valid for finite Δμa,IM in the IM is: The parameter Δμa,IM in Eq. (44) can be isolated to simplify its recovery from the measured ΔA : The first term is the same as the MBLL (Eq. (41)). The second term, which contains ΔA2 , is an order of magnitude smaller than the first term for a 10% Δμa and typical tissue optical properties μa = 0.1 cm−1 and μs′=10cm−1 . Equation (45) is the same as Eq. (54) in the Appendix but expressed in terms of DPF given in Eq. (42) rather than solely in terms of optical properties. Repeating the above approach for the SM geometry yields the following equation for the MBLL that is valid for finite Δμa,SM in the SM: A closed-form solution for Δμa,SM could not be obtained from Eq. (46), so nonlinear solvers should be used to recover Δμa,SM . Alternatively, the closed-form equation for Δμa,IM (Eq. (45)) can be used as a good approximation for recovering Δμa,SM , as later shown in Fig. 3. Equations (44) and (46) contain the baseline values of μa and μs′ . However, errors in these parameters lead to negligible errors in the recovered Δμa , as later shown in Fig. 2. Errors in the recovered Δμa due to errors in the assumed DPF value, shown for a range of errors in DPF (a) and for a range of Δμa values (b). Equations (44) and (46) are the derived finite-change equations of the modified Beer–Lambert law (Eq. (41)) for infinite medium (IM) and semi-infinite medium (SM) geometries, respectively. Errors in the recovered Δμa due to errors in the assumed baseline values of μa (top row) and μs′ (bottom row) using sensitivity factors for ΔA (a, d), Δm1 (b, e), or ΔV (c, f). Legends show the magnitude of Δμa relative to baseline. Accurate DPF value was used in Eq. (46). The MBLL is typically used for recovering changes in concentrations (ΔC) of chromophores, rather than Δμa . In NIRS, the main chromophores of interest are oxyhemoglobin (CHbO2) and deoxyhemoglobin (CHb) . Their concentration changes can be recovered using the relation: Δμa(λ)=εHbO2(λ)ΔCHbO2+εHb(λ)ΔCHb , where εHbO2 and εHb are the molar extinction coefficients of oxy- and deoxyhemoglobin, respectively, at wavelength λ [39]. Measurements at additional wavelengths enable recovery of changes in more chromophores, such as cytochrome-c-oxidase, which provides information on the rate of cellular oxygen consumption [4]. Results All results presented below were obtained for the SM geometry, and results for the IM geometry have negligible differences (not shown). The baseline optical properties of the medium are μa=0.1cm−1 , μs′=10cm−1 , n=1.4 , and r = 3 cm, unless stated otherwise. All results are presented as percentages relative to these values. Verification of equations To verify the sensitivity factor equations, we generated DTOFs using Eq. (2) for the baseline parameters and after parameter changes, and calculated ΔA,Δm1 , and ΔV from these DTOFs over the range from 0 to 7 ns. We recovered the parameter changes by solving the equations in Tables 3 to 7 using the fsolve function in MATLAB. Figure 1 presents the recovered Δμa,Δμs′ , and Δr . For an increase in μa (Fig. 1(a)), the recovered Δμa is underestimated when using SFaδ (Eqs. (11)–(13)), and the error increases with the magnitude of the change. The values of SFaδ decreases with increasing μa , whereas they are assumed constant in the recovery of Δμa . In contrast, SFaΔ (Eqs. (17)–(19)), is a function of Δμa and allows accurate recovery of Δμa of any magnitude, except for a slight deviation when using VSFaΔ , resulting in an error of approximately 2% (or 10%) for a 50% (or 100%) increase in μa . For Δμa larger than 100%, the error grows rapidly and Eq. (5) may be used instead of VSFaΔ . We found that these percentage errors have only a small dependence on the baseline optical properties and r. The equation for V (Eq. (5)) is accurate for any magnitude of Δμa , but it is challenging to invert it for Δμa and it is not well suited for stable numerical solution. For an increase in μs′ (Fig. 1(b)), the percentage errors in the recovered Δμs′ are similar to the percentage errors in the recovered Δμa using MPPaδ (Fig. 1(a)), as expected since SFsδ are proportional to μs′−1/2 and MPPaδ is proportional to μa−1/2 . MPPsδ has an additional term containing μs′−1 , but its value is much smaller than the term containing μs′−1/2 . The results presented in Fig. 1(c) are for two cases: using rav (corresponding to SFrΔ ) or setting rav=r (corresponding to SFrδ ) in Eqs. (35)–(37). MTSFrΔ and VSFrΔ in IM have no dependence on r, and there is a negligible dependence in SM. Therefore, the errors resulting from using r instead of rav are zero in IM and negligible in SM. Effect of errors in baseline values of μa,μs′ , or DPF The presented equations for parameter recovery require knowledge of the baseline μa,μs′ , and DPF. Here, we examine how errors in these parameters affect the recovery of Δμa . Figure 2 shows the errors in the recovered Δμa due to errors in the baseline μa or μs′ , for different magnitudes of Δμa (10, 20, and 30%, corresponding to 0.01, 0.02, and 0.03 cm−1, respectively). Consistent with the results in Fig. 1, the errors in recovered Δμa depend on the magnitude of Δμa when using SFaδ , but not when using SFaΔ . When μa is overestimated by 100% (Fig. 2(a)), the Δμa values recovered using MPPaΔ are overestimated by about 30%, corresponding to recovered values of approximately 0.013, 0.026, and 0.039 cm−1 for the three magnitudes of Δμa . When μs′ is overestimated by 100% (Fig. 2(d)), the Δμa values recovered using MPPaΔ are underestimated by about 30%. If the percentage errors in μa and μs′ are similar, the error in the recovered Δμa will be close to zero, as expected, since MPPaΔ (Eq. (17)) is nearly proportional to the ratio μs′/μa . When using MPPaΔ , the errors in the recovered Δμa remain nearly independent of the magnitude of Δμa despite large errors in the baseline μa and μs′ , unlike when using MPPaδ . The derived equations for the MBLL (Eqs. (44) and (46)) contain the baseline μa and μs′ , which cannot be estimated using only continuous-wave measurement [6]. However, large errors in these parameters have a negligible effect on the recovered Δμa when the value of DPF is known accurately (Fig. 2(a, d)). When using MTSFa (Fig. 2(b, e)) or VSFa (Fig. 2(c, f)), errors in the baseline μs′ and especially μa can lead to large errors in the recovered Δμa . For example, Δμa recovered using VSFaΔ is overestimated by approximately 60% when μa is overestimated by 20%. Therefore, when using MTSFa or VSFa , the baseline optical properties should be known accurately for reliable recovery of Δμa . Figure 3 shows the errors in the recovered Δμa due to errors in the assumed value of DPF when using the MBLL (Eq. (41)), or the presented Eqs. (44) and (46) for IM and SM, respectively. Figure 3(a) presents results for a range of errors in the DPF value, and Fig. 3(b) presents results for a range of Δμa magnitudes. A percentage error in DPF leads to a similar but negative percentage error in the recovered Δμa (Fig. 3(a)), as expected since DPF is a scaling factor. When DPF is overestimated by 20%, the recovered Δμa values are underestimated by approximately 18%. When using the MBLL, the percentage error in the recovered Δμa depends on the magnitude of Δμa , in contrast to using the derived Eqs. (44) and (46). Notably, the values of Δμa recovered using Eq. (45) are similar to those recovered using Eq. (46). For a 100%, 200%, and 300% Δμa in SM, Eq. (45) recovers Δμa as 104%, 212%, and 322%, respectively. Figure 3(b) shows the corresponding results up to 100%. For comparison, the corresponding values of Δμa recovered using the MBLL (Eq. (41)) are 85%, 153%, and 211%, respectively. Therefore, Eq. (45), which is derived for IM, may also be used for SM measurements with small errors, especially when DPF is known accurately (Fig. 3(b)). Spuriously recovered Δμa due to Δμs′ (or Δμs′ due to Δμa ) The sensitivity factors for recovery of Δμa or Δμs′ assume that only one optical property changes, since a single measurand cannot be used to recover changes in multiple parameters. Figure 4 shows the spuriously recovered Δμa or Δμs′ arising from unaccounted changes in the other optical property. Spuriously recovered Δμa due to Δμs′ (a), and Δμs′ due to Δμa (b). Recovery in (a) uses the SFaΔ (Eqs. (20)–(22)), and recovery in (b) uses the SFsΔ equations (Eqs. (32)–(34)). The ground-truth values are Δμa=0 in (a) and Δμs′=0 in (b). A 20% increase in μs′ leads to a spuriously recovered Δμa of approximately 30%, –25%, and –10% when recovered from ΔA,Δm1 , and ΔV , respectively (Fig. 4(a)). Similarly, a 20% increase in μa leads to a spuriously recovered Δμs′ of approximately 15%, –15%, and −35% when recovered from ΔA,Δm1 , and ΔV , respectively (Fig. 4(b)). These results were obtained for r = 3 cm. For r = 1.5 cm, the magnitudes of the spuriously recovered changes were approximately 25% lower than those obtained for r = 3 cm (results not shown). When both optical properties change, they can be recovered from changes in two or more measurands using Eq. (47) in the Discussion. Discussion In the present study, we extended previously established sensitivity factor equations for infinitesimal changes in μa(SFaδ) [25] to finite changes, yielding the SFaΔ equations. These equations can be used to recover Δμa in homogeneous media from changes in statistical moments of DTOFs, which are typically acquired with time-resolved NIRS. By correcting for finite changes, SFΔ improves recovery accuracy for Δμa and enables accurate recovery of large Δμa , as shown in Fig. 1. In addition, we derived sensitivity factor equations for changes in μs′ and r. Recovery using SFδ underestimates positive changes, as shown in Fig. 1, and overestimates negative changes (not shown). This forms the basis of the finite-change correction used to derive SFΔ . The SFΔ equations reduce to the SFδ equations in the limit of infinitesimal changes. The modified Beer-Lambert law (MBLL) is valid for infinitesimal changes in μa due to the assumption that DPF remains constant with respect to Δμa , which is not valid according to the equation for DPF (Eq. (42)). We present equations for the MBLL that are valid for finite changes (Eq. (44) for IM and Eq. (46) for SM), which can be applied in data analysis algorithms used for continuous-wave NIRS measurements. They require knowledge of the baseline values of μa and μs′ , in addition to DPF, and these parameters cannot be estimated from continuous-wave measurements with a single source-detector pair [6]. However, errors in the baseline μa or μs′ (up to 100%) have a negligible effect (a few percent) on the recovered Δμa (Fig. 2(a, d)). Therefore, baseline optical properties can be assumed within the range of values in literature for the measured tissue, e.g., μa = 0.1 cm−1 and μs′=10cm−1 for many biological tissues [13–15], resulting in only small percentage errors in the recovered Δμa . The value of DPF is typically taken from literature [38], and a percentage error in DPF leads to a similar percentage error in the recovered Δμa (Fig. 3), as expected since DPF is a scaling factor. The parameter recovery based on SFΔ can be done using nonlinear solvers (e.g., fsolve in MATLAB). Alternatively, for some of the SFaΔ equations we found analytical solutions for Δμa , which are presented in the Appendix and do not require nonlinear solvers. The SFΔ equations require knowledge of the baseline optical properties, and large errors in these parameters can lead to large errors in the recovered Δμa (Fig. 2). However, if the IRF is measured, the baseline μa and μs′ can be estimated using the moments approach [24] or by fitting a solution of the DE to the baseline DTOF [27] . Although this approach allows estimating μa and μs′ at each time point, analysis of changes in moments is computationally simpler and advantageous for monitoring dynamic changes. The baseline DTOF can be acquired over a longer time to estimate baseline μa and μs′ with higher precision, after which changes in moments can be used to recover subsequent dynamic changes in μa . The SFΔ introduce a finite-change correction to the established data analysis approach based on SFδ . Therefore, we expect their practical robustness to be similar. However, a separate stuty is needed to assess the effect of experimental factors including photon-count noise, DTOF cut-off region, instrument response function (IRF), and mismatch between the DE model and the measured medium. Theoretical uncertainty in DTOF moments due to photon noise, and the resulting uncertainty in recovered Δμa , has previously been analyzed for optimizing source-detector distance r [40]. A recent study showed the improved performance of time-domain moment techniques in comparison to conventional continuous-wave approaches in recovering hemodynamic responses [28]. A quantitative performance comparison using measured data showed that DTOF moments outperform measurands based on photon counts in time windows and their ratios [34]. All derived equations assume that the DTOF extends to infinite arrival time of photons. The calculated moments depend on the chosen cut-off noise region on the time axis, which is typically set at 1% of the maximum of the DTOF. A correction procedure was proposed to correct the values of moments due to a limited cut-off region [41]. In our previous study [42], in data analysis we compared changes in moments of the measured DTOFs to changes in moments of the analytically obtained DTOFs convolved with the IRF, such that both measured and analytically obtained moments are calculated using the same cut-off regions. DTOFs for homogeneous and multi-layered media can be obtained using a publicly available MATLAB script [42], which implements the solution of the DE presented by Liemert et al. [43]. This study considers a homogeneous medium, whereas measurements on layered tissue structures, such as the head, should use a multi-layered model [44,45]. Perturbation-based sensitivity factors SFδ can be obtained for any geometry and complex medium using Monte Carlo simulations, e.g., for MRI-based head models [46]. Equation (10) provides a general way to recover finite changes using SFδ calculated for both the baseline and perturbed states. When analytical expressions for SFδ are available, they can be substituted into Eq. (10) to derive SFΔ equations, as done in this study. When only numerical SFδ values are available, such as from simulations, a possible strategy is to use Eq. (10) iteratively: starting with Δμa=0 , estimate Δμa , recalculate SFδ for the perturbed state, and repeat until the estimated Δμa changes negligibly between iterations. The convergence and accuracy of such an approach would need to be verified for the specific geometry, medium, and measurand. In measurements on tissues, μs′ is typically assumed constant because separating Δμa from Δμs′ is challenging [47]. However, Δμs′ may be caused by changes in total hemoglobin concentration, which changes the number of scatterers within the volume of tissue interrogated by photons [48]. Studies carried out on rat brain slices have shown that Δμs′ can appear in line with Δμa [49–51]. In a number of studies, Δμs′ following fast neuronal activity have been reported in animals [52] and humans [53,54]. Changes in two or more measurands M, i.e., ΔA,Δm1 , and/or ΔV , allow recovering both Δμa and Δμs′ . Note that SFΔ equations were derived for changes in only one optical property while the other is assumed constant. The effects of changes in the other optical property are shown in Fig. 4. However, both Δμa and Δμs′ can be simultaneously recovered by using each SFΔ for a change in only one parameter. Time-resolved measurements allow estimating the pathlength traveled by detected photons from their arrival time. When the MBLL is applied to photon counts with the same arrival time, it remains valid for any magnitude of Δμa . The mean pathlength within a time channel is known as the time-dependent MPP (TMPP) [55]. This approach is also referred to as the microscopic BLL [17], which corresponds to applying the BLL to pathlength-resolved light intensity. This enables recovery of Δμa at multiple depths, since late photons, on average, probe deeper regions of the medium than early photons [55]. Similarly, sensitivity factors for higher-order moments have increased sensitivity to deeper compartments, allowing moments-based methods to recover Δμa in multiple layers [25,42,56]. The influence of changes in superficial layers is an inherent major challenge in brain functional studies [44]. The derived equations rely on the diffusion equations (DE) for IM and SM, which approximate the radiative transfer equation (RTE), which is a specific case of the more general transport theory applied to light propagation [1,57]. We did not introduce new assumptions and the derived SFΔ equations remain valid for any magnitude changes, as long as the DE remains valid (Eqs. (1) and (2)). The DE relies on the diffusion approximation, which assumes that μs′ is much greater than μa , and either the photon arrival time (for time-resolved measurements) or the source-detector distance r (for non-time-resolved measurements) is sufficiently large for light propagation to be diffusive [1]. In measurements on tissues, μs′ is typically two orders of magnitude higher than μa in the NIR range [13–15], hence the DE remains valid for large Δμa , provided that scattering continues to dominate absorption. The presented finite-change correction may also be extended to other geometries, such as slab (transmission geometry), or for multi-layer tissue models. Such extensions would need to be derived and validated separately. Conclusions We derived equations for sensitivity factors that accurately relate ΔA and changes in the statistical moments of DTOFs ( Δm1 and ΔV ) to finite (even large) Δμa,Δμs′ , and Δr in infinite medium IM and semi-infinite medium SM geometries. We verified the equations and showed the improvement in the accuracy of parameter recovery compared to the established perturbation-based approaches (Fig. 1). We derived equations for the modified Beer–Lambert law (MBLL) that are valid for finite Δμa in IM and SM (Eqs. (44) and (46)). These equations require knowledge of the baseline μa and μs′ , in addition to the differential pathlength factor DPF. However, even a 100% relative error in μa results in only a few percent error in the recovered Δμa , whereas errors in μs′ have an even smaller effect (Fig. 2). Therefore, typical values of μa and μs′ from literature can be assumed for tissue measurements. The sensitivity factor equations require knowledge of the baseline μa and μs′ , and large errors in these parameters can lead to large errors in the recovered parameter changes (Fig. 2). Therefore, optical properties at baseline need to be estimated, after which subsequent dynamic changes can be recovered using sensitivity factors. Changes in the other optical property lead to spurious changes in the recovered optical property (Fig. 4), in which case both Δμa and Δμs′ should be recovered from changes in two or more measurands (Eq. (47)). Future experimental work should assess the robustness of the presented equations to photon noise, instrumental effects, DTOF cut-off region, and model mismatch. However, the robustness is expected to be comparable to that of the established approach based on sensitivity factors for infinitesimal changes, since the presented equations only introduce a finite-change correction. The presented equations can improve estimation of changes in optical properties in NIRS. The derivation approach may also be useful for deriving finite-change corrections to other perturbation-based equations in diffuse optics, although they would need to be derived and validated separately. Appendix The sensitivity factor equations for finite absorption changes SFaΔ (Eqs. (17)–(22)) were derived by substituting the below expressions for SFa(μa) and SFa(μa∗) into Eq. (10). The two expressions used to derive MPPaΔ are the MPPaδ equation (Eq. (11)) for the baseline μa and the perturbed state μa∗ , where μa∗=μa+Δμa , and μe=3μaμs′ , and μe∗=3μa∗μs′ . The same approach was used to derive all SFsΔ equations for Δμs′ . To derive MTSFaΔ and VSFaΔ for IM, both expressions SFa(μa) and SFa(μa∗) contain both μa and μa∗ . To derive MPPaΔ for IM (Eq. (17)), the expressions used are: To derive MPPaΔ for SM (Eq. (20)), the expressions used are: To derive MTSFaΔ for IM (Eq. (18)), the expressions used are: To derive VSFaΔ for IM (Eq. (19)), the expressions used are: To derive MTSFaΔ and VSFaΔ for SM, their equations for IM are used together with the ratios that distinguish the moment equations (Eqs. (4)–(6)) for IM from those for SM. To derive MTSFaΔ for SM (Eq. (21)), the expressions used are: To derive VSFaΔ for SM (Eq. (22)), the SFa(μa) and SFa(μa∗) expressions used are: Nonlinear solvers, such as fsolve function in MATLAB, should be used to recover Δμa from the measured changes in moments using the corresponding SFaΔ equations (Eqs. (17)–(19)). Alternatively, for some of these equations we found analytical solutions for Δμa , which do not require nonlinear solvers. When using MPPaΔ for IM (Eq. (17)), the closed-form solution for Δμa,IM is: The first term is equivalent to using MPPaδ (Eq. (11)). The second term, which contains ΔA2 , is an order of magnitude smaller than the first term for a 10% Δμa and typical tissue optical properties μa=0.1cm−1 and μs′=10cm−1 . Equation (54) is the same as Eq. (45), but with DPF expressed in terms of optical properties. The MPPaΔ equation for SM (Eq. (20)) is too complex to find an exact closed-form solution for Δμa,SM . However, we found an approximate solution from numerical inversion, which is practically as accurate as the exact solution: When using MTSFaΔ for IM (Eq. (18)), the closed-form solution for Δμa,IM is: When using VSFaΔ for IM (Eq. (19)), we found an approximate solution from numerical inversion, which is practically as accurate as the exact solution: Equation (43) shows the relation between MPPaΔ and MPPaδ in IM, which was used to derive the finite-change equations for the MBLL. For completeness, we present the relation between MTSFaΔ and MTSFaδ (Eq. (58)), and between VSFaΔ and VSFaδ (Eq. (59)) in the IM. Equations (43), (58), and (59) have a common term (in the first brackets), and the equations for higher-order moments contain an additional term (in the second brackets). This additional term reflects the greater deviation from linearity of higher-order moment changes with respect to changes in μa . Funding National Science Centre https://ror.org/03ha2q922 ( 2020/39/D/ST7/03425, 2021/43/I/ST7/02900); Institute of Biocybernetics and Biomedical Engineering https://ror.org/05f8zcs55 ( ST263). Disclosures The authors declare no conflicts of interest. Data availability Data shown in this article are generated using analytical formulas as presented. Therefore, no data is available. References Associated Data Data Availability Statement Data shown in this article are generated using analytical formulas as presented. Therefore, no data is available.