WorksPhysics and Astronomy
Stochastic dose fluctuations in boron neutron capture therapy with heterogeneous boron distributions
Abstract
Dose fluctuations at the cellular scale in boron neutron capture therapy (BNCT) arise from stochastic neutron capture and spatially heterogeneous boron distributions. This study develops a stochastic framework combining energy-deposition responses from Geant4 simulations, a Poisson capture process, and spatially correlated lognormal boron fields to investigate nuclear dose variance and zero-hit probability. The results show that the variance contribution from stochastic neutron capture increases linearly with neutron fluence, whereas the contribution from boron heterogeneity increases quadratically. These distinct scaling behaviors produce a correlation-length-dependent crossover between the dominant sources of dose fluctuations. For a heterogeneity amplitude of and a correlation length of , the crossover occurs at approximately , while the ensemble relative dose variance approaches a plateau of at high fluence. Boron heterogeneity also increases the probability that a nucleus receives no energy deposition from alpha–lithium reaction products. At and , the zero-hit probability is higher than that for a homogeneous boron distribution with the same mean concentration. This study shows that spatial boron correlations and stochastic neutron capture influence nuclear dose fluctuations and zero-hit statistics beyond what can be characterized by mean boron concentration alone.
Introduction
Boron neutron capture therapy (BNCT) is a radiotherapy approach that combines the delivery of boron-containing compounds with neutron irradiation. Neutron capture by produces an alpha particle and a nucleus, which deposit their energy over distances of a few micrometers, comparable to the dimensions of cells and their nuclei. The short range of these particles provides the basis for the potential selectivity of BNCT, but also makes energy deposition within cellular structures highly dependent on the location of boron and the sites of neutron capture [5, 8]. Advances in BNCT technology, particularly accelerator-based neutron sources, boron delivery agents, and imaging techniques, have expanded its potential applications in cancer treatment [3, 4, 40]. Clinical progress, especially in the treatment of head and neck cancers, has further highlighted the need for dose calculations that account not only for energy deposition at the tissue level but also for its distribution at the cellular scale [13, 30, 42].
One of the main challenges in BNCT dosimetry is the heterogeneous distribution of boron. Delivery agents such as boronophenylalanine (BPA) and sodium borocaptate (BSH) can produce different boron concentrations across tissues, individual cells, and subcellular compartments [19, 6]. New boron delivery agents continue to be developed to improve boron accumulation and retention in target tissues [37, 39, 27]. Nevertheless, tumor heterogeneity and variations in boron localization remain important challenges when relating mean boron concentration to the dose received by cellular structures [43]. The influence of nonuniform boron distributions has been demonstrated in dose calculations based on -BPA positron emission tomography (PET), which show differences between heterogeneous and homogeneous boron assumptions [34, 14]. Recent PET-based modeling studies have also emphasized the importance of representing boron heterogeneity in treatment planning and in evaluating dose–toxicity relationships [33, 32]. However, tissue-level imaging cannot fully resolve variations in boron concentration at the subcellular scale. Developments in neutron autoradiography, particle track detectors, and boron isotope quantification have further highlighted the importance of characterizing boron microdistributions [36, 10, 28]. Beyond the magnitude of boron concentration variations, the spatial arrangement of boron is also important. Two boron distributions with the same mean concentration and heterogeneity amplitude may have different correlation lengths, leading to different degrees of spatial averaging as the reaction products deposit energy within the nucleus.
The effects of boron microdistribution and stochastic energy deposition have been investigated through several approaches to BNCT microsimetry. Santa Cruz and Zamenhof developed an analytical framework that accounts for boron localization and stochastic energy deposition at for microscopic scales [24]. Sato et al. subsequently developed stochastic microdosimetric kinetic models incorporating intra- and intercellular boron heterogeneity to estimate biological effectiveness [25, 26]. Particle transport simulations have also been used to investigate energy deposition in microscopic targets [12], while microdosimetric measurements at different site sizes have demonstrated the importance of spatial scale in characterizing energy deposition [29]. More recently, cellular-scale simulation approaches have been developed to relate energy-deposition statistics to biological response parameters in BNCT [31, 38]. Monte Carlo (MC) transport methods have also continued to advance for dose calculation and treatment planning [41, 21]. Other recent studies have combined boron microdistribution information with dosimetric modeling, including stochastic representations of boron concentrations in cellular compartments [16, 17]. Together, these developments show that mean absorbed dose alone may not adequately describe the stochastic nature of energy deposition at microscopic scales [23].
Although the importance of boron heterogeneity is well recognized, the relationship between its spatial correlation length and the relative contributions of different mechanisms to dose fluctuations requires further characterization. For a fixed boron distribution, dose fluctuations arise from the random number of neutron capture events and the energy deposited by each event. In contrast, variations between boron distributions produce differences in conditional mean dose that depend on their spatial correlation structure. These two mechanisms have different dependencies on neutron fluence, so their relative contributions can change even when the mean boron concentration remains fixed. Spatial correlations also affect the probability that a nucleus receives no energy deposition from the reaction products. A statistical framework is therefore needed to connect boron correlation length with dose variance decomposition and zero-hit probability.
In this study, the boron distribution is represented by a spatially correlated lognormal random field, while neutron capture events are modeled as a Poisson process conditional on the boron field. Variations in capture intensity between field realizations give rise to a Cox process [18]. Single-event energy-deposition responses, including the mean deposited energy, its second raw moment, and the hit probability, are obtained from Geant4 simulations [1, 2] and combined with the stochastic formulation to determine nuclear dose statistics. This approach allows the total dose variance to be separated into contributions from stochastic neutron capture and spatial boron heterogeneity, with their dependence on neutron fluence determined analytically. The correlation length is varied while keeping the mean boron concentration and heterogeneity amplitude fixed to determine the fluence at which the two variance contributions become equal and the limiting relative variance at high fluence. The zero-hit probability is also calculated and compared with the homogeneous-boron reference. Analytical predictions are compared with MC sampling to assess the consistency of the formulation and its numerical implementation.
The contribution of this study is a quantitative characterization of how boron correlation length influences the mechanisms of dose fluctuations and the zero-hit probability at the nuclear scale. Unlike approaches that directly predict biological responses, the present work focuses on the physical statistics of energy deposition by alpha–lithium reaction products. The analysis is restricted to a spherical nucleus in a water-equivalent medium with a spatially uniform neutron field and electromagnetic transport of the reaction products. The zero-hit probability is not interpreted as a cell survival probability, since other BNCT radiation components and cellular biological responses are not included in the model.
Model and Methods
Stochastic boron deposition
The model describes how spatial boron heterogeneity influences the stochastic nature of neutron capture and energy deposition in cellular targets, as previously investigated in BNCT microdosimetry [24, 25, 17]. The surrounding medium is assumed to be water-equivalent, with fixed composition and density, while the distribution of is allowed to vary spatially. These variations are assumed not to affect the transport properties of the medium or the neutron field. The number density of atoms is represented by a spatially correlated lognormal random field,
where and is a stationary Gaussian random field with zero mean and unit variance. The lognormal representation ensures that the local boron concentration remains positive and provides a random intensity field suitable for a Cox process [18]. In the neutron capture equations, denotes the number of atoms per unit volume. The mass concentration used as a numerical input is converted into this quantity.
The Gaussian correlation function is chosen as , where and is the Gaussian correlation-length parameter. The heterogeneity amplitude is defined by , giving . The boron concentration covariance, , and its normalized correlation function are given by
for . The boron correlation length is defined through . This definition applies to the concentration field after the lognormal transformation, so and are generally different. The Gaussian correlation parameter satisfying this definition is obtained from
This construction allows the heterogeneity amplitude and the correlation length of the boron concentration field to be specified independently. For , the boron field becomes homogeneous, and a correlation length is no longer required.
For a given realization of the boron field, neutron capture locations follow an inhomogeneous Poisson process with the time-integrated capture intensity
where is a dimensionless capture factor. Here, denotes the differential neutron fluence integrated over the irradiation period, and is the cross section for the reaction. Under the thermal-neutron approximation used in the numerical calculations, . The quantity represents the expected number of neutron captures per unit volume. Accordingly, the number of captures within a volume element , conditional on , follows
In the voxel-based discretization, this integral is approximated by the capture intensity at the voxel center multiplied by its volume. Capture counts in disjoint volumes are independent when conditioned on a fixed boron field [15]. Correlations between different regions arise when variations in the boron field are included. The model neglects boron depletion, changes in its spatial distribution during irradiation, and modifications of the neutron field caused by boron heterogeneity.
Each neutron capture produces an alpha–lithium pair, which is treated as a single combined event. The energy deposited by both reaction products is summed within each history to preserve the correlations associated with their emission directions and energy deposition. For a target with mass and source region , the total deposited energy and absorbed dose are defined as
where is the capture location and includes the reaction branch, emission directions, and transport of the reaction products and their secondaries. The deposited energy is zero when an event does not deposit any energy in the target. All histories, including those with zero energy deposition, are included when calculating the response moments,
Here, is the second raw moment, whereas the variance of the energy deposited by a single event at a fixed location is . For a spherical target in a homogeneous medium, rotational symmetry allows the responses to be expressed as functions of distance from the target center, denoted by and . Deposited energy is converted into joules and target mass into kilograms when the absorbed dose is expressed in gray.
Nuclear dose statistics
Particle transport histories from different capture events are assumed to be conditionally independent for a given boron field. By treating the transport outcomes as marks of the Poisson process, the moment formulas for Poisson integrals give [15]
The conditional mean dose is determined by the mean energy-deposition response weighted by the local capture intensity. The conditional variance involves the second raw moment because it accounts for fluctuations in both the number of capture events and the energy deposited by each event. The mean response alone is therefore insufficient to determine the dose variance.
Averaging over the ensemble of boron fields gives
For the same mean boron concentration and transport response, the ensemble mean dose is identical to that obtained under the homogeneous-boron assumption. The effect of boron heterogeneity instead appears in the dose fluctuations. Applying the law of total variance yields
where
The stochastic contribution represents the mean conditional variance for a fixed boron field, whereas the heterogeneity contribution measures the variation in conditional mean dose between different field realizations. Thus, this decomposition applies to an ensemble that includes both variations in boron distributions and repeated capture events. For repeated irradiation of the same fixed boron field, only the conditional variance in Eq. (9) contributes to the fluctuations.
If increasing the neutron fluence does not change the boron statistics, neutron energy spectrum, or transport response, the two variance contributions scale as and . Since the mean dose is proportional to , the relative dose variance can be written as
where the fluence-independent coefficients are
The term decreases with increasing neutron fluence, whereas represents the limiting ensemble relative variance at high fluence. Its value depends on the spatial structure of the boron distribution and the target response function. This limiting value does not imply a persistent relative variance when irradiation is repeated for an unchanged boron distribution.
The relative importance of the two variance contributions is measured by
Values of indicate that stochastic capture fluctuations dominate, whereas indicates that boron heterogeneity is the dominant contribution. For , the condition defines the crossover fluence, . This condition corresponds to equal contributions from the two variance components. For a homogeneous boron field, , and no finite crossover fluence exists.
To characterize hit statistics, the probability that a single capture event at position deposits energy above a threshold is defined as . In this study, , so a hit occurs whenever an alpha–lithium event deposits a positive amount of energy in the target. Thinning the Poisson process according to gives the conditional mean number of hits [15],
For a given boron field, the number of hits is Poisson distributed with mean . Averaging the conditional
zero-hit probability over boron field realizations gives . Since the exponential function is convex, Jensen’s inequality gives . The right-hand side is the zero-hit probability for a homogeneous boron distribution with the same mean concentration, neutron fluence, and transport response. This inequality compares heterogeneous and homogeneous boron distributions; the dependence on heterogeneity amplitude and correlation length is evaluated for the specified random-field model. The quantity accounts only for hits from alpha–lithium reaction products and must not be interpreted as a cell survival probability. Such an interpretation requires additional biological response modeling, as considered in previous microdosimetric studies [25, 31, 38].
Geant4-based numerical methods
Monte Carlo methods have been widely used in BNCT dosimetry, including the development of the COMPASS dose calculation engine and GATE-based simulations [41, 21]. In the present study, the energy-deposition response from a single neutron capture event is calculated using Geant4 [1, 2, 7]. The target is modeled as a spherical nucleus with radius embedded in a water-equivalent medium with a density of . Each history begins with the generation of an alpha particle and a nucleus at the same location, emitted in opposite directions with an isotropically distributed orientation. Two reaction branches are considered, with probabilities of 93.7% and 6.3% [25, 12]. The kinetic energies of the alpha particle and lithium nuclei are 1.47 and 0.84 MeV, respectively, for the dominant branch, and 1.78 and 1.01 MeV for the other branch. Particle transport is simulated using G4EmStandardPhysics_option4 and G4StepLimiterPhysics [11]. This configuration includes energy loss and electromagnetic interactions of charged particles without enabling inelastic hadronic nuclear processes. Neutron capture is modeled stochastically outside Geant4, so neutrons are not explicitly transported. Gamma photons from the capture reaction, as well as other neutron and photon dose components, are excluded from the calculations.
The response is evaluated at 40 source distances between and 18 from the nucleus center. At each distance, histories are simulated in each of five independent runs. The transport settings use a maximum step length of 0.10 and a production cut of 0.05 . The latter specifies a range-based threshold for secondary particle production, rather than an energy threshold for terminating the transport of an existing particle. The energy deposited in the nucleus is recorded for every history, including those with zero energy deposition, to calculate the mean response , the second raw moment , and the hit probability . A hit is defined as an event that deposits a positive amount of energy within the target. Uncertainties in and are estimated from the standard errors of the five independent runs, while 95% confidence intervals for are calculated using the Wilson method [35]. The response statistics are checked against the inequalities and , together with the energy balance of the reaction products. The radial responses are linearly interpolated for spatial integration, while contributions beyond are neglected in the main calculations.
The boron field is generated in a cubic domain with side length and a baseline resolution of voxels, corresponding to a voxel size of . A stationary Gaussian field is generated using a Fourier-based spectral method with periodic boundary conditions and subsequently transformed into a lognormal field according to Eq. (1). The Gaussian correlation parameter is chosen so that the concentration field after transformation satisfies the prescribed correlation-length definition . The mean and variance are specified at the ensemble level, without separately normalizing each field realization. The mean boron concentration is set to of per gram of medium and converted into atomic number density using , where , is Avogadro’s constant, and is the molar mass of . The resulting number density is approximately . Thermal-neutron capture is calculated using a cross section of barn [22]. Changing the neutron fluence affects only the capture intensity, while the boron field statistics and particle transport responses remain fixed.
The dose variance decomposition is examined using 2000 independent boron field realizations, with eight stochastic repetitions for each field. Analytical predictions are obtained from Eqs. (12) and (13). In the MC sampling, the number of events producing hits in each volume element follows a Poisson distribution conditional on the local boron concentration and the response . The energy deposited by each hit is sampled from a bounded surrogate distribution matched to the Geant4 response moments. For , the first and second moments of the deposited energy conditional on a hit are and , respectively, whereas events without hits are assigned zero deposited energy. This construction preserves the hit probability and the two energy moments required to determine the mean dose and its variance, but does not reproduce the complete Geant4 energy-deposition distribution for individual histories.
The stochastic variance contribution is estimated by averaging the variances across repetitions within each fixed boron field. The heterogeneity contribution is obtained from the variance of the field-averaged doses after correcting for the stochastic fluctuations associated with the finite number of repetitions. This correction is given by the estimated stochastic variance divided by . The estimator is described in more detail in Appendix § 8.1. Accordingly, the comparison between MC results and analytical predictions tests the consistency of the moment formulation and the stochastic sampling procedure, rather than providing an independent validation of the full energy-deposition spectrum from Geant4.
The nuclear zero-hit probability is calculated by averaging over 2000 boron field realizations and is compared with Poisson hit-count sampling using 16 repetitions per field. To reduce sampling uncertainty across field realizations, a control variate estimator is used, based on the mean hit intensity of the homogeneous-boron reference model [20]. Statistical uncertainty is evaluated using 1000 bootstrap resamples at the field-realization level [9]. All Poisson repetitions associated with the same boron field are retained as a group during resampling, rather than treated as independent field realizations. Consistency checks include probability bounds, Jensen’s inequality, and agreement between the averaged conditional Poisson probabilities and the sampled hit counts. Since both approaches use hit intensities obtained from the same Geant4 response, their comparison validates Poisson sampling within the stochastic model, rather than providing an independent Geant4 transport validation. The probability estimators and control variate correction are detailed in Appendix § 8.2.
Additional numerical checks include comparisons of Monte Carlo variances with analytical predictions, tests of the Geant4 response cutoff over –, and spatial grid sensitivity tests using resolutions of , , , and at a fixed domain size. As shown in Appendix § 8.3, the relative difference in zero-hit probability from the reference calculation at decreases to less than 1% when the response is retained up to and becomes very small beyond for the tested configuration. The asymptotic relative variance coefficient and zero-hit probability also vary by approximately 1% or less between the baseline grid and the reference grid, accounting for sampling uncertainty. These results support using the Geant4 response up to and an grid in the main calculations, although they do not replace convergence tests with respect to domain size or transport parameters. The analysis is limited to a spherical nucleus, a water-equivalent medium, a spatially uniform neutron field, and energy deposition from alpha–lithium reaction products within the electromagnetic transport model considered.
Results and Discussion
Dose variance decomposition
To characterize stochastic energy deposition at the cellular scale, we first examine the response from a single neutron capture event using Geant4 simulations. Each capture produces an alpha particle and a nucleus. The energy deposited in the nucleus depends on the capture location, emission direction, and particle trajectories through the surrounding medium. This dependence on geometry and particle transport is an important aspect of BNCT microdosimetry [24, 12]. The nucleus is modeled as a sphere with radius embedded in a water-equivalent medium. The responses are calculated at 40 source distances between and , while the range is shown in Figure 1 to illustrate their variation near the nucleus. Three response quantities are considered: the mean deposited energy , the second energy moment , and the hit probability . These quantities are used to determine the mean dose, dose variance, and probability that the nucleus receives no hit.
Figure 1. Single-event energy-deposition responses from neutron capture in a spherical nucleus of radius embedded in a water-equivalent medium, calculated using Geant4 with electromagnetic transport. (a) Mean deposited energy, ; (b) second raw energy moment, ; and (c) hit probability, , as functions of the capture-site distance from the nucleus center. The displayed range is –, while the spatial calculations use responses extending to . The vertical dashed line indicates the nuclear boundary at . Symbols represent simulation results, and lines serve as guides to the eye. Error bars indicate the standard errors from five independent runs for and , and 95% Wilson confidence intervals for .
As shown in Figure 1(a) and (Figure 1b), both the mean deposited energy and its second raw moment decrease with increasing distance between the capture site and the nucleus center. At , the mean deposited energy is approximately 1.66 MeV, while the second raw moment is about 2.76 MeV. Both quantities decrease as the source approaches the nuclear surface and fall more rapidly for , where only a limited range of emission directions and particle trajectories can contribute to energy deposition in the target. Although and exhibit similar spatial trends, they describe different statistical properties. The mean response determines the average contribution of a capture event to the dose, whereas is the second raw moment required to calculate the variance of a sum of Poisson-distributed events, as given by Eq. (12). The variance of the energy deposited by a single event at a fixed source location is . However, replacing with this variance would omit fluctuations in the number of neutron capture events. The importance of target size in BNCT microdosimetry has also been demonstrated experimentally at different site sizes [29], although those measurements are not used as direct validation of the present Geant4 responses.
The hit probability in Figure 1(c) illustrates how nuclear geometry and particle range affect the likelihood of energy deposition. For capture sites inside the nucleus, is close to unity and decreases as the source moves outside the target. At , the Geant4 response gives and MeV. The hit probability decreases to approximately at , while no hits were observed in the sampled histories for . The absence of observed hits within the available number of histories does not imply that the true hit probability is exactly zero. The different behaviors of and also show that the mean deposited energy alone cannot determine whether a capture event reaches the target. The response is therefore treated separately when calculating the zero-hit probability in Subsection 3.2.
These Geant4 responses are then used to calculate dose statistics for spatially correlated lognormal boron fields. According to Eq. (11), the total dose variance is separated into the stochastic capture contribution and the spatial boron heterogeneity contribution . The baseline calculations use a mean concentration of 30 /g, a heterogeneity amplitude of , and a correlation length of . The neutron fluence is varied from to . Monte Carlo checks are performed using 2000 boron field realizations, with eight stochastic repetitions per field. The variance decomposition, relative dose variance, and changes in the dominant fluctuation mechanism are presented in Figure 2.
Figure 2. Decomposition of nuclear boron dose variance as a function of neutron fluence . (a) Contributions from stochastic neutron capture (), boron heterogeneity (), and the total variance. (b) Relative dose variance for homogeneous and heterogeneous boron distributions, together with the high-fluence limit arising from boron heterogeneity. (c) Dominance ratio for correlation lengths , 3, and 6 . The condition marks the crossover fluence at which the two variance contributions are equal. Lines show analytical predictions, while symbols in (a) and (b) represent MC results. Symbols in (c) mark the analytically determined crossover points. The calculations use Geant4 responses with electromagnetic transport, , , and . Monte Carlo checks use 2000 boron field realizations with eight stochastic repetitions per field.
Figure 2(a) shows that the two variance contributions have different dependencies on neutron fluence. The stochastic contribution increases linearly, , whereas the heterogeneity contribution increases quadratically, . At low fluence, the total variance is mainly determined by fluctuations in the number of capture events and the energy deposited by each event. As fluence increases, the heterogeneity contribution grows more rapidly and eventually becomes dominant. For the baseline configuration, the two contributions become equal at . This crossover identifies a change in the dominant source of dose fluctuations within the model, rather than a phase transition or a threshold for biological effectiveness. The MC results follow the analytical predictions over the fluence range examined.
The difference between the two mechanisms becomes clearer when the variance is normalized by the square of the mean dose, as shown in Figure 2(b). For a homogeneous boron distribution, the relative variance decreases inversely with neutron fluence, because only stochastic capture fluctuations contribute. For heterogeneous boron fields, the relative variance instead approaches a constant at high fluence as the heterogeneity contribution becomes dominant. For the baseline parameters, Eq. (14) gives the numerical relation
The first term represents the relative fluctuations arising from stochastic neutron capture, which decrease with increasing fluence. The second term determines the limiting relative variance at high fluence. The plateau value depends on the heterogeneity amplitude, the spatial correlation structure of the boron field, and the target response, and is therefore not universal. Importantly, this limit describes variations between boron field realizations, rather than a persistent relative variance under repeated irradiation of the same fixed distribution. The importance of boron microdistribution and stochastic energy deposition has been discussed in previous microdosimetric models [25, 17]. Here, we specifically examine how the relative contributions of these two variance mechanisms change with neutron fluence and boron correlation length, without modeling biological responses.
The influence of correlation length on the crossover is shown in Figure 2(c). The dominance ratio increases linearly with neutron fluence, but the position of depends on the boron correlation length. For , the crossover fluences are , , and for , 3, and 6 , respectively. Increasing the correlation length from 1.5 to 6 therefore reduces the crossover fluence by approximately one order of magnitude. This behavior arises because boron concentration variations correlated over longer distances undergo less spatial averaging through the energy-deposition response. As a result, the differences in conditional mean dose between field realizations become larger, allowing boron heterogeneity to dominate at lower fluence. The crossover points in panel (c) are obtained analytically rather than from independent Monte Carlo estimates.
Quantitative agreement between MC sampling and the analytical predictions is assessed using the ratio of the two results for each variance component. With 2000 boron field realizations, the Monte Carlo-to-theory ratios range from 0.972 to 1.028 for the stochastic contribution, 0.980 to 1.101 for the heterogeneity contribution, and 0.995 to 1.034 for the total variance. As a separate check, the variance of the conditional mean doses across field realizations gives a ratio of 1.020 relative to the analytical heterogeneity prediction. These results show that the variance decomposition is reproduced consistently by stochastic sampling, although the heterogeneity component has a larger statistical uncertainty. Additional sensitivity tests involving the Geant4 response cutoff and spatial grid resolution are presented in Appendix § 8.3.
These results show that nuclear dose fluctuations arise from two mechanisms with distinct fluence dependencies. The boron correlation length controls the magnitude of the heterogeneity contribution and the fluence at which it becomes dominant, even when the mean boron concentration remains fixed. Thus, mean absorbed dose alone does not adequately characterize dose variations at the nuclear scale. These findings are limited to alpha–lithium energy deposition, the target geometry, and the statistical properties of the boron fields considered here.
Nuclear zero-hit probability
In addition to affecting dose variance, spatial boron heterogeneity also influences the probability that a nucleus receives no energy deposition from neutron capture reaction products. The zero-hit probability, , is defined as the probability that no event deposits a positive amount of energy within the nucleus during irradiation. In the Poisson model, this probability depends on the neutron capture intensity and the single-event hit probability obtained from the Geant4 simulations in Figure 1(c). For a given boron field realization, the zero-hit probability is , while the ensemble probability is obtained by averaging this quantity over the correlated lognormal boron fields. The calculations use Geant4 responses with electromagnetic transport, a mean concentration of , and a spherical nucleus of radius . The effects of neutron fluence, heterogeneity amplitude , and correlation length on the zero-hit probability are presented in Figure 3.
Figure 3. Nuclear zero-hit probability () associated with alpha–lithium reaction products in heterogeneous boron fields. (a) Dependence of on neutron fluence for different heterogeneity amplitudes at . (b) Variation of with at and , with . (c) Zero-hit probability ratio as a function of the normalized correlation length , where is an external reference length scale and . Curves represent the ensemble averages of the conditional Poisson zero-hit probabilities, while open symbols in (a) and (b) show MC sampling of Poisson hit counts. Shaded bands in (c) indicate 95% bootstrap confidence intervals. The calculations use Geant4 responses with electromagnetic transport, , , and 2000 boron field realizations with 16 Poisson repetitions per field.
Figure 3(a) shows that the zero-hit probability decreases monotonically with increasing neutron fluence. This decrease occurs because higher fluence increases the expected number of neutron captures capable of depositing energy within the nucleus. At low fluence, the curves for homogeneous and heterogeneous boron distributions nearly overlap, whereas their differences become more pronounced as fluence increases. At a given fluence, larger heterogeneity amplitudes produce higher zero-hit probabilities over the parameter range examined. Although all zero-hit probabilities decrease at high fluence, differences between homogeneous and heterogeneous fields remain visible on a relative scale. Thus, increasing neutron fluence reduces the likelihood of receiving no hit, but does not eliminate the influence of spatial boron heterogeneity on hit statistics.
The effect of heterogeneity amplitude is shown more clearly in Figure 3(b), which compares at two neutron fluences for a fixed correlation length of . For and , the zero-hit probability increases from 0.5323 for homogeneous boron to 0.5416 for the heterogeneous distribution, corresponding to a relative increase of approximately 1.75%. At , the probability increases from 0.1508 to 0.1732, equivalent to a relative increase of 14.87%. These results show that the influence of heterogeneity on zero-hit probability becomes more pronounced at higher fluence within the range studied, even though the absolute probabilities are smaller.
Variations in boron concentration broaden the distribution of hit intensities across field realizations. Realizations with lower hit intensities make a relatively larger contribution to the ensemble zero-hit probability because of the exponential dependence in the Poisson model. This behavior is consistent with Jensen’s inequality, , for the same mean boron concentration, neutron fluence, and transport response. However, this inequality does not by itself guarantee that increases monotonically with every heterogeneity parameter. The trends observed in Figure 3(b) are specific to the random-field model and parameter range considered here.
In addition to heterogeneity amplitude, the boron correlation length has a substantial effect on the zero-hit probability, as shown in Figure 3(c). The correlation length is expressed as , where is a reference length scale adopted from an earlier energy-deposition response study. This value is not derived from the hit probability in Figure 1 and is used only to normalize the horizontal axis, rather than to define a universal physical threshold.
For and , the ratio increases from 1.0295 at to 1.1487 at , reaching 1.4049 at . These values correspond to relative increases in zero-hit probability of 2.95%, 14.87%, and 40.49%, respectively. For , the 95% bootstrap confidence interval for the relative increase is 38.20–42.81%. At the lower fluence of , the effect of correlation length remains visible, but the relative enhancement is considerably smaller. These findings show that heterogeneity amplitude alone is insufficient to characterize hit statistics, because the spatial extent of regions with different boron concentrations also affects the distribution of effective hit intensities within the nucleus. The role of correlation length can be understood in terms of spatial averaging through the hit response. When is relatively small, the effective hit intensity receives contributions from many regions with different boron concentrations, so local concentration fluctuations tend to average out. At larger correlation lengths, boron-rich and boron-poor regions extend over greater distances, and neighboring concentration variations become more strongly correlated. This increases fluctuations in hit intensity across field realizations and, through the exponential dependence of the Poisson zero-hit probability, raises the ensemble-averaged . This mechanism is consistent with the dose variance results in Figure 2, where longer correlation lengths also increase the heterogeneity contribution and shift the crossover fluence to lower values.
The numerical consistency of the zero-hit calculations is examined by comparing the ensemble average of the conditional Poisson probabilities with direct Poisson hit-count sampling on the same boron fields. Using 2000 field realizations and 16 repetitions per field, the maximum absolute difference between the two estimates is approximately 0.0057. Confidence intervals are obtained by bootstrap resampling at the field realization level. The agreement supports the implementation of the stochastic model, but does not constitute an independent validation of Geant4 particle transport. The hit response is taken from Geant4 calculations extending to . Sensitivity to the response cutoff and spatial grid resolution is examined separately in Appendix § 8.3. These tests support the numerical stability of the results for the configurations considered, although the effects of domain size and transport parameters remain limitations of the analysis.
The results show that the nuclear zero-hit probability depends not only on neutron fluence and mean boron concentration, but also on the amplitude and correlation length of boron heterogeneity. Boron fields with the same mean concentration can produce different zero-hit probabilities because their distributions of capture intensity at the micrometer scale are not identical. This result complements the dose variance analysis by quantifying the probability that alpha–lithium reaction products deposit no energy within the nucleus. However, the calculated does not represent a cell survival probability, because the model does not include all BNCT dose components or cellular biological responses. Predicting cell survival requires additional biological modeling, as considered in previous microdosimetric studies [25, 31, 38].
Conclusions
This study investigated how stochastic neutron capture and spatial boron heterogeneity affect nuclear dose statistics in boron neutron capture therapy (BNCT). By combining Geant4 energy-deposition responses, a Poisson capture process, and spatially correlated lognormal boron fields, the dose variance was separated into contributions from stochastic capture fluctuations and boron heterogeneity. These contributions increase linearly and quadratically with neutron fluence, respectively, leading to a correlation-length-dependent crossover between the dominant sources of dose fluctuations. For and , the crossover occurs at approximately . Increasing the correlation length from 1.5 to 6 reduces the crossover fluence from to . When normalized by the square of the mean dose, the stochastic contribution decreases with fluence, whereas boron heterogeneity produces a limiting relative variance of 0.160 for the baseline configuration.
Boron heterogeneity also increases the probability that a nucleus receives no hit from alpha–lithium reaction products, even when the mean boron concentration remains unchanged. For , , and , the zero-hit probability is 40.49% higher than that for the homogeneous boron reference, with a 95% bootstrap confidence interval of 38.20–42.81% for the relative increase. These findings show that both the amplitude and correlation length of boron heterogeneity affect not only the magnitude of dose fluctuations, but also the probability of no energy deposition within the nucleus. Thus, mean boron concentration alone is insufficient to characterize dose statistics at the micrometer scale.
This study provides a quantitative framework for understanding how spatial boron heterogeneity affects nuclear dose fluctuations beyond conventional mean-dose-based dosimetry. The agreement between analytical predictions and MC sampling, supported by numerical sensitivity tests, demonstrates the consistency of the proposed approach. By explicitly accounting for stochastic neutron capture and spatial boron correlations, the framework reveals how these mechanisms influence dose variance and zero-hit probability at the cellular scale. Although the model focuses on energy deposition from alpha–lithium reaction products, the findings provide a useful basis for assessing the effects of boron microdistribution and the limitations of homogeneous-boron assumptions in BNCT dosimetry.
Author Contributions
R. Fardela: Writing, Formal analysis, Validation, Review & editing. R. Muslim: Conceptualization, Methodology, Writing, Software, Formal analysis, Validation, Visualisation, Review & editing. T. A. I. Kusuma: Writing, Formal analysis, Validation, Review & editing. All authors read and reviewed the paper.
Declaration of Interests
The contributors declare that they have no apparent competing business or personal connections that might have appeared to have influenced the reported work.
Acknowledgments
R. Muslim was supported by the YST Program of the Asia Pacific Center for Theoretical Physics (APCTP), funded by the Science and Technology Promotion Fund and Lottery Fund of the Korean Government, and by the Management Talent Program of the National Research and Innovation Agency of Indonesia (BRIN).
Appendix
Boron dose variance decomposition
For the lognormal boron field defined in Eq. (1), let and . Using the joint moment properties of Gaussian random fields, we obtain
Here, denotes the covariance of boron concentrations at two positions. Since , the normalized boron concentration correlation function for is
Defining the boron correlation length through , the Gaussian correlation parameter is determined by
In the limit , we have and . For , the boron field becomes homogeneous, so the normalized correlation function is no longer defined and the correlation length does not affect the results.
To derive the dose statistics, the total deposited energy in Eq. (6) is treated as a sum of marks associated with a Poisson process. Assuming that individual capture histories are conditionally independent for a given boron field, the conditional cumulant-generating function is
Differentiating with respect to at gives the first two conditional cumulants,
The first cumulant is the conditional mean deposited energy, while the second is its conditional variance. Thus, the variance of the total deposited energy depends on the second raw moment , rather than the single-event energy variance . This distinction arises because the Poisson process accounts for fluctuations in both the number of capture events and the energy deposited by each event.
Converting the deposited energy into absorbed dose through and applying the law of total variance yields
To make the dependence on the capture factor explicit, we define three coefficients,
Substituting into the total variance relation gives
These coefficients are independent of neutron fluence as long as the boron field statistics and transport responses remain unchanged. The relative dose variance and the dominance ratio are therefore
Under the thermal-neutron approximation, . The condition then defines the crossover fluence,
provided that . For a homogeneous boron distribution, and , so no finite crossover fluence exists. These relations also show that the high-fluence limit of the relative dose variance is determined by , which represents variations in the conditional mean dose across boron field realizations.
For the MC consistency checks, let denote the absorbed dose for boron field realization and stochastic repetition . Let be the mean dose over repetitions for field , and let be the corresponding sample variance. The two variance contributions are estimated as
where is the sample variance of the field-averaged doses across the field realizations. The second term is subtracted because the mean over a finite number of repetitions still contains stochastic capture fluctuations. For independent field realizations and unbiased sample variance estimates, this procedure separates the two contributions in expectation.
Zero-hit probability estimation
For a fixed boron field, each neutron capture at position produces a hit in the nucleus with probability . By the thinning property of Poisson processes, the conditional number of hits also follows a Poisson distribution with mean . Its probability-generating function is
where is defined in Eq. (18). Setting and averaging over the ensemble of boron fields gives
For thermal neutrons, the mean hit intensity can be expressed as , where
Since , the zero-hit probability for a homogeneous boron distribution with the same mean concentration is . It follows that
This inequality follows from the convexity of the exponential function and Jensen’s inequality. It guarantees that the ensemble zero-hit probability for a heterogeneous boron distribution is no smaller than that of the homogeneous reference with the same mean hit intensity. However, it does not generally guarantee a monotonic dependence on every heterogeneity parameter. The effects of and depend on the statistical properties of the boron field considered.
The zero-hit probability is evaluated using independent boron field realizations with hit-intensity coefficients . To reduce sampling uncertainty across fields, we use a control variate estimator based on the analytically known mean intensity [20],
The correction term is based on a linear approximation to the fluctuations in . It has zero expectation because . The estimator therefore remains unbiased for the ensemble zero-hit probability , while reducing the variation associated with fluctuations in the hit intensity.
As a separate consistency check of Poisson sampling, hit counts are generated according to for repetitions . Using the same control variate, the MC estimator is defined as
Because , the conditional expectation of the MC estimator, for a given set of boron fields, equals the corresponding conditional Poisson estimator. Both approaches therefore have the same ensemble expectation, although their sampling uncertainties differ. This comparison tests the consistency of Poisson hit-count sampling and does not constitute an independent Geant4 transport simulation.
The 95% confidence intervals are estimated by bootstrap resampling [9] at the field-realization level. All Poisson repetitions belonging to the same boron field are retained as a group, preserving the dependence between outcomes generated from that field. The main zero-hit calculations use boron field realizations and Poisson repetitions per field, with 1000 bootstrap resamples used to estimate statistical uncertainty.
Numerical validation
The reliability of the numerical calculations is assessed by comparing MC results with analytical predictions and examining the sensitivity to the Geant4 response cutoff and spatial grid resolution, as shown in Figure 4. In Figure 4(a), the MC-to-theory variance ratios range from 0.972 to 1.028 for the stochastic contribution, 0.980 to 1.101 for the heterogeneity contribution, and 0.995 to 1.034 for the total variance. These comparisons use 2000 boron field realizations with eight stochastic repetitions per field. The maximum deviation in the total variance is approximately 3.4%, while the heterogeneity component shows greater statistical variation. These results support the consistency of the variance decomposition at the level of the response moments used in the model.
Figure 4. Numerical validation and sensitivity analysis of the stochastic boron dose model. (a) Ratios of MC variances to analytical predictions for stochastic capture fluctuations, boron heterogeneity, and total dose variance as functions of neutron fluence. (b) Relative difference in zero-hit probability as a function of the maximum source distance , using as the reference. (c) Sensitivity of the asymptotic relative variance coefficient and the zero-hit probability to spatial grid resolution, normalized by the results obtained with the grid. Calculations use Geant4 responses with electromagnetic transport, , , , and . Panels (b) and (c) use . Error bars in (a) and (c) represent 95% bootstrap confidence intervals.
The effect of truncating the radial response is shown in Figure 4(b). Relative to the calculation using responses up to , the difference in is approximately 14% when the response is truncated at . This difference decreases to less than 1% at and becomes very small when the cutoff is extended to . These results indicate that the contribution from the response tail becomes small for the configuration examined, although they do not establish that the physical hit probability is exactly zero beyond the cutoff.
Figure 4(c) shows the sensitivity to spatial grid resolutions of , , , and within the same computational domain. The coefficient changes only slightly, while the difference in between the baseline grid and the reference grid is on the order of 1% or less, with some variation attributable to field-sampling uncertainty. The sensitivity tests in panels (b) and (c) use 1000 boron field realizations. These results support the use of Geant4 responses extending to and an spatial grid for the main calculations over the parameter range examined.
References
- [1]Agostinelli, S., et al., 2003. Geant4—a simulation toolkit. Nuclear Instruments and Methods in Physics Research Section A 506, 250–303. doi:10.1016/S0168-9002(03)01368-8.DOI
- [2]Allison, J., et al., 2016. Recent developments in Geant4. Nuclear Instruments and Methods in Physics Research Section A 835, 186–225. doi:10.1016/j.nima.2016.06.125.DOI
- [3]Barth, R.F., Vicente, M.G.H., Harling, O.K., Kiger, W.S., Riley, K.J., Binns, P.J., Wagner, F.M., Suzuki, M., Aihara, T., Kato, I., Kawabata, S., 2012. Current status of boron neutron capture therapy of high grade gliomas and recurrent head and neck cancer. Radiation Oncology 7, 146. doi:10.1186/1748-717X-7-146.
- [4]Barth, R.F., Wu, G., Vicente, M.d.G.H., Grecula, J.C., Gupta, N., 2024a. Boron neutron capture therapy of cancer: Where do we stand now? Cancer Communications 44, 889–892. doi:10.1002/cac2.12581.DOI
- [5]Barth, R.F., et al., 2005. Current status and future prospects. Clinical Cancer Research 11, 3987–4002. doi:10.1158/1078-0432.CCR-05-0035.DOI
- [6]Barth, R.F., et al., 2024b. Evaluation of sodium borocaptate (BSH) and boronophenylalanine (BPA) as boron delivery agents for neutron capture therapy (NCT) of cancer: An update and a guide for the future clinical evaluation of new boron delivery agents for NCT. Cancer Communications 44, 893–909. doi:10.1002/cac2.12582.
- [7]Charef, K., Ahlaf, F.Z., Khoulaki, Y., Benchekroun, D., Harrass, H., Fathi, I., Makhkhas, Y., 2025. Analysis of alpha and lithium-7 particle energy deposition in BNCT using Geant4 simulation. Atom Indonesia 51, 229–241. URL: https://doi.org/10.55981/aij.2025.1501, doi:10.55981/aij.2025.1501.
- [8]Dymova, M.A., et al., 2020. Boron neutron capture therapy: Current status and future perspectives. Cancer Communications 40, 406–421. doi:10.1002/cac2.12089.DOI
- [9]Efron, B., 1979. Bootstrap methods: Another look at the jackknife. The Annals of Statistics 7, 1–26. doi:10.1214/aos/1176344552.DOI
- [10]Galuzzi, L., Parisi, G., Pascali, V., Niklas, M., Bortot, D., Protti, N., Altieri, S., 2025. Fluorescent neutron track detectors for Boron-10 microdistribution measurement in BNCT: A feasibility study. Materials 18, 621. URL: https://doi.org/10.3390/ma18030621, doi:10.3390/ma18030621.
- [11]Geant4 Collaboration, 2025. Geant4 Physics Reference Manual. Geant4 Collaboration. URL: https://geant4.web.cern.ch/documentation/dev/prm_html/PhysicsReferenceManual/index.html. release 11.4, revision 5 December 2025; accessed 1 November 2026.
- [12]Han, Y., et al., 2023. Microdosimetric analysis for boron neutron capture therapy via Monte Carlo track structure simulation with modified lithium cross-sections. Radiation Physics and Chemistry 209, 110956. doi:10.1016/j.radphyschem.2023.110956.DOI
- [13]Hirose, K., et al., 2021. Boron neutron capture therapy using cyclotron-based epithermal neutron source and borofalan (10B) for recurrent or locally advanced head and neck cancer (JHN002): An open-label phase II trial. Radiotherapy and Oncology 155, 182–187. doi:10.1016/j.radonc.2020.11.001.DOI
- [14]Kobayashi, Y., et al., 2025. Comparison of dose distribution with and without reflecting heterogeneous boron distribution using 18F-BPA positron emission tomography in boron neutron capture therapy. Applied Radiation and Isotopes 219, 111720. doi:10.1016/j.apradiso.2025.111720.DOI
- [15]Last, G., Penrose, M., 2017. Lectures on the Poisson Process, volume 7 of Institute of Mathematical Statistics Textbooks. Cambridge University Press. doi:10.1017/978131604477.DOI
- [16]Lin, Z., Lin, Z., Zhang, Z., Wu, G., Zhu, Y., Liu, Y., Dai, Y., Dai, Z., 2025. Analysis of the influence of boron distribution on compound biological effectiveness values in boron neutron capture therapy. International Journal of Radiation Oncology, Biology, Physics 122, 502–509. URL: https://doi.org/10.1016/j.ijrobp.2025.02.005, doi:10.1016/j.ijrobp.2025.02.005.
- [17]Marcaccio, B., et al., 2026. Unraveling the role of boron microdistribution in BNCT dosimetry of glioblastoma multiforme: Combined theoretical and experimental insights. Physics in Medicine and Biology 71, 085004. doi:10.1088/1361-6560/ae54fd.DOI
- [18]Møller, J., Syversveen, A.R., Waagepetersen, R.P., 1998. Log Gaussian Cox processes. Scandinavian Journal of Statistics 25, 451–482. doi:10.1111/1467-9469.00115.DOI
- [19]Monti Hughes, A., Hu, N., 2023. Optimizing boron neutron capture therapy (BNCT) to treat cancer: An updated review on the latest developments on boron compounds and strategies. Cancers 15, 4091. doi:10.3390/cancers15164091.DOI
- [20]Owen, A.B., 2013. Monte Carlo Theory, Methods and Examples. URL: https://artowen.com/domains/mc/. online textbook.
- [21]Pistone, D., Bortolussi, S., Fatemi, S., Marcaccio, B., Bagnale, L., Pezzi, C., Paganelli, M., Ramos, R.L., Formicola, E., Sica, R., Buompane, R., Porzio, G., Manti, L., Gianalella, L., Vercesi, V., Postuma, I., 2025. A GATE monte carlo study on ICRP110 phantoms for BNCT dosimetry evaluation. Applied Radiation and Isotopes 220, 111724. doi:10.1016/j.apradiso.2025.111724.DOI
- [22]Prosdocimi, A., Deruytter, A.J., 1963. A precise determination of the thermal neutron absorption cross section of 10B and natural boron by time of flight. Journal of Nuclear Energy. Parts A/B. Reactor Science and Technology 17, 83–87. URL: https://doi.org/10.1016/0368-3230(63)90027–2, doi:10.1016/0368-3230(63)90027–2.
- [23]Santa Cruz, G.A., 2026. Microdosimetry in BNCT. Precision Radiation Oncology doi:10.1002/pro6.70063.
- [24]Santa Cruz, G.A., Zamenhof, R.G., 2004. The microdosimetry of the 10B reaction in boron neutron capture therapy: A new generalized theory. Radiation Research 162, 702–710. doi:10.1667/RR3257.DOI
- [25]Sato, T., Masunaga, S.I., Kumada, H., Hamada, N., 2018. Microdosimetric modeling of biological effectiveness for boron neutron capture therapy considering intra- and intercellular heterogeneity in 10B distribution. Scientific Reports 8, 988. doi:10.1038/s41598-017-18871-0.DOI
- [26]Sato, T., et al., 2021. Implementation of simplified stochastic microdosimetric kinetic models into PHITS for application to radiation treatment planning. International Journal of Radiation Biology 97, 1450–1460. doi:10.1080/09553002.2021.1956003.DOI
- [27]Selg, C., Hey-Hawkins, E., 2026. Boron neutron capture therapy at a crossroads: Translational gap and emerging delivery agents. Chemistry – A European Journal 32, e03533. doi:10.1002/chem.202503533.DOI
- [28]Selg, C., Sauerwein, W.A.G., Hey-Hawkins, E., 2026. Quantification of 10B and 11B in biological samples in translational science and clinical applications for boron neutron capture therapy (BNCT). Medical Research Reviews 46, 1362–1389. URL: https://doi.org/10.1002/med.70060, doi:10.1002/med.70060.
- [29]Selva, A., Bianchi, A., Bellan, L., Fagotti, E., Pisent, A., Conte, V., 2024. Microdosimetry for BNCT: First measurements at different site sizes. Radiation Measurements 178, 107298. URL: https://doi.org/10.1016/j.radmeas.2024.107298, doi:10.1016/j.radmeas.2024.107298.
- [30]Shen, S., et al., 2024. A clinician’s perspective on boron neutron capture therapy: Promising advances, ongoing trials, and future outlook. International Journal of Radiation Biology 100, 1126–1142. doi:10.1080/09553002.2024.2373746.DOI
- [31]Shigehira, T., Watanabe, T., Suzuki, M., Hirata, Y., Ogawa, T., Fujimura, A., Sakurai, Y., Sato, T., 2026. Development of linear interpolation system for SMK model parameters evaluated from cellular-scale simulation (LISMECC) and its application to BNCT dosimetry. Journal of Radiation Research 67, 170–181. doi:10.1093/jrr/rraf075. author name corrected in doi:10.1093/jrr/rraq027.DOI
- [32]Teng, Y.C., Cao, Y., Zhu, X., Huo, L., Hong, J., Lin, X., Lai, Y., Zhou, P., Bian, Y., Chen, J., Zhong, W., Ping, J., Lin, Q., Pan, J., Liu, Y.H., 2026a. PET-guided modeling of mucosal boron heterogeneity improves dose-toxicity prediction in BNCT. International Journal of Radiation Oncology, Biology, Physics doi:10.1016/j.ijrobp.2026.05.056. published online June 11, 2026.DOI
- [33]Teng, Y.C., Chen, J., Zhong, W.B., Lin, Q., Liu, Y.H., 2026b. Hypo-segmentation of 18F-BPA PET-based boron distribution: A pragmatic approach for efficient BNCT treatment planning. EJNMMI Research 16, 18. doi:10.1186/s13550-025-01341-4. first published online December 2025; volume assignment 2026.DOI
- [34]Teng, Y.C., et al., 2023. Correcting for the heterogeneous boron distribution in a tumor for BNCT dose calculation. Scientific Reports 13, 15741. doi:10.1038/s41598-023-42284-x.DOI
- [35]Wilson, E.B., 1927. Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association 22, 209–212. doi:10.1080/01621459.1927.10502953.DOI
- [36]Wu, Y., Shu, D., Geng, C., Postuma, I., Tang, X., Liu, Y.H., 2025. Optimization of subcellular boron distribution measurement using UV-C imprint and neutron autoradiography in boron neutron capture therapy. Radiation Measurements 181, 107351. URL: https://doi.org/10.1016/j.radmeas.2024.107351, doi:10.1016/j.radmeas.2024.107351.
- [37]Xu, H., Liu, J., Li, R., Lin, J., Gui, L., Wang, Y., Jin, Z., Xia, W., Liu, Y., Cheng, S., Yuan, Z., 2024. Novel promising boron agents for boron neutron capture therapy: Current status and outlook on the future. Coordination Chemistry Reviews 511, 215795. doi:10.1016/j.ccr.2024.215795.DOI
- [38]Yamazaki, R., Hu, N., Takata, T., Nojiri, M., Zhao, L., Tanaka, H., 2026. Development of a microdosimetry-based method to derive cell survival rates for evaluating the biological effects of BNCT. Biomedical Physics and Engineering Express 12, 025014. URL: https://doi.org/10.1088/2057-1976/ae44a2, doi:10.1088/2057-1976/ae44a2.
- [39]Yang, R., Shi, M., Zhang, Y., Ma, H., Chen, X., Xu, X., 2026. Achievements, challenges, and perspectives in the design of boron delivery agents for neutron capture therapy. Coordination Chemistry Reviews 553, 217555. doi:10.1016/j.ccr.2025.217555.DOI
- [40]Zheng, D., Han, G., Lemus, O.D.M., Podgorsak, A., Webster, M., Li, F., Zhou, Y., Jung, H., Yoon, J., 2026. Boron neutron capture therapy: A technology-driven renaissance. Cancers 18, 498. doi:10.3390/cancers18030498.DOI
- [41]Zhong, W.B., Chen, J., Teng, Y.C., Liu, Y.H., 2023. Introduction to the monte carlo dose engine COMPASS for BNCT. Scientific Reports 13, 11965. doi:10.1038/s41598-023-38648-y.DOI
- [42]Zhou, L., Chen, Y., Peng, M., Yin, X., Liang, H., Mo, J., Zhang, W., Liu, Z., 2025a. Clinical application of boron neutron capture therapy for cancer treatment: A systematic review. Precision Radiation Oncology 9, 304–313. doi:10.1002/pro6.70031.DOI
- [43]Zhou, R., You, Z., Liu, L., 2025b. Boron neutron capture therapy in the context of tumor heterogeneity: Progress, challenges, and future perspectives. Frontiers in Oncology 15, 1601013. doi:10.3389/fonc.2025.1601013.DOI



