Feasibility of convection‐enhanced delivery of unlabeled 225Ac for glioblastoma: Transport, dosimetry, and clonogenic survival modeling Authors: Maram El Sabri, Leyla Moghaddasi, Puthenparampil Wilson, Frank Saran, Eva Bezak Published: 2026 ABSTRACT AbstractBackgroundGlioblastoma (GB) is an aggressive primary brain tumor with local recurrence rates exceeding 90% and a five‐year survival rate of approximately 5%. Standard treatments such as surgical resection, external beam radiotherapy (EBRT), and chemotherapy often leave residual tumor cells that are resistant to therapy. Local targeted alpha therapy (TAT) with actinium‐225 (225Ac) offers high‐linear energy transfer (LET) cytotoxicity but is limited by short diffusion range and elevated intratumoral pressure. Convection‐enhanced delivery (CED), which applies mild hydraulic pressure to drive bulk flow, may improve radionuclide penetration.PurposeThis study presents a computational model simulating the distribution of unlabeled 225Ac delivered via CED. The full 2 2 5Ac decay chain (parent + daughters) is included in the dosimetric calculations. By characterizing transport dynamics, full decay‐chain dose deposition, and resultant clonogenic survival, the model explores both the therapeutic potential and limitations of CED‐based alpha therapy.MethodsA Python‐based CED model was developed incorporating diffusion (diffusion coefficient = 1.0 × 10−5 cm2/s), bulk flow (velocity in surgical cavity = 7.0 × 10−5 cm/s; velocity in brain tissue = 5.0 × 10−5 cm/s), and radioactive decay in a radially symmetric brain geometry. Time‐dependent activity profiles were imported into the Monte Carlo simulation platform TOol for PArticle Simulation (TOPAS) to calculate voxel‐level dose distributions. Simulated dose per decay was converted to clinically relevant administered activity (MBq) and integrated over 50 days, adjusted for a relative biological effectiveness (RBE) of 5 (sensitivity range: 3, 7, and 10), and applied via the linear‐quadratic model with tumor repopulation parameters (tumor cell doubling time T p = 14–50 days; repopulation onset T k = 0–50 days) to a probabilistic GB cell map.ResultsCED achieved therapeutic doses (≥ 2 Gy (RBE)/ MBq) to ∼1.43 cm from injection sites. A five‐injection plan delivering a total of 10 MBq reduced the global survival fraction (SF) from 100% (no alpha therapy) to ∼3.4%, with further reductions to ∼2.0% and ∼1.4% at 20 MBq and 30 MBq, respectively. Residual clonogenic tumor cells were mainly located > 2 cm from injection sites. Sensitivity analysis identified injection site placement and total activity as the dominant efficacy factors, whereas varying the RBE had minimal effect. Early tumor repopulation (T k ≤ 14 days) substantially reduced control in fast‐doubling tumors.ConclusionsThis study supports the feasibility of CED of 225Ac for GB, highlighting the need for optimized catheter placement, dosing, and scheduling to address its limited therapeutic radius and tumor repopulation. While results are promising, model assumptions require refinement, and experimental validation is essential to guide clinical translation.Keywords: alpha therapy, computational modeling, convection‐enhanced delivery, glioblastoma, Monte‐Carlo simulation, radionuclide therapy INTRODUCTION Glioblastoma (GB) is a highly malignant and aggressive brain tumor characterized by local recurrence rates exceeding 90% and a five‐year survival rate of approximately 5%. 1 , 2 Current standard of care consists of maximal safe surgical resection followed by external beam radiotherapy (EBRT) with concurrent and adjuvant temozolomide (TMZ) chemotherapy. 3 However, the invasive nature of GB almost invariably results in microscopic residual disease within the normal brain parenchyma beyond the primary surgical resection margin, even after radiologically confirmed complete macroscopic resection. These tumor cells become resistant to standard therapeutic approaches and form the primary substrate leading to local recurrence and subsequent death. Very little progress has been made in expanding the currently available effective therapeutic options for the majority of affected patients and continues to pose a significant clinical challenge highlighting the need for innovative treatments. Given the proven efficacy of radiotherapy to measurably delay local recurrences and improve overall survival targeted alpha therapy (TAT) has emerged as a promising therapeutic strategy due to the distinctive properties of alpha‐emitting radionuclides, which deliver high linear energy transfer (LET) radiation capable of inducing potent, localized cell damage while sparing surrounding healthy tissues. 4 Early phase clinical studies have demonstrated that intracavity administration of TAT can enhance local tumor control. 5 , 6 However, the efficacy of this approach is limited by elevated intratumoral pressure and the inherently slow diffusion of therapeutic agents within brain tissue, significantly constraining their penetration depth and therapeutic reach. 7 , 8 To overcome these limitations, convection‐enhanced delivery (CED) has been proposed as an advanced method to deliver therapeutic agents directly into either macroscopic residual disease or the adjacent surgical cavity (SC) brain's interstitial space. 9 By employing pressure‐driven infusion through stereotactically placed catheters, CED effectively bypasses the blood‐brain barrier, facilitating deeper and more uniform distribution of the therapeutic agents like alpha‐emitting compounds compared to passive diffusion alone. This study presents a computational model simulating the distribution of Actinium‐225 (2 2 5Ac), an alpha‐emitting radionuclide, delivered via CED for the postoperative treatment of primary GBs. By characterizing transport dynamics, dose deposition, and resultant clonogenic survival, the model explores both the therapeutic potential and limitations of CED‐based alpha therapy. These insights offer a theoretical foundation to support future experimental validation and inform clinical translation of TAT in routine clinical management of GBs. METHOD Development of the CED model for 225Ac local distribution Transport mechanisms In this work, “unlabeled 2 2 5Ac” refers to 2 2 5Ac administered as a simple chemical form (2 2 5Ac chloride) without conjugation to a molecular targeting ligand. The dosimetry therefore reflects the physical transport and decay of 2 2 5Ac and its radioactive daughters, rather than receptor‐mediated uptake or biological targeting. 2 2 5Ac decays through a series of seven daughter radionuclides, the parent 2 2 5Ac has a physical half‐life of 9.92 days, whereas all daughters have substantially shorter half‐lives (Figure 1). 10 , 11 Accordingly, dosimetry should include emissions from the complete 225Ac decay chain without assuming local secular equilibrium. The complete decay chain was modeled using the appropriate radioactive‐decay physics lists, so that the absorbed dose includes contributions from 2 2 5Ac and all its daughters rather than from the parent nuclide alone. Decay chain of 225Ac illustrating all daughter radionuclides included in the dosimetric calculations. 11 To predict this distribution, a CED simulation tailored to unlabeled 225Ac was developed. In the extracellular space of the central nervous system (CNS), the movement of therapeutic agents occurs through two primary mechanisms, diffusion and bulk flow. Diffusion relies on the concentration gradient (∇C) of the therapeutic agent, as described by Fick's First Law 12 : where, J represents the diffusive flux, D is the tissue diffusivity, and ∇C denotes the spatial concentration gradient. The value of D is highly dependent on the molecular weight and physical properties of the therapeutic molecule. However, for macromolecular therapeutic agents diffusion is relatively slow due to their size, and effective delivery requires steep concentration gradients near the source to achieve meaningful penetration into tissue. 13 Bulk flow, on the other hand, involves the extracellular movement of fluid driven by pressure gradients. This process can be modeled by treating the brain tissue as a porous medium, comprising a fluid phase (interconnected interstitial spaces) and a solid phase (cells, vasculature, and extracellular matrix). The relationship between fluid velocity (v) and pressure gradient (∇p) is described by Darcy's Law 14 : In this equation, K represents the hydraulic conductivity, indicating the tissue's ability to transmit fluid. Unlike diffusion, bulk flow is less influenced by the molecular weight of the therapeutic agent, making it a more effective in delivering radionuclides such as 225Ac to the CNS. In CED, a cannula is inserted into the interstitial spaces of the CNS, with its tip positioned at the target site. Controlled infusions generate pressure gradients that drive bulk fluid flow, enabling therapeutic agents to distribute over larger volumes within the extracellular space compared to diffusion alone. This technique achieves penetration depths spanning multiple centimeters, bypassing the limitations of diffusion. 8 During CED, drug transport is influenced by the interplay between convection (bulk flow), molecular diffusion, and clearance. The relative contribution of each mechanism determines the extent of drug distribution and the resulting therapeutic efficacy. In the present implementation, transport was modeled using convection and diffusion with physical decay of the parent 2 2 5Ac, biological clearance was not included (Section 4). To simulate these processes and evaluate therapeutic coverage, a multistep modeling workflow was implemented. Figure 2 illustrates the simulation pipeline from CED model set‐up to spatial dose calculation and estimation of cell survival. Modeling pipeline of CED of unlabeled 225Ac and calculation of therapeutic effect. Governing transport equation The computation of the spatial and temporal distribution of unlabeled 225Ac within the CNS was based on a mathematical framework adapted from the foundational work on drug distribution in CED by Morrison and Dedrick. 15 This framework describes the time‐dependent changes in the concentration of the therapeutic agent, C, by incorporating three primary mechanisms discussed, expressed as: where, ∂C/∂t​ is the rate of change of concentration over time, D is the diffusion coefficient, ∂2C∂x2 represents the curvature of the concentration profile that governs diffusive spreading, v is the bulk flow velocity, ∂C∂x represents the spatial concentration gradient and, λ is decay constant of 225Ac. This equation integrates the effects of diffusion, advection, and 225Ac decay, providing a comprehensive model for predicting the spatiotemporal distribution of 225Ac in brain tissue. Geometry and computational model setup To model 225Ac transport in brain tissue, a radially symmetric polar geometry was implemented in Python 3.13.5. The geometry consists of a spherical post‐SC (radius 1.50 cm) surrounded by brain parenchyma extending to 4.50 cm (Figure 3). The SC radius was chosen as a representative postoperative cavity size (set with input from the clinical co‐author), and the 4.50 cm outer radius was selected to extend beyond reported microscopic GB infiltration distances (up to ∼4 cm) to avoid boundary effects. 16 Radial symmetry reduced the 2D transport equations (diffusion, advection, decay) to a 1D radial form, improving computational efficiency. All radial profiles and penetration depths reported in this work are defined with respect to distance from the injection tip (i.e., radially averaged about the catheter), rather than distance from the cavity center. Although the catheter is off‐center relative to the cavity, under the homogeneous/isotropic assumptions the transport solution can be represented as a catheter‐centered radial profile; verification showed that peri‐cavitary placements produced nearly identical radial concentration profiles (maximum mean radial difference ∼0.1 mm over 15 h). Diagram representing the modeled geometry. The post SC (CSF/serosanguinous mix) and its adjacent edematous rim were treated as a single fluid‐dominated region with low hydraulic resistance for diffusion, while region‐specific advection velocities were assigned. The domain was discretized into uniform radial segments of width Δr = 50 µm and 360 angular points (1° resolution) for isotropic angular integration. Because each ring at radius r corresponds to a volume proportional to r2Δr, the initial 2 2 5Ac count was normalized in each bin by its shell volume. This ensured the agent's concentration (atoms per unit volume) remained uniform across all radii, avoiding artificial inflation of activity in the outer bins. As highlighted by Yang et al., 17 drug distribution during CED can be compromised by leakage into the low‐pressure resection cavity, thereby reducing effective parenchymal coverage. To reflect clinically relevant peri‐cavitary catheter placement, a baseline injection distance of 10 mm beyond the SC boundary was selected for the primary model configuration. Given the cavity radius of 1.50 cm, this corresponds to a radial injection position of r = 2.50 cm from the center. Five discrete injection sites were arranged symmetrically around the cavity at this baseline distance, with equal angular spacing of 72°, consistent with clinical CED practice and supported by a sensitivity analysis examining the influence of injection number on survival probability (SP) and survival fraction (SF) (Section S1 and Figure S1 and S2). To examine sensitivity to catheter placement, additional injection distances of 0 mm (at the cavity wall), 1 mm, and 5 mm beyond the SC boundary were evaluated as discrete sensitivity cases, while maintaining the same five‐injections geometry and total administered activity. Injection distance was therefore treated as a discrete parameter rather than a continuous variable, and no optimization of injection distance was performed. Due to computational constraints, only a single bolus injection geometry was explicitly simulated using Monte Carlo methods, corresponding to the baseline configuration at r = 2.50 cm. This modeling choice was supported by verification of the CED transport solution, which demonstrated that the spatial distributions of 2 2 5Ac within the SC and surrounding brain parenchyma were highly similar across peri‐cavitary injection locations, with a maximum mean radial difference of ∼0.1 mm over the simulated transport period (see Results, Section 5.1). The resulting dose distribution was therefore translated to the multi‐catheter configuration and to alternative injection distances under the assumptions of radial symmetry, identical infusion conditions, and homogeneous tissue properties. We acknowledge that in vivo variability in tissue resistance, local permeability, and injection pressure could lead to non‐uniform delivery, and this simplification represents a limitation of the current model. Parameter selection Diffusion was modeled using a coefficient D of 1.0 × 10−5 cm2/s. The nominal value was selected as a plausible upper‐range starting point for an initial feasibility study, consistent with the broad range used in published CNS CED modeling, 18 , 19 , 20 , 21 and enabling a computationally tractable baseline evaluation within our multi‐stage pipeline. Because unlabeled 2 2 5Ac is administered as AcCl3, D is treated here as an effective diffusivity for the transported actinium species in the simplified model. Uncertainty in D (and in the assumed average velocity v) was evaluated via a sensitivity analysis (Results Section 5.4). As no direct experimental measurements of the diffusion coefficient of 225Ac in aqueous or biological media are currently available, this value was selected based on reported diffusion coefficients for species of comparable size and transport regime. Published CED and convection–diffusion models report diffusion coefficients in brain tissue on the order of ∼10−5−10−7cm2/s for macromolecules and contrast agents, whereas small ions typically exhibit higher diffusivities, in the range of approximately ∼2×10−5cm2/s. 18 , 19 , 20 , 21 (see Section S3, and Table S1). Heavy monoatomic species such as radon‐220 (220Rn) exhibit diffusion coefficients on the order of 10−5cm2/s, providing a physically relevant reference point for heavy alpha‐emitting species in water. 22 While radon is not a chemical analogue of actinium, ionic species such as 225Ac3 + are expected to diffuse more slowly due to hydration and electrostatic interactions, but within the same order of magnitude. The same diffusion coefficient was applied to both the SC (Dcavity) and surrounding brain tissue (Dtissue). This approach is supported by Haar et al., 23 which show that the diffusion coefficient is largely independent of the extracellular environment for medium‐sized molecules unless there are significant changes in extracellular fraction or tortuosity. In both tissue environments, diffusion is governed primarily by the molecular properties of the agent rather than regional variations in tissue structure. The diffusion of unlabeled 225Ac was continuously modeled over 15‐h, comprising a 6‐h CED infusion period and a 9‐h post‐infusion phase. Advection velocity, v, was set to vcavity at 7.0 × 10−5 cm/s in the SC and vtissue at 5.0 × 10−5 cm/s for surrounding brain tissue, consistent with observed values for CED modeled in the brain. 23 , 24 These velocities were assigned based on tissue region (surgical cavity versus surrounding brain tissue) and applied during the 6‐h infusion period. The chosen values align with the expected range of interstitial fluid velocities in normal brain tissue, as reported by Zhan et al., 24 who modeled drug transport under CED conditions and observed velocities of 2 ∼ 8 × 10−5 cm/s, depending on the infusion rate and site location. Similarly, Noh et al. 25 used bromophenol blue solutions in brain surrogates to experimentally validate the interplay of advection and diffusion. Their experiments, conducted at injection rates ranging from 0.25 to 4 µL/min, measured velocities between 1 × 10−5 to 1 × 10−4 cm/s, further supporting the selected range. The velocity difference between normal tissue and SC regions was derived from Haar et al., 23 who observed a 40% increase in interstitial velocity within edematous tissues, reflecting altered fluid dynamics due to increased porosity and reduced tissue resistance. To simplify the computational model, average velocities under CED were used, although in reality interstitial velocities vary based on pressure gradients created by the injection rate and the radial distance from the infusion point. The radioactive decay of 225Ac was modeled using its physical half‐life (T1/ 2 = 9.92 d), corresponding to a parent decay constant λ=ln(2)9.92days. This decay constant was applied exclusively to the parent 225Ac activity in the transport equation, which models only the diffusion and advection of the parent radionuclide. Absorbed‐dose calculations, however, include energy deposition from 225Ac and all radioactive progeny, implemented through the Geant4 radioactive‐decay physics in TOPAS. In this framework, the parent and all daughter radionuclides are explicitly generated and tracked until they come to rest, ensuring that the absorbed dose reflects the complete 2 2 5Ac decay chain, consistent with the decay‐chain description in Section 2.1.1. A summary of model parameters is provided in Table 1. TABLE 1Baseline parameter values used in simulation.ParameterBaseline valueReference D (cm2/s)1.0 × 10−5 14 vbrain (cm/s)5.0 × 10−5 15, 16 vcavity (cm/s)7.0 × 10−5 15, 16 rcavity (cm)1.5– rbrain (cm)4.5–No. of peri‐tumoral injections5 13 Radial placement (mm)10 13 Atoms per bolus106 –Advection duration (hours)6–Diffusion duration (hours)15– Baseline parameter values used in simulation. Dose calculation: Model implementation into tool for particle simulation (TOPAS) Dose distributions were calculated using the TOol for PArticle Simulation (TOPAS, version 3.9.0), a publicly available Monte Carlo (MC) platform built on Geant4, designed for high precision dosimetric modeling. 26 The dynamic diffusion model developed in Python (as described in the previous section) was used to simulate the spread of 225Ac following a single bolus injection containing 106 atoms over a 15‐h period. The 225Ac distribution was recorded at 1‐h intervals, generating hourly concentration profiles. Each hourly profile was imported into TOPAS as the source distribution for the corresponding 1‐h interval, allowing the dose delivered during that interval to be calculated. Each run was performed in a water phantom with a spatial resolution of 50 µm, with a relative statistical uncertainty of ∼ 0.1% under Poisson statistics. The scoring geometry for dose‐to‐medium calculations was configured into cubic voxels measuring 20 × 20 × 20 µm3 reflecting the average size of GB cells, 27 , 28 thus enabling voxel‐level dosimetry representative of cellular dose deposition. To ensure accurate particle transport, appropriate Geant4 physics lists were explicitly defined in the TOPAS macro file. The selected Physics Lists were: −g4em‐standard_opt4, g4deca g4em‐standard_opt4, g4deca −g4radioactivedecay g4radioactivedecay −g4h‐phy_QGSP_BIC_AllHP g4h‐phy_QGSP_BIC_AllHP −g4em‐livermore g4em‐livermore These lists facilitate precise modeling of electromagnetic (EM) physics, including low‐energy EM interactions, as well as radioactive decay processes and Hadronic physics. Accordingly, the scored absorbed dose includes contributions from 2 2 5Ac and all daughter radionuclides (full decay chain). Additionally, simulated alpha‐particle energies and corresponding particle ranges generated from 225Ac decays were validated against National Institute of Standards and Technology (NIST) reference data, 29 serving as a quality assurance measure. All Monte Carlo simulations were executed on the Phoenix HPC cluster (Adelaide University) using 72 CPU cores, over a total run‐time of 13.5 days. To make the 20 × 20 × 20 µm3 voxel resolution computationally feasible, only a single 5° angular wedge (0°–5°) of the spherical domain was explicitly simulated, with a slightly extended rectangular scoring volume to ensure edge‑to‑edge dose coverage. The dose distribution from this wedge was azimuthally replicated over 360° (72 rotations) to reconstruct the full spherical dose distribution, as illustrated in Section S2 and Figure S3. Under the assumption of radial and azimuthal symmetry with homogeneous tissue properties, this approach also enabled replication of the single simulated injection dose distribution across multiple catheter sites. While diffusion into the resection cavity was modeled, the dose distribution for the cavity medium was not recomputed, dosimetry and outcome analyses were restricted to surviving cells in parenchyma external to the SC. This approach significantly reduced computational demands while maintaining dosimetric accuracy. Translation of simulated dose to clinically relevant administered activity Spatial and temporal adjustment TOPAS outputs absorbed dose per 106 decays in units of Gray (Gy). To convert these values into clinically relevant doses, a dose‐scaling procedure was applied. The total number of decays within each hourly interval was calculated using the radioactive decay law: where A0 is the initial activity (1 MBq) and λ is the decay constant of 225Ac and Ninterval represents the total number of radioactive decays occurring within the time interval t start​ and tend​. As each Monte Carlo run simulated a fixed number of decays, the absorbed‐dose output for each interval was scaled using an activity scaling factor defined as: where Nsim=106 is the number of decays simulated in TOPAS for that interval. The cumulative dose over all intervals up to time 𝑡𝑚 for the simulated 5° wedge was then obtained using a discrete numerical integration (Riemann sum): where Dsim(ti) is the absorbed‐dose distribution scored by TOPAS during the i‐th interval. To reconstruct the full 3D spherical dose distribution, the wedge dose was azimuthally replicated over 360° (72 rotations). In addition, a volume correction was applied to account for the difference between the simulated 2D slice and the full spherical volume. For a voxel width w = 0.002 cm and a sphere of radius r = 2.5 cm (from injection point), the full‐to‐slice volume ratio is: The final cumulative dose was therefore calculated as: RBE adjustment α‐particles are substantially more biologically effective per unit absorbed dose than low‐LET radiation (photons/β‐particles). Published in‐vitro and in‐vivo data place the therapeutic RBE of α‐particles between 3 and 7. 30 Consistent with TAT clinical practice, 31 , 32 , 33 a central RBE value of 5 was adopted when converting the absorbed dose‐to‐medium (Gy/MBq) to an RBE‐weighted dose (Gy(RBE)/MBq). A sensitivity analysis was performed using four discrete RBE values (3, 5, 7, and 10) to capture potential variability in α‐particle radiobiology. SF estimates were calculated for RBE values of 3, 5, 7, and 10. Integration over time The activity and dose were calculated for up to 50 days (∼5 half‐lives of 225Ac), with the integration period aligned to that used by Tronchin et al., 34 who demonstrated that organ doses plateau near 50 days. Monte Carlo transport, however, was modeled over 15 h of physical decay and diffusion. For direct comparison, an analytical decay‐count conversion was introduced to preserve the total number of decay events between the two time windows: Considering passive diffusion of large therapeutic agents is spatially limited (∼1–2 mm), 17 , 35 , 36 the spatial distribution obtained at t = 15 h was assumed to remain fixed for the remainder of the integration period. For times t > 15 h up to 50 days, voxel doses were scaled in magnitude only to account for radioactive decay, with no additional spatial redistribution assumed. Accordingly, each voxel dose was scaled as: SP and SF calculation The GB model developed by Moghaddasi, Bezak, and Harriss‐Phillips 37 was adapted in this study. This model takes into account cellular oxygen status, radiation sensitivity, and the microscopic extension probability (MEP) models for GB. In this study the irregular MEP model was used, which uses a probability‐density function (PDF) based on published clinical data regarding the recurrence pattern of GB. The PDF assigns each cell with a probability of being a tumor clonogenic cell as a function of radial distance from the gross tumor volume, GTV, based on published clinical recurrence patterns. Because the central region is a post‐SC rather than solid tumor, the model was modified so that the density of clonogenic cells within the cavity reflects the sparse residual‐cell population expected after resection. Specifically, the volumetric density was set to 107 cells/cm3, 38 with a gradient that decreases toward the cavity center and increases toward the cavity wall (Figure 4). Modified irregular‐MEP distribution. The dosimetry matrices converted to clinically relevant administered activities and integrated over 50 days were combined with the irregular‐GB model to obtain individual SP of each cell and cell‐based SF. SP was modeled using the linear quadratic (LQ) model, 39 incorporating the Lea–Catcheside dose‐protraction factor (G) to account for the protracted dose delivery and sub‐lethal DNA damage repair typically observed in targeted radionuclide therapy (TRT). 40 The G factor was calculated using: where, the effective decay constant of 225Ac was assumed equal to the physical decay constant due to the assumption of no biological clearance in the current simulation. The double strand break (DSB) repair rate (μ) was set to 0.047 ± 0.006 h− 1, based on GB cell lines LN18 and U87 irradiated with 6 MeV alpha particles. 41 Tumor repopulation dynamics were incorporated by introducing an offset (lag) time Tk, representing the onset of accelerated tumor repopulation following treatment. Repopulation dynamics were incorporated by adding a growth penalty term to the SP equation, accounting for accelerated proliferation after a specified repopulation initiation time (T k). 40 The modified survival equation becomes: where Tp is the proliferation doubling time (14, 20, 30, 40 50 days), 42 , 43 T is the overall treatment duration (50 days), and Tk is the offset time for repopulation initiation, before which no repopulation penalty is applied and after which accelerated exponential tumor regrowth is assumed. Based on radiobiological evidence from EBRT, T k was set at 37 ± 8.5 days, with sensitivity checks at 0, 14, 20, and 50 days reflecting realistic tumor repopulation scenarios observed clinically. 44 , 45 , 46 The radiation sensitivity parameters (α, β) distributions were generated by Moghaddasi and Bezak 16 using a Gaussian‐weighted distribution based on Taghian et al.’s 47 in vitro study of 21 malignant glioma cell lines. The global SF for cells beyond the SC was computed using: Uncertainty in SP and SF calculations The voxel‐wise SP and global SF were calculated using an exponential model incorporating the absorbed dose (D), DNA double‐strand break repair kinetics (G), α/β ratio, tumor cell doubling time (T p), and repopulation onset (T k). The uncertainty analysis presented here is a first‐order approximation, correlations and higher‐order interactions between input parameters were not propagated. While TOPAS accounts for recoil energy from 225Ac daughter nuclei in the absorbed dose, no explicit spatial migration of daughter radionuclides was modeled, and any associated dose redistribution was not included in the uncertainty analysis. The α/β ratio was voxel‐wise sampled from literature‐derived empirical distributions; thus, explicit uncertainty bars were omitted. The α/β ratio was voxel‐wise sampled from literature‐derived empirical distributions; thus, explicit uncertainty bars were omitted. A sensitivity analysis accounted for inter‐patient variability in T p (range: 14–50 days), rather than using explicit uncertainties. A sensitivity analysis accounted for inter‐patient variability in T p (range: 14–50 days), rather than using explicit uncertainties. Explicit uncertainties for D, G, and T k are summarized in Table 2 and combined via root‐sum‐square RSS, yielding an overall voxel‐wise SP uncertainty of ± 26.8%. Explicit uncertainties for D, G, and T k are summarized in Table 2 and combined via root‐sum‐square RSS, yielding an overall voxel‐wise SP uncertainty of ± 26.8%. TABLE 2Parameter and their uncertainties.ParameterDescriptionUncertainty/methodReferenceα/β valuesVoxel‐wise sampled from empirical distributionStochastic sampling 38 Absorbed dose (D)Dose matrix (Monte Carlo)±0.1%TOPASDSB repair rate (G)GB cell‐line DNA repair kinetics±13.7% 31 Tumor repopulation onset (T k)GB repopulation onset during EBRT±23% 32, 33 Tumor doubling time (T p)Phenotype/genotype‐dependent; broad range (−42 to ∼3000 days)Sensitivity analysis (14–50 days) 34, 35, 36 Combined SP and SF uncertainty D, G, Tk combined via RSS ±26.8% Calculated Parameter and their uncertainties. VALIDATION AND SENSITIVITY STUDY Monte Carlo dose‐distribution validation Alpha‐particle energy/range check Simulated α‐particle energies and ranges were compared with NIST (Stopping‐Power and Range Tables) reference data. Agreement was within < 5 %, confirming appropriate physics‐list selection and particle tracking (Table 3). TABLE 3Validation of alpha particle ranges in TOPAS against NIST reference data.Energies of α‐particles emitted during 225Ac decay (MeV)TOPAS range (µm)NIST range (µm)Δ%5.8344.147.24.56.3453.253.91.37.0763.664.00.68.3882.384.32.4 Validation of alpha particle ranges in TOPAS against NIST reference data. PHITS cross‐verification The validity of the decay‐based conversion to clinically relevant administered activity (Section 2.1.6) was benchmarked against independent PHITS (Particle and Heavy Ion Transport code System) simulations by Stephen Tronchin (Table 4). The PHITS simulations used initial conditions of 1 MBq of 225Ac uniformly distributed within a 1 cm3 water cube, scored at a voxel resolution of 20 µm, and recorded doses (Gy/MBq) at 15 h, 24 h, and 50 days. The TOPAS simulations were configured identically regarding geometry, voxel size, and time points, employing 107 225Ac decays per simulation run. TOPAS outputs were then converted according to Equation 6 to produce absorbed doses expressed in Gy per MBq, ensuring direct comparability to PHITS results. Mean voxel doses from both methods demonstrated close agreement at all evaluated time points, differing by <10%. This good correspondence validates the implemented conversion approach. Slightly higher doses reported by TOPAS relative to PHITS can be attributed to TOPAS's inclusion of dose contributions from all particle emissions, while the PHITS calculations considered α‐particle contributions exclusively. TABLE 4Mean‐voxel dose comparison (Gy/MBq).TimePHITS mean dose (Gy/MBq)TOPAS mean dose (Gy/MBq)Δ%15 h225.39240.766.8224 h359.91380.255.6550 d5297.15463.13.14 Mean‐voxel dose comparison (Gy/MBq). Sensitivity analyses Seven parameter groups were systematically varied to quantify their impact on the radial 225Ac spread and SF, as detailed in Table 5. TABLE 5Perturbation matrix.ParameterNominal valuePerturbation factorsOutput quantities monitored Diffusion coefficient (D) 1.0 × 10− 5 cm2/s0.25×, 0.5×, 2×, 5×Radial profiles of 225Ac spread Advection velocity (v) a vcavity = 7.0 × 10− 5 cm/s vtissue = 5.0 × 10− 5 cm/s 0.25×, 0.5×, 2×, 5×Radial profiles of 225Ac spreadRBE5 (baseline)3, 7, 10 RBESP maps; SF rangesInjection distance from SC10 mm0, 1, 5, 12 mmSP maps; SF rangesTotal activity10 MBq20, 30 MBqSP maps; SF rangesTumor cell doubling time T p 30 days14, 20, 40, 50SF rangesOffset time for repopulation initiation T k 37 ± 8.5 days0, 14, 20, 50 daysSF ranges a Velocity scaling applied to both vcavity and vtissue. Perturbation matrix. Diffusion coefficient (D) Advection velocity (v) a vcavity = 7.0 × 10− 5 cm/s vtissue = 5.0 × 10− 5 cm/s Velocity scaling applied to both vcavity and vtissue. MODEL ASSUMPTIONS AND LIMITATIONS To keep the present study computationally tractable and to establish a baseline framework for 225Ac transport and dosimetry, we adopt the following simplifying assumptions. Tissue and transport. Brain tissue is treated as a homogeneous porous medium with a spatially constant diffusion coefficient. Bulk flow velocity and pressure gradients are assumed steady‐state during infusion. Diffusion is modeled as isotropic, equal in all directions. These assumptions enable efficient simulation but do not capture heterogeneity (necrosis or regional extracellular matrix differences), time‐varying infusion‐induced changes (pressure fluctuations or deformation), or anisotropic transport along structured pathways such as white‐matter tracts. The model also neglects cannula backflow, which can alter the near‐field distribution in practice. Source representation and injection layout. Injections are approximated as point sources, rather than explicitly modeling catheter geometry (finite diameter, outlet ports, or directional features) that can influence the initial concentration field and promote asymmetry. For multipoint delivery, the five injection sites are placed within a single 2D polar slice (equatorial plane) at equal angular spacing (72°). The resulting 2D dose map is then revolved to construct a 3D spherical distribution, implying azimuthal symmetry and locally identical tissue properties. Clinically, optimal injection coordinates depend on patient‐specific resection geometry, cavity orientation, and proximity to critical structures. Radioactive decay and daughters. 225Ac and its daughter isotopes are included in the dose calculation (via TOPAS). However, transport of daughter isotopes (diffusion/migration) is not modeled, which may underestimate dose spread if daughters exhibit appreciable mobility relative to the parent distribution. These assumptions ensure computational efficiency and provide a foundational framework for modeling 225Ac transport. However, they also highlight the need for refinement and experimental validation. Future iterations of the model could incorporate dynamic boundary conditions, anisotropic tissue properties, and cellular‐scale interactions to enhance realism and predictive accuracy. Addressing these limitations would provide a more robust representation of therapeutic agent transport in complex tissue environments such in GB. RESULTS 225Ac distribution using CED The spatiotemporal distribution profiles of 225Ac over the 15‐h simulated period for a single injection site are illustrated in Figure 5a. During the initial 6 h, both diffusion and advection contributed to radionuclide transport, resulting in concentration profiles with distinct shoulders indicative of bulk flow effects. Additionally, the 225Ac distribution within the cavity and brain parenchyma domain were relatively similar with maximum mean radial difference of only ∼ 0.1 mm. Spatiotemporal 225Ac concentration profiles within the brain parenchyma:(a) Radial concentration distribution from the injection site over the simulated 15‐h period; (b) concentration at the injection center as a function of time; (c) maximum penetration distance of 225Ac front as a function of time (1% threshold). The concentration at the injection site (Figure 5b) significantly decreased during the first 6 h due to advection, with an average reduction rate of approximately 1556 atoms bin−1 hour−1. Following this period, the rate of concentration decrease slowed substantially to about 199 atoms bin−1 hour−1, reflecting passive diffusion and radioactive decay processes. The radial penetration depth (1% concentration threshold) of 225Ac is shown in Figure 5c. The penetration depth reached approximately 2 cm within the initial 6‐h period. After transitioning to passive diffusion, additional penetration was limited, extending by less than 1 cm over the next 9 h. Penetration rates decreased from 0.24 cm/h (advection‐dominated phase) to 0.12 cm/h (diffusion‐only phase), consistent with established diffusion behavior for large molecules. 25 , 35 The corresponding two‐dimensional (2D) spatial distributions of 225Ac for each hourly interval are presented in Figure 6. These maps highlight the progressive spread of the radionuclide from the injection site, visually reinforcing the marked difference in radial spread between advection‐active and diffusion‐only phases. Hourly 2D concentration maps (1–15 h) for a single 2 2 5Ac injection placed 10 mm beyond the SC boundary. SC contoured in white. Monte Carlo dose distribution and clinical translation Cumulative dose distributions from Monte Carlo simulations, scaled to clinically used units (Gy (RBE)/MBq), are shown in Figure 7. Single‐injection dose accumulation over 15 h highlights rapid dose escalation at the injection site, rising from approximately 1.49 × 103 Gy (RBE)/MBq at 1 h to 9.42 × 103 Gy (RBE)/MBq at 15 h. Radial dose distribution sharply declined with distance, limiting clinically significant doses (≥ 2 Gy (RBE)/MBq) to within 1.5 cm from the source, with negligible doses beyond 3.0 cm. Cumulative dose (Gy (RBE)/MBq) distributions at 1, 6, 10, and 15 h following a single injection. Dose maps are displayed using a logarithmic color scale to enhance visualization of low‐dose regions and radial dose gradients at increasing distances from the injection site. Long‐term cumulative doses up to 50 days (∼5 half‐lives of 225Ac, corresponding to 1200 h) approached a plateau of 2.14 × 105 Gy (RBE)/MBq at the injection center, decreasing sharply with radial distance to approximately 52.2 Gy (RBE)/MBq at 1.5 cm and about 0.37 Gy (RBE)/MBq at 3.0 cm (Table 6). TABLE 6Absorbed dose (Gy (RBE)/MBq) at specific radial distances from the injection point over various simulation intervals. Dose values correspond to RBE = 5 and activity = 2 MBq per injection.Hour0 cm0.5 cm1.0 cm1.5 cm2.0 cm2.5 cm3.0 cm4.0 cm11.49E+031.69E+004.00E‐030.00E+000.00E+000.00E+000.00E+000.00E+0065.15E+032.01E+012.19E+002.19E‐011.30E‐020.00E+000.00E+000.00E+00107.18E+033.37E+015.46E+009.58E‐011.20E‐018.00E‐032.00E‐030.00E+00159.42E+034.89E+019.78E+002.30E+004.26E‐014.40E‐021.60E‐021.00E‐0312002.14E+051.11E+032.22E+025.22E+019.67E+001.00E+003.67E‐011.30E‐021200 a 5.02E+001.62E+015.83E+012.26E+021.11E+032.14E+051.11E+035.23E+01 a Dose distribution referenced from the center of the surgical cavity for the five‐injection configuration, illustrating cumulative dose contributions from multiple peri‐cavity injections. Absorbed dose (Gy (RBE)/MBq) at specific radial distances from the injection point over various simulation intervals. Dose values correspond to RBE = 5 and activity = 2 MBq per injection. Dose distribution referenced from the center of the surgical cavity for the five‐injection configuration, illustrating cumulative dose contributions from multiple peri‐cavity injections. To relate physical dose deposition to biological effectiveness, an effective distance from the injection point was defined using the radially averaged SP. Starting from the injection radius (2.5 cm), the effective distance deff was identified as the location where the mean SP first exceeded SP = 0.01, corresponding to 99% predicted clonogenic cell kill. This occurred at reff=3.93cm from the cavity center, equivalent to deff=1.43cm from the injection point, representing the average radial extent of biologically effective cell kill around an injection site. Figure 8 illustrates dose distribution from five peritumoral injections (each delivering 2 MBq, total 10 MBq) placed 10 mm from the SC. This multi‐injection strategy provided enhanced peripheral dose coverage, effectively targeting peripheral tumor cell populations located further from the SC boundary. Spatial distribution of cumulative dose (Gy (RBE)) after 50 days from five peri‐cavity injections (2 MBq each, total 10 MBq; RBE = 5). Dose values are displayed on a logarithmic scale (log₁₀). The cavity boundary (radius = 1.5 cm) is shown as a circular contour, and injection sites are indicated by crosses. Distances are referenced to the SC center. SP and SF calculation The cell‐based SP and global SF results, computed for cumulative doses over 50 days, are depicted in Figures 9 and 10. A total administered activity of 10 MBq distributed evenly across five peritumoral injection sites (each 10 mm from the SC edge) yielded a global SF of approximately 3.4% ± 0.9%, using a DSB repair factor G of 0.047 and T k of 30 days. Increasing the total administered activity to 20 MBq and 30 MBq further decreased the global SF to 2.0% ± 0.5% and 1.4% ± 0.4%, respectively, demonstrating a distinct dose‐response correlation. Figure 9 shows the SF with varying T p and activity. SF beyond the SC versus doubling time (Tp) for three activity levels (10, 20, and 30 MBq). Error bars reflect propagated uncertainties from dose, G, and Tk. Spatial SP maps at 10 MBq, 20 MBq, and 30 MBq, showing higher SP values at the periphery of the tumor domain. The spatial SP distribution (Figure 10) highlights surviving clonogenic tumor cells predominantly at peripheral locations distant from the injection sites, corresponding to regions receiving lower absorbed doses (> 2 cm from injection sites). This pattern emphasizes the restricted radial coverage of alpha‐emitting radionuclides and suggests strategic optimization of injection sites, increased administered activities, or combined therapeutic approaches to enhance efficacy against peripheral tumor populations. Sensitivity analysis outcomes A one‐at‐a‐time sensitivity analysis was performed for D and v by scaling each parameter by factors of 0.25, 0.5, 2, and 5 relatives to the nominal case (Figure 11). Using a 1% of peak concentration threshold at the end of infusion (t = 6 h), penetration depth was highly sensitive to D (1.67 cm at 0.25D; 2.29 cm at 0.5D; 3.19 cm at D; and ≥4.5 cm for D ≥ 2D, where the 1% front reached the 4.5 cm domain boundary), but only weakly sensitive to v (3.16–3.37 cm across 0.25v–5v). Transport‐parameter sensitivity of 2 2 5Ac radial profiles at t = 6 h. Relative atoms (%) versus radius, normalized to the nominal peak (100%). Additionally, the sensitivity analysis revealed that injection distance and total administered activity had the most significant impact on therapeutic efficacy, while changes in RBE produced comparatively modest effects (Figure 12). Specifically, increasing the RBE from 3 to 10 only reduced the SF from 4.9% ± 1.2% to 2.0% ± 0.5%, suggesting that RBE variations within this range do not strongly influence treatment outcomes. In contrast, altering the injection distance from the SC had a pronounced effect: SF rose from 3.4% ± 0.9% at 10 mm (baseline) to 14.5% ± 3.6% when injections were placed directly at the cavity wall (0 mm), indicating reduced control of clonogenic cells. This reflects the steep fall‐off of α‐particle dose with distance, which requires close spatial overlap between the high‐dose region and the highest clonogenic cell density. In the model, clonogenic cells are concentrated in the peri‐cavity rim and decrease with distance from the SC edge (Figure 4). At 0 mm, a larger fraction of activity lies within/adjacent to the cavity, reducing coverage of clonogenic cells further away and increasing SF, shifting injections outward into tissue (10 mm) improves overlap with the peri‐cavity rim. Distances beyond 10 mm further compromised tumor coverage, however, with distances > 10 mm the clonogenic cells within the SC were observed thus for the model distance of 10 mm was used. Increasing the overall activity from 10 MBq to 30 MBq substantially lowered SF from to 1.4% ± 0.4% and enabled effective tumor control even when injections were placed up to 12 mm from the SC. These findings highlight that optimizing injection placement and activity levels is more critical for therapeutic success than accounting for small uncertainties in RBE. A full T p–T k sweep (Figure 13) reveals that SF beyond the SC is also highly sensitive to the timing of tumor repopulation, particularly for fast‐doubling cells. For the shortest doubling time (T p = 14 d), SF drops from ∼30% when repopulation begins immediately (T k = 0 d) to < 1% by T k = 50 d. In contrast, for slow‐growing cells (T p = 50 d), the SF only falls from ∼6% to ∼1% over the same T k window. This demonstrates that early repopulation severely undermines treatment efficacy in rapidly proliferating tumors, whereas longer doubling times buffer against delayed regrowth. Clinically, it suggests that adjusting CED scheduling to precede the expected T k (e.g. ≤ 14 d) may substantially improve tumor control in aggressive GBs. The roles of injection distance, and RBE on SF. Effects of the repopulation dynamics T p (tumor cell doubling time) and T k (repopulation onset) on SF. DISCUSSION The computational model confirms that CED of 225Ac delivers high, spatially confined α‐particle doses. Effective coverage extended to approximately 1.43 cm from each injection tip, highlighting the critical role of both the activity loaded per injection and the three‐dimensional arrangement of cannulas as primary dose‐limiting factors. The model confirms that during CED convection dominates diffusion by nearly an order of magnitude, producing the “shouldered” concentration fronts previously observed experimentally. 17 , 25 This highlights the limitations of passive diffusion alone for radionuclide delivery, such as 225Ac, aligning with experimentally observed diffusion depths of approximately 2 mm in brain tissue. 35 The selected interstitial velocities (5–7 × 10− 5 cm s− 1) fall within the ranges measured in mathematical 24 and physical brain surrogates, 25 supporting physiological plausibility. Monte Carlo cellular‐level dosimetry further emphasizes the importance of catheter placement. Our simulations reveal steep α‐particle dose gradients around peri‐cavitary injection sites, wherein even a 5 mm shift towards the cavity can halve the dose delivered to adjacent brain tissue. This aligns with pre‐clinical findings in the F98 rat brain tumor model, where placement of CED catheters slightly beyond the SC, rather than directly within it, significantly improves the distribution volume achieved by multi‐injection arrays. 17 In clinical practice, post‐resection delivery presents challenges such as cavity collapse, oedema, and disrupted CSF boundaries, which can reduce infusate retention and distribution volume through leakage. 17 , 48 To mitigate these effects, catheter placements should be guided by updated postoperative MRI, ensuring trajectories terminate a few millimeters beyond the SC, avoiding CSF spaces, and maintaining safe distances from cortical surfaces and ventricles. 17 , 48 Aligning infusions with white matter tracts can further enhance distribution in anisotropic brain tissues, as demonstrated by Yang et al. 48 Postoperative cavities are often irregular and non‐spherical, which limits the applicability of the symmetry assumptions in this model, moreover, the predominantly 2D representation used here may under‐represent 3D heterogeneity relevant to catheter planning and clinical translation. Further simulation studies incorporating patient‐specific datasets should refine catheter positioning by explicitly accounting for leakage, backflow, and biological clearance of 225Ac and its daughter isotopes. Moreover, current simulations show a clear activity‐dependent reduction in clonogenic survival, aligning with clinical findings from Krolicki et al. 49 who reported that dose escalation from 10 MBq to 30 MBq per cycle for intracavitary 225Ac‐DOTA‐Substance P improved local retention but also induced neurological adverse events above 10 MBq, including worsening hemiparesis, aphasia, and severe perifocal oedema consistent with radiation‐induced injury. This highlights the narrow therapeutic window for single‐cycle treatments. The residual cologens persist at the tumor periphery, beyond the high‐dose core, irrespective of further activity escalation. These observations emphasize that geometric optimization, precise catheter placement, the addition of extra peri‐margin catheters or combined therapy, may yield a greater therapeutic outcome than simply increasing activity, which is limited by dose‐related neuro‐toxicity. Additionally, only one injection cycle (i.e. one dose) was modeled, incorporating repopulation dynamics via tumor doubling time penalties after T k. However, published clinical studies typically involved two to six administrations spaced 6–8 weeks apart, 49 , 50 , 51 , 52 allowing normal tissue recovery and repeated irradiation of repopulated clonogens. Given the Linear‐Quadratic formalism's dose additivity, repeated cycles could further reduce SFs, assuming consistent catheter placement and minimal biological clearance. Future work will therefore extend simulations to cumulative dosing across multiple cycles and its impact on clonogenic cell kill. The sensitivity analysis further indicates that rapid repopulation of GB cells significantly erodes cytoreduction between cycles, supporting current clinical practice favoring fractionated CED schedules over single, large‐dose infusions. Collectively, these findings suggest an optimal treatment strategy combining moderate per‐cycle activity, image‐guided catheter trajectories, margin‐focused infusion, repeated dosing, and patient‐specific dosimetry to achieve sustained local control with manageable risk. Additionally, sensitivity analyses highlighted injection distance and total administered activity as critical determinants of therapeutic success, whereas variations in RBE within clinically relevant ranges produced only minor effects, indicating robustness of outcomes to moderate RBE uncertainties. Despite these promising results, clinical translation will require extending the framework to capture effects not represented in the present implementation (see Section 4), including patient‐specific anisotropy along white‐matter tracts, post‐resection evolution (oedema, cavity changes, and permeability shifts), and catheter‐related phenomena such as finite catheter geometry and backflow. Future work should incorporate anisotropic transport informed by diffusion‐tensor imaging, time‐dependent boundary conditions, explicit catheter/backflow modeling, and more detailed handling of daughter‐isotope migration, supported by experimental validation and imaging (MRI‐based catheter verification and monitoring of infusion fronts). Additionally, emerging evidence suggests EBRT–TAT combination strategies may improve the therapeutic ratio by increasing tumor dose while limiting normal‐tissue exposure and mitigating hypoxia‐driven radioresistance. 53 CONCLUSION This study demonstrated the feasibility and potential therapeutic outcome of CED of 225Ac for treating GB. The simulation findings highlight that while α‐particle therapy can achieve highly localized dose deposition, its effective therapeutic radius remains limited, necessitating optimization of catheter placement, activity and need for combined therapies. Additionally, the simulations highlight the importance of accounting for tumor repopulation dynamics in scheduling treatments to maximize therapeutic outcomes. Despite promising results, the simplifying assumptions regarding tissue homogeneity and isotropic diffusion necessitate future model refinements. This is a theoretical modeling platform intended as a first‐order approximation to inform and guide future clinical development. Overall, our data provide a foundation for optimizing TAT strategies and guiding future experimental validations and clinical implementation. CONFLICT OF INTEREST STATEMENT The authors declare no conflicts of interest. Supporting information Supporting Information: mp70566‐sup‐0001‐SuppMat.docx