PULSAR: Personalized Ultra-Fractionated Radiotherapy

• PULSAR is a radiotherapy paradigm that delivers high-dose pulses separated by weeks, enabling adaptive modifications based on tumor response and imaging data.
• It integrates stereotactic precision with adaptive planning, allowing treatment schedules to be tailored through re-evaluation of tumor anatomy, dosimetry, and immune dynamics.
• Emerging multiomics, digital-twin, and delta-radiomic techniques in PULSAR research offer actionable insights while highlighting the need for standardized protocols and optimal pulse spacing.

Personalized Ultra-Fractionated Stereotactic Adaptive Radiotherapy (PULSAR) is a radiotherapy paradigm in which a patient is given a large dose or pulse of radiation a couple of weeks apart rather than daily small doses, and tumor response is monitored to determine when the subsequent pulse should be given; in a complementary formulation, it is the adaptation of stereotactic ablative radiotherapy toward personalized cancer management (Rouf et al., 2024, Peng et al., 2024). Across the current literature, its defining elements are stereotactic dose delivery, temporally separated pulses, interval imaging, and the possibility of modifying later treatment on the basis of evolving anatomy, dosimetry, or inferred biology rather than adhering to a fixed daily schedule (Yu et al., 9 Sep 2025, Zhang et al., 2024).

1. Conceptual position within stereotactic and adaptive radiotherapy

PULSAR is best understood as an extension of the technical substrate established by stereotactic body radiation therapy (SBRT) and adaptive radiotherapy (ART). SBRT is characterized by few fractions, high dose per fraction, steep dose gradients, and strong reliance on immobilization, image guidance, motion management, and conformal delivery. In the SBRT overview literature, the canonical pattern is 1–5 treatments with 8–30 Gy per fraction, short overall treatment time, and a workflow centered on geometric precision; ART is described there as observing changes during the first few treatments and modifying the plan accordingly, historically with substantial emphasis on setup error and position change (Zong et al., 2024). PULSAR preserves the stereotactic requirement for high geometric accuracy, but shifts the adaptive locus from early-course geometric correction toward inter-pulse personalization under intentionally prolonged spacing. This suggests that PULSAR is not merely “hypofractionation,” but a reorganization of the temporal structure of stereotactic treatment around reassessment opportunities.

Methodological precursors also came from robust adaptive radiotherapy rather than from PULSAR-labelled clinical series. In a stochastic minimax framework on a one-dimensional phantom, robust initial planning combined with selective between-fraction adaptation improved organ-at-risk protection, and the best-performing strategy was not simply margin inflation or global conservativeness adjustment, but patient-specific updating of the underlying uncertainty model (Böck et al., 2016). That logic is structurally aligned with PULSAR: not every patient or fraction needs full replanning, but adaptation should be triggered when delivered dose quality or observed change makes the original plan inadequate.

A separate translational precursor came from mathematical oncology rather than stereotactic clinical workflow. An in silico prostate model that did not mention PULSAR explicitly nonetheless tested larger fractions, longer intervals, and nonuniform weekly dosing; it found that larger fractions with longer spacing reduced final tumor volume by more than half in PC3 but only five percent in DU145, indicating that pulse-like fractionation benefit can be strongly tumor-type dependent (Alvarez et al., 2021). That result is central to later PULSAR thinking because it argues against treating interval extension as universally advantageous.

2. Radiobiological and immunologic rationale

The PULSAR literature repeatedly treats timing as a biological variable rather than a scheduling convenience. One discrete-time mechanistic model couples tumor volume, accumulated radiosensitivity, and two T-cell compartments—intratumoral resident T cells and newly infiltrating T cells—through the updates

$T_{n+1} = S_n\, T_n\, e^{\mu - Z_n},$

$$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$

$$R_{n+1} = \min\bigl(\tau R_n + (1 - S_n),\, 1\bigr),$$

with anti-PD-L1-gated immune killing and separate radiation sensitivities for resident and newly infiltrating T cells (Xing et al., 2023). In that formulation, short intervals preserve accumulated radiosensitivity and can improve RT-alone control, whereas longer intervals permit recruitment, infiltration, and conversion of more radiosensitive newly infiltrating T cells into more radio-resistant intratumoral effectors before the next pulse. The fitted parameters $\phi_1 = 0.05205$ and $\phi_2 = 0.964$ make this asymmetry explicit (Xing et al., 2023).

A related ODE-based PULSAR-effect model introduced a timing-dependent weight function,

$F(t)=\tanh(t),$

to modulate radiation-induced recruitment of newly infiltrating T cells and delayed activation of intratumoral T cells, with the delay constrained to $5 \le \tau \le 9$ days and fitted at $\tau=5$ (Rouf et al., 2024). In that framework, pulse spacings of 9 days or more may induce a PULSAR effect, whereas 1-day spacing leaves the delayed immune-enhancing term too small to generate the same synergy (Rouf et al., 2024). A transformer-based murine analysis reached a similar qualitative conclusion by a different route: in the Lewis Lung Carcinoma model, radiation-only contribution peaked around day 4, whereas the combined radiation–anti-PD-L1 interaction peaked around day 10, consistent with a delayed rather than instantaneous synergy (Peng et al., 2024).

Systemic extension of this rationale remains provisional. A 2026 interaction-picture framework treated the abscopal effect as a continuous, stochastic phenomenon rather than a binary event, separating intrinsic growth from treatment-induced local and systemic perturbations in bilateral tumor models (Peng et al., 7 Mar 2026). The same work explicitly called the abscopal effect controversial and did not claim that PULSAR had already been shown to enhance it clinically (Peng et al., 7 Mar 2026). The current state of evidence therefore supports a biologically plausible timing-sensitive radioimmunologic mechanism, but not a settled universal rule for optimal pulse spacing across diseases.

3. Clinical implementations and disease-specific workflows

The most explicit clinical PULSAR implementation in the provided literature is a UT Southwestern brain-metastasis program using Gamma Knife Icon. In that workflow, 39 patients with 69 brain metastases treated between November 1, 2021 and May 1, 2023 received an initial course of three fractions or pulses of 5–6 Gy with a two-day interval between fractions, then a second treatment cycle approximately three weeks later guided by a second MRI; related reports from the same cohort describe the intervening interval as 2–4 weeks and emphasize that treatment adjustments could be made on the basis of tumor volume changes or vasogenic edema (Peng et al., 21 Jun 2025, Yu et al., 9 Sep 2025). The lesion-level endpoint used repeatedly in this body of work was short-term volumetric response, typically whether follow-up volume had decreased by more than 20% relative to baseline (Peng et al., 21 Jun 2025, Zhang et al., 2024).

This brain-metastasis implementation also clarifies what “adaptive” means in current PULSAR practice. The second course may be reduced to a smaller target, maintained, intensified, or occasionally omitted if the lesion is no longer visible; the key point is that the temporal gap is intentionally used to reassess the lesion rather than to continue a fixed plan (Zhang et al., 2024). Current practice in these studies is criticized for relying heavily on gross volume change and clinician visual judgment, which motivates the development of richer predictive and spatially localized biomarkers (Peng et al., 21 Jun 2025).

A different, anatomy-driven realization of PULSAR-like logic appears in proton prostate SBRT. Digital-twin studies on two-fraction regimens derived from 2STAR and 2SMART used one-week spacing between fractions, 26 Gy to the CTV, and boost prescriptions up to 32 Gy to dominant intraprostatic lesions, with same-day CBCT or corrected CBCT used for fraction-specific plan selection or reoptimization (Chang et al., 2024, Chang et al., 17 Jun 2025). These prostate studies are best read as operational analogues rather than full biologically guided PULSAR: they implement ultra-hypofractionation, between-fraction reassessment, and patient-specific modification, but the adaptation driver is principally anatomy and dosimetry rather than interval tumor biology.

4. Imaging biomarkers, multiomics, and predictive intelligence

A major branch of PULSAR research seeks to make the adaptive window quantitatively actionable. In brain metastases, a multiomics classification study compared six single-scenario models with an ensemble feature selection model integrating pretreatment radiomics and dosiomics, intra-treatment radiomics, and delta-radiomics; the ensemble model achieved an AUC of 0.979, accuracy of 0.917, sensitivity of 0.907, specificity of 0.920, precision of 0.786, and $F_1 = 0.821$ (Zhang et al., 2024). The central message was that intra-treatment and delta features carry more response information than baseline-only features, which is precisely what a PULSAR workflow would predict if the interval MRI is biologically informative.

A regression formulation pushed the same idea further by predicting continuous relative gross tumor volume change rather than a binary responder label. In that study, a multi-omics support vector regression model integrating radiomics, dosiomics, and delta features achieved $R^2 = 0.743$ and $$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$0, with delta-radiomic features forming the majority of selected predictors in the best-performing scenario (Yu et al., 9 Sep 2025). The distinction between classification and regression is consequential in PULSAR, because estimating the magnitude of change is more directly relevant to dose adjustment, replanning, or altered pulse timing than a simple thresholded response class.

A separate line of work addresses spatial localization. In the same Gamma Knife brain-metastasis cohort, a CNN plus CAM framework trained on first-MRI data outperformed radiomics and gradient-based models on a temporally separated test set, with the pixel-wise CAM model reaching sensitivity $$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$1, specificity $$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$2, accuracy $$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$3, AUC $$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$4, precision $$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$5, and $$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$6 (Peng et al., 21 Jun 2025). The key methodological claim is not only higher classification performance but the generation of lesion-specific saliency maps that vary across tumors rather than collapsing to a fixed core-or-rim heuristic. The paper explicitly states that hotspots in non-responders may indicate radioresistant subregions, but also explicitly treats that as a hypothesis rather than proof (Peng et al., 21 Jun 2025).

Generative modeling extends this ambition from response scoring to response simulation. A diffusion-model study in brain metastases trained Denoising Diffusion Implicit Models to map pre-treatment MRI to post-treatment MRI and associated drift maps through

$$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$7

with a forward process

$$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$8

and $$S_n = e^{-(\alpha d_n + \beta d_n^2)} \; e^{-\gamma R_n (\alpha d_n + \beta d_n^2)},$$9 diffusion steps (Peng et al., 20 Jun 2025). The authors report qualitative success in synthesizing post-treatment appearance and lesion-specific drift patterns, but they also state that no quantitative evaluation had yet been conducted because of limited sample size (Peng et al., 20 Jun 2025). This places the work in the category of promising adaptive-intelligence infrastructure rather than validated clinical decision support.

Adjacent evidence from outside named PULSAR studies strengthens the serial-imaging premise. In MR-guided lung SBRT, only 47 of 107 radiomic features were stable under both temporal and spatial perturbation tests, and delta-radiomic skewness was associated with locoregional failure by ANCOVA ($$R_{n+1} = \min\bigl(\tau R_n + (1 - S_n),\, 1\bigr),$$0); tumor diameter and volume were not associated with local recurrence-free survival (Zha et al., 2024). This suggests that early texture evolution can carry information not captured by gross size alone, a proposition that is directly relevant to PULSAR even though the lung study itself was not a PULSAR trial.

5. Digital twins, uncertainty management, and adaptive planning infrastructure

If imaging biomarkers supply the “why” of adaptation, digital-twin and robust-planning studies supply much of the “how.” A CBCT-guided digital-twin framework for adaptive proton SBRT in prostate cancer used patient-specific uncertainty prediction and fraction-specific plan selection from a personalized plan bank. Averaged across 10 patients, the framework preserved CTV $$R_{n+1} = \min\bigl(\tau R_n + (1 - S_n),\, 1\bigr),$$1 at 98.8% versus 99.0% in clinical plans, reduced hotspots from 106.0% to 105.1%, reduced bladder neck $$R_{n+1} = \min\bigl(\tau R_n + (1 - S_n),\, 1\bigr),$$2 from 29.6% to 14.0%, reduced bladder $$R_{n+1} = \min\bigl(\tau R_n + (1 - S_n),\, 1\bigr),$$3 Gy from 12.0 cc to 9.5 cc, and improved plan score by +2.0 to +15.5 points (Chang et al., 2024). The same study showed that 50% of optimal second-fraction plans came from the original planning-CT-derived candidate set and 50% from updated CBCT-derived plans, implying that adaptive libraries need not be regenerated de novo at every fraction (Chang et al., 2024).

A related digital-twin proton study restricted online reoptimization to 10 minutes and compared high-similarity and low-similarity synthetic CT priors. Daily CBCT evaluation showed that clinical plans often violated OAR constraints and that DIL $$R_{n+1} = \min\bigl(\tau R_n + (1 - S_n),\, 1\bigr),$$4 occurred in two patients, indicating simultaneous integrated focal boost failure; DT-H plans achieved better or comparable DIL/CTV coverage and lower OAR doses, and DT-H-REopt-A scores ranged from 154.3 to 165.9 within the time limit (Chang et al., 17 Jun 2025). The strongest practical lesson was that high-quality anatomy matching matters most when online adaptation time is severely constrained: low-similarity priors could sometimes catch up, but only with more optimization time (Chang et al., 17 Jun 2025).

Foundational adaptive-planning methodology points in the same direction. In a robust adaptive IMRT simulation, patient-specific updating of uncertainty distributions outperformed simple conservativeness adjustment or margin adaptation, and weekly-like reassessment was more effective than replanning before every fraction because overly frequent adaptation could overreact to noise (Böck et al., 2016). A different operational bottleneck—segmentation—has been addressed by patient-specific fine-tuning of deep models across fractions. In prostate CT, fine-tuning a baseline CNN on previous scans from the same patient significantly improved contour quality across prostate, seminal vesicles, bladder, and rectum compared with a fixed baseline model, supporting the broader adaptive principle that segmentation models, not only treatment plans, can become personalized over time (Elmahdy et al., 2020).

6. Limitations, controversies, and future directions

The current PULSAR evidence base is technically rich but clinically immature. Most studies are retrospective, single-institution, and small: the principal brain-metastasis cohort contains 39 patients and 69 lesions, the multiomics regression analysis is lesion-based and therefore subject to within-patient dependence concerns, and several generative or digital-twin studies are proof-of-concept demonstrations rather than outcome trials (Peng et al., 21 Jun 2025, Yu et al., 9 Sep 2025, Zhang et al., 2024). Endpoints are often short-term volumetric surrogates rather than local control, progression-free survival, radionecrosis, or overall survival, and external validation is generally absent (Peng et al., 21 Jun 2025, Yu et al., 9 Sep 2025).

Several recurrent misconceptions are directly addressed by the literature. First, PULSAR is not equivalent to any hypofractionated or staged stereotactic regimen; some closely related studies explicitly do not mention PULSAR and should be interpreted as conceptual precursors rather than implementations (Alvarez et al., 2021, Zha et al., 2024). Second, longer spacing is not automatically better: one prostate model found marked benefit in PC3 and only minimal benefit in DU145, and radioimmunologic models predict an optimal intermediate window rather than monotonic improvement with delay (Alvarez et al., 2021, Xing et al., 2023). Third, image-defined hotspots or saliency maps do not prove radioresistant biology; the CAM literature treats such interpretations as testable hypotheses requiring biological anchoring (Peng et al., 21 Jun 2025). Fourth, modality generalization remains largely conceptual. Brain-metastasis CAM authors state that their findings may guide photon and particle therapies, but the actual data are entirely from Gamma Knife radiosurgery, so extension to proton or carbon therapy remains unvalidated (Peng et al., 21 Jun 2025).

Future directions are correspondingly multidisciplinary. Imaging-defined subregions are proposed for alignment with CODEX multiplexed tissue imaging, spectroscopic MRI, and radiopathomic registration in order to test whether CAM-localized niches or delta-radiomic signatures correspond to hypoxia, metabolic stress, immune composition, or other resistant phenotypes (Peng et al., 21 Jun 2025). Predictive modeling studies call for larger independent cohorts, grouped validation designs, broader regression and classification benchmarks, and movement beyond short-term volumetric endpoints (Yu et al., 9 Sep 2025, Zhang et al., 2024). Preclinical PULSAR-effect studies call for direct immune measurements, richer compartmental biology, and schedule optimization in additional tumor models, while the interaction-picture abscopal framework argues for standardized reporting of systemic effect magnitude rather than binary responder language (Peng et al., 2024, Xing et al., 2023, Peng et al., 7 Mar 2026).

Taken together, the literature supports a precise but limited conclusion. PULSAR has emerged as a stereotactic adaptive radiotherapy paradigm in which the interval between high-dose pulses is itself a source of personalization. The strongest current evidence concerns workflow architecture, imaging-derived response prediction, and timing-sensitive radioimmunologic plausibility. What has not yet been established is that any specific biomarker, saliency map, digital twin, or pulse-spacing rule is universally valid, biologically definitive, or prospectively outcome-improving across institutions, disease sites, and treatment modalities.