Digital Twins in Oncology: From Multimodal Data Integration to Precision Clinical Practice Authors: Annamalai Vairavan, Rupsa Bhattacharjee, Bagyam Raghavan Published: 2026 ABSTRACT Oncology digital twins are patient-specific computational models that are built by combining electronic health records, multiomics genomic data, and diagnostic imaging to simulate individual tumor biology and predict multiple treatment-related outcomes. Conceptually originated from aerospace engineering, it has matured clinically through convergent advances in radiomics, mechanistic tumor modeling, pharmacokinetic- pharmacodynamic systems, federated machine learning, and, most recently, large language model (LLM)-based clinical interfaces and agentic artificial intelligence (AI). For a practicing radiologist, digital twins offer a transformative role: imaging-derived quantitative features serve as the primary data, positioning diagnostic imaging at the center of these personalized oncology workflows. Key clinical applications especially in oncology span from chemotherapy response prediction, immunotherapy patient selection, personalized radiation planning, tumor progression modeling, and treatment toxicity forecasting. Several of these applications are achievable with current technology without any significant infrastructure investment. Substantial challenges include imaging data standards, absence of prospective validation, algorithmic bias in underrepresented populations, and regulatory uncertainty for continuously self-updating AI. This narrative review provides radiologists with a balanced, comprehensive overview of digital twin architecture, advanced enabling technologies, current clinical evidence, a practical roadmap for implementation, and a candid appraisal of barriers to adoption. Introduction Cancer affects around 19.3 million people each year, continues to rise globally with significant morbidity and mortality, and remains one of the most biologically diverse diseases seen in clinical practice. 1 Despite major advances in genomics-guided targeted therapy and immune checkpoint inhibitors, treatment selection remains largely derived from population-level data—clinical trial outcomes in thousands of patients—while individual patients' tumor biology, immune microenvironment, pharmacogenomic profile, and comorbidity burden may differ fundamentally from the trial cohort. The consequence is substantial variation in clinical response rates, typically 30 to 60% even in biomarker-selected populations, and a pattern of empirical treatment failure, subsequent regimen changes, and cumulative toxicity that erodes both survival and quality of life. 2 The oncology digital twin is a patient-specific, continuously updated computational representation of an individual patient that integrates heterogeneous, real-world data—imaging, genomics, pathology, treatment response, and clinical data over time—into a single simulation framework. The twin can dynamically model an individual's cancer biology, project how that patient will respond to a specific intervention before it is administered, and detect early signals of resistance or progression in real time. For radiologists, the digital twin is directly relevant because diagnostic imaging is the only modality capable of three-dimensional, serial, spatially resolved phenotyping of the tumor and its microenvironment throughout the treatment course. Quantitative and longitudinal imaging analysis—radiomics, functional magnetic resonance imaging (MRI), and metabolic positron emission tomography (PET) parameters—forms the update signal that keeps the twin synchronized with the patient's evolving biological state. Radiologists who understand the architecture and clinical utility of digital twins are uniquely positioned to enhance their contribution to multidisciplinary cancer care. 3 4 5 This review provides a comprehensive, balanced, and practically oriented overview of oncology digital twins for a radiology readership. We describe the technical architecture and key enabling technologies; the specific role of imaging, radiomics, and digital pathology; clinical applications stratified by evidentiary maturity; and the barriers that must be addressed before widespread adoption. Basic Architecture of the Oncology Digital Twin An oncology digital twin is distinguished from a conventional predictive model by three properties ( Table 1 ). First, bidirectional synchronization: the twin is continuously updated with patient data and generates outputs that shape clinical decisions, which in turn generate new patient data. Second, multiscale biological integration: molecular, cellular, tissue, and whole-patient levels are modeled simultaneously. Third, prospective simulation: the twin can project future states under proposed interventions. 3 6 This architecture comprises five functional layers ( Fig. 1 ). Table 1Basic differences between conventional AI models vs. oncology digital twinConventional AIDigital twinRelies on single modality inputs. Focused data streamsData integrationIntegrates multimodal data sourcesComprehensive patient profileOperates on static snapshotsPoint in time analysisTemporal perspectiveTracks longitudinal patient progressionOngoing temporal monitoringUtilizes task-specific trainingFixed objective functionsLearning methodologyEmploys continuously updated modelsAdaptive learning loopsGenerates specific predictionsOutcome forecastingFunctional outputProvides interactive simulationsScenario testingFunctions as a decision aidDiagnostic tool supportClinical applicationActs as a virtual patient modelPersonalized medicine platformAbbreviation: AI, artificial intelligence. Basic differences between conventional AI models vs. oncology digital twin Abbreviation: AI, artificial intelligence. Five-layer architecture of the oncology digital twin. The data ingestion layer standardizes heterogeneous inputs via FHIR/HL7 interfaces. The feature engineering layer applies modality-specific pipelines—NLP for clinical notes, PyRadiomics for imaging, and variant annotation for genomics. The multimodal fusion layer integrates feature vectors using transformer networks or graph neural networks. The simulation engine layer uses mechanistic models (ODE, ABM, PK/PD, linear-quadratic radiation modeling) to simulate tumor dynamics. The clinical interface layer, increasingly mediated by LLMs and agentic AI, generates actionable, uncertainty-quantified clinical recommendations. ABM, agent-based model; EHR, electronic health record; FHIR, Fast Healthcare Interoperability Resources; LLM, large language model; NLP, natural language processing; ODE, ordinary differential equation; PK/PD, pharmacokinetic-pharmacodynamic; WSI, whole-slide image. These five functional layers are the operational machinery through which the three defining properties above are achieved in practice—the two frameworks are complementary rather than separate. Each layer corresponds to a step in the twin's working cycle: data acquisition and harmonization, multimodal fusion into a shared patient representation, mechanistic or data-driven modeling of tumor biology, simulation of a candidate intervention, and feedback of model outputs into the clinical workflow. We can therefore think of the three properties as what a digital twin does, and the five layers as how it does it. Table 2 summarizes this mapping. Table 2Mapping the five functional layers to the three defining propertiesFunctional layerOperational roleDefining property primarily served1. Data acquisition and harmonizationStandardized capture of imaging, genomic, pathology, and clinical data at a shared patient anchorBidirectional synchronization2. Multimodal fusionSpatial, semantic, and latent-space integration into a unified patient representationMultiscale biological integration3. Biological modelingMechanistic and/or data-driven modeling of tumor behaviorMultiscale biological integration4. SimulationForward projection of candidate interventions before administrationProspective simulation5. Clinical feedbackTranslation of outputs into decisions, generating new data that reenters layer 1Bidirectional synchronization Mapping the five functional layers to the three defining properties Methods: Search Strategy and Selection Criteria This narrative review synthesizes the current state of oncology digital twin technology for a radiology audience, with deliberate emphasis on the intersection of imaging, radiomics, and multimodal data fusion rather than on general oncology or genomics literature in isolation. PubMed/MEDLINE, Embase, Scopus, and Google Scholar were searched for English-language articles published between January 2015 and May 2026, using combinations of the terms “digital twin,” “oncology,” “radiomics,” “multimodal data fusion,” “precision oncology,” “artificial intelligence,” “radiology,” “mechanistic tumor model,” and “federated learning.” Reference lists of retrieved articles and recent systematic reviews were hand-searched for additional relevant studies. Priority was given to peer-reviewed original research, systematic reviews, and consensus or methodological guidance documents (e.g., Image Biomarker Standardization Initiative [IBSI], Checklist for Artificial Intelligence in Medical Imaging [CLAIM], TRIPOD + AI [Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence]) directly addressing imaging-derived data streams within a digital twin or closely related predictive-modeling framework. Conference abstracts without peer-reviewed full-text publication and non-English-language articles were excluded. As this is a narrative rather than a systematic review, no formal quality-scoring or PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses)-style selection flow diagram was applied. Advanced Enabling Technologies Five distinct but complementary technologies constitute the technical foundation of the oncology digital twin. Radiomics, mechanistic modeling, and multiomic integration form the twin's core biological components; federated learning and large language model (LLM)-based systems are enabling technologies that scale the data available to train those components and interface the twin's outputs with clinicians. Understanding each, and their integration, is essential for radiologists engaging with twin-based clinical systems. Radiomics and Quantitative Imaging Analysis Radiomics is the extraction of quantitative features from Digital Imaging and Communications in Medicine images. 4 Numerous features can be extracted from a single computed tomography (CT) lesion segmentation, capturing shape characteristics, first-order intensity statistics, and texture features derived from gray-level cooccurrence matrices, run-length matrices, and wavelet transforms. The IBSI provides a validated reference framework for feature reproducibility across scanner types and acquisition protocols—a prerequisite for multi-institutional harmonization. 7 Delta-radiomics is the quantification of radiomic feature changes between serial imaging examinations. This is a particularly powerful twin update signal because it captures dynamic tumor biology rather than a single static data point. Studies in nonsmall cell lung cancer (NSCLC), breast cancer, and HER2-targeted neoadjuvant response have demonstrated that radiomic and delta-radiomic signatures can predict pathological complete response to neoadjuvant chemotherapy with the area under the curve (AUC) values in the range of 0.85 to 0.93 in the primary reports cited here. 8 9 10 These estimates derive from retrospective, largely single- or few-institution cohorts and should be interpreted accordingly. Mechanistic Tumor Growth and Treatment Response Models Unlike purely data-driven machine learning/deep learning (DL) models, mechanistic simulation engines encode biological first principles into the twin, providing physically interpretable predictions with uncertainty bounds even when training data are limited. 11 Reaction-diffusion partial differential equation systems model tumor growth as a spatially heterogeneous process governed by cell proliferation, death, and diffusion through tissue, calibrated with MRI-derived cellularity and vascularity parameters, and were originally pioneered in glioblastoma. 12 The approach has since been extended to breast cancer: a biology-based mechanistic model calibrated to serial MRI in 105 triple-negative breast cancer (TNBC) patients predicted pathological complete response to neoadjuvant chemotherapy with an AUC of 0.82, and patient-specific in silico optimization of the chemotherapy schedule was associated with an absolute improvement in predicted pathological complete response of approximately 21 to 25 percentage points compared with the standard regimen. 13 Agent-based models simulate individual tumor cell behaviors and immune-tumor interactions at cellular resolution, capturing emergent phenomena such as immunological synapse formation and treatment-induced immunogenic cell death that cannot be represented by continuum equations. 14 Multiomic Data Integration and Genomic Profiling The twin's molecular identity is defined by its genomic modules. Next-generation sequencing panel data, whole-exome sequencing, and ribonucleic acid sequencing provide somatic mutation profiles, copy number variations, gene fusions, and transcriptomic pathway activation states that anchor the twin's treatment vulnerability predictions. 15 Tumor mutational burden, microsatellite instability status, and programmed death-ligand 1 expression scores populate the immunotherapy selection module. Germline pharmacogenomic profiling (e.g., DPYD, UGT1A1, CYP2D6, TPMT) provides individual metabolizer status that calibrates the pharmacokinetic/pharmacodynamic toxicity model; several such tests are already in routine clinical use as standalone risk factors and represent among the most immediately implementable components of the twin framework. 16 The two technologies below are not themselves biological models of the tumor. Rather, they scale the data available to train the core components above, and they interface the twin's outputs with clinicians. Federated Learning and Privacy-Preserving Model Training The predictive performance of any digital twin model is fundamentally constrained by the size, diversity, and quality of the data sets on which it is trained. Single-institution oncology data sets, even at major academic centers, are typically insufficient to train models that generalize robustly across different populations, scanner types, genomic sequencing platforms, and treatment protocols. Federated learning addresses this constraint by training models across multiple institutions without centralizing patient data: each participating site trains the model on its local data and shares only model parameter gradients, not patient-level records with the central aggregator. 17 18 Large Language Models and Agentic Artificial Intelligence LLMs trained on biomedical literature, clinical notes, and structured medical data have demonstrated clinical reasoning capabilities that are transforming the human-artificial intelligence (AI) interface within digital twin platforms. 19 20 In the oncology twin context, LLMs fulfill three principal functions. First, they act as natural language extraction interfaces that convert unstructured clinical notes, radiology reports, and pathology texts into structured data fields populating the twin's longitudinal record. 21 Second, LLMs generate synthesized, guideline-referenced clinical summaries from the twin's multimodal predictions, formatted for tumor board presentation in plain clinical language with uncertainty quantification. Third, they serve as natural-language query interfaces. Agentic AI extends these capabilities by equipping LLMs with tools that execute code, query databases, and invoke specialized models, enabling autonomous multistep clinical reasoning chains. 22 A practical example: upon upload of a new CT examination, an agentic system can automatically (1) extract lesion measurements and radiomic features, (2) retrieve the patient's current twin state, (3) calculate RECIST response category, (4) compare delta-radiomics to the predicted trajectory, (5) cross-reference the latest circulating tumor deoxyribonucleic acid (ctDNA) result, and (6) generate a structured clinical summary with ranked treatment recommendations—all before the radiologist views the study, with every step auditable and the radiologist retaining final interpretive authority. Data Modalities, Processing Pipelines, and the Twin Update Cycle Fig. 2 illustrates the principal data streams feeding the oncology digital twin, their processing pipelines, the specific features they contribute, and the twin modules they populate. The update cycle is bidirectional: each clinical event generates new data that recalibrates the twin's model parameters, and the twin's updated predictions inform the next clinical decision, which in turn generates further data. Multimodal data integration pipeline. Seven data streams (left) undergo modality-specific processing (center left) to yield quantitative features (center right) that populate specific twin modules (right, digital twin). The bidirectional arrow indicates that clinical actions triggered by twin outputs generate new data, which recalibrate the model—the defining characteristic of a true digital twin as distinct from a static predictive model. ADC, apparent diffusion coefficient; ctDNA, circulating tumor deoxyribonucleic acid; DCE, dynamic contrast-enhanced; IHC, immunohistochemistry; NGS, next-generation sequencing; NLP, natural language processing; PGx, pharmacogenomics; PRO, patient-reported outcome; SMI, skeletal muscle index; TIL, tumor-infiltrating lymphocyte; VAF, variant allele frequency; WSI, whole-slide image. Multimodal Data Fusion: Methods for Integrating Heterogeneous Data Streams Collecting data is necessary but not sufficient for an oncology digital twin; the deeper challenge is ensuring the data are used correctly. This process mathematically combines imaging, genomic, clinical, and temporal data into a unified representation in which each modality enhances the information value of the others rather than existing as isolated silos. Without principled fusion, a digital twin is merely a data warehouse. Multiple strategies are appropriate to different data types and clinical objectives, each governed by a fundamental requirement: a shared patient anchor that links every data stream unambiguously to the same individual at the same biological moment. Spatial Alignment: Multiscale Coregistration The most direct form of fusion applies when two data modalities occupy physical space and can be mapped onto a common geometric reference frame before their information can be combined. Multiscale coregistration is the technical solution: a computational process that aligns images acquired at different resolutions, time points, and/or with different physical contrast mechanisms into a single spatial coordinate system. The most clinically common and relevant example is PET-CT coregistration, in which metabolic activity maps from fluorodeoxyglucose PET are over laid onto anatomical CT image to assign precise spatial coordinates to show glucose uptake. Multiscale coregistration also applies within imaging modalities, for example, radiology–histopathology image alignment. Serial CT examinations acquired at different time points or phases must be coregistered before delta-radiomics features can be meaningfully extracted; the same applies to motion correction in MRI sequences and to registration of on-treatment cone-beam CT images to the planning CT before adaptive radiotherapy replanning can proceed. The Common Coordinate Framework: A Mathematical Shared Reference A common coordinate framework (CCF) extends the spatial alignment concept from pairwise image registration to a universal reference space in which all data modalities are expressed in comparable mathematical coordinates. In oncology, frameworks are emerging at the organ level for prostate, brain, and breast cancer, enabling systematic spatial comparison of radiomics features, genomic alterations, and pathological findings across patients by mapping each data set onto a standardized organ template rather than patient-specific native coordinates. 23 For the digital twin, a CCF provides three practical benefits. First, it enables population-level learning: radiomic features extracted from all patients and expressed in CCF coordinates can be compared across patients to identify spatially recurrent patterns of treatment resistance or metastatic spread that would be invisible in patient-specific coordinate systems. Second, it provides a natural framework for integrating multiresolution data: cellular-resolution histological data and millimeter-resolution CT data exist at inherently different scales, but both can be expressed as density maps projected onto the CCF, enabling quantitative cross-scale comparison. Third, it supports longitudinal tracking: as a tumor evolves through treatment, its changing radiomic and molecular phenotype can be tracked in CCF space, distinguishing true biological change from apparent change caused by variation in tumor position, orientation, or segmentation boundary. Semantic and Temporal Alignment Not all fusion problems are spatial. Electronic health record data, genomic reports, radiology reports, and pathology reports are structurally heterogeneous: they represent different concepts, use different terminologies, and are collected at different clinical time points. Semantic alignment is the process of mapping these heterogeneous data into a shared conceptual vocabulary so that the twin can reason across them consistently. Temporal alignment is the process of anchoring each data element to a common clinical timeline so that the twin correctly understands which data were available when, and which followed from which. The Systematized Nomenclature of Medicine-Clinical Terms provides a reference ontology for clinical findings and diagnoses; the Human Phenotype Ontology encodes phenotypic abnormalities; the Gene Ontology organizes gene functions; and the Logical Observation Identifiers Names and Codes system standardizes laboratory test names. When clinical notes, laboratory results, and genomic reports are mapped onto these shared, publicly maintained ontologies, disparate data sources become computationally comparable. The Latent Space Bridge: Deep Learning Fusion Spatial coregistration and semantic alignment operate in the native representation spaces of each modality. DL models instead convert imaging, pathology, genomics, and clinical data into a shared “latent space” where biologically similar tumors cluster together despite differences in data type or format—for example, a breast tumor with aggressive MRI features, HER2 amplification, and high-grade pathology may group together because all reflect the same underlying tumor biology. Variational autoencoders, contrastive learning, and transformer-based models are commonly used for this multimodal fusion. 24 For radiologists, this means radiomic imaging features become part of an integrated biological model rather than isolated image biomarkers. Knowledge Graphs and Probability Distribution Comparison Knowledge graphs organize known biological relationships between genes, proteins, drugs, pathways, and tumor behavior using curated databases such as OncoKB, KEGG, and PharmGKB—of which DrugBank is a widely used exemplar, aggregating drug-target-pathway relationships that support this kind of biological cross-referencing. 25 This helps the digital twin interpret even rare mutations using established biological pathways: when imaging, genomics, and ctDNA findings agree, prediction confidence increases; when results are discordant—for example, stable imaging but rising ctDNA—the system should flag uncertainty and prompt closer clinical review rather than a premature treatment change. Properly designed fusion combines not point estimates but full probability distributions from each modality, though robust clinical implementations of this specific approach remain limited in the current literature. Digital Shadow versus Digital Twin, and the Shared Anchor Requirement A digital shadow passively collects patient data and reflects what has already happened. A true digital twin continuously updates itself using new clinical, imaging, and molecular data, while also influencing future treatment decisions. 6 Most current oncology systems remain digital shadows rather than fully adaptive digital twins. A rare example of a system approaching genuine bidirectional twin behavior is a mechanistic model calibrated to serial breast MRI in the ARTEMIS neoadjuvant chemotherapy cohort, in which the calibrated twin was subsequently used prospectively, in silico, to test alternative chemotherapy schedules for each patient and identify a schedule predicted to improve response—illustrating the “prospective simulation” property in a patient-specific setting rather than a purely retrospective one. 13 All fusion methods also require a reliable shared patient anchor like an unique ID, accurate timestamps, and correct clinical context. Without this, data from different sources may be incorrectly combined, reducing real-world reliability; models that appear valid in retrospective evaluation may still fail in prospective clinical deployment for this reason. Simulation Engine and Its Operating Cycle The simulation engine is what makes a digital twin different from a conventional AI model: instead of giving a one-time prediction, the twin continuously learns and updates itself using new patient data. 26 It works in four repeating steps: Initialize—load the patient's current imaging, genomics, labs, and treatment history; Simulate—predict tumor response under the current therapy; Compare—match predictions against new follow-up data; Recalibrate—update the model if the actual response differs from the prediction. In practical terms, what the engine is simulating is the tumor's future trajectory under a given treatment. A simple way to picture this: the decision tree (DT) model represents the tumor as a population of cells whose numbers grow according to a proliferation rate and shrink according to a death rate, both estimated from that patient's own imaging, where data on how densely packed tumor cells appear to be on diffusion MRI, or how much blood supply a region shows on contrast-enhanced imaging are taken. Given a proposed chemotherapy dose and schedule, the model plays this growth-versus-death process forward in time, correlated with thousands and thousands of data points, which were fed into the system previously, and the DT model comes with an output prediction before the patient has actually received it. This predicted output is again correlated with the actual outcome, which updates this patient-specific DT ( Fig. 3 ). This figure shows the simple cycle that makes a digital twin engine simulating the tumor's future growth trajectory under a given treatment. Cyclical Alteration of the Digital Twin by Real-Time Data The power of this cyclical architecture lies in the fact that the twin becomes progressively more accurate for a specific patient the longer it runs, because each new data event adds evidence that constrains the biological parameter space and reduces the range of plausible disease trajectories ( Fig. 4 ). At enrolment, the twin is calibrated with population-level priors reflecting average tumor biology for that cancer type and molecular subtype. As patient-specific data accumulate, these priors are continuously overwritten by posterior estimates that reflect the individual's actual biological behavior—the conceptual basis on which twin-guided, theranostically informed personalized treatment is intended to build. 26 Role of imaging across the digital twin lifecycle: Imaging → Radiomics → Fusion → Simulation → Response Prediction → Recalibration, forms a continuous feedback loop updating the digital twin, which differentiates from the routine machine learning (ML) model which is static. Clinical Applications: From Evidence to the Bedside ( Tables 3 and 4 ) Table 3Current single-output predictive models in onco-imaging with high potential for conversion into patient-specific digital twinsCurrent predictive component modelsCurrent outputPotential areas for digital twin conversionResponse to chemotherapyUsually binary (responder/nonresponder)Can be extended into mechanistic models by linking imaging biomarkers with tumor perfusion and proliferation dynamics. Binary predictions can evolve into continuous response trajectories, enabling adaptive treatment planning and dose optimizationDelta-radiomics (serial texture change)Feature vectors from early on-treatment CT/MRI capturing dynamic biologyDirectly quantifies tumor kinetics for predicting pCR in cancers such as NSCLC and breast cancer. Imaging features can be incorporated into reaction–diffusion PDE models for tumor growth calibration and integrated into mechanistic digital twins using IBSI-standardized radiomic featuresCT body composition (skeletal muscle index at L3)Automated quantification (CNN-based U-Net)Readily deployable using existing AI tools. Predicts chemotherapy toxicity and treatment discontinuation risk across multiple tumor types. Can be integrated into digital twin toxicity modules to support personalized dose modification and early nutritional interventionPerfusion/blood flow parameters (DCE-MRI, arterial spin labeling)Quantitative maps of blood flow, permeability, and transit time with high spatial resolutionProvides physiologically meaningful inputs for mechanistic PDE-based vascular models. Imaging-derived perfusion parameters can serve as boundary conditions for modeling treatment-induced vascular changes and predicting radiotherapy or immunotherapy responsectDNA kinetics and variant allele frequencySerial measurements reflecting tumor burden (doubling time) and emergence of resistanceCan be integrated with longitudinal imaging trajectories using mathematical kinetic models. Discordance between ctDNA dynamics and delta-radiomic changes can identify biological uncertainty, prompting closer surveillance. Enables multimodal digital twins by combining genomic evolution with imaging phenotypes for adaptive therapy decisionsAbbreviations: AI, artificial intelligence; CNN, convolutional neural network; CT, computed tomography; ctDNA, circulating tumor deoxyribonucleic acid; DCE, dynamic contrast-enhanced; IBSI, Image Biomarker Standardization Initiative; MRI, magnetic resonance imaging; NSCLC, nonsmall cell lung cancer; pCR, pathological complete response; PDE, partial differential equation. Current single-output predictive models in onco-imaging with high potential for conversion into patient-specific digital twins Abbreviations: AI, artificial intelligence; CNN, convolutional neural network; CT, computed tomography; ctDNA, circulating tumor deoxyribonucleic acid; DCE, dynamic contrast-enhanced; IBSI, Image Biomarker Standardization Initiative; MRI, magnetic resonance imaging; NSCLC, nonsmall cell lung cancer; pCR, pathological complete response; PDE, partial differential equation. Table 4Clinical applications of the oncology digital twin, by category of evidentiary maturityCategoryApplicationData sourceClinical benefitRefsCurrentAutomated RECIST assessmentCT/MRIReduced interreader variability 28 29 CurrentCT body composition (sarcopenia)CTPersonalized supportive care 30 CurrentEquipment predictive maintenanceScanner IoT sensorsReduced unplanned downtime (vendor-reported) 27 EmergingRadiomic response biomarkersCT/MRIEarly response prediction 31 32 EmergingMechanistic breast cancer twin (ARTEMIS)Serial MRIpCR prediction; in silico schedule optimization 13 33 EmergingHead and neck treatment-selection twinCT/MRI, clinicalSequencing decision support 34 EmergingctDNA kinetics monitoringLiquid biopsyEarlier detection of resistance 35 36 FutureMechanistic radiotherapy-optimization twin (glioma)Serial MRIIn silico fractionation optimization (not yet patient-tested) 37 FutureImmunotherapy multibiomarker twinsMultimodalPrecision patient selection (investigational) 12 38 Abbreviations: CT, computed tomography; ctDNA, circulating tumor deoxyribonucleic acid; IoT, Internet of things; MRI, magnetic resonance imaging. Clinical applications of the oncology digital twin, by category of evidentiary maturity Abbreviations: CT, computed tomography; ctDNA, circulating tumor deoxyribonucleic acid; IoT, Internet of things; MRI, magnetic resonance imaging. As noted above, most currently available evidence supports single-output predictive components rather than fully bidirectional, continuously updating twins. Applications below are therefore stratified into three categories that reflect this distinction: applications currently used in routine clinical practice, emerging applications undergoing validation, and future digital twin applications requiring continuous recalibration and prospective simulation. Applications Currently Used in Routine Practice Equipment-Level Digital Twins (Device Predictive Maintenance) It is worth distinguishing patient-facing digital twins from a separate, already-commercialized category: device-level digital twins used for imaging-equipment reliability. GE HealthCare's OnWatch Predict, for example, builds a continuously updated virtual model of an installed MRI system from real-time sensor data that can forecast component failure and can schedule a maintenance preemptively reporting reductions in unplanned downtime of up to 60% in internal analyses. 27 This is a genuine, deployed digital twin architecture—continuous data ingestion, simulation, and forward prediction—but it is not patient-specific. Automated RECIST-Based Response Assessment AI tools can automatically detect, measure, and track tumors on serial CT and MRI studies, with a recent systematic review of image- and report-derived approaches finding that automated pipelines can achieve good agreement with radiologist-assigned RECIST 1.1 response categories, though performance varies considerably across studies and reference standards. 28 Commercial systems such as Siemens Healthineers AI-Rad Companion and Arterys Oncology already support automated lesion tracking. Reinforcement-learning approaches are increasingly explored for automated landmark and lesion detection in medical imaging more broadly. 29 CT-Derived Body Composition Analysis Automated quantification of skeletal muscle index (SMI) at the L3 vertebral level from routine staging and response CT examinations provides an actionable biomarker for sarcopenia (defined as SMI < 52.4 cm 2 /m 2 in men, < 38.5 cm 2 /m 2 in women) that is independently associated with chemotherapy toxicity, treatment discontinuation, and overall survival across multiple tumor types. 30 Fully automated convolutional neural network (CNN) architectures such as 2D/3D U-Nets are typically trained on multi-institutional data sets to approach human-level accuracy for SMI measurement and can be integrated into twin outcome modules to generate individualized survival probability estimates. Patients flagged with sarcopenia may benefit from prospective dietitian referral, dose modification guidance, and structured exercise prescription. Emerging Applications Undergoing Validation Radiomic Biomarkers for Treatment Response Prediction Radiomic feature extraction from pretreatment and early on-treatment CT and MRI examinations can help predict pathological complete response to neoadjuvant chemoradiotherapy, using LASSO (least absolute shrinkage and selection operator) or random forest feature selection to identify the most stable and predictive texture features. 31 32 Deep-radiomics approaches employ CNNs or vision transformers to learn spatial features automatically, and delta-models using recurrent neural networks or long short-term memory to model feature change over time have, in selected cohorts, achieved higher predictive power than static baseline features alone. Patient-Specific Mechanistic Digital Twin—Triple-Negative Breast Cancer The most mature example of a genuine bidirectional oncology digital twin in the current literature comes from the ARTEMIS trial cohort. A biology-based mathematical model of tumor cell proliferation and treatment-induced death, calibrated to serial multiparametric MRI acquired before treatment and after two and four cycles of chemotherapy in 56 patients with TNBC, achieved an AUC of 0.89 for predicting pathological complete response to neoadjuvant chemotherapy. 33 In a related study using an expanded ARTEMIS cohort, the calibrated twin was used prospectively, in silico, to test alternative anthracycline/taxane scheduling options per patient, identifying patient-specific schedules associated with improved predicted pathological complete response rates in retrospective validation—a demonstration of the prospective-simulation property that distinguishes a genuine twin from a static predictive model. 13 Treatment-Selection Digital Twin Dyad—Head and Neck Cancer A coupled “patient–physician digital twin dyad” for oropharyngeal squamous cell carcinoma used deep Q-learning, trained on 536 retrospective patients, to recommend optimal sequencing between induction chemotherapy, radiotherapy, and concurrent chemoradiation, achieving 87 to 91% accuracy in predicting 5-year survival and toxicity outcomes and identifying treatment policies associated with improved predicted survival in retrospective testing. 34 As with the TNBC example above, this qualifies as an early-stage genuine twin (multistep decision optimization with outcome feedback) rather than a single-output classifier, though it remains retrospectively validated and requires prospective testing before clinical adoption. ctDNA Kinetics for Early Resistance Monitoring Serial ctDNA monitoring, combined with mathematical modeling of variant allele frequency kinetics calibrated to an individual's tumor doubling time, can detect emergent resistance in advance of radiological progression in epidermal growth factor receptor-mutated NSCLC and estrogen receptor-positive breast cancer. 35 36 When ctDNA progression and delta-radiomics are concordant, earlier treatment modification may be considered; discordant signals should prompt closer follow-up rather than immediate action. Future Digital Twin Applications Requiring Continuous Recalibration and Prospective Simulation Mechanistic Radiotherapy-Optimization Twin—High-Grade Glioma A predictive digital twin for high-grade glioma has been developed that calibrates a mechanistic tumor-growth model to serial MRI using Bayesian model calibration, then solves a risk-aware optimization problem to propose personalized radiotherapy fractionation schedules. In an in silico cohort of 100 simulated patients, twin-optimized regimens achieved either a modest increase in predicted time-to-progression at equivalent radiation dose, or an equivalent predicted tumor-control probability at reduced dose. 37 This work has not yet been tested in actual patients and represents a research-stage proof of concept rather than a clinically validated tool. Immunotherapy selection using multibiomarker digital twin models and mechanistic tumor growth modeling for broader prognosis and planning remain high-priority applications requiring prospective validation before clinical deployment; longitudinal immune-signature analyses in checkpoint-blockade cohorts illustrate the biological rationale for the former. 12 38 Additional research-stage concepts—including virtual population trials for immunotherapy-beyond-progression decisions, federated pediatric-oncology risk-stratification platforms, metabolically calibrated preclinical twins, and detailed physics-based PET/CT scanner twins for synthetic data generation—are under active development but are not included here as discrete cited applications, since specific primary publications independently verifying these systems could not be confirmed at the time of writing. Challenges and Limitations Imaging Data Harmonization Radiomic models developed at one institution or center may not perform accurately at another without proper standardization and quality control, as radiomic features vary across scanner manufacturers, field strengths, reconstruction kernels, and acquisition protocols. 7 39 ComBat statistical harmonization and DL harmonization techniques can reduce some of this variability, but they require large multi-institutional data sets and are not yet fully reliable. Initiatives such as the IBSI and the Quantitative Imaging Biomarkers Alliance are developing frameworks to improve reproducibility, but this requires consistent imaging protocols, phantom calibration, and ongoing quality-assurance infrastructure. Prospective Validation Gap Most published oncology digital twin studies are retrospective, single-center studies evaluated mainly using statistical metrics such as AUC rather than real clinical outcome measures. 40 Robust evaluation additionally requires appropriate internal, internal–external, and external validation stages before a model's performance claims can be trusted across settings. 41 Strong performance in academic data sets may not always translate into meaningful clinical benefit across different patient populations and institutions, and prospective trials comparing twin-guided care with standard treatment remain limited. Studies such as the I-SPY2 breast cancer trial provide an important model for the type of prospective, adaptive validation that digital twin oncology now requires. 42 Algorithmic Bias and Health Equity Digital twin training data sets predominantly derive from academic medical centers in high-income countries, encoding demographic and socioeconomic biases that may substantially degrade model performance in underrepresented groups. 43 Body composition sarcopenia thresholds derived from North American cohorts, genomic variant pathogenicity classifications reflecting European ancestry, and radiomic models trained on specific scanner types may produce clinically harmful predictions when applied to patients from different ethnic or geographic backgrounds, particularly in low- and middle-income countries (LMICs). Mandatory subgroup performance audits stratified by ethnicity, sex, age, and socioeconomic status should be a condition of any clinical deployment, alongside prospective bias-monitoring protocols. Federated learning across geographically diverse institutions is the most promising structural remedy to this challenge. Digital Twins in LMIC Oncology Digital twin implementation in LMICs faces additional challenges beyond algorithmic bias and the amplification of health care inequities, including fragmented Picture Archiving and Communication System infrastructure, variable imaging protocols, limited genomic testing access, and a lack of large longitudinal data sets. At the same time, LMICs may particularly benefit from scalable digital twin applications such as automated RECIST assessment, CT-based body composition analysis, AI-assisted radiomics, and cloud-based federated learning networks that enable multi-institutional collaboration without direct patient data sharing. 17 18 Standardized imaging protocols, structured reporting, and low-cost interoperable digital infrastructure will be critical for meaningful implementation in LMIC oncology practice. Regulatory and Liability Framework The classification of continuously self-updating AI models as medical devices presents a novel regulatory challenge. The U.S. Food and Drug Administration's Predetermined Change Control Plan pathway and the EU AI Act's high-risk AI provisions offer emerging frameworks, but neither fully addresses a twin that updates its own parameters in response to new patient data without a discrete software update event. 44 The chain of liability when a twin-generated recommendation contributes to an adverse outcome—distributed across model developers, deploying institutions, and treating clinicians—remains legally unresolved. Radiologists should be aware that institutional deployment of digital twin tools may require institutional review board oversight, prospective performance monitoring protocols, and formal legal review of liability frameworks before clinical integration. Future Directions Foundation models are large AI systems trained on massive multimodal medical data sets that can support digital twin development across radiology, pathology, and genomics. Models such as CONCH for pathology and RAD-DINO for radiology can be adapted with smaller local data sets, reducing the need for extensive local training data. 45 46 Wearable sensors may transform digital twins from episodic, clinic-based tools into continuously updated patient companions; feasibility studies during radiotherapy have already demonstrated that smartwatch-derived heart rate and activity data can be collected longitudinally from patients undergoing treatment, raising the prospect—not yet clinically validated—that such signals could help flag treatment toxicity or disease progression earlier than clinical presentation. 47 Emerging technologies such as spatial transcriptomics and single-cell multiomics will allow more detailed characterization of the tumor microenvironment, improving biological precision within digital twin models. For radiology, this evolution will require structured reporting standards, reproducible radiomic workflows, robust quality assurance, and active radiologist involvement in clinical implementation. The Evolving Role of the Radiologist in the Digital Twin Era Radiologists will play a central role in the clinical implementation of oncology digital twins. Imaging is uniquely suited for longitudinal, whole-body, noninvasive assessment of tumor biology, making radiology the primary real-time update signal within the twin framework. 4 5 The radiologist's role will increasingly expand from visual interpretation to quantitative biomarker analysis. Radiomics, functional imaging, PET-derived metrics, and delta-radiomics provide dynamic information on tumor heterogeneity, treatment response, and early progression that cannot be captured by static imaging assessment alone. 7 8 9 10 Integration of imaging with genomics, pathology, ctDNA, and clinical data will require radiologists to participate actively in multimodal data interpretation and imaging-genomic correlation. Realizing this potential in routine radiology practice depends on infrastructure standards that sit squarely within radiology's domain, not just algorithm performance. Imaging protocol standardization, anchored to IBSI-compliant feature definitions, is a prerequisite, since scanner- and vendor-specific acquisition differences alone can shift radiomic feature values; heterogeneity of acquisition and segmentation protocols across institutions compounds the lack of multicenter validation in radiomics research. 7 39 An AI application already being explored here is the automated protocol-compliance checker that flags nonharmonized acquisitions before a scan enters a twin pipeline. Lesion segmentation must also be reproducible, since delta-radiomic and volumetric inputs are only as reliable as the boundary drawn; validated DL autosegmentation tools, such as the self-configuring nnU-Net framework, offer a practical route to this and should be paired with built-in interreader variability reporting. 48 Structured reporting should replace free-text narrative with discrete, machine-ingestible fields, an area where natural-language-processing and LLM-based structured-data extraction tools may allow existing narrative reports to be converted retrospectively into twin-compatible fields. Quality control of radiomics pipelines should draw on existing appraisal frameworks rather than be reinvented: the Radiomics Quality Score assesses the radiomics workflow across domains spanning image protocol, feature extraction, data analysis and statistics, model validation, clinical validity, and open science, 4 while the CLAIM checklist and the TRIPOD + AI statement provide complementary, imaging-specific, and general prediction-model reporting standards against which any twin-integrated radiomics or AI component should be benchmarked before clinical adoption. 49 50 Automated RQS/CLAIM-compliance auditing tools are now being piloted for exactly this kind of pipeline review. Reproducibility across scanners and institutions remains one of the most significant unresolved barriers and should be explicitly tested, using harmonization algorithms such as ComBat or DL style-transfer methods, rather than assumed. Because twin outputs are probabilistic rather than deterministic, radiologists also need working familiarity with uncertainty reporting: calibrated confidence intervals or uncertainty scores allow a radiologist to distinguish a high-confidence twin prediction from one requiring closer scrutiny, and radiology-specific frameworks for uncertainty quantification in DL—including Bayesian DL and conformal-prediction methods that output a confidence interval alongside every twin prediction—are now available to guide this practice. 51 Finally, radiologists will increasingly be responsible for communicating twin-derived information at multidisciplinary tumor boards, translating probabilistic, model-generated outputs into clinically actionable language for oncologists and surgeons while making explicit which components of a twin's prediction rest on validated evidence versus investigational modeling; LLM-based summarization tools that convert multimodal twin outputs into structured, guideline-referenced tumor board summaries with uncertainty flagged are an early example of AI support for this task. Without attention to these radiology-specific implementation requirements, the technical promise of oncology digital twins described elsewhere in this review will not translate into reliable, reproducible clinical practice. Conclusion Oncology digital twins represent a clinically maturing framework for patient-specific, data-driven cancer care, though the majority of published evidence to date supports individual predictive components—what this review terms digital shadows—rather than fully realized, bidirectional twins. The radiologist has a pivotal role in this framework: quantitative imaging analysis—radiomics, functional MRI, PET metrics, and digital pathology—constitutes the primary real-time update signal that keeps the twin synchronized with the patient's evolving biological state. The framework rests on three core biological components—radiomics and quantitative imaging analysis, mechanistic tumor growth and treatment response modeling, and multiomic data integration—supported by two enabling technologies, federated learning for multi-institutional model training and LLM-based clinical interfaces with agentic AI reasoning capabilities. These technologies sit at varying stages of clinical readiness. CT body composition analysis and automated RECIST assessment are implementable now with available tools and evidence, albeit with performance that varies across settings; radiomic response biomarkers and early mechanistic twins (as in the ARTEMIS breast cancer and oropharyngeal cancer examples) are entering early prospective and retrospective multi-institutional validation; and immunotherapy multibiomarker selection, mechanistic radiotherapy-optimization twins, and broader tumor modeling remain in research-stage validation, largely untested in real patients. Readers should treat claims of near-term “clinical readiness” for any specific application with appropriate caution and verify the underlying evidence tier before altering practice. Barriers that remain—prospective validation gaps, imaging standardization, algorithmic bias risks, and regulatory uncertainty—are real and require active attention from radiologists, oncologists, and informaticists working in concert. The practical roadmap is clear: begin with the entry-point applications that require no new infrastructure, build the imaging data quality and standardization foundation that radiomic models require, contribute to prospective validation studies, and engage proactively in the governance and quality-reporting frameworks (IBSI, CLAIM, TRIPOD + AI, RQS) that will define clinical deployment. Radiologists who develop fluency in this domain will be positioned to contribute meaningfully to the next generation of cancer care.