Evolving and novel applications of artificial intelligence in interventional oncology

Abstract

Artificial intelligence (AI) is increasingly influencing clinical oncology, yet its role in interventional oncology (IO) has not advanced as rapidly, despite IO’s heavy reliance on imaging, precision, and real-time decision-making. While AI has demonstrated utility in other medical disciplines such as diagnostic radiology and medical oncology, its adoption in IO is still emerging and varies across applications. This review provides a comprehensive overview of current and emerging applications of AI across the IO workflow, including procedural planning, real-time navigation and treatment delivery, and post-procedural evaluation. It also explores system-level innovations and patient-centered tools designed to enhance workflow efficiency, documentation, education, and follow-up. These technologies hold promise for improving consistency, personalizing treatment, and enhancing safety. Although existing applications demonstrate clear clinical value, broader adoption will depend on overcoming technical, ethical, and infrastructure barriers. Addressing these challenges will require prospective validation, infrastructure development, and sustained multidisciplinary collaboration. As AI tools evolve, their potential to connect fragmented stages of care and enable adaptive, data-driven interventions may redefine IO practice. This review clarifies the current landscape, identifies key implementation barriers, and outlines priorities for future research and clinical integration.

Key Words: Artificial intelligence; Interventional oncology; Image-guided therapy; Machine learning; Deep learning; Radiomics; Procedural planning; Real-time navigation; Neural network; Natural language processing

Core Tip: Artificial intelligence is transforming interventional oncology by enhancing precision, streamlining workflows, and enabling personalized care across all procedural stages. This review highlights emerging applications and outlines how artificial intelligence integration is reshaping practice, improving outcomes, and setting the foundation for data-driven, adaptive interventions in image-guided oncology.

INTRODUCTION

Interventional oncology (IO) is a rapidly advancing subspecialty that employs image-guided, minimally invasive procedures for diagnosing and treating cancer. These procedures require exceptional precision, real-time decision-making, and seamless integration of multimodal data. As artificial intelligence (AI) becomes increasingly embedded in modern clinical medicine, its potential to enhance diagnostic accuracy, streamline workflows, and support personalized treatment strategies is actively being explored across multiple oncologic disciplines[1-4]. In oncology, AI has already proven valuable in tumor detection, risk stratification, and treatment planning, particularly within the fields of diagnostic radiology and medical oncology[5].

Despite IO’s reliance on advanced imaging and technical precision, AI adoption in the field remains comparatively underdeveloped. Variability in procedural execution, workflow inefficiencies, and imaging limitations can compromise procedural accuracy and patient outcomes[1]. Nonetheless, these challenges position IO to benefit from AI solutions that automate time-intensive tasks, improve image quality, and integrate multimodal data to support consistent, evidence-based decision-making.

This review organizes and contextualizes current and emerging applications of AI across the IO continuum, from how procedures are planned, executed, and evaluated, with both clinician and patient perspectives in mind. It highlights clinical opportunities, implementation challenges, and priorities for future research and adoption, aiming to advance a more adaptive, efficient, and cost-effective IO practice[6].

CURRENT LANDSCAPE OF AI IN IO

AI in IO has experienced rapid research growth with select tools, especially those focused on image processing, entering early clinical use. However, most remain in pilot or investigational stages due to limited prospective validation, lack of standardized protocols, incomplete regulatory frameworks, and infrastructure gaps. These factors collectively limit adoption into routine practice, despite promising technical performance and growing clinical interest[6-9]. To date, relatively few AI tools have obtained United States Food and Drug Administration (FDA) clearance for procedural use. Most FDA-authorized AI devices are diagnostic in nature. Many AI models in IO remain investigational, and those that are authorized are generally classified as ‘software as a medical device’. This designation reflects ongoing uncertainty surrounding the regulation of adaptive AI systems. Historically, FDA authorizations treated AI models as static, requiring re-clearance for any updates. Recent policy changes now permit pre-approved updates via predetermined change control plans, allowing for more flexible model evolution without full resubmission[10-13].

In contrast, the European Union (EU) offers a more structured regulatory pathway through the Conformité Européene (CE) mark. AI-enabled medical devices must undergo a conformity assessment, including audits of the manufacturer’s quality system and a review of technical documentation to ensure safety and performance. For certain high-risk devices, notified bodies are required to consult expert panels, and in some cases, from the European Medicines Agency itself. This process is governed by the Medical Devices Regulation (EU 2017/745) and complemented by the AI Act (EU 2024/1689), which adds lifecycle monitoring and risk management requirements for high-risk AI systems[14-17]. Differences in post-market surveillance, validation standards, and liability frameworks between the FDA and CE-mark processes continue to influence global scalability, investment decisions, and clinician trust in AI deployment. While the European AI Act marks progress toward harmonized regulation, its implementation remains challenged by inconsistent national resources, overlapping frameworks, and gaps in enforcement. As highlighted by Vardas et al[18], these unresolved challenges contribute to persistent uncertainty in clinical adoption and risk management. While national society guidelines, such as those from the American Society of Clinical Oncology, are beginning to address the integration of AI into oncology by issuing principles on transparency, equity, accountability, and human-centered application[19], IO-specific guidance is still lacking.

Within the oncologic procedural workflow, pre-procedural tools such as lesion segmentation and image enhancement are the most developed, with several models demonstrating high accuracy and clinical utility[7,20]. Intra-procedural technologies, including real-time needle tracking, motion compensation, and robotic navigation, largely remain experimental, facing challenges from latency and integration with existing workflows[21,22]. Tools for predicting recurrence, evaluating margins, and standardizing reports have shown promise in retrospective studies but are not yet standard practice[23,24]. At the system level, applications aimed at improving workflow efficiency or enhancing patient experience are still in early development but represent emerging areas of interest[25,26]. As AI capabilities expand, opportunities lie in rigorous validation, scalable infrastructure, and seamless integration across procedural phases. The transition from isolated AI applications to integrated systems presents an opportunity to unify pre-, intra-, and post-procedural phases, thereby improving procedural consistency, care quality, and efficiency. Table 1 summarizes current and emerging AI applications in IO discussed throughout the text.

Table 1 Artificial intelligence applications by procedural phase in interventional oncology.

IO phase	AI application
Pre-procedural	Lesion segmentation (CNNs, radiomics)
	Needle path planning (CT-guided)
	Outcome prediction (e.g., TACE response, recurrence risk)
	Imaging enhancement (denoising, SNR/CNR improvement)
Intra-procedural	Real-time motion correction. Image fusion (CT/US)
	Needle tracking and trajectory optimization
	Multimodal image registration
	Robotic assistance
Post-procedural	Structured reporting (NLP, LLMs)
	Margin assessment (3D modeling, deformable registration)
	Recurrence prediction (radiomics, XGBoost)
	Longitudinal lesion tracking
System-level/patient-centered	Predictive maintenance (equipment)
	Workflow optimization (scheduling, triage)
	Patient education (AR tours, chatbots)
	Research support (trial matching, literature mining)

IO: Interventional oncology; AI: Artificial intelligence; CNN: Convolutional neural network; CT: Computed tomography; TACE: Transarterial catheter embolization; SNR: Signal-to-noise ratio; CNR: Contrast-to-noise ratio; US: Ultrasound; NLP: Natural language processing; LLMs: Large language models; 3D: Three-dimensional; XGBoost: Extreme gradient boosting; AR: Augmented reality.

AI-ASSISTED PRE-PROCEDURAL PLANNING

IO’s pre-procedural planning is foundational to procedural accuracy and patient safety, relying heavily on imaging to guide lesion targeting, procedural strategy, and risk assessment. Recent literature highlights AI’s growing role in supporting this phase, offering tools for lesion segmentation, needle trajectory planning, and treatment response prediction[6,27]. Early clinical applications of AI have enhanced pre-procedural imaging across different organ systems, including thoracic and abdominal oncology, improving diagnostic clarity and refining lesion localization and procedural strategies[28,29]. These advances also generate outputs that directly feed into intra- and post-procedural decision making[30,31].

Lesion segmentation & image enhancement

Accurate lesion delineation and high-quality imaging are essential for successful IO procedures. Segmentation is among AI’s most widely validated applications, with machine learning (ML) models achieving high Dice similarity scores across modalities, particularly in liver and lung oncology[7,8,32]. In colorectal liver metastases (CRLM), AI-assisted segmentation significantly reduced operator interaction time compared to manual methods (P < 0.001), while preserving accuracy[20].

AI-based denoising and reconstruction pipelines improve lesion visibility, shorten scan times, and enable substantial dose reductions without compromising image quality[9,33,34], all of which are especially pertinent to oncology populations, who are frequently subjected to repeated imaging and consequently face a heightened risk for radiation-induced complications. Deep learning reconstruction (DLR) has improved signal-to-noise ratio and contrast-to-noise ratio in both magnetic resonance imaging (MRI) and computed tomography (CT) across multiple organs, including prostate MRI[35] and liver imaging[36-38]. These advances enhance lesion visualization even at reduced radiation doses[38,39], in MRI, can shorten breath-hold times[36]. For focal hepatic lesions, DLR has even outperformed traditional iterative reconstruction, offering clearer lesion characterization and potentially better targeting[40]. Whether through suppressing artifacts or increasing sharpness, improved resolution enhances image quality and can translate into more successful procedures. These image optimization tools support more confident and accurate targeting, serving as the basis for downstream applications such as ablation zone simulation[21,22] and catheter navigation[41-44].

Procedural path & catheter planning

Beyond imaging, AI facilitates optimal patient-specific needle, probe, and catheter trajectories. In CT-guided lung biopsies, convolutional neural networks (CNNs) have produced clinically viable paths with reduced needle insertion time and depth[21,22,45]. Similar models applied to thermal ablation predict optimal probe placement, achieving up to 78% accuracy with minimal manual adjustment and reduced preparation time[46]. While geometric models are advancing to support automated treatment planning for thermal ablation procedures[47], many current approaches still struggle to account for the complex thermal interactions that occur in tissue when multiple needles are used. The interactions driven by overlapping heat diffusion from neighboring probes can lead to suboptimal treatment coverage if not properly modeled. To address this, simulation platforms using GPU-accelerated bioheat models allow real-time visualization of ablation zones, improving planning precision and potentially reducing incomplete treatments[48]. Although these models do not yet fully capture multi-probe thermal dynamics, AI-driven approaches have significantly improved the precision of ablation planning. For example, computational modeling for catheter placement optimization and dynamic energy delivery enables customized ablation shapes tailored to tumor geometry, as demonstrated in automated systems that modulate probe power and trajectory to match tumor contours[47]. These models support more precise and individualized treatment strategies by simulating ablation zones to maximize tumor coverage while sparing adjacent structures[41-43].

For catheter-based therapies such as transarterial chemoembolization (TACE) and transarterial radioembolization with Yttrium-90, multiple studies have explored models for segmenting hepatic vasculature and informing catheter planning, with potential to enhance precision[49-53]. A radiomics-based model developed by Masthoff et al[54] predicts the most effective TACE techniques, suggesting that in nearly two-thirds of cases, an alternative TACE method might have been more appropriate. This finding underscores the critical importance of individualized therapy selection, which can significantly improve clinical outcomes and reduce unnecessary interventions. Integration of vascular modeling with hemodynamic simulations can guide microcatheter placement, optimizing embolic distribution while reducing the likelihood of non-target deposition[55,56]. Complementing AI-guided catheter planning strategies with innovations such as tumor vessel-adaptable adhesive microspheres can allow these microspheres to be deployed more effectively, aligning embolic behavior with patient-specific vascular territories[57]. These anatomy-informed strategies streamline intra-procedural execution and reduce procedural variability.

Patient selection & responder prediction

Determining which patients will benefit the most from a given IO procedure remains a clinical challenge. Predictive algorithms trained on imaging, radiomics, and clinical data are being integrated to aid in selecting candidates for TACE, ablation, and surgical resection. Incorporating radiomic features into planning tools further enhances prediction of tumor behavior, including nodule shape, histologic subtype, and genetic profile[58,59]. For hepatocellular carcinoma (HCC), prognostic modeling has demonstrated reliable performance, with models effectively distinguishing between patients likely to respond vs those at higher risk of poor outcomes with ablation[60]. Among other liver-directed therapies, such as the ML CNN models described by Morshid et al[61], integrated CT imaging features with clinical data to predict HCC response to TACE, improving Barcelona Clinic Liver Cancer staging accuracy from 62.9% to 74.2%. Similar radiomics-based models have been applied to predict tumor response, and in some cases, stratify patients between less invasive approaches such as microwave ablation (MWA) vs surgical resection[62,63]. Ding et al[63] designed a treatment selection model to guide decisions between laparoscopic hepatectomy and MWA for HCC, resulting in a reduction of early recurrence (ER) rates by over 12%. Notably, predicted ER rates closely aligned with observed outcomes, and adherence to the model’s recommendations was associated with significantly lower ER rates (P = 0.042 for laparoscopic hepatectomy; P = 0.048 for MWA)[63].

While most pre-procedural prognostication models have concentrated on liver-directed therapies[64-66], similar frameworks are emerging for adrenal, lung, and other metastatic disease sites, accurately predicting local control and survival outcomes. For example, Daye et al[67] trained a support vector machine to predict ablation response, local progression, and survival in patients with adrenal metastases, with radiomic features improving predictive accuracy over clinical data alone[67]. In lung IO, deep learning (DL) models have shown strong performance in predicting outcomes of procedures when trained on pre-procedural imaging and clinical data[41]. Prognostic DL models combining clinical, genomic, and radiomic features have further improved prediction of treatment response with radiomics-based mutation prediction, highlighting AI’s comprehensive treatment planning ability in a more tailored approach[32,68]. Similarly, computational modeling tools such as TumorScope simulate tumor microenvironments and drug delivery kinetics using patient-specific data to inform systemic planning and predict therapeutic response and dosing strategies[69].

Early identification of likely non-responders allows clinicians to redirect patients toward more appropriate alternative therapies, minimizing unnecessary procedures and their associated risks. The pre-procedural AI inputs can be used synergistically to directly inform intra-procedural guidance systems and post-procedural monitoring. By integrating these capabilities into a continuous workflow, AI transforms pre-procedural planning from a static task into a dynamic, data-driven foundation for precision oncology in IO (Figure 1).

Figure 1 Schematic overview of artificial intelligence integration in interventional oncology, with inputs and outputs directly and indirectly guiding procedural planning, execution, and evaluation. AI: Artificial intelligence; CT: Computed tomography; MRI: Magnetic resonance imaging; PET: Positron emission tomography; US: Ultrasound.

AI-ENABLED INTRA-PROCEDURAL GUIDANCE

Intra-procedural guidance represents a critical inflection point in IO, where precision, adaptability, and real-time multimodal data integration directly influence therapeutic outcomes. AI technologies are being developed to support proceduralists through live imaging analysis, needle tracking, safety monitoring, and adaptive treatment delivery. These promising tools process streaming data, compensate for anatomical shifts, and integrate multiple modalities to enhance spatial precision and reduce variability[21,70,71]. As a result, proceduralists can respond dynamically to anatomical variation, evolving procedural conditions, and procedural variables that previously compromised precision, such as respiratory motion. Although many of these tools remain in early-stage or investigational use, preliminary applications demonstrate measurable utility in efficiency, targeting, and safety.

Real-time imaging analysis & fusion

Integrating DL with multimodal imaging platforms has accelerated safer and more effective techniques with fusion-guided interventions. DLR algorithms produce high-fidelity CT images while enabling substantially lower radiation doses[72-74]. Concurrently, AI-driven artifact reduction techniques have also improved needle and lesion visibility. In CT-guided cryoablation, DLR-based metal-artifact suppression improved radiologist visual image quality scores by 34%-46%[75]. These approaches support real-time fusion of intra-procedural imaging with pre-procedural scans. Automated segmentation of organs and tumors facilitates synchronization across modalities, such as live CT with diagnostic CT, MRI, or positron emission tomography, creating real-time overlays that enhance targeting precision[76,77]. In practice, this reduces repeat scans and radiation exposure while improving localization and treatment efficacy.

Needle tracking & motion compensation

While pre-procedural AI tools simulate optimal needle trajectories using static imaging, intra-procedural navigation focuses on real-time adjustments[45]. These models process live image streams to predict needle paths and detect deviations. Current systems, still under active investigation, deliver immediate feedback on needle position and trajectory. In complex procedures requiring precise targeting, AI-assisted navigation has demonstrated high spatial precision across multiple imaging modalities, supporting more efficient interventions. For example, in MRI-guided prostate biopsies, a Mask R-CNN model achieved sub-millimeter needle tip localization with a median error of 0.71 mm[78]. A separate study employing a three-dimensional (3D) fully CNN reported median localization and trajectory errors of 2.80 mm and 3.00 mm, respectively, comparable to expert performance while reducing localization time by nearly half[79]. In ultrasound-guided procedures, deep convolutional networks have successfully localized steeply inserted needles in phantom studies, achieving tip errors as low as 0.23 ± 0.05 mm and operating in real time at 25 frames per second[80].

Building on AI-assisted navigation, robotic integration further enhances guidance. Arapi et al[44] trained a detection transformer neural network integrated with a robotic ultrasound scanner to autonomously optimize probe positioning, visualize both needle and target, and compute viable insertion angles and paths, achieving a mean average precision of 74%-77% in vitro using 3D phantoms. Similarly, more recent CNN-based needle trajectory generators for CT-guided lung biopsies matched or outperformed expert-selected paths in over 85% of cases, demonstrating high concordance with historical biopsy trajectories[21].

AI-driven motion correction tools further improve spatial alignment and targeting. DL models applied to cone-beam CT have significantly increased spatial congruence between predicted and actual vascular structures during hepatic ablation, with median Dice scores improving by up to 0.29 units and notable gains in vessel sharpness[71]. Similarly, DL-based image fusion improved registration precision of CT-ultrasound, outperforming manual alignment and reducing targeting error[81].

In addition to image alignment, AI models are now capable of segmenting arteries and catheters in real time, potentially enabling immediate recognition and correction of vessel displacement or tissue deformation[71]. This reduces repositioning attempts and mitigates the risk of incomplete ablation. Multimodal integration of CT, MRI, and ultrasound enables identification of subtle anatomical variants and tumor margins that are often missed with single-modality imaging. Investigational tools have shown promise in flagging catheter or needle deviation before non-target embolization, enhancing intra-procedural vigilance while minimizing complications, though these tools are not yet standard in IO practice[70,81]. Early preclinical and clinical data suggest more reliable targeting and procedural safety, with measurable reductions in procedure time, radiation exposure, fewer repositioning attempts, and higher first-attempt success rates during biopsies and ablations compared to conventional methods[21,70,81].

Safety monitoring

In addition to these safer techniques, AI offers new models that incorporate advanced safety monitoring features to detect complications in real time. CNNs trained on angiographic data have achieved receiver operating characteristic area under the curve values of 0.85 or higher in detecting arterial extravasation[82,83]. Similarly, spatiotemporal models applied to digital subtraction angiography during thrombectomy have identified vessel perforations with high precision and recall[83]. Although primarily developed outside of IO, these models are directly applicable for detecting complications such as bleeding or vascular injury during embolization or ablation. In IO-specific use, neural networks for photon-counting CT thermometry have achieved temperature measurement errors within 1.8 °C to 4.0 °C, enabling tighter energy regulation and reducing collateral injury during ablation[84,85]. These real-time monitoring capabilities enhance procedural safety and lay the foundation for dynamic, response-adaptive treatment delivery.

Treatment delivery optimization

Recent advances in AI are enabling more dynamic, patient-specific modulation of treatment parameters during IO procedures. While AI applications in clinical oncology already play a strong role in digital pathology, biomarker development, and dose modulation[86], evidence for AI optimizing real-time dosage and delivery kinetics remains limited. Adaptive platforms such as CURATE.AI, originally developed for systemic therapy, personalize dosing based on treatment response profiles. In early studies, these platforms have demonstrated feasibility in minimizing overall drug exposure by adjusting doses over time, rather than adhering to a fixed regimen[87,88]. Although not yet widely available in IO, these adaptive dosing strategies offer a conceptual framework for real-time, patient-specific modulation of therapeutic delivery - an approach increasingly reflected in emerging IO techniques.

In TACE, emerging DL algorithms can segment hepatic vasculature, identify tumor-feeding vessels, and predict treatment response from digital subtraction angiography in real-time[89], while comparative studies of different TACE techniques demonstrate that tailoring embolic load to tumor characteristics is associated with differences in both tumor response and adverse events[90]. Similarly, individualized dosimetry and dose-monitoring with transarterial radioembolization allow for optimization of radiation delivery, reducing the risk of toxicity while maintaining therapeutic efficacy[91]. Together, these approaches illustrate how real-time, data-driven treatment adjustments could be integrated into IO, forming a foundation for adaptive, response-guided interventions akin to the dynamic dosing strategies explored earlier in systemic therapy platforms. Future integration of AI with intra-procedure imaging, perfusion dynamics, and vascular architecture could enable tailored interventions, adjusting embolic volume, delivery speed, and drug concentration to align with therapeutic intent, whether curative, ablative, or catheter-based treatments[92,93].

AI-driven applications in IO could enable pre-emptive adjustments that maximize on-target delivery while minimizing off-target effects, allowing proceduralists to modify treatment plans in response to perfusion dynamics, incomplete embolization, or other intra-procedural factors. By integrating quantitative radiomics metrics and simulation-based guidance, these systems facilitate earlier therapeutic decisions, standardize selection of embolic agents, refine catheter positioning, and adapt treatment for suboptimal responders sooner. Identifying patients most likely to benefit from locoregional therapies helps avoid ineffective or potentially harmful procedures and optimizes resource utilization by prioritizing the most effective treatment regimens for each patient[6,90,91,94]. Overall, AI supports more individualized, response-adaptive interventions that align procedural parameters with lesion biology and dynamic treatment response.

AI IN THE POST-PROCEDURAL PHASE

Post-procedural care in IO encompasses a range of critical tasks, such as outcome monitoring, documentation, quality assurance, and longitudinal follow-up. These processes are often time-intensive and susceptible to variability, making them ideal targets for AI-driven optimization. These technologies enhance the accuracy, efficiency, and standardization of post-procedural workflows, from documentation to recurrence prediction, alleviating clinician workload while improving data quality and consistency[95]. AI technologies can also be used to reengineer the quality control of procedures or any documentation created thereafter. Robust post-procedural analysis is essential for predicting treatment success, recurrence, and complications[6]. AI systems enhance the completeness and accuracy of many post-procedural operations while providing actionable feedback that can inform future procedural planning. Together, these AI-driven tools represent a shift toward intelligent, data-driven post-procedural monitoring that strengthens safety and standardization in IO.

Structured reporting & documentation

Accurate and timely documentation in IO is essential for continuity of care, multidisciplinary coordination, and longitudinal treatment planning. Traditional documentation methods, which rely heavily on manual input and clinician recall, are time-consuming and prone to inconsistency and omissions. AI-powered structured reporting systems mitigate these issues by auto-populating standardized templates using procedural metadata, imaging findings, and electronic medical record (EMR) integration[96,97]. These systems improve report completeness, reduce variability, and streamline documentation workflows[95,98].

Recent advances in DL have enabled the generation of free-text radiology reports directly from imaging data, with uncertainty estimations to distinguish between confident and ambiguous findings[99]. This capability enhances interpretability and supports more informed clinical decision-making, particularly in complex post-procedural assessments. Natural language processing (NLP) and large language models (LLMs) are also being piloted to convert free-text procedural notes and voice dictation into structured, codified reports, improving completeness and uniformity while reducing documentation burden[24,99-101]. These systems also support real-time tagging of procedural steps, automated coding for billing and registries, and integration with tumor board workflows[100,101]. However, these applications remain in early-stage development, and their clinical validation and regulatory approval are still evolving, requiring cautious integration into practice.

In oncology-specific contexts, structured templates have demonstrated utility in improving staging accuracy and surgical planning, particularly in complex malignancies such as pancreatic cancer[102]. Adopting AI-enhanced documentation platforms represent a foundational step toward scalable, data-driven IO practice. However, these advances depend on robust infrastructures, including standardized templates and interoperable data pipelines to support downstream analytics[103,104].

Quality assurance & technical evaluation

Maintaining technical consistency in IO procedures remains a challenge, particularly in high-volume or time-constrained settings. Algorithms have been applied to retrospectively assess procedural accuracy, including assessment of ablation margins or tracking needle trajectories, and treatment overlap[95,105,106]. In some cases, these algorithms can flag issues such as absent or insufficient ablation margins or mismatched tumor laterality - errors that may be overlooked in real time. 3D modeling tools and deformable image registration techniques enable retrospective assessment of needle trajectories and treatment margins, providing post-procedural feedback that may refine technique and reduce variability across operators[6,107].

Unlike manual segmentation, AI-based volumetric tools provide rapid and reproducible margin assessments in ablation workflows. These tools, including those modeled after the COVER-ALL protocol, assess tumor coverage and residual disease in real time, while also identifying gaps in documentation[6,107]. Post-processing platforms, such as those used by Laimer et al[108], can quantify peri-ablational margins and highlight spatial mismatches between planned and delivered ablation zones[6].

Initially developed for intra-procedural use, trajectory analysis tools are now being applied retrospectively to simulate alternative ablation paths while accounting for anatomical constraints. Robot-assisted CT-guided targeting systems using image-only navigation have demonstrated high precision in stereotactic needle insertion without the need for positional sensors, offering a modern framework for trajectory planning and evaluation[109]. Complementing these retrospective tools, AI-enhanced fusion imaging and other advanced imaging techniques are also transforming immediate post-procedural evaluation by enabling precise and timely margin assessment before discharge. A systematic review of 22 studies reported technical success rates of 81%-100% for fusion imaging, with many enabling same-session re-ablation[110]. Supporting this, a prospective study using 3D contrast-enhanced ultrasound fusion identified residual lesions in real time. The residual lesions requiring re-ablation achieved a final technical effectiveness of 98.8%[110]. These findings underscore the versatile utility of fusion techniques in IO, from identifying missing information to actively optimizing treatment outcomes at all phases of therapy.

Outcome prediction, response monitoring, & longitudinal tracking

AI applications in the post-procedural phase extend to predicting treatment outcomes, monitoring recurrence, and tracking lesion evolution[111-113]. For instance, Yin et al[23] developed two radiomics-based ML classifiers - lasso regression and extreme gradient boosting (XGBoost) - to detect ablation site recurrence in post-thermal ablation CT scans of patients with HCC or metastases from other organs. These models achieved receiver operating characteristic area under the curves of 0.97 and 0.93, with corresponding accuracies of 92.73% and 89.09%, respectively[23].

AI has also significantly improved the reliability of radiological treatment response classification, offering more nuanced and timely insights into therapeutic efficacy. For example, an automated DL model developed to detect longitudinal changes in liver metastases on contrast-enhanced abdominal CT scans outperformed the standard RECIST 1.1 criteria, improving disease classification in 34.5% of cases[114]. This improved disease classification enables earlier treatment adjustments and supports more personalized care pathways[115]. Consistent lesion tracking is also crucial for accurate response assessment and guiding follow-up. The deep lesion tracker, developed by Cai et al[116], offers a novel approach to longitudinal lesion tracking by leveraging temporal and spatial imaging features, accurately identifying lesions across serial imaging, reducing the likelihood of mismatched lesions, and improving tracking fidelity. Deep lesion tracker outperformed leading registration algorithms in both accuracy and speed, while attaining comparable accuracy results to expert radiologists when integrated into expert tumor monitoring workflows[116].

Post-procedural outcome prediction also includes identifying patients at risk of complications. For instance, Xu et al[117] trained an artificial neural network to preoperatively assess pleural effusion risk in patients with lung cancer undergoing MWA, using parameters such as ablation power, puncture count, and proximity to the pleura. The model demonstrated high concordance in both training and external validation cohorts. While outcome prediction is still in early development, its potential to be used in conjunction with response prediction tools and customize post-procedural follow-up makes it a promising research avenue. These tools can inform risk mitigation strategies and procedural adjustments in future cases, improving overall safety and effectiveness.

SYSTEM-LEVEL AND PATIENT-CENTERED APPLICATIONS IN IO

The complexity of IO demands a high degree of coordination, precision, and personalization, often requiring the integration of multiple data streams across clinical systems. Unlike fields such as diagnostic radiology or surgical oncology, IO procedures are often performed in real time under image guidance, with outcomes influenced by anatomical variability, tumor biology, and procedural expertise. These challenges are compounded by limited access to real-time analytics, variability in procedural outcomes, and patient anxiety surrounding cancer therapies. While often considered ancillary to other domains, system-level and patient-centered applications of AI are also essential to the advancement of IO. AI technologies are increasingly being deployed to optimize infrastructure, streamline workflows, enhance multidisciplinary collaboration, personalize patient care, and support research - all of which are critical to improving safety, efficiency, and outcomes in IO[118].

AI in multidisciplinary tumor boards

Multidisciplinary tumor boards (MTBs) are integral to decision-making in IO, particularly in complex cases such as HCC, CRLM, and renal tumors[119-122]. These collaborative platforms integrate radiologists, oncologists, surgeons, and interventional specialists to guide treatment selection, especially when image-guided therapies such as TACE, thermal ablation, or stereotactic body radiotherapy are under consideration[123]. In HCC, MTBs help align treatment strategies with Barcelona Clinic Liver Cancer staging, while in CRLM and renal tumors, MTBs facilitate consensus-based decisions regarding resectability, systemic therapy sequencing, and eligibility for interventional procedures[124,125].

AI systems are increasingly being integrated into MTBs to synthesize multimodal data, including imaging, pathology, genomics, and clinical history, and generate structured, evidence-based recommendations[120]. NLP tools can extract relevant clinical information from EMRs, while ML models stratify patients by risk or predict treatment response. These technologies streamline case preparation, improve consistency across reviewers, and allow more time for nuanced decision-making, particularly in high-volume oncology centers where time constraints may limit comprehensive case review.

Workflow optimization & clinical operations

To address the operational complexity of IO practice, AI-driven systems are being developed to streamline scheduling, resource allocation, and administrative workflows. These tools optimize case distribution by accounting for procedural duration, equipment availability, and staff expertise[25]. Predictive models can also estimate patient no-show risk[126] and dynamically adjust scheduling to improve throughput and minimize idle time[26]. However, a more recent model found that predictions were less reliable for oncology patients, highlighting the unique and complex nature of their appointment patterns and clinical workflows[127]. In oncology, ensuring patients attend appointments is especially crucial, as treatments are often time sensitive. Missed visits can delay care, negatively impact outcomes, and contribute to increased downstream healthcare costs. Beyond scheduling, AI integration with radiology information systems (RIS) and picture archiving and communication systems (PACS) enables automation of routine documentation tasks, including structured reporting and follow-up recommendations[24].

In high-volume IO centers, AI-assisted triage systems could be utilized to prioritize patients for urgent procedures such as TACE or ablation, using severity scores derived from imaging and clinical data. These tools may support operational efficiency while ensuring that high-risk patients receive timely care; selecting the most appropriate candidates and triaging for IO procedures is critical to optimizing outcomes and resource allocation. For instance, in prostate MRI interpretation, ML models have effectively triaged patients without clinically significant cancer, reducing radiologist workload while maintaining diagnostic accuracy[128]. Such approaches are increasingly relevant in IO, where timely intervention can significantly influence clinical outcomes.

Patient-centered outcomes & personalized follow-up

As IO procedures become more personalized, so too does much of the post-procedural care that follows, which is supported by AI-enabled platforms monitoring patient-reported outcomes and personalized follow-up care in IO. Mobile health platforms equipped with AI-driven symptom tracking and feedback algorithms enable real-time recovery monitoring, allowing for early detection of complications and dynamic adjustment of care plans[23,129]. These systems integrate lesion characteristics, procedural data, and comorbidities to tailor follow-up schedules and self-care instructions, improving adherence and reducing readmissions. Additionally, AI-powered tools can analyze longitudinal imaging and clinical data to assess treatment response and recurrence risk, supporting more precise and individualized surveillance strategies[130].

Patient experience & comprehension enhancement

Patient-facing education platforms using NLP make medical information more accessible by tailoring content to patients’ literacy levels and individual needs[131,132]. In oncology and interventional care settings, these tools help explain diagnoses and procedures in ways patients can understand. AI is reshaping patient education, orientation, engagement, and recovery in IO by tailoring tools to each patient’s context.

While much of this work is still emerging, proposed applications include interactive tools and virtual interfaces that adapt to individual clinical contexts. For example, AI-driven 360-degree tours allow patients to explore the interventional suite and receive step-by-step explanations of their upcoming procedure, aiming to reduce anxiety and improve informed consent. These immersive strategies build on prior work in radiology education using virtual reality (VR) and augmented reality (AR)[133]. These tours can be captured via computer-aided design modeling, panoramic videos, or digital composition, each offering varying levels of interactivity.

Beyond standalone education materials, AI-powered conversational agents can extract structured data from EMRs to deliver personalized preparation instructions and clinician-facing summaries. These tools integrate key patient variables such as renal function, anticoagulation status, and allergy history to generate individualized pre-procedural plans, reducing instruction errors and enhancing patient safety that could otherwise occur in complex oncology patients[134].

Post-procedural recovery is also being enhanced through smartphone-based systems. For example, structured daily symptom reporting with algorithmic feedback has improved recovery trajectories in perioperative settings[129]. Applied to IO, these apps can deliver tailored post-ablation monitoring schedules, prompt self-care tasks, and escalate recommendations based on lesion characteristics and comorbidity profiles. Examples include encouraging hydration after contrast administration, temperature monitoring for infection risk, and dynamically adjusting care recommendations. These capabilities support safe outpatient management of increasingly complex procedures[129].

Technical support infrastructure & predictive maintenance

AI-based predictive maintenance systems are being deployed to monitor the performance and reliability of interventional equipment, including CT scanners and related imaging platforms[135,136]. These tools use operational data to anticipate hardware failures, reduce downtime, and improve system readiness by analyzing sensor data and detecting early signs of malfunction, such as abnormal thermal patterns or mechanical vibrations. In IO, where procedural success depends on uninterrupted access to high-performance imaging and therapeutic devices, predictive maintenance reduces unplanned downtime and enhances reliability[136].

Education & simulation-based training

AI is advancing IO physician training through simulation-based learning, AR/VR technologies, and personalized education platforms. High-fidelity simulators replicate diagnostic challenges and diagnostic errors[137], while AR/VR systems enable risk-free practice of complex procedures. These platforms complement traditional training by offering real-time feedback, error tracking, and skill progression metrics, and have been shown to improve technical proficiency, particularly among early-career practitioners[137-139]. Broader educational frameworks also emphasize the importance of AI-related competencies for future healthcare professionals[140]. When integrated with AI-based feedback mechanisms, these tools offer performance metrics, such as needle trajectory accuracy, radiation exposure, and complication rates, while tailoring case difficulty and feedback to the user’s performance for objective skill assessment[141].

AI-driven clinical decision support systems also serve as valuable educational tools, offering interactive, case-based modules with immediate feedback and evidence-based recommendations to help trainees and practicing clinicians stay current with evolving procedural standards and improve decision-making consistency in IO practice[142,143]. Personalized education platforms identify skill gaps and anatomical challenges, adapting content and feedback accordingly[141]. Radiology trainees have rated AI-integrated MRI simulations as more educationally valuable than traditional modules[144], and AI-assisted interpretation of chest radiographs has improved sensitivity and specificity for lung nodule detection among junior radiologists[145]. As explored in other specialties, AI may also support credentialing and proctoring, potentially streamlining training assessments, monitoring procedural competency, and personalizing continuing education pathways[146].

Research facilitation & data science integration

AI is also reshaping how research is conducted in IO, helping address longstanding challenges such as procedural heterogeneity, limited case volumes, and fragmented datasets. These issues often complicate trial design, comparative studies, and multicenter data aggregation. AI can now support various stages of the research process, including automating literature reviews, streamlining clinical trial enrollment, and improving data standardization, which are all key components of high-quality IO research[147-149].

NLP models have been employed to identify research gaps, synthesize large datasets, and generate structured evidence summaries[98,147]. These tools accelerate manuscript drafting, citation management, and systematic reviews, enabling IO investigators to focus on higher-order tasks such as hypothesis generation and study design[98,147]. AI tools can also parse structured and unstructured EMR data to assess trial eligibility, improving recruitment efficiency and reducing manual workload. LLMs such as OncoLLM have been applied to identify candidates, improve trial matching accuracy, and reduce manual screening time[150]. By streamlining eligibility screening, this approach expedites treatment delivery. Chatbot-based systems have even achieved consent rates comparable to human-led processes, reducing administrative burden and allowing research teams to focus on protocol development and patient interaction[134]. While LLM-based trial matching tools have demonstrated improved efficiency, they are not yet fully autonomous or clinically validated for unsupervised use. Human oversight remains essential to ensure eligibility accuracy and prevent misclassification, particularly in complex or rare clinical scenarios.

Ontology-based tools further enhance research efficiency by extracting key information from scientific articles and linking findings to standardized medical ontologies. This improves reproducibility, which is particularly valuable in IO, where procedural variability complicates cross-study comparisons and limited case volumes pose challenges to traditional research methods. AI-enabled platforms now support meta-analyses across institutions and standardize outcome reporting, such as ablation margin assessment using deformable image registration and radiomics. Tools like those evaluated by Boeken[151] have demonstrated improved precision in margin evaluation and the potential to reduce local recurrence.

Federated learning (FL) approaches, which allow multi-institutional collaboration without sharing raw data, are increasingly used to overcome data privacy concerns and limited case volumes in IO[152]. Nasajpour et al[153] provide a comprehensive review of FL’s role in oncology, highlighting its potential to improve generalizability and model robustness across diverse datasets.

By automating labor-intensive components of the research process and enabling cross-institutional collaboration, AI is accelerating innovation and improving scientific rigor in IO research, expediting hypothesis testing and study implementation. These tools facilitate systematic reviews and comparative analyses by automating repetitive tasks and improving the consistency of evidence synthesis[147]. AI also supports cross-institutional meta-analyses and standardizes outcome reporting, such as ablation margin assessment, thereby advancing reproducibility and regulatory validation. These tools support a broader transition toward adaptive, patient-centered cancer care (Table 2).

Table 2 Current and potential patient-focused artificial intelligence applications and validation gaps in interventional oncology.

Application	Description	Highest level of clinical evidence currently available	Most significant validation gap	Ref.
Patient selection & response prediction	Analyzes clinical data to predict therapeutic response, enabling patient-specific treatment recommendations while accounting for potential risks	Retrospective studies, simulation platforms, some prospective modeling	Lack of prospective trials and integration into real-time clinical decision-making	[41,54,55,60-63,67,111,115]
Patient triage	Prioritizes patients for urgent interventional oncology procedures by evaluating imaging and clinical indicators of disease severity	Retrospective validation in diagnostic settings	Limited IO-specific validation and real-time deployment in procedural prioritization	[128]
Personalized post-procedural follow-up and long-term management	Integrates clinical, procedural, and imaging data to customize follow-up schedules and management plans, improving patient outcomes over time	Retrospective studies, mobile health feasibility trials	Lack of prospective trials with integration with EMRs and imaging systems	[23,111-113,115,116,129,130]
Improved patient experience	Delivers personalized education, procedural tours, and recovery support to foster a more comfortable and informed perioperative experience	Feasibility studies and pilot implementations, RCT	Limited usability testing and lack of standardized patient outcome metrics	[129,133,134]
Comprehension enhancement	Provides clear, individualized instructions, interactive virtual tours, and tailored recovery feedback to improve patient understanding and adherence	Systematic reviews and NLP-based readability studies	Lack of prospective validation and integration into clinical education workflows	[129,131-134]

IO: Interventional oncology; EMR: Electronic medical record; RCT: Randomized controlled trial; NLP: Natural language processing.

LIMITATIONS

Despite the growing promise of AI in IO, several critical limitations hinder its widespread clinical adoption. These challenges span technical, ethical, infrastructural, and regulatory domains. Key barriers include algorithmic bias, limited generalizability, lack of model interpretability, data privacy concerns, medico-legal ambiguity, and integration difficulties within clinical workflows. Addressing these limitations is essential to ensure safe, equitable, and effective deployment of AI in IO.

Data limitations

The performance and reliability of AI models are intrinsically linked to the quality, diversity, and representativeness of their training datasets[152,154,155]. Many current models are trained on curated or institution-specific datasets that often lack demographic diversity, leading to biased outputs and reduced generalizability[154]. The underrepresentation of minority populations, socioeconomically disadvantaged groups, and rare disease phenotypes is particularly concerning, as it can exacerbate health disparities when such models are deployed in clinical practice[156,157]. For instance, Larrazabal et al[158] demonstrated that gender-imbalanced datasets significantly impaired model performance, underscoring the need for inclusive data curation.

Additionally, variability in imaging protocols, annotation standards, and procedural techniques across institutions further limits external validity[148]. Models validated in controlled research settings may underperform in real-world clinical environments. These challenges are particularly pronounced in IO, where the field’s relative novelty and procedural variability limit the availability of large, annotated datasets. As dataset size and diversity increase, the statistical power of AI models improves, helping to mitigate the effects of data inaccuracies, sampling bias, and demographic underrepresentation. This, in turn, supports more equitable, generalizable, and reliable model performance.

Cost & accessibility

The widespread adoption of AI in IO has been slower than expected, largely due to financial and infrastructural barriers. Implementing AI technologies in IO often requires substantial financial investment, including acquiring specialized hardware, proprietary software, and integration with existing hospital information systems[10,136]. Additional hidden costs such as workflow disruption, clinician training, and medico-legal liability further complicate adoption. These costs pose significant barriers, particularly for low- and middle-income countries and smaller community hospitals with limited information technology (IT) infrastructure and technical support[159-161]. Even in well-resourced institutions, AI tools are often not fully integrated into interventional suites, which limits their utility during procedures.

Regulatory ambiguity compounds these challenges. The absence of formal FDA classification complicates billing and reimbursement[162], as many AI tools lack dedicated Current Procedural Terminology codes and receive limited payer recognition[163-165]. This undermines financial viability and discourages institutional invest in AI platforms[166,167]. Payer skepticism also persists, with insurers requiring strong clinical evidence before approving reimbursement for AI-assisted procedures[168]. However, generating such evidence is itself resource-intensive, and the high costs of implementation can limit access to the very tools needed for prospective validation. The interdependency between adoption and validation forms a cycle that delays clinical integration, as limited adoption slows evidence generation and limited evidence restricts adoption.

Additionally, many models are developed using data from academic centers, which may not reflect the operational realities of non-academic settings, and may not translate well without significant adaptation[154,158]. This mismatch often necessitates extensive customization and validation, further delaying implementation and escalating costs. While broader oncology-focused analyses suggest that AI may improve outcomes and reduce long-term costs, no formal cost-benefit analyses specific to IO have been published to date[169]. This gap complicates institutional decision-making, particularly in resource-constrained settings where financial justification is essential for adoption and long-term sustainability.

Model explainability (‘black box’ problem)

A major barrier to clinical trust and adoption is the opacity of many AI models, particularly DL architectures[170]. These systems often function as ‘black boxes’, generating outputs without explainable reasoning for how decisions were made. This lack of explainability, distinct from interpretability and transparency, can restrict their clinical utility[170-172]. In procedural contexts where decisions carry high clinical risk, clinicians must be able to understand and evaluate the rationale behind AI-generated recommendations[173]. As Abgrall et al[174] emphasize, although interpretability, transparency, and explainability are related, the terms are often conflated. Interpretability refers to the degree to which the internal mechanics can be understood, transparency reflects openness about its design, data sources, training process, and limitations. Explainability, on the other hand, addresses why specific outputs are produced, typically through post hoc methods, in a way that is meaningful to end users.

This lack of explainability complicates clinical decision-making and makes it difficult to evaluate model performance, especially when AI outputs conflict with established protocols or personal clinical judgment[11,120]. Without insight into how a model arrives at its conclusions, clinicians may struggle to determine when to rely on its output or override it, limiting AI’s safe and effective use in procedural decision-making[120]. It also complicates regulatory approval, medico-legal accountability, and integration into clinical workflows. Although these related concerns are addressed in subsequent sections, explainability remains a prerequisite for addressing each. Without interpretable outputs, AI systems cannot be critically assessed or responsibly deployed in IO[175].

Privacy, data ownership, & confidentiality

The use of patient data to train and deploy AI systems raises complex ethical and legal concerns. Current regulatory frameworks often lack clear definitions regarding data ownership, consent, and accountability[176]. While compliance with the Health Insurance Portability and Accountability Act (HIPAA) is mandatory in the United States[10], the need for large, diverse datasets inherently increases the risk of data breaches and misuse. Nonetheless, effective AI training requires access to shared patient data, inherently creating tension between model development and privacy protection[177]. Transparent governance models, robust de-identification protocols, and secure data-sharing infrastructures are critical to maintaining public trust and ethical integrity in AI development.

Medico-legal liability

The integration of AI into clinical decision-making introduces unresolved medico-legal questions. Traditional negligence frameworks do not adequately address liability in cases where AI contributes to adverse outcomes. Uncertainty persists regarding who bears responsibility when AI-generated recommendations result in harm - the clinician, the institution, or the software developer[11]. Though regulatory frameworks are evolving to address AI used as medical tools, Drabiak et al[12] explain that the United States FDA classifies many AI tools as ‘software as a medical device’, but regulatory guidance on liability, especially in dynamic procedural environments like IO, remains limited[11].

HIPAA compliance adds another layer of complexity when protected health information is accessed by AI developers or third-party vendors[175,178,179]. In the absence of clear legal precedents and regulatory standards, medico-legal ambiguity remains a significant barrier to AI adoption in clinical practice. To address these challenges, shared liability models and clinical oversight frameworks are actively being explored. These include multi-level governance structures that distribute accountability among developers, clinicians, and institutions, as well as proposals for localized monitoring and certification to ensure safe and equitable use[180,181].

Integration into clinical workflow

Successfully deploying AI in live IO practice requires overcoming significant workflow and infrastructure hurdles. Most current imaging and hospital IT systems are not designed for real-time AI analytics. When imaging data must be routed through external servers, latency and operational complexity can arise, potentially disrupting procedural flow and diminishing clinical utility[182]. Clinician acceptance and usability are other critical factors. Interventional teams are accustomed to established workflows, and the introduction of AI tools inevitably introduce learning curves and workflow adjustments.

Poorly designed systems can fragment attention between imaging, devices, and AI dashboards. This divided focus in conjunction with multiple outputs can risk increasing cognitive load and may lead to decision fatigue, especially during high-stakes or prolonged cases[183]. While AI might be seen as a potential solution to reduce this burden by adding features to automate certain tasks or offering decision support, it is not a substitute for cognitive effort and clinical cognition. Overreliance on AI recommendations, known as automation bias, may lead clinicians to defer to algorithmic recommendations and undervalue their own judgment, especially when the systems lacks transparency or fails to communicate confidence levels effectively[184]. In such scenarios, AI may inadvertently heighten decision uncertainty and further intensify cognitive burden rather than alleviate them.

Ergonomics shortcomings also limit usability. Displays positioned outside the natural line of sight, hardware obstructing C-arm movement, or excess cable clutter in high-traffic areas can compromise efficiency and safety[185]. For AI to be effective in IO, outputs must be delivered in near real-time, remain easily interpretable, and integrate smoothly with existing platforms such as PACS, RIS, and EMRs because even modest computational delays can negate the benefit of guidance[186]. Systems co-designed with end-users, tested iteratively in simulated procedural environments, and embedded seamlessly within the ergonomics of the interventional suite are most likely to succeed[187].

Real-world deployment: A convergence of challenges

Overall, the limitations outlined above highlight the multifaceted challenges of deploying AI in real-world IO settings. Models that perform well in controlled environments may falter in clinical practice due to hidden stratification. This is a phenomenon in which models excel on common cases but underperform on rare or atypical presentations, which can lead to clinically significant errors and complications[188]. This is particularly concerning in IO, where anatomical and pathological variability is high.

Poor generalization to novel clinical environments is frequently rooted in limited and biased training datasets[29]. As IO remains an emerging subspecialty, the scarcity of large, annotated datasets limits robust model development. Moreover, institutional reluctance to share data, sometimes termed ‘institutional xenophobia’, further impedes collaborative model progress[177,189]. These challenges, combined with the high upfront and maintenance costs of AI systems, contribute to slow and uneven adoption across healthcare settings.

Incorporating AI into clinical settings requires substantial upfront investment, including costs for purchasing the AI software, upgrading hardware to host it, and training healthcare professionals on new AI models. Ongoing expenses such as maintenance, the workforce support, and any subscription fees compound the financial burden[189,190]. These costs, coupled with an uncertain return on investment, can slow adoption, especially in resource-constrained settings[190]. Collaboration involving patient data must balance ethical considerations, financial implications, and the continuously evolving nature of AI technologies[176].

FUTURE DIRECTIONS

While current limitations in AI applications for IO, including data heterogeneity, limited generalizability, and workflow integration challenges, pose significant barriers to clinical adoption, they also define clear priorities for future research and development. Addressing these gaps will require shifting from retrospective, proof-of-concept studies to prospective, clinically embedded, and system-integrated solutions. As AI continues to evolve, its successful integration into IO will depend on addressing key translational challenges. These include improving model generalizability, ensuring clinical interoperability, validating performance in prospective trials, and aligning AI tools with real-world procedural workflows. The following directions outline strategic areas for advancing AI to meet the specific procedural, operational, and patient-centered needs of IO (Figure 2).

Figure 2 Summary of opportunities and challenges in artificial intelligence adoption within interventional oncology. AI: Artificial intelligence; IO: Interventional oncology.

Foundational infrastructure

Data infrastructure and generalizability: The development of robust, generalizable AI models depends on access to large, diverse, and well-annotated datasets. Current limitations in dataset size, demographic representation, and institutional variability hinder model performance and compromise equity. To address these challenges, future efforts must prioritize the development of multi-institutional, demographically inclusive datasets that reflect real-world clinical diversity. FL frameworks offer a promising solution by enabling collaborative model training across institutions while preserving data privacy[152,153,191,192]. Standardization will also be essential to ensure reproducibility and facilitate regulatory approval. This includes harmonizing imaging protocols, adopting structured reporting, and establishing consensus-based annotation guidelines[158,193]. The American College of Radiology Data Science Institute is actively working to improve the representativeness of datasets used by AI algorithms in radiology to support safe and effective implementation in clinical care for diverse clinical populations[194]. Expanding initiatives similar to those endorsed by the American College of Radiology Data Science Institute can help address the current data limitations in AI algorithms for IO, ensuring equitable tools across diverse clinical populations.

Prospective validation and clinical trials: Most current studies are retrospective or pre-clinical and technically focused, which limits their clinical applicability. Future research should prioritize prospective, multicenter trials that evaluate AI tools in real-world IO settings. Trial designs should compare AI-assisted approaches with standard-of-care methods across key domains, including ablation planning, catheter navigation, and post-procedural assessment. In addition to technical performance, longitudinal studies are needed to assess the impact of AI on clinically meaningful outcomes, such as recurrence rates, overall survival, and patient-reported quality of life. Establishing standardized trial designs and outcome metrics - building on the foundational standardization of imaging and data protocols - is essential for securing regulatory approval, facilitating clinical adoption, and enabling reimbursement. Future research should align with evolving regulatory frameworks, such as those established by the FDA and other international regulatory bodies, to facilitate clinical adoption and ensure safe, effective deployment of AI tools in IO.

Reliability and interoperability: For AI to be adopted in routine IO practice, it must be seamlessly integrated into existing clinical infrastructure, including PACS, RIS, electronic health records, and interventional imaging systems. Outputs must be interpretable, actionable, and available instantly. This requires incorporating explainability features such as saliency maps, uncertainty scores that flag low-confidence predictions, feature attribution, counterfactual simulations and rule-based logic that contextualize outputs against known procedural protocols to support clinical decision-making[11,120,171,195]. Similar to how radiation therapy planning software adjusts dose maps based on tumor geometry, counterfactuals in IO can simulate alternative probe placements or embolization paths simulate how small changes in lesion size or location alter AI recommendations[196]. In addition to explainability and interpretability, transparency regarding model provenance, training data, and known limitations is essential to ensure regulatory compliance and clinician trust.

Table 3 outlines the features necessary to maintain reliability and situational awareness during live procedures. These forms of explainability must be delivered in real time, integrated into procedural consoles, and presented in intuitive formats that allow clinicians to validate, adjust, or reject AI outputs as needed[197,198]. In the high-stakes environment of IO, interoperability, cybersecurity, and intuitive user interface design will be essential to ensure safe and effective deployment. However, whether these features meaningfully enhance clinician trust remains an open question[199,200]. As Dietrich and Patlas[201] emphasize, addressing adversarial AI threats requires strong cybersecurity measures to preserve model integrity and maintain clinician trust[201]. Prioritizing these elements is central to safe, reliable, and clinically acceptable deployment in IO.

Table 3 Human-centered artificial intelligence features for enhancing clinician trust and ensuring safe deployment.

Feature	Purpose	Example	Deployment considerations
Feature attribution	Identifies the imaging or clinical features that most influenced the AI’s recommendation	In ablation planning, feature attribution can highlight lesion boundaries, proximity to critical structures, or perfusion metrics that guided probe placement	Should be integrated into procedural consoles with toggleable overlays for real-time validation
Uncertainty quantification	Provides confidence scores or probability distributions to help clinicians assess risk and determine whether to rely on or override the output	During catheter navigation, an AI system might suggest a path with 92% confidence, giving the proceduralist a quantifiable basis for trust	Must be displayed in plain language (e.g., “low confidence”) and updated dynamically during the procedure
Saliency maps or visual overlays	Highlight relevant anatomical regions on live imaging by overlying AI-derived insights (e.g., tumor margins, vessel segmentation) to support real-time targeting	Enhances targeting precision in ultrasound- or CT-guided procedures by showing which regions the AI model considers most relevant	Requires seamless integration with imaging feeds and adjustable settings
Counterfactual examples	Illustrate how small changes in input (e.g., lesion size or location) would alter the AI’s recommendation, helping assess model robustness	Could be used pre-procedurally to simulate alternative probe placements or embolization strategies	Should be available pre-procedurally for simulation and intra-procedurally for real-time adjustment

AI: Artificial intelligence; CT: Computed tomography.

Improving usability and ergonomics: AI systems must be designed with human-centered principles that prioritize clarity, control, and contextual relevance aligned with the cognitive and physical demands of IO practice[202]. Interfaces should minimize distraction and support rapid decision-making. Context-sensitive alerting systems can reduce alarm fatigue by distinguishing background processes from time-critical events. Tiered alerting mechanisms, where low-priority features operate silently and high-salience cues (i.e., graphics and/or sounds) reserved for urgent actions, may improve situational awareness without overwhelming the operator[203,204]. Outputs must be explainable at a glance; calibrated, plain-language labels such as ‘low confidence - verify vessel origin’ are more effective than dense statistical metrics that require prolonged interpretation, especially under time pressure[205,206]. Interfaces should include clear physical override options and quick ‘AI off’ toggles to preserve clinician control and confidence.

Displays should be positioned within the operator’s natural line of sight, ideally adjacent to primary imaging monitors, and be adaptable to low-light environments. Hardware mounts must avoid obstructing imaging equipment, and cable management should be optimized to reduce clutter in high-traffic areas[185]. Hands-free interaction methods, such as voice control or foot pedals, can further enhance procedural efficiency while maintaining sterility[207]. Overlays should never obstruct key imaging views or require extra manual steps to reconcile with existing guidance systems. Functionality should be linked to specific case phases, such as planning, access, navigation, therapy, verification, documentation - surfacing only the relevant AI functions needed at each stage. Progressive disclosure techniques, in which simple recommendations are presented first, with detailed information available on demand, can further support usability.

To ensure safety and reliability, systems should undergo rigorous validation in simulation labs that replicate fluoroscopy conditions and procedural time constraints. Structured micro-drills can train clinicians on voice commands, override mechanisms, and uncertainty interpretation, while structured checklists can guide readiness assessment, including latency benchmarking with live feedback, overlay stability testing under typical motion, alert tier reviews to ensure relevance, and configuration of sterile, hands-free interaction pathways[44]. Override mechanisms and mode indicators should be tested for reliability and ease of access. Following deployment, continuous monitoring of both procedural metrics and human factors is essential to ensure that AI tools enhance safety and efficiency without introducing unintended risks. Impact should be assessed through performance indicators, such as procedure dose, alert rate, and override frequency, as well as usability metrics, including gaze metrics per minute, time with overlays enabled, and error recovery time. Successful and safe integration must address human factors in addition to technical and software compatibility, with attention to ergonomic design and cognitive workflow alignment to ensure systems meet clinician needs and domain requirements[202].

Translating research into clinical practice: As AI applications in IO mature, translating validated tools into real-time clinical workflows becomes a critical next step. Promising applications include real-time needle trajectory optimization in CT-guided lung biopsies[21], dynamic ultrasound fusion during liver ablation[81], and automated vessel segmentation to support embolization procedures[208]. Integration of such tools into angiography consoles and robotic platforms may reduce operator variability and improve procedural safety[209,210]. Additionally, pairing AI with advanced imaging modalities such as probe-based confocal laser endomicroscopy, hyperspectral imaging, and optical coherence tomography could enable immediate, non-invasive tissue characterization[211]. Applying implementation science frameworks may facilitate the effective and safe integration of AI in routine clinical practice by identifying barriers, optimizing workflows, and supporting sustainable adoption. These translational efforts are essential for bridging the gap between innovation and implementation in IO.

Clinical intelligence & personalization

Outcome prediction and risk stratification: AI-based outcome prediction represents a high-impact yet underutilized opportunity in IO. Recent models have demonstrated improved prognostic accuracy in treatment response and ER following TACE, particularly when radiomic features are combined with clinical and laboratory data[54,62,64]. Comparable advances have been reported in ablation and Yttrium-90 radioembolization, where AI-driven simulations of ablation zones enhance individualized planning[41]. Voxel-based dosimetry, while not AI-based, provides high-resolution mapping of radiation dose distribution across tissue volumes[212]. This precision modeling could complement AI approaches by providing detailed spatial data for training and validation, ultimately supporting more personalized strategies. These tools lay the groundwork for more precise patient selection, procedural strategies, and follow-up protocols.

While technical advances in outcome prediction are accelerating, their clinical utility will depend on how effectively they can be integrated into real-time decision-making. Tools that stratify recurrence risk or predict treatment response must be designed to dynamically adapt to evolving patient data and procedural variables. Digital twin technologies, which are essentially virtual representations of individual patients, may further enhance this adaptability by simulating patient-specific procedural outcomes and informing personalized care strategies[213]. Embedding such tools into IO workflows will be essential for realizing their full potential in precision oncology.

Personalized treatment planning: Alongside prognostic tools, the refinement of real-time, personalized procedural planning remains an important area of ongoing development. In IO, this includes tailoring procedural parameters such as embolic dose, particle size, infusion rate, and delivery volume, based on tumor vascularity, perfusion imaging, and patient-specific hemodynamics. Approaches that integrate imaging, clinical, and molecular data are already being applied in other domains to inform patient-specific strategies and hold promise for dynamic intra-procedural adaptation.

Radiogenomics, which links imaging features with underlying tumor biology, offers a promising avenue for therapy selection and risk stratification[214]. Models trained on such data may assist in identifying patients who are more likely to benefit from specific interventions or targeted therapies. In parallel, IO is expected to see an increase in the use of targeted therapies that disrupt aberrant oncogenic pathways while minimizing collateral damage to healthy tissue[215].

To realize this vision, future work should emphasize the integration of multimodal data - including imaging, genomics, and treatment history - into clinically usable AI frameworks that can guide dose and delivery strategies in real time. Digital twin models, though still in early development, may eventually support procedural planning by simulating treatment response and toxicity based on patient-specific variables[213]. These models must be interpretable, seamlessly integrated into clinical workflows, and aligned with existing decision-making pathways. Such capabilities will be essential to translating personalized treatment planning from conceptual innovation to routine clinical practice in IO.

Enhancing procedural precision, scalability & training

Quantitative margin assessment: Initial models required manual annotation, but current systems now achieve Dice scores exceeding 90% with automated pipelines[7,8]. Establishing standardized evaluation metrics for AI-assisted ablation margin assessment should be a near-term priority to enhance reproducibility and clinical reliability. While accurate assessment of ablation margins is essential for achieving durable oncologic control, future models must better reflect real-world procedural dynamics, such as overlapping thermal diffusion from multiple devices that can distort margin interpretation. These complexities are only beginning to be addressed in simulation platforms, such as those employing GPU-accelerated bioheat modeling, which offer a more realistic representation of thermal interactions and may help reduce incomplete treatments[48].

AI-based deformable image registration and fusion imaging techniques have shown early promise in quantifying ablation margins and detecting residual disease[110]. These tools have the potential to standardize margin assessment, reduce inter-operator variability, and enable same-session re-intervention. However, prospective validation is needed to determine their impact on technical success rates, recurrence, and long-term outcomes. Future efforts should also prioritize integrating these tools into intraoperative workflows and assessing their performance across a range of tumor types and anatomical locations to ensure broad clinical applicability.

Device tracking and navigation: AI-enhanced tracking of interventional tools such as catheters, guidewires, and needles in fluoroscopy and cone-beam CT has the potential to improve procedural efficiency, targeting accuracy, and radiation safety. DL models capable of identifying device tips and trajectories with high spatial accuracy have been developed[79], but their integration into clinical workflows remains limited. Future work should focus on real-time deployment, incorporating operator-in-the-loop validation, and ensuring compatibility with diverse imaging platforms and procedural contexts.

Simulation and training applications: Next-generation procedural platforms that combine AI with AR and robotics are poised to transform procedural education and execution. Patient-facing tools (e.g., educational platforms, recovery monitoring) are still in early development and require usability testing. For clinicians, AR-guided interventions have been shown to reduce needle passes, procedure time, and radiation exposure in preclinical models[209]. These systems overlay segmented anatomical structures onto the clinician’s visual field, enhancing spatial orientation, depth perception, and procedural confidence[216]. When integrated with AI-driven segmentation and tumor detection, AR platforms may further improve targeting precision and reduce variability.

AI-powered simulation environments also offer opportunities for procedural rehearsal, skill acquisition, and competency assessment. These platforms can analyze procedural video or simulated performance to identify technical errors, provide individualized feedback, and track skill progression. Their incorporation into IO training programs may accelerate learning curves, support standardized evaluation or competency assessment, and improve procedural readiness, particularly for complex or high-risk interventions.

To prepare future clinicians for AI-integrated practice, training curricula should be expanded to include core principles of data science, ML, and radiomics, along with AI literacy to equip physicians for evolving clinical environments[217,218]. Instruction should also cover model interpretability, bias mitigation, and ethical AI use[219,220]. Trainees must be competent in HIPAA compliance, data stewardship, and medico-legal implications of AI deployment[221]. The Association of American Medical Colleges underscores the importance of preparing learners for the use of AI in the delivery of high-quality care and ensuring educators and staff are appropriately prepared to teach and facilitate learning of AI-enabled, patient-centered care[222]. Case-based learning can be used to illustrate real-world challenges such as algorithmic bias, liability, and decision-making in AI-assisted procedures[223]. Introducing these concepts early in training may improve clinicians’ ability to critically evaluate AI outputs and recognize when human judgment should override automated recommendations[224].

STRATEGIC INTEGRATION AND INSTITUTIONAL IMPERATIVES

As AI technologies mature, hospitals have the opportunity to lead in precision medicine by investing in AI-integrated IO platforms. This section outlines key considerations for hospital administrators, clinical leaders, and strategic planners. While the long-term benefits are compelling, they must be weighed against real-world barriers such as upfront costs, ongoing maintenance requirements, and regulatory complexities (Table 4).

Table 4 Current business case for hospital investment in artificial intelligence-interventional oncology platforms.

Strengths of AI	Supporting claim
Clinical efficacy & cost savings	Reduced procedure time
	Fewer complications with better precision
	Optimized resource use
	Cost savings
Improved outcomes & quality metrics	Higher technical success rates
	Personalized treatment
	Better documentation
Strategic positioning & innovation	Competitive differentiation
	Academic and research value
	Patient experience
Scalability & long-term ROI	Data-driven learning systems
	Alignment with value-based care models

AI: Artificial intelligence; ROI: Return of investment.

The current business case for hospital investment in AI-IO platforms hinges on their potential to improve procedural accuracy, reduce complications, and streamline workflows, all of which can translate into long-term cost savings and better patient outcomes. However, adoption remains constrained by several socioeconomic barriers. Reimbursement is a major hurdle: Many AI tools lack Current Procedural Terminology codes and payer recognition, making return on investment difficult to quantify. Regulatory clearance also varies - CE mark approval in Europe is generally faster and less costly than FDA clearance in the United States, which requires more extensive validation and can delay deployment. Total cost of ownership includes not only upfront expenses for hardware and software, but also integration with hospital IT systems, clinician training, and continuous model monitoring[10]. These factors collectively impact scalability, especially in resource-constrained settings. Institutions must weigh these challenges against the strategic advantages of early adoption, including improved care delivery, research leadership, and alignment with value-based care models.

Synthesizing value: Rationale & pathways for hospital investment in AI-IO platforms

The integration of AI into IO offers value on multiple levels, combining technological innovation and healthcare transformation. While the clinical utility of AI in enhancing procedural accuracy and consistency is increasingly supported by empirical evidence, its broader institutional value lies in its capacity to advance precision medicine, optimize resource utilization, and align with evolving care delivery models. From a clinical standpoint, AI facilitates more individualized treatment planning and standardized assessment of ablation margins, contributing to improved long-term outcomes. AI-assisted planning and navigation can shorten procedure duration, lowering operating room utilization and associated costs, while enhancing patient throughput. Improved targeting and real-time monitoring reduce variability and adverse events, potentially minimizing downstream costs with extended hospital stays or readmissions. Operational efficiencies arise from automation of imaging tasks such as segmentation and fusion, which streamline procedural workflows and reduce manual burden. These improvements extend to documentation practices, where structured reporting enhances coding accuracy and regulatory compliance. When integrated with scheduling and triage systems, AI can support dynamic resource allocation, enabling more efficient use of staff, equipment, and procedural suites.

Strategically, early adoption places institutions at the forefront as leaders in precision medicine. This leadership may confer advantages in attracting high-caliber clinical and research talent, securing competitive funding, and increasing patient referrals. AI platforms facilitate clinical trials, data mining, and publication opportunities, enhancing institutional visibility and scholarly output. AI-driven education and recovery tools improve patient comprehension, engagement, and satisfaction, strengthening institutional reputation. Scalable AI platforms align with evolving value-based care and reimbursement models by enabling outcome tracking, risk stratification, and personalized follow-up. Predictive models can help select optimal therapies by predicting treatment response and guiding treatment selection, improving long-term outcomes and reducing ineffective procedures over time. As AI platforms continuously improve with increased data input, their performance and value enhance over time, creating sustained efficiency gains and long-term cost savings.

A path forward

Given the accelerating pace of AI development and its transformative potential in IO, leading medical institutions must move even quicker to shape its integration responsibly. Rather than awaiting regulatory mandates, hospitals should take an active role in establishing institutional oversight to ensure safe, ethical, and scalable deployment. This includes forming multidisciplinary AI governance boards comprising clinicians, data scientists, ethicists, and legal advisors to oversee model validation and performance monitoring, and retraining protocols. Designating clinical champions to guide implementation and maintain feedback loops will further enhance institutional readiness and reduce operational risk. By acting proactively, hospitals can set the standard for AI adoption in IO at their own pace, reinforcing their role as innovators in precision medicine. Together, these strategic, clinical, and operational benefits underscore the broader themes discussed throughout this review, namely, that AI is not just a technological advancement, but a transformative force in the future of IO.

CONCLUSION

AI is reshaping IO, with validated applications spanning pre-procedural planning, intra-procedural guidance, and post-procedural evaluation (Table 4). These technologies enhance lesion detection, procedural precision, treatment personalization, and workflow efficiency, while also shifting from isolated, task-specific tools to integrated, data-informed systems that unify the procedural continuum. By allowing data generated before, during, and after a procedure to inform one another, AI supports adaptive learning systems that continuously refine decision-making and clinical outcomes. Figure 3 visually depicts this intersection, highlighting AI’s potential to create a closed-loop framework for IO.

Figure 3 Performance of artificial intelligence models across the interventional oncology workflow. AI: Artificial intelligence; NLP: Natural language processing.

Widespread clinical adoption remains limited by challenges including model explainability, data heterogeneity, and workflow integration. Overcoming these barriers will require more than technical refinement - it will demand robust infrastructure, cross-institutional data sharing, prospective validation frameworks, integration with existing non-AI-based frameworks, and formalized pathways for clinician involvement in algorithm development. Ultimately, the success of AI in IO will depend not on algorithmic sophistication alone, but on the ability of clinicians to thoughtfully and ethically integrate these tools into practice.

As AI begins to bridge fragmented stages of care into adaptive, data-driven systems, interventional oncologists must pair procedural expertise with digital fluency. The convergence of these domains will define the future of IO, driving more personalized, efficient, and outcome-focused care (Table 5). Achieving this vision requires a broader perspective on how AI integration affects future clinicians and patients. Hospitals should invest in AI to attract top talent, train the next generation, and differentiate themselves in oncology care. In practical terms, this requires updating training programs to include cover concepts on data science, informatics, and algorithm oversight, and developing governance structures that embed clinician input into model deployment and monitoring. To realize this vision, the role of the interventional oncologist must evolve. As AI becomes integral to procedural medicine, interventional oncologists must assume a leadership role in collaborative, interdisciplinary efforts to ensure that adaptive AI systems improve patient outcomes safely and equitably. Institutions, in turn, should establish multidisciplinary oversight committees and implement continuous monitoring with protocols for retraining and auditing AI models. By embracing these changes, IO can lead the way in integrating AI into procedural medicine, setting new standards for precision, safety, and patient-centered care.

Table 5 Perioperative artificial intelligence applications in interventional oncology and barriers to clinical adoption.

Application	Description	Highest level of clinical evidence currently available	Most significant validation gap	Ref.
Lesion segmentation1	Improves accuracy in delineating tumors for precise targeting	Retrospective studies, phantom trials	Limited prospective validation and generalizability across modalities and institutions	[7,8,20,21,34,44]
Procedural path planning1	Generates patient-specific needle or probe trajectories that, reducing preparation time and improving procedural accuracy	Retrospective studies, phantom trial	Fails to integrate real-time procedural variables and thermal interactions, especially in multi-needle procedures	[21,22,41-43,45,46]
Radiomics integration1	Incorporates radiomic features into planning to predict tumor characteristics and genetic profiles, enabling personalized treatment strategies	Retrospective studies	Limited prospective validation, lack of standardized radiomic pipelines, and poor reproducibility across institutions and imaging platforms	[32,54,59,60,63]
Catheter planning1	Analyzes vascular anatomy and perfusion patterns to provide individualized catheter placement recommendations, improving efficiency and accuracy of transarterial therapies	Retrospective studies, simulation models	Insufficient real-time validation and integration with hemodynamic data	[31,49-55,89,90]
Personalized treatment1 planning	Uses imaging and clinical data to tailor treatments and avoid unnecessary procedures. Digital twin simulations model patient-specific procedural outcomes, aiding in decision-making	Retrospective studies, simulation models	Lack of prospective trials and real-time clinical deployment	[41,61-63,67,69,213]
Imaging analysis2	Enhances image fusion to overlay of intra- and pre-procedural imaging in real time, improving precise lesion localization	Retrospective studies, phantom trials	Latency and lack of seamless fusion across modalities	[70,71,74-77,81]
Needle tracking2	Provides real-time needle localization and trajectory prediction, reducing procedure time and improving first-attempt success rates	Retrospective studies, phantom trials	Limited clinical validation and integration with robotic systems	[21,44,78-80]
Motion correction2	Maintains spatial alignment and alerts to tool deviation, enhancing procedural safety and efficiency	Retrospective studies, phantom trials, simulation models	Lack of real-time deployment and anatomical variability handling	[21,70,71,81]
Safety monitoring2	Detects intra-procedural risks, such as hemorrhage, vascular injury, or thermal injury, alerting clinicians in real time	Retrospective studies, preclinical models	Limited IO-specific validation and standardization of margin assessment	[82-84]
Treatment delivery & dosing optimization2	Optimizes dosing, dose mapping, and targeted therapy delivery using real-time imaging features to improve safety and precision	Feasibility trials in systemic therapy	Lack of IO-specific prospective trials and adaptive dosing platforms	[69,86,88,90,91]
Quality assurance3	Evaluates documentation and ablation margins to ensure procedural consistency	Retrospective studies	Limited prospective validation and standardization of margin assessment	[95,105-108,151]
Retrospective trajectory analysis3	Simulates alternative procedural approaches using image navigation and fusion, accounting for anatomical constraints	Retrospective studies, phantom trials	Lack of integration into intraoperative workflows	[109,110]
Treatment outcome prediction3	Predicts survival, recurrence risk, treatment outcomes, and complications, enabling proactive risk mitigation and individualized adjustments for future procedures	Retrospective studies, systematic reviews	Need for prospective validation and integration into decision-making	[62,111,113,117]
Response monitoring3	Evaluates lesion response and/or recurrence following treatment through imaging features and radiomics	Retrospective studies	Limited real-time deployment and standardization of response metrics	[23,112,114,115]
Longitudinal lesion tracking3	Tracks lesions AI across serial imaging for accurate identification and consistent follow-up guidance	Retrospective studies, algorithm benchmarking	Limited clinical integration and validation across imaging platforms	[116]

¹Pre-procedural.

²Intra-procedural.

³Post-procedural.

IO: Interventional oncology; AI: Artificial intelligence.