In surgical oncology, clinical translation denotes the pathway by which AI and ML methods move from retrospective proof-of-concept toward deployment as workflow-integrated tools that measurably improve patient care under routine conditions.^5,13 This translational interface is dominated by clinical decision support (CDS), defined here as computer-based systems that provide patient-specific assessments or recommendations intended to aid clinical decision-making.^{14, 15} Decisions such as operative candidacy, extent and timing of resection, perioperative optimization, postoperative triage, and surveillance planning directly shape both short-term morbidity and long-term oncologic benefit.^{5, 16, 17} At this stage, model outputs must be not only accurate but also actionable, interpretable, and deliverable at the moment a decision is made, properties repeatedly associated with CDS effectiveness in broader health systems research and explicitly emphasized in surgical risk calculator design.^{14, 18}

Contemporary perioperative decision-making is anchored by traditional risk stratification tools (eg, comorbidity indices and general surgical risk calculators), yet these instruments can be imperfectly calibrated to specific oncologic procedures, patient populations, and evolving perioperative pathways.^18–21 ML-based CDS has therefore concentrated on three clinically actionable tasks: (1) preoperative prediction of postoperative morbidity, mortality, and resource use (eg, prolonged length of stay); (2) perioperative process-of-care decisions (eg, escalation of monitoring intensity or thromboprophylaxis); and (3) oncologic outcome prediction (eg, early recurrence, metastasis, survival) intended to tailor adjuvant therapy counseling and surveillance strategies.^{5, 17, 22–24}

Decision-support systems in surgical oncology are typically trained on multimodal clinical data rather than imaging alone.^{22, 25, 26} Common substrates include structured EHR variables (demographics, comorbidities, laboratory values, medications), perioperative process data (procedure type, operative duration, transfusion, enhanced recovery pathway adherence), and longitudinal physiologic measurements.^{22, 27, 28} Unstructured text has particular value because key prognostic and procedural descriptors are often documented in operative notes, pathology reports, and radiology narratives.^{25, 26} In ovarian cancer surgery, for example, natural language processing (NLP) applied to preoperative computed tomography (CT) reports improved prediction of morbidity and mortality beyond structured variables, illustrating how extraction of clinically salient detail from free text can enrich perioperative risk models without requiring new data collection.²⁶

Patient-generated health data are an emerging input modality for post-discharge decision support.²⁹ Remote telemonitoring platforms capturing symptoms and physiologic parameters have been evaluated for the prediction of postoperative complications after cancer surgery, addressing a recognized vulnerability in surgical pathways: many complications manifest after discharge, when traditional in-hospital monitoring has ended, and clinical contact is intermittent.²⁹ These studies suggest that integration of patient-reported and sensor-derived data may extend the temporal window of actionable risk detection, but they also raise implementation requirements (device adherence, data governance, and clinical response capacity) that must be incorporated into translational evaluation rather than treated as downstream operational details.

Lastly, the notion of regulatory readiness for perioperative AI now extends to include a lifecycle plan for controlled updates and performance management, not merely pre-market performance and accuracy. In the case of the United States, the Food and Drug Administration (FDA) has issued guidance for AI-enabled medical devices that supports iterative modification via the Predetermined Change Control Plan (PCCP), which addresses what changes can occur and how they can be assessed to provide reasonable assurance of safety and effectiveness. At the same time, FDA cybersecurity guidance also addresses security-by-design within the quality system and the integration of cybersecurity documents within the device submission package for those devices with cybersecurity risk. In the EU, the Medical Device Regulation (MDR 2017/745) has also placed greater responsibility on manufacturers for post-market surveillance (Articles 83–86) and for software IT security within general safety and performance requirements. In practice, we would advocate the alignment of OR-related quality and cybersecurity processes (secure health software lifecycle processes IEC 81001-5-1) and the establishment of triggers for escalation and rollback based upon statistically significant performance degradation on previously determined clinical endpoints (including calibration), sustained increases in out-of-distribution (OOD) rates and tracking and registration failures beyond control limits, safety events plausibly linked to the AI model performance, or significant cybersecurity events; mitigation should include rollback to the last known good model performance, silent mode revalidation, and auditable change history consistent with the PCCP.^{15, 29}

Methodologically, most decision-support applications remain supervised prediction models trained on retrospective cohorts, benchmarking logistic regression or regularized regression against tree-based ensembles and, less commonly, neural networks.^{22, 30–32} Several comparative evaluations suggest that increased model complexity does not guarantee improved performance.^{22, 30} For example, in a population-based analysis of mortality prediction using administrative diagnosis codes, boosted trees did not outperform logistic regression, emphasizing that the added computational and interpretability costs of complex models should be justified by demonstrable gains in calibration, discrimination, or clinical utility relevant to the intended decision.^{14, 22, 33}

A central translational premise of perioperative CDS is to identify individuals at elevated risk for adverse outcomes and to trigger targeted pathways that reduce preventable morbidity without disproportionate resource consumption.^{5, 14} In colorectal cancer surgery, Rosen et al. reported a registry-based ML model predicting 1-year mortality, deployed as a clinical decision support tool coupled to personalized perioperative care pathways; the implementation was associated with reductions in postoperative complications, improved “textbook” outcomes, shorter length of stay, and lower costs, illustrating an end-to-end translational pattern from prediction to workflow-embedded intervention and measurable system-level impact.⁵ This kind of coupling between risk prediction and standardized, protocolized responses is also consistent with broader CDS literature showing that systems are more likely to change practice when they deliver recommendations (or pathways) within routine workflow rather than standalone risk estimates.^{14, 15}

Across oncologic procedures, ML-based models have been developed to predict composite postoperative morbidity as well as specific complications that motivate discrete prophylactic or monitoring interventions.^{21, 28, 31, 33, 34} Examples include models for postoperative complications after liver surgery, early postoperative complications after radical gastrectomy, and complication risk stratification after lung cancer surgery.^{21, 31, 33} In a thoracic oncology cohort, an elastic net model predicting postoperative complications after lung cancer surgery achieved good discrimination and improved risk prediction compared with the Charlson Comorbidity Index, demonstrating that procedure-specific perioperative models can outperform general comorbidity summaries when appropriately developed and temporally validated.^{19, 21} Similarly, in head and neck cancer surgery, ML models have been used to predict prolonged length of stay and were compared against the ACS-NSQIP risk calculator and conventional statistical models, highlighting a recurring translational theme: models must be evaluated not only in isolation but also against the clinical baselines that actually shape care decisions.^{18, 20}

Risk prediction for procedure-specific events illustrates both the opportunity and the translational constraints of CDS. Anastomotic complications after esophagectomy (leak, stricture) are high-impact outcomes with potential for algorithmically guided monitoring and early intervention.^{23, 35, 36} Recent studies have developed models combining clinical variables with imaging-derived features to predict anastomotic leakage, reporting discrimination and calibration metrics, as well as threshold-based evaluation intended to support clinical decision-making.^{23, 35} In parallel, automated ML approaches have been applied to predict post-esophagectomy strictures, suggesting that algorithmic pipelines can be designed for scalability when feature extraction and model selection are automated, although automation does not obviate the need for external validation, drift monitoring, and transparent reporting.^{36, 37}

Dynamic decision support aligned to perioperative trajectories is increasingly emphasized because physiologic and laboratory changes often precede clinically recognized deterioration.^{27, 28} In a multicenter gastrectomy cohort, time-sequential ML models using serial laboratory and vital-sign measurements improved early prediction of postoperative complications compared with baseline models, supporting the concept that CDS should be designed as a longitudinal process rather than a single preoperative score.²⁸ Complementary evidence comes from dynamic modeling of postoperative venous thromboembolism risk after colorectal cancer surgery, where repeated updates to risk estimates reflected evolving clinical states and supported individualized prophylaxis and monitoring decisions in a multicenter context.²⁷

Prospective validation is particularly important for perioperative CDS because apparent performance in retrospective data can be inflated by outcome misclassification, missing data mechanisms, and site-specific practice patterns.^{13, 37, 38} Notably, prospective evaluation has begun to appear in some surgical oncology prediction studies.^{28, 34} In lung cancer surgery, Chen et al. developed and prospectively validated an explainable ML model for postoperative pulmonary complications, reporting discrimination, calibration, decision curve analysis, and feature attribution via SHapley Additive exPlanations (SHAP), an unusually comprehensive suite of translational evaluation components for a single procedure-focused model.³⁴ Together, these studies illustrate a maturation of perioperative CDS from static, internally validated scores toward longitudinal, interpretable, and prospectively assessed systems aligned with intervention thresholds and workflow integration.^{27, 34, 37}

Decision support in surgical oncology must also address endpoints that determine cancer control and long-term survivorship, particularly early recurrence and metastasis after ostensibly curative resection.^{17, 39} In esophageal cancer, Rahman et al. developed ML models to predict early recurrence after surgery using multinational data and applied internal–external validation across centers, providing an instructive template for transportability assessment when multicenter external validation is not yet feasible.³⁹ In pancreatic cancer surgery, an interpretable ML model was developed to predict early liver metastasis after resection and was externally validated with calibration and decision-curve analyses; notably, the model was also implemented as an accessible application, reflecting an explicit translation intent. These examples underscore that oncologic CDS often targets events (recurrence, metastasis) that are clinically meaningful precisely because they can change downstream actions such as adjuvant therapy selection, surveillance intensity, and enrollment into clinical trials.^{15, 39}

In glioblastoma, for instance, a quintessential conundrum in oncology, the same reasoning applies, in that despite ‘maximal safe’ resection, early failure on imaging and failure patterns at the edge of resection are frequent and directly inform intensification of treatment, trial participation, and follow-up intensity.^{40, 41} A CDS approach using preoperative magnetic resonance imaging (MRI) radiomics, intraoperative factors such as extent of resection, and molecular factors such as MGMT promoter methylation or IDH mutation could risk-stratify for early failure and inform postoperative treatment and follow-up.⁴² Notably, in glioblastoma, this CDS system would need to directly address issues related to uncertainties in the differentiation of true failure from treatment-associated radiation changes and pseudoprogression.

Certain CDS applications bridge the boundary between decision support and intraoperative perception by producing staging or resectability assessments that directly influence operative strategy.^{24, 32} The artificial intelligence laparoscopic exploration system (AiLES) is an AI system designed to recognize intra-abdominal metastasis during laparoscopic gastric cancer surgery using real-time video, with evaluation against surgeon assessments; this form of intraoperative decision support can alter the surgical plan and downstream therapy selection when occult metastatic disease is detected.²⁴ In parallel, perioperative personalized decision support reviews have emphasized the need to integrate educational and feedback components with prediction outputs so that CDS improves not only prediction but also clinical action and decision quality.^{15, 32}

A consistent finding across systematic reviews is that much of the surgical oncology AI literature remains concentrated in retrospective development studies, with heterogeneous endpoints and limited prospective evaluation demonstrating improved patient outcomes or workflow efficiency.^{13, 37} Bektaş et al., in a systematic review of ML applications in upper gastrointestinal cancer surgery, noted substantial heterogeneity in outcomes and methodological approaches, limiting comparative synthesis and emphasizing the need for standardized reporting and validation.³⁷ Similarly, in rectal cancer surgery, a systematic review of AI models predicting surgical difficulty highlighted variability in definitions, imaging features, and validation strategies, illustrating that even seemingly bounded decision support targets (eg, “difficulty”) can suffer from endpoint ambiguity that impairs translation.³⁷

Generalizability across institutions and time is a core translational constraint in perioperative CDS because case mix, perioperative protocols, coding practices, and follow-up capture differ across sites and evolve over time.¹³ Several surgical oncology studies have begun to address this through explicit external validation designs.^{6, 27, 29, 39} Dal Cero et al. performed an international external validation of a machine learning model for 90-day mortality after gastrectomy, providing direct evidence that transportability must be tested under clinical and geographic shifts rather than assumed from internal performance alone.⁶ Complementary internal–external validation strategies, as used in multinational recurrence prediction after esophageal cancer surgery, similarly represent pragmatic approaches for early transportability assessment when full external validation cohorts are not yet available.^{37, 39}

From a clinical translation perspective, discrimination is necessary but insufficient.^{43, 44} Poorly calibrated models can misestimate absolute risk and therefore misdirect threshold-based decisions (eg, intensive monitoring, admission to higher-acuity units, or initiation of prophylaxis).^{18, 43} Consequently, calibration assessment and decision-analytic evaluation should be considered minimum requirements for CDS systems intended to change clinical actions.^{43, 44} This emphasis is reflected both in methodological guidance and in several recent surgical oncology studies reporting calibration curves, Brier scores, and decision curve analysis (DCA) alongside discrimination.^{23, 34, 35, 40} The reporting of DCA is particularly relevant to decision support because it quantifies net benefit across thresholds, connecting model outputs to the clinical consequences of false positives and false negatives rather than relying solely on rank-based discrimination.^{34, 44}

Even when predictive performance is favorable, CDS adoption depends on usability, clinician trust, and integration at the time and location of decision-making.^{14, 45} A systematic review of clinical decision support systems identified features associated with improved clinical practice, including automatic provision within workflow and delivery of recommendations rather than assessments alone, design principles that are directly relevant to AI/ML tools intended for perioperative oncology care.¹⁴ Qualitative research on decision support interventions targeting shared decision-making similarly underscores clinician concerns about appropriateness of recommendations, time costs, and alignment with clinical judgment, anticipating barriers to AI/ML CDS if outputs are poorly contextualized or not coupled to feasible actions.⁴⁵ For patient-centered surgical decisions, evidence from randomized evaluation of conversation aids in breast cancer surgery further illustrates that the quality of decisions depends on aligning clinical options with patient values and understanding; prediction outputs are therefore only one component of the translational goal of improving decision quality.^{15, 46}

Equity considerations are central to translational readiness because prediction models trained on historical healthcare data may encode structural inequities in access, treatment, and documentation.^{13, 47} In a widely cited example outside oncology, Obermeyer et al. demonstrated substantial racial bias in a commercial risk prediction algorithm arising from the use of healthcare costs as a proxy for health needs; analogous proxy-label and documentation biases are plausible in surgical oncology when outcomes, comorbidities, and follow-up are differentially captured across groups.^{25, 37, 47} Systematic review evidence in cancer pathways also indicates inconsistent reporting of subgroup performance and limited attention to equity in prospective evaluations, reinforcing the need to treat fairness assessment as a core translational requirement rather than an optional post hoc analysis.^{13, 38, 42}

Finally, reporting and evaluation standards are prerequisites for reliable translation.^{42, 48–50} TRIPOD and the updated TRIPOD+AI statement provide minimum reporting items for prediction model studies using regression or ML methods, including a transparent description of data sources, outcome definitions, handling of missingness, and full reporting of performance metrics.^{48, 49} PROBAST and the updated PROBAST+AI tool support structured assessment of risk of bias and applicability, offering a shared framework for judging whether a model is likely to generalize and whether reported performance is credible.^{42, 50} For prospective and interventional evaluation of AI-enabled decision support, CONSORT-AI and SPIRIT-AI extend trial reporting and protocol guidance, while DECIDE-AI provides a reporting guideline for early-stage clinical evaluation of decision support systems driven by AI, an appropriate framework for perioperative CDS that may be introduced first in limited workflows before broader rollout.^{13, 15, 51}

Human factors readiness of perioperative AI systems for use will demand interface design elements that support safety in situations of uncertainty, drift, and incorrectness. For example, at the UI level, uncertainty-first design principles must be implemented through quantification of probabilities and explicit threshold bands with prespecified actions, insufficient evidence states when probabilities are too low, drifting detection and notification, and hard constraints against providing confident guidance through overlays and trajectories when upstream uncertainty bands are breached. At the workflow level, each AI system will need a compact failure mode playbook that maps common types of failures (eg, detection miss, segmentation bias, registration offset, tracking drift/OOD, cybersecurity issues) to standard responses such as confirmation steps, escalation protocols, fail-safe transitions to manual or alternative modes of imaging, documentation requirements, and review triggers after failure events. In parallel, equity assessment will need to be addressed as a translational deliverable through a simple audit template that reports stratified performance and calibration across clinically meaningful subgroups and settings (eg, by sex, age, race/ethnicity if available, comorbidity burden, language and insurance status if applicable), subgroup-specific net benefit analysis through DCA, and prespecified mitigation approaches such as targeted data enrichment and reweighting and domain adaptation, subgroup-dependent thresholds with clinically justified trade-offs and monitoring with rollback criteria if disparate errors are seen (Figure 1).

Radiologic staging and anatomic definition remain central determinants of operability and approach selection; accordingly, AI/ML methods have been applied to extract structured staging signals from imaging, radiology narratives, and complementary tissue sources.^{1, 59–62} Radiomics-based strategies, whether relying on engineered feature families or learned representations, have been positioned as a means to quantify tumor phenotype and context in ways that may complement conventional human interpretation, while also amplifying sensitivity to imaging protocol variability and preprocessing choices.¹ In breast oncology, convolutional neural network analysis of multiparametric MRI has been evaluated for preoperative prediction of axillary lymph node metastasis, directly targeting a staging component that may influence the extent of axillary surgery and systemic therapy sequencing.⁵⁹ In gastric cancer, multimodal approaches that combine CT-derived information with digital pathology features (whole-slide imaging [WSI]) have been reported for staging-related prediction, illustrating how “preoperative planning” increasingly extends beyond imaging alone when pre-treatment tissue is available (eg, endoscopic biopsy).⁶⁰

Natural language processing (NLP) has also been explored to convert narrative radiology into structured, decision-relevant staging information. Automated esophageal cancer staging from free-text radiology reports has been evaluated using large language models (LLMs), supporting the feasibility of extracting standardized staging descriptors from routine documentation and potentially reducing manual abstraction burden in multidisciplinary workflows.⁶¹ Because many planning decisions hinge on a synthesis of imaging, histology, and clinical narrative, LLM-centered approaches may be attractive for workflow integration; however, their translational value depends on rigorous evaluation under realistic variability in report style, missingness, and institutional terminology, as well as safeguards against unsupported inference, and values that integrate restorative ventures for at risk social groups in healthcare.^{61, 63}

Planning outputs extend beyond the categorical stage to intermediate, decision-relevant targets such as anticipated technical difficulty, required exposure, and margin risk.^{41, 62} In rectal cancer surgery, AI models applied to preoperative MRI have been studied for predicting surgical difficulty, an endpoint that can influence the choice of minimally invasive vs open approach, need for specialized instrumentation, and staffing/experience requirements.⁴¹ In oral squamous cell carcinoma, pathomics-based models have been studied for predicting positive surgical margins, illustrating an emerging paradigm in which pre-treatment or perioperative tissue-derived features are translated into margin-focused planning inputs.⁶² These examples underscore a broader trend: planning systems often seek to forecast “actionable constraints” (eg, limited working space, high-risk dissection planes) rather than only long-term outcomes, but such targets require careful operationalization and validation because they can be sensitive to surgeon factors, institutional practice patterns, and differences in operative technique.^{1, 41}

Most preoperative planning workflows depend on delineation of tumors, organs-at-risk, and critical structures, making anatomic segmentation an enabling task for 3D reconstruction, simulation, and downstream guidance.^{2, 54–56, 58, 64–67} In thoracic oncology, semantic segmentation of chest CT has been used to recognize patient-specific pulmonary vessel variants relevant to resection planning, illustrating how segmentation can be directed toward clinically consequential anatomic variability rather than organ boundary extraction alone.⁵⁸ In rectal cancer, 3D reconstruction has been described as a means to improve diagnosis and surgical planning, reflecting sustained interest in translating pelvic MRI into spatially coherent representations that can support operative strategy and multidisciplinary discussion.⁵⁵

Segmentation also supports simulation-oriented planning that seeks to anticipate intraoperative views and dissection trajectories.^64–66 An AI-driven 3D simulation system for gastric cancer surgery has been reported as a retrospective validation study, consistent with an early translational pathway in which automated model construction and anatomy recognition are first evaluated for feasibility and concordance with expert interpretation before the impact on procedural decisions is tested prospectively.⁶⁴ Similarly, AI-based technology to generate a 3D model for rectal cancer surgery planning from MRI has been reported as a step toward preoperative simulation, highlighting an approach in which imaging-derived models are used for rehearsal and spatial understanding rather than direct automation of operative steps.⁶⁵ At a more granular anatomic scale, automated segmentation of male pelvic floor soft tissues have been proposed for preoperative simulation and morphologic assessment in lower rectal cancer surgery, reflecting the planning value of extracting patient-specific pelvic anatomy that may influence exposure and dissection strategy.⁶⁶

Planning applications have also been described in hepatobiliary contexts using intelligent image segmentation approaches, aligning with long-standing clinical needs for patient-specific visualization of vascular and biliary anatomy and for volumetric assessment in resection planning.⁶⁷ Although organ-specific reviews emphasize potential benefits of AI/ML for liver cancer surgery, the evidentiary standard for preoperative planning systems remains high because errors in anatomic delineation or spatial relationships can directly affect operative decisions.⁵²

Quantitative biomarkers derived from segmentation can inform preoperative optimization and approach selection when measurement performance is adequate, and the clinical mapping from measurement to action is explicit.^68–70 Automated body composition assessment from routine CT has been evaluated as a scalable method to measure muscle and adipose compartments, including validation-focused studies of deep learning derived measurements from standard-of-care CT examinations.^{68, 69} Related work has linked imaging-derived sarcopenic obesity with prognostic outcomes, supporting the premise that quantitative imaging may identify phenotypes relevant to prehabilitation, nutritional optimization, and candidacy assessments.⁷⁰ These approaches may be particularly attractive because they leverage imaging already obtained for staging, but translational use requires transparent reporting of measurement error, robustness across scanners and protocols, and clear articulation of how biomarker thresholds would change management.^{1, 68, 70}

Preoperative planning frequently requires integration across modalities (eg, multiphase CT and MRI), across timepoints, and across coordinate frames that differ from the intraoperative configuration; therefore, image registration and fusion are recurrent bottlenecks for constructing coherent 3D models that can be inspected by clinicians and, in some settings, transferred to guidance systems.^{2, 53, 54, 56} Reviews of 3D reconstruction and organ modeling emphasize that static preoperative models can be undermined by soft-tissue deformation and physiologic motion (eg, respiration), motivating deformation-aware planning and, where feasible, updating or reconciling models with intraoperative sensing.^{2, 54} The urologic robotics literature similarly situates preoperative modeling within a broader imaging-robotics ecosystem, highlighting both the promise of patient-specific reconstructions and the technical challenges of aligning preoperative images with intraoperative anatomy.⁵³

Because a preoperative plan must be interpretable and actionable for surgeons and operative teams, many planning systems emphasize operator-facing 3D visualization and rehearsal rather than prediction outputs alone.^{2, 54–56, 58, 64, 65} Work on 3D lung model development for minimally invasive lung cancer surgery has explicitly framed a progression from static reconstructions toward real-time or dynamic modeling, consistent with the clinical observation that thoracoscopic and robotic approaches increase dependence on accurate spatial representations when tactile feedback is limited.⁵⁴ This emphasis on visualization is also apparent in gastrointestinal oncology planning efforts that use AI-enabled reconstruction to support preoperative simulation in gastric and rectal cancer operations.^{64, 65}

Virtual reality and 3D-printed models have been systematically reviewed as planning tools in image-guided, robot-assisted nephron-sparing surgery, providing evidence synthesis on how patient-specific 3D representations may support surgeon understanding, rehearsal, and intra-team communication.⁵⁷ Complementary discussions of robotic partial nephrectomy in the era of 3D virtual reconstructions similarly reflect continued interest in operationalizing preoperative 3D models as part of routine planning.⁷¹ While these technologies are conceptually aligned with surgical oncology goals, improving spatial understanding and reducing unanticipated anatomy, the translational question is whether such tools measurably improve decision quality or outcomes when implemented at scale, accounting for the time and expertise required for model generation and review.^{57, 71}

Preoperative planning also includes selection and orchestration of adjunct intraoperative sensing modalities and tracers, with plans often specifying timing of administration, expected contrast behavior, and decision thresholds.^{56, 72, 73} Reviews of precision cancer surgery describe fluorescence-guided approaches that seek to improve tumor margin identification and lymphatic mapping, while emphasizing constraints imposed by probe pharmacokinetics, tissue optical properties, and imaging hardware.⁷² A randomized translational study of protein- and peptide-based probes illustrates that probe selection and deployment strategy are empirically tractable components of the planning process, but also highlights that clinical utility depends on matching probe behavior to a defined decision task (eg, margin assessment or lesion localization).⁷³ Robotic- and image-guided workflows have likewise been described for fluorescence detection, reflecting ongoing efforts to integrate targeted imaging into operative systems.⁷⁴

Lymphatic mapping provides a complementary example in which tracer workflow and imaging protocol determine intraoperative localization targets. Indocyanine green (ICG)-based lymph node mapping has been described in cancer surgery contexts, and preoperative sentinel node mapping work in gynecologic oncology demonstrates how imaging protocol design and tracer deployment can be codified into planning pathways that influence operative navigation targets.^{75, 76} Optical coherence tomography (OCT) adds a further dimension by providing microstructural imaging that can, in principle, support margin-focused decision-making; systematic and scoping reviews describe OCT in intraoperative tissue assessment and AI-enabled OCT for disease detection and diagnosis, respectively, emphasizing both potential and the need for clinically grounded validation.^77–79

For AI/ML-enabled preoperative planning, translational readiness depends on demonstrating that outputs are reliable under routine imaging heterogeneity, that errors are bounded and detectable, and that any claimed benefit is linked to decision-relevant endpoints rather than surrogate technical metrics alone.^{1, 2, 54, 56, 68} Validation studies in segmentation and quantitative imaging (eg, body composition measurement) illustrate the importance of reporting agreement with reference standards and clarifying conditions under which manual correction remains necessary.^{68, 69} Because planning models often function as upstream dependencies for downstream navigation, simulation, or robotic workflows, failure modes should be characterized not only at the level of segmentation overlap but also in terms of clinically meaningful spatial errors (eg, mislocalization relative to vessels, planes, or margins).^{2, 54, 56, 58}

Surgeon-facing generative systems and automated staging tools introduce additional requirements beyond conventional imaging models, including provenance-aware synthesis, safeguards against hallucinated or non–evidence-based assertions, and evaluation under realistic documentation variability.⁶¹ Current single-center retrospective planning studies, whether focused on 3D reconstruction, simulation, or prompt-driven assistance, should therefore be interpreted as feasibility signals; subsequent work must incorporate external validation, prospectively defined endpoints, and, where possible, decision-impact study designs that quantify how planning outputs change operative choices and whether those changes improve patient-centered outcomes.^{1, 54, 57, 61–65}

Across robotic oncologic surgery, the dominant substrate for intraoperative ML is surgical video, reflecting the ubiquity and standardization of endoscopic and robotic camera feeds and the relative ease of rendering outputs as overlays within existing display pipelines.^{3, 53, 80} Proposed enabling architectures for AI-enhanced processing of da Vinci robot-assisted video emphasize that translational feasibility is conditioned by system-level constraints, deterministic latency, robust video ingestion, and reliable handling of large data streams, which can be as decisive as model accuracy for operating room deployment.⁸²

Video-based perception has been used to recognize anatomic structures directly tied to oncologic safety and postoperative function. Deep learning approaches for vessel and tissue recognition during third-space endoscopy illustrate a generalizable pattern: models are trained to detect structures that may be visually subtle yet clinically consequential when injured, with outputs intended to support safer dissection in minimally invasive contexts.⁸³ In urologic oncology, an AI-based intraoperative nerve recognition system has been described for nerve sparing during robotic prostatectomy, exemplifying a navigation-adjacent use case in which “recognize-and-overlay” can provide moment-to-moment spatial cues during technically demanding dissection.⁸⁴ These approaches align with broader reviews of robotic oncologic surgery that highlight computer vision-based recognition of anatomy and “danger zones” as a near-term, clinically plausible tier of intraoperative AI support.^{3, 53, 80, 81}

Intraoperative perception has also been extended to targets that are explicitly oncologic rather than purely anatomic. In laparoscopic gastric cancer surgery, the AiLES was developed for real-time recognition of intra-abdominal metastasis during exploration, directly addressing a clinically consequential failure mode in which small or occult lesions are overlooked, and staging is thereby misclassified.²⁴ Reviews of intraoperative imaging tools in gastrointestinal oncology likewise emphasize that the clinical value of intraoperative perception increases when outputs are tied to specific intraoperative actions (eg, escalation to biopsy/frozen section, targeted inspection, or altered resection strategy) rather than solely to descriptive visualization.⁵⁶

Functional state estimation provides a complementary intraoperative objective because tissue perfusion and viability influence anastomotic strategy and can interact with oncologic extent through resection planning.^{56, 85, 86} AI-based real-time microcirculation imaging has been evaluated for assessing colonic perfusion status, representing a class of intraoperative systems that infer physiologic states from subtle imaging cues that are difficult to quantify reliably by human vision alone.⁸⁵ In addition, intraoperative near-infrared (NIR) functional imaging has been studied in colorectal cancer settings, reflecting efforts to quantify perfusion-related and function-related signals that could be integrated into navigation workflows and assessed against endpoints such as leak risk, complication rates, or re-intervention.⁸⁶

Navigation systems must establish spatial relationships between recognized structures, instruments, and (when applicable) preoperative models, often under conditions of soft-tissue deformation, camera motion, and variable insufflation.^{3, 53, 80} Real-time vascular anatomical image navigation has been demonstrated for laparoscopic surgery as an approach to aligning vascular anatomy with intraoperative views to support orientation and safer dissection around critical vessels.⁸ In rectosigmoid oncology, the feasibility of optical stereotactic navigation supported by deep learning based 3D modeling has been reported, highlighting the clinical interest in “coordinate-consistent” guidance for tumor localization and margin-conscious dissection within anatomically constrained pelvic spaces.⁹

Evidence from thoracic oncology similarly indicates active development of geometry-aware navigation strategies. A prospective application study of non-contact, AI-assisted intraoperative 3D navigation in lung cancer surgery demonstrates a pragmatic approach in which spatial information is acquired and reconstructed intraoperatively to mitigate the mismatch between preoperative imaging and intraoperative anatomy.¹² Across these demonstrations, the translational challenge is not merely producing a geometric solution but maintaining accuracy under clinically routine perturbations (eg, deformation, retractors, motion, variable insufflation), and predefining failure detection and escalation workflows when navigation assumptions no longer hold.^{3, 9, 12, 53}

Visualization is a central component of intraoperative navigation because guidance must be interpretable at the point of action. Reviews of extended reality and “metaverse” concepts in surgery discuss how augmented visualization environments could provide spatial context and training ecosystems, while emphasizing limited clinical evidence for outcome benefit and the nontrivial challenges of integrating head-mounted or mixed reality interfaces into sterile workflows.⁸⁷ In robotic radical prostatectomy, specialty-focused reviews similarly frame advanced visualization (including potential augmented guidance) as an emerging direction that will require rigorous clinical validation and human factors evaluation before routine adoption.⁸¹

Optical contrast agents and intraoperative imaging can function as “signal amplifiers” for tumor, lymphatics, or function, and ML methods are increasingly used to standardize interpretation and generate overlays that can be integrated into robotic visualization pipelines.^{53, 56} Robotic assistance can stabilize imaging geometry and enable instrument-mounted sensing, as illustrated by robotic “click-on” fluorescence detection using surgical instruments to characterize molecular tissue aspects, a design pattern that can support localized assessment within an otherwise video-dominated workflow.⁷⁴

Deep learning enabled fluorescence imaging has been developed for surgical guidance tasks aligned with oncologic decision-making. For oral cancer, in silico trained deep learning methods have been described for fluorescence-based depth quantification and for fluorescence-guided margin classification in preclinical models, representing an approach that seeks to convert fluorescence signal into explicitly interpretable, decision-relevant outputs rather than qualitative visualization alone.^{88, 89} More broadly, reviews of precision cancer surgery emphasize that the translational value of imaging-augmented guidance depends on linking the imaging signal (and any ML-derived interpretation) to prespecified intraoperative decisions (eg, additional resection, targeted sampling, or altered dissection plane) and demonstrating benefit against clinically meaningful endpoints.⁵⁶

Near-infrared fluorescence and functional imaging also support navigation decisions beyond margin assessment, including lymphatic mapping and perfusion-oriented guidance.^{56, 86, 90} In upper gastrointestinal oncology, indocyanine green (ICG) lymph node mapping has been described for cancer surgery, and prospective mapping of lymphatic drainage patterns with NIR fluorescence has been studied during robotic-assisted minimally invasive Ivor Lewis esophagectomy.^{75, 91} These studies illustrate that intraoperative guidance is often probabilistic and pathway-dependent (eg, drainage patterns and node basins) and therefore benefits from clear protocols describing how intraoperative findings are intended to modify the operative plan.⁵⁶

Label-free tissue characterization provides an alternative approach to intraoperative guidance by aiming to classify tissue state directly from biophysical signatures, potentially enabling margin assessment without exogenous tracers.^{11, 92, 93} Perioperative tissue assessment using combined mass spectrometry and histopathology imaging has been described as a framework for integrating molecular and morphologic information to inform intraoperative decisions in cancer surgery.⁹³ Within this domain, uncertainty estimation for surgical margin detection using mass spectrometry has been reported as a strategy to calibrate confidence and support decision thresholds suited to high-stakes intraoperative use.¹¹

Representation-learning strategies have also been applied to reduce dependence on exhaustive labeling while preserving clinically relevant performance, exemplified by image-driven self-supervised learning for mass spectrometry-based tissue assessment.⁹² Although these methodological advances may improve scalability, systematic syntheses in robotic cancer surgery converge on the need for external validation and explicit specification of action policies when model outputs are uncertain, particularly when outputs are used to trigger irreversible actions such as widening margins, converting operative plans, or escalating to additional resection.^{3, 53} Accordingly, translational protocols should predefine how uncertainty will be communicated to the surgical team, how clinicians should respond to uncertain outputs (eg, confirmatory pathology), and how performance will be monitored over time once deployed.^{3, 11, 53, 92}

Microscopic intraoperative imaging modalities such as OCT have been evaluated as tools for margin-focused guidance and tissue characterization when macroscopic imaging is insufficient to resolve microstructural features at the resection interface.^{77, 79} A systematic review of OCT in oncologic surgery summarizes evidence for intraoperative tissue assessment and highlights translational barriers, including limited standardization, variable acquisition conditions, and uncertainty about how best to integrate OCT-derived information into operative workflows.⁷⁹ A scoping review focused on AI in advancing OCT further reinforces that algorithmic interpretation is a key enabler for converting high-dimensional OCT data into actionable intraoperative outputs, while also underscoring the need for robust validation and clinically grounded study designs.⁷⁷

Hybrid systems that fuse OCT with ML inference have been reported for breast cancer surgical margin visualization, illustrating a design pattern in which high-resolution imaging is paired with automated classification to generate decision-supportive outputs intraoperatively.⁷⁸ Translation of such systems into routine oncologic practice requires demonstration of robustness to blood, motion, and specular artifacts; prespecified uncertainty-aware operating points; and empirical evidence that OCT-based guidance improves margin-related outcomes without disproportionate resection or operative burden.^77–79

Robotic platforms create opportunities for continuous performance monitoring and algorithmic safety layers that detect and mitigate hazards, but the evidentiary threshold increases substantially when systems move from passive guidance to action-constraining control.^{3, 53, 80} AI-based hazard detection in robotic-assisted single-incision oncologic surgery exemplifies a safety-oriented direction by targeting the identification of hazardous states in a modality characterized by constrained instrument motion, limited triangulation, and heightened collision risk.¹⁰

Robotic data streams can also be used to evaluate and standardize operative performance, with potential downstream implications for oncologic outcomes and complication risk. A study associating skill and errors with outcomes in robotic rectal cancer surgery illustrates how intraoperative performance metrics can be linked to clinical endpoints, supporting the concept that intraoperative analytics may provide actionable feedback for quality improvement.^{80, 94} At the same time, outcome-linked analytics emphasize the need to avoid overinterpreting surrogate performance measures unless they are validated against patient-centered endpoints and contextualized by case complexity, anatomy, and oncologic extent.^{3, 53, 80, 81, 94, 95}

Engineering frameworks for AI-enhanced processing of da Vinci robot-assisted video reinforce that translation requires not only accurate models but also reliable system-level integration, including secure data handling, deterministic latency, and compatibility with the robotic platform’s data interfaces.⁸² Systematic synthesis of AI integration into robotic cancer surgery indicates that most published systems remain at the level of feasibility or retrospective validation, with limited evidence from prospective studies and limited clarity regarding how ML outputs are incorporated into standardized intraoperative actions.^{3, 53} This evidence profile supports a conservative near-term posture: prioritize decision support, safety monitoring, and uncertainty-aware guidance before pursuing control-constraining behaviors that would require substantially higher safety assurance.^{3, 10, 53}

Across intraoperative navigation and robotic assistance/control, translational readiness depends on demonstrating reliable real-time operation under routine variability, explicit handling of uncertainty, and measurable benefit in decision-relevant endpoints.^{3, 53, 56} Feasibility studies in vascular and stereotactic navigation demonstrate that geometry-aware guidance can be operationalized in oncologic procedures, while prospective application of AI-assisted 3D navigation in lung surgery shows that intraoperative navigation can be evaluated in forward-looking designs rather than solely retrospective settings.^{8, 9, 12} Prospective demonstration of real-time metastasis recognition during exploration further illustrates how intraoperative perception can be tied to immediate oncologic decisions, providing a model for future studies that specify workflow integration and decision thresholds a priori.²⁴

For imaging-enhanced guidance (eg, fluorescence and NIR functional imaging), translation requires separating device- and protocol-specific variability from biologically meaningful signal and validating ML-derived overlays across acquisition settings, patient factors, and institutions.^{56, 86, 88, 89} For label-free modalities (eg, mass spectrometry and OCT), uncertainty-aware inference and robustness to intraoperative artifacts are central because both false reassurance and false alarms can produce clinically meaningful harm (eg, inadequate margins vs unnecessary tissue loss).^{11, 77–79} These requirements are emphasized across systematic and specialty-focused reviews, which report that the clinical evidence for many AI-enabled intraoperative systems remains dominated by retrospective analyses and early feasibility work, with relatively few prospective evaluations tied to patient-centered outcomes.^{3, 53, 80, 81, 95} Prospective application studies in navigation (eg, non-contact 3D guidance) and real-time disease detection (eg, AiLES) provide useful templates for study designs that connect real-time inference to standardized intraoperative actions and measurable outcomes.^{12, 24} Finally, because robotic surgery continues to expand across complex oncologic procedures (including pancreatic surgery), the bar for safe, clinically effective integration of intraoperative AI should remain anchored in workflow-aware validation, transparent uncertainty handling, and rigorous prospective evaluation rather than technical performance alone.^{3, 10, 53}

In addition, outside the perioperative realm, there have been equal advancements in AI for psychiatry and mental health applications, including risk assessment (suicide attempts, relapse, and acute decompensation), symptom profiling, and treatment recommendation with multimodal data sources like longitudinal electronic health records, wearable and smartphone-based digital phenotyping, speech/language processing, and neuroimaging.^96–98 In precision oncology, the combination of radiology, digital pathology, and genomic profiling enabled by AI has facilitated tumor profiling, prediction of treatment response (including immunotherapy), and computational triage for molecular analyses, thereby propelling the transition of high-dimensional biomarkers into decision-making.^99–101 These parallel applications collectively point to a shift towards data-driven, personalized decision support systems, while also pointing out common issues in fairness, generalization, explainability, and validation that are just as applicable in surgical oncology.