New IDEAL-AI Framework Brings Surgical-Style Accountability to Medical AI

Artificial intelligence has moved from research curiosity to clinical force in radiology and neurosurgery, yet the systems meant to ensure its safety and accountability have not kept pace. Now a team of Toronto-based researchers has proposed a solution: adapting a framework originally designed to evaluate surgical innovations, called IDEAL, to create a stage-gated pathway for clinical AI. The proposal, published in Frontiers in Radiology, argues that AI should be judged not by benchmark accuracy or adoption speed but by reproducible clinical benefit for patients, clinicians, and health systems.

The researchers, led by Jasleen Saini from the University of Toronto’s Division of Neurosurgery, note that AI algorithms now delineate tumor margins, classify molecular signatures, and forecast outcomes before a single incision is made. In neurosurgery, AI supports preoperative planning, image registration, augmented-reality overlays, intraoperative navigation, and real-time scene interpretation. But these applications introduce risks not captured by image-only benchmarks, including registration error, brain shift during surgery, latency, device interoperability issues, and the consequences of erroneous guidance delivered in real time.

Models that perform strongly in controlled studies may show reduced performance across new institutions, scanners, workflows, or patient populations. Scanner calibration, institutional heterogeneity, and underrepresentation of vulnerable populations continue to limit generalizability. The team argues that existing reporting guidance, medical-device standards, regulatory requirements, and machine learning operations each address important parts of the problem but were not designed to serve as a unified, stage-gated clinical-evidence pathway across the full AI lifecycle.

The proposed IDEAL-AI framework organizes evidence generation across five stages—Idea, Development, Exploration, Assessment, and Long-term study—and three concurrent domains: validation; governance, regulation, and ethics; and patient and system outcomes. At the Idea stage, AI development should begin with a defined clinical need and intended use rather than a technical challenge in search of an application. Technical, fairness, safety, human-factors, and clinical endpoints should be specified before model tuning begins.

A critical distinction the framework draws is between retrospective model development and prospective live exploration. Development covers model refinement using retrospective data, requiring separated training, tuning, and locked test sets along with version control. Exploration begins prospective live clinical use across increasing site and clinician diversity, focusing on generalizability, workflow integration, and safety under real conditions. This separation prevents broad clinical claims from being inferred from retrospective development alone.

The Assessment phase requires adequately powered, multi-institutional comparative evaluation of a locked system with a prospectively specified protocol. Randomization is preferred when feasible. The Long-term phase ensures oversight continues beyond authorization, using real-world dashboards to track calibration, subgroup performance, drift, and safety signals over time. Predefined alert, stop, rollback, and recalibration rules should be linked to change-control processes such as the FDA’s Predetermined Change Control Plans.

The team illustrates the framework with a worked example: glioma segmentation for surgical planning. Despite strong Dice similarity coefficients reported in systematic reviews, the field often stops at early stages due to frequent reuse of curated public datasets and limited prospective evaluation. Under IDEAL-AI, a segmentation tool would need to demonstrate not just technical overlap but how the contour changes planning, what errors matter near eloquent brain structures, and which downstream outcomes are plausibly affected.

The framework also confronts automation bias—the tendency for clinicians to over-rely on AI advice, particularly dangerous in high-stakes settings like neurosurgical navigation where false precision can lead to irreversible harm. The researchers argue that mitigation requires human-factors design, prospective evaluation, clinician training, usable uncertainty displays, and clear escalation pathways. Systems should actively support rather than merely permit clinician challenge.

IDEAL-AI is designed as an orchestration layer rather than a competitor to existing frameworks. It complements facility-level governance programs, registry infrastructure for post-deployment monitoring, reporting guidelines, and operational pipelines for reproducible deployment. The researchers emphasize that a technically compliant system may still lack evidence of workflow fit or patient benefit, while a clinically promising pilot may remain unready for scale because of inadequate documentation or subgroup evaluation.

The measure of success, the team concludes, will not be how quickly AI is implemented but how rigorously it is evaluated and governed. Clinicians, regulators, and developers share a collective duty to ensure that accuracy, equity, and accountability evolve in tandem. The next era of AI in medicine will be defined not by who adopts it first, but by how effectively it improves patient outcomes.

The IDEAL framework that underpins this proposal has a notable history in surgery. First articulated more than a decade ago by collaborative groups of surgical researchers, it was designed to bring order to the evaluation of new procedures and devices, where early enthusiasm and small single-center series often outpaced rigorous evidence. Its adaptation to AI reflects a recognition that machine-learning tools share key features with surgical innovations: they evolve iteratively, their real-world behavior can differ from controlled testing, and their harms may only become visible after widespread adoption. The authors argue that this lineage gives IDEAL-AI a practical advantage over frameworks built solely around software, because it was conceived for technologies whose safety depends on both the artifact itself and the skill of the person deploying it.

The distinction between the Exploration and Assessment stages deserves particular attention. In many current deployment pathways, a tool that performs acceptably at one institution is rapidly scaled across a health system without a formal comparative evaluation. IDEAL-AI instead requires that live exploration with increasing numbers of sites and clinicians precede any adequately powered multi-institutional comparison, and that the comparison itself use a locked version of the system with a prospectively specified protocol. This sequencing mirrors how new surgical techniques traditionally earn acceptance, and it creates natural decision points where expansion can be paused if safety signals or performance degradation emerge.

The framework’s emphasis on prespecified endpoints addresses a well-documented weakness in the AI literature. Retrospective studies frequently report favorable results because the metrics, thresholds, and subgroups analyzed were chosen after the results were seen. By requiring that technical, fairness, human-factors, and clinical endpoints be fixed before model tuning begins, IDEAL-AI reduces the flexibility that enables such selective reporting. The inclusion of fairness metrics as a required domain rather than an optional consideration reflects growing evidence that performance can vary substantially across patient subgroups defined by age, sex, ethnicity, comorbidity, or imaging characteristics, with underrepresented populations bearing the greatest risk of silent failure.

The glioma segmentation example illustrates why downstream consequences matter more than headline metrics. A segmentation model might achieve an impressive overlap score on a public benchmark while producing contour errors that are clinically trivial in some brain regions and catastrophic near eloquent cortex. The framework’s insistence on asking how a contour changes surgical planning, which errors matter, and which outcomes are plausibly affected shifts evaluation from abstract similarity measures toward the decisions that actually shape patient care. This logic applies equally to radiology, where a detection tool’s false-negative rate in a specific subgroup may matter far more than its average sensitivity.

Post-market surveillance receives particular emphasis because AI systems can degrade silently. Unlike a static surgical implant, a deployed model is exposed to evolving scanner fleets, changing acquisition protocols, shifting referral patterns, and updates to upstream software. Drift may occur gradually and without any alarm, which is why the authors call for real-world dashboards tracking calibration, subgroup performance, and safety signals, with predefined alert, stop, rollback, and recalibration rules. Linking these operational triggers to regulatory mechanisms such as Predetermined Change Control Plans provides a pathway for manufacturers to update models while preserving oversight, rather than forcing a choice between frozen tools and uncontrolled change.

The proposal also engages with the practical reality that most healthcare AI governance today is fragmented. Adoption decisions are often shaped by local procurement, and critical judgments about validation, safety, and oversight fall to individual organizations with widely varying resources and expertise. The authors suggest that multidisciplinary Clinical AI Oversight Boards, including clinicians, data scientists, ethicists, quality leaders, patient representatives, and regulatory experts, can anchor accountability at the institutional level, while linked health-data systems and federated evaluation extend performance tracking across institutions without centralizing patient-level data. This layered approach acknowledges that no single actor—regulator, vendor, hospital, or clinician—can carry the full burden of accountability alone.

Ultimately, the framework’s most consequential claim may be its redefinition of success. By judging AI not on adoption speed or benchmark accuracy but on reproducible clinical benefit, IDEAL-AI reframes the incentive structure for developers, purchasers, and clinicians alike. A tool that cannot demonstrate benefit across diverse sites, patient populations, and workflow conditions would fail to progress, regardless of how impressive its test-set performance appears. Whether the field embraces such discipline, or continues to reward rapid deployment, will shape the safety and trustworthiness of clinical AI for years to come.

Subject of Research: From scanner to scalpel: adapting IDEAL for accountable AI in radiology and image-guided neurosurgery

Article Title: From scanner to scalpel: adapting IDEAL for accountable AI in radiology and image-guided neurosurgery

Article References: Saini, J., Jassar, S., Adams, S. J., Alli, S. B., Colak, E., & Wilson, J. R. (2026). From scanner to scalpel: adapting IDEAL for accountable AI in radiology and image-guided neurosurgery. Frontiers in Radiology, 6, Article 1979076. https://doi.org/10.3389/fradi.2026.1979076

Image Credits: AI Generated

DOI: 10.3389/fradi.2026.1979076

Keywords: scanner, scalpel, adapting, IDEAL, accountable, radiology, image-guided, neurosurgery, scientific research