The Study You Deleted May Have Been the One That Mattered: Radiology Retention Law in the Age of AI

Every radiology archive manager has run this calculus at some point: storage costs money, retention schedules exist, the law permits deletion after a defined period, and the legal exposure of keeping data past the minimum window may actually exceed the exposure of destroying it. The case for deletion is tidy. The math is usually straightforward. The institutional comfort level is high.

What that calculus did not account for — and what is now impossible to ignore — is that AI can retroactively read the studies you're about to delete, and find things that were missed the first time. Things that, had they been found when the image was acquired, might have led to treatment that would have changed a patient's outcome. Or saved their life.

The legal permission to delete a radiology study and the ethical wisdom of deleting it are no longer the same question. AI has permanently changed what it means to call an old study "no longer clinically relevant."

This is not a theoretical concern. The clinical evidence is accumulating. The legal landscape is shifting. And the global AI training data question — which connects directly to the value of historical imaging archives — adds a second dimension that the industry has barely begun to grapple with.

What U.S. Law Actually Says — And What It Doesn't

There is no single federal standard governing how long radiology images must be retained. HIPAA establishes privacy and documentation obligations but defers to states for most medical record retention periods. The result is a patchwork of state-by-state requirements that varies dramatically — and creates a compliance floor that many organizations treat as the ceiling.

The American College of Radiology's guidance — that institutions should retain records for whichever is longer, the applicable statute of limitations or the state's prescribed retention period — reflects the traditional liability calculus: keep records long enough to defend against malpractice claims, then destroy them to reduce legal exposure.

That framework made sense in an era when old images were essentially inert once the interpretive report was filed. A ten-year-old chest X-ray had no more clinical utility than the film it was printed on. Today, that assumption is wrong.

What AI Can Find in Studies You've Already Read — and Filed

The evidence that AI can identify clinically significant findings in prior imaging that human readers missed is no longer preliminary. It is substantial, and it is growing rapidly.

A 2025 retrospective study from the Netherlands, analyzing over 42,000 mammograms from women screened between 2016 and 2018 with outcomes tracked through the Netherlands Cancer Registry for up to 52 months, offered one of the most comprehensive looks yet at what AI finds when it re-reads prior studies. The rationale was precise: to determine whether AI could identify cancers that human radiologists missed — and whether those missed cases were clinically significant or likely to progress.

The findings were sobering. About 30% of tumors are missed during screening and later appear as interval cancers. Retrospective analysis shows that 20–25% of those missed cancers were already visible on the original mammogram — present, interpretable, and unrecognized. These are not subtle findings identified only in hindsight with the knowledge that cancer developed. They were detectable at the time. They just weren't detected.

A 2024 study published in Scientific Reports went further, evaluating two state-of-the-art AI models against a dataset of 729 cancer cases in which a panel of three expert radiologists retrospectively identified prior missed cancers. In total, 24.82% of the cancers had been labeled as missed at the original read. AI achieved sensitivity of 83% on those missed cases — meaning that for the majority of studies where a trained double-reading panel had already failed, AI was still able to identify the finding.

A 2025 retrospective study published in European Radiology, evaluating an AI system against interval breast cancer mammograms from a Swiss screening program, found that 31.1% of interval breast cancers had retrospective abnormalities identifiable on prior studies. Nearly half of all interval breast cancer cases received an AI suspicion score above a meaningful threshold — meaning the AI, reading the old study, would have elevated these patients for additional review.

These findings extend beyond mammography. Retrospective AI analysis of pulmonary nodules on prior CT studies, of early vascular changes in cardiac imaging, and of subtle pathological patterns across multiple modalities all point toward the same fundamental reality: the interpretive value of a radiology study does not expire when the report is filed. With AI, it may extend indefinitely.

The Ethical Calculus That the Retention Schedule Doesn't Capture

The traditional argument for routine study deletion rests on two pillars: storage cost and legal risk reduction. Both are real. Neither is adequate to the current moment.

The ethical dimension is the one that existing retention frameworks don't address at all: if a patient is alive today, and a study exists from eight years ago that AI could now read to identify a finding that may be affecting their health outcome, and that study is scheduled for deletion this quarter — what is the institution's obligation?

This question does not yet have a legal answer in most jurisdictions. But it is no longer purely hypothetical. As AI retrospective analysis tools become clinically validated and commercially available, the window in which this remains an abstract governance question is closing.

The Global Training Dimension: Who AI Is Being Built For — and Who Gets Left Out

There is a second argument for preserving historical imaging archives that receives almost no attention in the retention discussion: the value of older imaging data as training material for AI systems that will be deployed in environments where those imaging modalities are still the primary clinical tool.

In the United States, computed radiography (CR) has largely been displaced by direct digital radiography (DR). Early-generation CT scanners have been replaced by multi-detector systems. Older ultrasound equipment has been retired. The imaging infrastructure of U.S. healthcare has advanced continuously, and the studies produced by the equipment we've retired live in archives that we're actively scheduled to destroy.

But globally, that retired equipment is still in active clinical use. The data documenting this disparity is stark:

When U.S. healthcare organizations delete historical studies produced on older imaging modalities, they are not just deleting locally obsolete data. They are destroying training material that could make AI systems substantially more useful in the clinical environments where most of the world's patients receive care.

This is a governance and data equity problem that sits entirely outside current retention frameworks. Those frameworks were designed to manage liability and storage costs within a single jurisdiction. They were not designed to account for the global value of longitudinal imaging data as an input to AI systems that will be deployed across dramatically different clinical infrastructures.

AI training datasets for medical imaging are consistently identified as insufficiently diverse — not just demographically, but in terms of the imaging equipment, protocols, and clinical environments they represent. A 2024 review of publicly available radiology AI datasets found that most lack adequate demographic, geographic, and disease representation. The historical archives that U.S. health systems are routinely destroying contain exactly the kind of diverse, longitudinal, multi-modality imaging data that researchers and AI developers consistently say they need and cannot easily acquire.

What a Governance-First Approach to Retention Actually Requires

None of this means that every radiology study should be retained indefinitely. Storage has real costs, data governance has real complexity, and some studies have genuinely diminished utility over time. The argument here is not that deletion is always wrong. It is that the existing frameworks for making that decision were designed without accounting for two realities that now exist:

1. AI can retroactively find clinically significant findings in old studies that were read as normal — and the percentage of prior studies containing missed findings that AI can identify is non-trivial.
2. The global training value of longitudinal imaging data, including data produced on modalities that are no longer in widespread U.S. use, is real, unquantified, and not recoverable once deletion occurs.

A governance-first approach to this question requires at minimum the following:

Re-examine retention schedules in light of AI capability, not just legal minimums

Most institutional retention policies were last reviewed in an era when archived studies had no meaningful clinical utility after a defined window. That assumption is no longer valid. The policy review cycle for retention schedules should be tied to AI capability milestones, not just regulatory updates.

Separate the legal compliance question from the data strategy question

Legal minimum retention and optimal data retention are increasingly different calculations. Healthcare institutions need governance frameworks that treat them separately — with legal counsel addressing the compliance floor and data governance leadership addressing the strategic value of what exists above that floor.

Establish de-identification and research use frameworks before deletion occurs

The most productive path forward is not indefinite retention of identifiable clinical images, but structured de-identification and governed research use of historical imaging data before it is destroyed. This requires HIPAA-compliant de-identification workflows, institutional review frameworks, and data partnership structures — exactly the kind of infrastructure that most health systems have not built. Building it takes time. Deletion is irreversible. The order of operations matters.

Engage with the global AI training data equity question proactively

If U.S. health systems want AI that works across the full range of clinical environments — for patients everywhere, not just patients in well-resourced settings — then the historical imaging data those health systems hold is a strategic asset, not just a compliance liability. The question of how to govern, de-identify, and make that data available for legitimate research and AI development deserves deliberate institutional attention, not a default to scheduled deletion.

A Decision That Can't Be Undone

I've spent over two decades managing the systems that store radiology images — watching the evolution from film to CR to DR to cloud-native VNA. The storage conversation has always been a cost conversation. And I understand the operational pressures that make deletion feel like the responsible choice.

But the emergence of AI has fundamentally changed the value equation for archived imaging data. A study that was unremarkable when it was filed may now be the study that explains a patient's current clinical presentation. A collection of historical images produced on equipment we retired years ago may be the training data that makes an AI system clinically useful for a hospital in sub-Saharan Africa that cannot afford a new scanner.

The compliance frameworks governing radiology image retention were not written with any of this in mind. They permit deletion because they were designed in a world where deletion was, in most cases, genuinely safe. That world no longer exists. And unlike most governance failures, this one is not recoverable. Once a study is deleted, the potential it represented — for retroactive diagnosis, for patient notification, for AI training, for research — is gone permanently.

The question facing healthcare organizations now is not "Are we allowed to delete this?" The law answers that clearly. The question is "Should we?" And on that question, the law offers no guidance at all.

Jim Cook

President, CEO & CAITO, Radiant AI Health Data, Inc. | Founder & CEO, Health AI Data, Inc.

Jim Cook is a healthcare data executive and Senior PACS Administrator with more than 28 years in healthcare IT, including over two decades of direct operational experience managing DICOM systems, enterprise PACS, and clinical imaging infrastructure. He writes on AI governance, data readiness, and the operational realities of deploying AI in healthcare environments. Questions or thoughts? info@radiantaihealthdata.com · linkedin.com/in/jim-cook-haid