IPM Take
Medical AI has a data problem that is easy to overlook.
Hospitals generate enormous volumes of clinical images, but living patients cannot be scanned, sectioned and annotated at microscopic resolution simply to create better training data. Donated bodies can.
A 2026 Clinical Anatomy review describes body-donor-derived datasets as a potential source of highly structured anatomical information for AI, 3D modelling, education, surgery and forensic applications. Techniques such as micro-CT, photogrammetry and cryosectioning can preserve spatial relationships and anatomical variation at resolutions that would be impossible to obtain routinely in living patients.
But this creates a new kind of governance problem.
A person may have consented to donate their body for education or research without imagining that their anatomy could later become a permanent digital dataset, be shared internationally, incorporated into an AI model or contribute to a commercial product.
The question is no longer only whether donor data can improve medical AI. It is whether consent frameworks are precise enough for what digital anatomy now makes possible.
Executive Summary
The Clinical Anatomy review by Sun and colleagues argues that body-donor-derived anatomical data could become an important foundation for medical AI because they offer clear structural boundaries, stable spatial relationships and fine-grained detail. Once digitised, these resources can be annotated, reused and incorporated into AI training or validation.
This could be particularly useful for systems that need to understand anatomy spatially rather than simply classify an image. A donated organ can be imaged repeatedly, sectioned, reconstructed in three dimensions and linked with histology without the movement, radiation constraints or clinical limitations present in living patients.
The obvious limitation is that a cadaver is not a living patient. Blood flow, muscle tone, pressure, respiration and movement are absent, while fixation can alter tissue appearance and relationships. The Medscape report therefore describes the likely future as hybrid, combining detailed donor-derived structural data with much larger clinical datasets from living patients.
Current AI performance also shows how far the field has to go. In a separate Clinical Anatomy study, ChatGPT-4o correctly identified only 22.26% of 265 labelled structures in cadaveric photographs within three attempts. Accuracy ranged from 64.71% for osteological structures to just 8.82% for isolated thoracic organs, and the model occasionally generated nonexistent anatomical terms.
The technical opportunity is therefore real, but still early.
Why it matters
- HTA bodies: AI systems trained on enhanced anatomical datasets may eventually support imaging, surgical planning or clinical decision support, but HTA will need evidence that richer training data translate into measurable improvements in accuracy, safety, workflow or outcomes rather than simply producing more sophisticated models.
- Payers: Donor-derived datasets could improve the performance of diagnostic and procedural AI, but they also create new development and governance costs. Reimbursement should depend on demonstrated clinical value, not on the novelty or resolution of the training data.
- Industry / innovation partners: Body-donor-derived datasets could become strategically valuable because high-quality, expertly annotated anatomical data are scarce. Companies using them will need defensible consent, provenance, access controls and rules around secondary use and commercialisation if they want those datasets to remain trustworthy.
Artificial intelligence is often described as hungry for data.
In medicine, the problem is not always quantity. Sometimes it is the quality and structure of the data available.
Clinical imaging captures the living body, which is essential for diagnosis and treatment. But it also comes with practical limits. Patients move. Organs change shape. Resolution is constrained. Histology and high-resolution sectional anatomy cannot usually be acquired across an entire living body.
Donated bodies offer something different.
They can be scanned at very high resolution, dissected, photographed, sectioned and reconstructed repeatedly. Structures can be labelled by anatomists and aligned across imaging, gross anatomy and histology. The resulting dataset can become a digital anatomical reference rather than a single image.
That could help AI learn where structures are expected to be and how anatomical variation appears.
In principle, this could support surgical navigation, imaging interpretation, educational platforms and anatomical digital twins.
But “digital twin” needs qualification.
A reconstructed donor heart may capture anatomy at extraordinary resolution while lacking everything that makes a living heart dynamic: pressure, contraction, blood flow and electrical activity. The Medscape source puts the limitation clearly: anatomically detailed does not mean physiologically complete.
The ethical problem starts when the body becomes data
Digitisation changes the nature of donation.
A physical specimen exists in one place and can ultimately be disposed of according to donor-program procedures. A digital representation can be copied indefinitely, transferred across institutions and potentially incorporated into commercial systems.
Professional guidance is beginning to catch up.
The American Association for Anatomy’s 2025 best-practice report recommends that consent address the duration of use or permanent retention of digitised body data, possible domestic or international transfer, image collection, and whether donor material or data could contribute to commercial products.
International Federation of Associations of Anatomists guidance similarly recommends that anatomical imaging be covered by informed consent appropriate to the intended use. Where images could be used commercially, more specific consent may be required.
Those principles become even more important when AI enters the picture.
Training a model on a donor-derived dataset can create outputs far removed from the original anatomical laboratory. Data may be reused repeatedly, integrated into proprietary systems or distributed across jurisdictions.
That creates questions that conventional body-donation forms were not necessarily designed to answer.
Who owns a digitised anatomical dataset? Can it be licensed to a technology company? Can the donor’s data remain inside an AI model after the underlying files are withdrawn? What does meaningful consent look like when future uses cannot yet be predicted?
There are not yet universal answers.
Better data will not automatically produce trustworthy AI
The other risk is technological overconfidence.
The ChatGPT-4o anatomy study is a useful reminder that modern AI can perform impressively on many medical tasks while struggling badly with detailed real-world anatomical recognition.
An overall accuracy of 22.26% in cadaveric structure identification is nowhere near a level that would justify relying on the model for practical anatomy recognition.
Better donor-derived datasets may help.
But the path from high-resolution anatomy to safe clinical AI will still require expert annotation, multimodal training, external validation and testing in living patients.
Donated bodies could become an unusually valuable source of anatomical ground truth.
They should not become an unusually convenient source of ungoverned data.
If medical AI is going to learn from people after death, the rules governing that learning need to be as precise as the anatomy itself.
