Surgical Video Datasets for AI: Consent, HIPAA and How Hospitals Share Footage

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why public surgical video benchmarks rarely fit a commercial model

Public surgical datasets are useful for prototyping, but their size, scope and licenses usually rule them out for a shipped product. Cholec80, the common reference set for laparoscopic phase recognition, contains only tens of cholecystectomy videos from one center, labeled for phases and tool presence. Academic surgical video sets are commonly distributed under non-commercial terms, often Creative Commons NC or NC-SA variants, and segmentation or annotation sets derived from them generally carry the same restrictions. Read the actual license file and the data request form for each set rather than a summary on an aggregator site.

The practical gaps go beyond the license. Most public sets cover one procedure, one or a few centers and one tower vendor, so a model trained on them meets new optics, white balance, smoke patterns and port placement in deployment. Share-alike terms also create a problem for proprietary weights and derived annotations. Treat public benchmarks as an evaluation reference and plan to license primary footage for training, as the video data hub describes for other footage types.

Which kinds of surgical video exist and who holds them

Surgical video falls into three capture types, and each carries different identifiers and different holders. Internal (intracorporeal) views come from laparoscopic, endoscopic, arthroscopic and robotic camera systems; they show anatomy and instruments, rarely a face, but carry overlays and metadata. Microscope and exoscope feeds in ophthalmology, neurosurgery and ENT behave similarly. Room cameras, sometimes called operating room black boxes, capture staff, the patient's body and often faces and audio.

Holders are usually hospitals and ambulatory surgery centers, sometimes the surgeons' practice group, and the recordings often live on the vendor's capture appliance or cloud archive. Ownership and access are unsettled in practice, including whether a recording is part of the designated medical record, and hospital policies differ. A buyer should never assume the hospital, the surgeon and the platform vendor agree on who may license a file.

• Laparoscopic and robotic: phase recognition, instrument detection and tracking, critical view of safety, skill assessment.
• Flexible endoscopy: polyp detection, withdrawal-time quality, landmark recognition; often stored next to pathology results.
• Microscope and exoscope: step recognition in cataract and neurosurgical work.
• Room and wide-angle: workflow timing, turnover, team activity; the hardest category to de-identify.

How hospitals can share surgical footage under HIPAA

A hospital that is a covered entity can release footage for commercial AI training most cleanly once it is de-identified under 45 CFR 164.514 [2]. Safe Harbor requires removing 18 identifier types, including full-face photographs and any comparable images, all date elements except year, medical record numbers and device identifiers and serial numbers, with no actual knowledge that the rest could identify the patient [1]. Expert Determination instead has a qualified expert document that re-identification risk is very small, which can let a team keep procedure dates or timing that Safe Harbor would strip [1]. Market commentary often frames AI training on health data as exactly this trade-off between a simple checklist and a documented, detail-preserving expert review [4].

A limited data set is a different path: it removes 16 direct identifiers, can keep dates and some geography, and requires a data use agreement for research, public health or health care operations [2]. It is not a route to a general commercial training license, and buyers should not treat it as one. Disclosing identifiable footage for payment would generally require patient authorization, which most archived surgical video does not have. For the method choice, see HIPAA Safe Harbor vs Expert Determination for AI training and what a HIPAA-compliant label must mean.

Research consent is the other trap. Footage collected under an IRB protocol or a research consent form was approved for that study, not for licensing to a third party for model training. Ask whether the release relies on de-identification, on a patient authorization that names commercial use, or on study consent, and get the document.

Where identifiers hide in surgical video files

Surgical video leaks identity through pixels, audio and containers, not just the obvious face. The failure modes below come up repeatedly when footage is reviewed frame by frame, and each maps to a Safe Harbor category [1].

• Burned-in on-screen display: capture towers often render patient name, medical record number, date of birth, procedure date and surgeon name into the first frames or a persistent corner banner. Because it is in the pixels, deleting metadata does not remove it; it needs OCR detection plus masking or cropping.
• Out-of-body frames: when the scope is withdrawn through the trocar for cleaning, the camera briefly films the room, staff faces, wristbands and whiteboards. These segments need detection and removal or blurring.
• Container and DICOM metadata: MP4 or MOV creation_time, device serial numbers, and DICOM tags such as PatientName, PatientID, StudyDate and InstitutionName survive transcoding unless stripped.
• Audio tracks: staff say names, ages and bed numbers during time-outs; most training uses drop audio entirely.
• Filenames and folder paths: archives are often organized by MRN or case date.
• Room cameras: faces of patients and staff, plus case boards. Recording staff faces also raises biometric questions; Texas, for example, treats face geometry as a biometric identifier and requires notice and consent before commercial capture [3]. See licensing video that contains faces and recording employees on video.

What to specify in a surgical video request

A useful request describes the footage and labels precisely so a holder can tell quickly whether it has a match. Name the procedure (for example laparoscopic cholecystectomy or colonoscopy), capture source (internal view only or room camera), resolution and frame rate, tower or robot platform if it matters, and case mix such as conversions, complications and difficult anatomy.

Labels matter as much as footage. Phase, step and instrument annotations, segmentation masks and skill ratings such as GOALS or OSATS scores are separate work products with their own rights; the surgical video annotation guide covers the schemas. Ask for de-identification evidence at the frame level, not a statement.

Illustrative example: invented to show structure; it does not describe an available dataset.

Diligence checklist before licensing surgical footage

Rights and de-identification diligence for surgical video should be done per file collection, not per vendor relationship. Use this list with counsel and your privacy reviewer.

1. Who holds the recordings and who controls them: hospital, physician group, or a capture-platform account whose terms limit export or secondary use?
2. What is the release basis: Safe Harbor, Expert Determination report, or patient authorization naming commercial use [1][2]?
3. Were any cases collected under an IRB study, and does the study consent restrict use?
4. How were burned-in overlays, out-of-body segments, audio and container metadata handled, and is there a per-case log?
5. Was a random sample of frames reviewed by a person after automated masking, and what did the reviewer find?
6. Are annotations licensed together with footage, and who annotated them?
7. Does the license define the records, allowed uses, term and delivery, and cover derived weights and labels?
8. Will delivery run through an access-controlled transfer rather than shared drives or email? See dataset delivery formats and transfer.

Sourcing surgical video datasets through SourceX

SourceX sources operational datasets from US companies on request and handles licensing; data is not held in stock, so a request does not guarantee a match. Health records must be HIPAA de-identified through Safe Harbor or Expert Determination, every dataset is rights-reviewed and licensed with defined records, uses, term and delivery, and each release is approved by the supplying organization. Describe the surgical video you need on the SourceX buyers page. For context on medical data, see healthcare buyers, do AI labs buy medical data, licensing medical records and licensing as a HIPAA-covered entity.

Describe the surgical video you need

SourceX finds US businesses that hold the data you describe, reviews rights and licensing permissions, and agrees pricing and allowed uses in a license before anything is delivered. Nothing is contracted until the supplier agrees. Start your request at https://sourcex.si/buyers.