OralSurgery/dental-implant-surgery-sample

Dental Implant Surgery — Multimodal Annotated Video (Sample Case)

A public sample from one complete All-on-4 full-arch mandibular dental implant surgery: two synchronised camera angles, the operating surgeon narrating while he works, and six layers of structured clinical annotation (L0–L5) tied frame by frame to what he said.

This is a showcase slice, not the whole case. What is here is enough to judge the structure, the annotation quality and the honesty of the documentation.

Read LIMITATIONS before building anything on this. The video is edited, the material has had no face processing, and there are no instrument bounding boxes. None of that is hidden and none of it is cosmetic.

---

What is in this repository

Path	What it is	Files	Size
demo/	6 min 58 s annotated demo video, 1080 × 1540, no audio stream	1	186 MB
keyframes_annotated/	54 keyframes, each as an untouched original/ and an annotated/ render, plus their index	109	13 MB
annotations/	the L0–L5 annotation sample, the phase timeline, the closed vocabularies, whole-case statistics	4	459 KB
transcript/	de-identified surgeon narration, JSON + SRT	2	98 KB
lerobot_v2_1/	nine episodes in LeRobot v2.1 layout: meta/, 9 parquet tables, 18 video files, and its own README	32	189 MB
examples/load_dataset.py	loads and cross-checks every part of this repository	1	6 KB
README.md, LICENSE, SHA256SUMS.txt, .gitattributes	this card, CC BY 4.0 in full, a checksum of every file, and Git-LFS tracking for .mp4 / .jpg / .parquet	4	60 KB

Everything in this repository is on one clock: seconds in the delivered 54:53 case video, which is what the burned-in timer in the demo shows. Any timestamp here can be checked by eye against that timer.

The case

Procedure	All-on-4 / All-on-X, mandible, 5 implants
Delivered case video	3293.023 s = 54:53, 98 692 frames @ 30000/1001, 1080 × 1276, H.264, yuv420p
Layout	two camera panels stacked — overhead 1080 × 608 on top, lateral 1080 × 608 below, timer strip between
Cameras	two fixed cameras, 4K native, delivered as a 1080-class composite
Sensors beyond those two cameras and the narration	none
Keyframe images in the case	1050 — 1009 annotated, 41 auto-labelled as having no usable view
Annotated frames by camera	554 overhead, 455 lateral
Surgical phases	9, contiguous, covering 12 s → 3293 s

Annotation layers

Per annotated keyframe, counted over the whole delivered case:

Layer	What it holds	Volume
L0	verbatim timestamped surgeon statements, with a category	1070
L1	surgical phase — one of 9 normalised labels, plus free-text detail	1009 frames
L2	instruments present, each with a confidence and a provenance (visual / doctor_mentioned / both), plus the anatomy visible	4578 observations, closed vocabulary of 39 terms
L3	(instrument, action, target) triplets	1403
L4	reasoning — why the frame reads the way it does	969 frames
L5	clinical context around the frame	969 frames

L2 confidence splits 3531 certain / 1047 probable. L2 provenance splits 4337 seen in the picture / 227 both seen and named / 14 named by the surgeon only. Anatomy vocabulary: 21 terms. L2 carries no coordinates — see Limitations.

The annotation sample in this repository covers 184 annotated frames — every annotated keyframe inside the nine LeRobot episode windows, plus the 54 frames of the keyframe showcase — with 234 L0 statements and 848 L2 instrument observations, all six layers complete.

Phase timeline

Times are seconds in the delivered case video.

#	Phase	Start	End	Duration	Annotated frames
1	Incision and flap reflection	12	299	287 s	55	intact
2	Ridge smoothing and reduction (alveoloplasty)	299	552	253 s	77	intact
3	Osteotomy preparation	552	805	253 s	56	intact
4	Implant placement	805	1174	369 s	83	intact
5	Multi-unit abutments	1174	1788	614 s	270	intact
6	Cap placement and case summary	1788	1902	114 s	40	intact
7	Suturing	1902	2552	650 s	265	intact
8	Photogrammetric implant capture (PIC) scan	2552	2801	249 s	68	shortened by the edit from 630 s
9	Impressions, bite and registration	2801	3293	492 s	95	shortened by the edit from 688 s

Seven of the nine phases are complete and continuous — the whole surgical procedure from the first incision through suturing. Two later phases were condensed by the edit. A phase-duration model trained on this case without accounting for that would learn a wrong prior.

The demo video

demo/case_002_mandible_demo_7min.mp4 — 6:58, 45 seconds from each of the nine phases in order, with a title over the picture at the start of each.

Drawn on the picture: the L1 phase label and its number, the L2 instrument list of the current annotated keyframe (naming which camera's keyframe it belongs to, because the two cameras are annotated independently and interleave), and the surgeon's narration as subtitles. No bounding boxes are drawn — this case has no instrument coordinates, and drawing boxes would mean inventing data.

On the resolution. The file is 1080 × 1540, not 1920 × 1080. That is the native geometry of the material: a 1080 × 1276 two-panel composite plus the annotation bands. It is 1080-class — each camera panel is 1080 px wide — but it is portrait, not 16:9. Fitting the same frame inside 1920 × 1080 would mean scaling it to 757 px wide, throwing away 30 % of the linear resolution of the surgical field to gain nothing but an aspect ratio. It is shipped at native size instead.

The video has no audio stream at all — not muted, no stream. The narration is delivered as text.

Annotated keyframes

keyframes_annotated/ holds 54 keyframes as 108 images — each frame both as the untouched original and as an annotated render.

Selection: six frames per surgical phase, three from each camera, spread evenly in time across the phase. A frame qualifies only if it is labelled image quality ok, shows a surgical field, carries at least two L2 instruments, names the anatomy visible, and has an L4 reasoning layer written for it — so every image in the showcase has something to show and something to read against it. All nine phases and both cameras are covered, 18 phase × camera cells with three frames each.

Overlay: phase, phase number, camera, video timestamp and the complete L2 instrument list with confidence and provenance are drawn on bands added above and below the picture, so no annotation ever covers the surgical field. The first L0 statement attached to the frame is quoted underneath. keyframes_annotated/keyframes_sample_index.json carries the same content as data.

LeRobot v2.1 sample — and an honest note about the format

lerobot_v2_1/ holds 9 episodes, one per surgical phase, 45.0 s each at 30 fps = 1350 frames per episode, 12 150 rows in total, with the two camera panels split into their own video streams as the format requires: 18 video files, 1080 × 608 each.

Two things you should know before you load it:

1. This is human surgery, not robot teleoperation. LeRobot's data model is built around (observation, action) pairs from a robot. There is no robot here, no kinematics, no joint states, no gripper commands. Rather than fabricate an action vector, this dataset declares no action and no observation.state feature at all. The observations are the two camera streams; the labels are the surgical phase and the join keys back into the annotation corpus. A policy cannot be trained on it as-is, and it is not offered for that. It is offered as a format-compatible carrier for surgical video and its annotation.
2. The layout follows the published v2.1 specification but was not run through the reference validator. lerobot_v2_1/README.md states exactly what was checked, what was not, and the two commands that would close the gap.

Transcript

transcript/or_transcript_deid.json and .srt — 204 segments, English.

This is not raw model output. It is a whisper.cpp large-v3 automatic transcription of the operating-room narration that was then de-identified and then reviewed term by term against the audio by the operating surgeon. Known recognition errors that remain are listed in the file rather than silently patched, and low-activity segments are flagged rather than deleted.

Loading

No third-party packages are needed to read anything here:

python examples/load_dataset.py            # loads and cross-checks every part

Reading the LeRobot slice with the library, once it is installed:

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("<org>/<repo>", root="lerobot_v2_1")
print(ds.meta.info["codebase_version"], ds.num_episodes, ds.num_frames)

Reading only the per-frame tables, with datasets:

from datasets import load_dataset
d = load_dataset("<org>/<repo>")          # the configs block points at lerobot_v2_1/data

Joining an annotation to its picture:

import json
ann = json.load(open("annotations/annotations_sample_L0_L5.json"))
row = ann["frames"][0]
row["keyframe_id"]            # stable identity, appears in every file here
row["video_timestamp_sec"]    # seconds in the 54:53 case video, = the burned-in timer

The keyframe filenames in keyframes_annotated/ encode the phase and the video timestamp. The keyframe_id is the stable join key and contains no time at all.

---

Limitations

Every one of these is a property of the material. They are stated here so they are not discovered later.

No instrument bounding boxes. L2 is a confirmed list per frame with confidence and provenance. There are zero coordinates in this case. Nothing in this repository draws a box, and any document that promises boxes for this case is wrong.

No depth stream, no navigation telemetry, no robot kinematics, no force or pose data. Two RGB cameras and the narration. That is the whole sensor set.

The video is edited. Six stretches were removed from the recording and the remaining six pieces concatenated: 4370.9 s of source became 3293.023 s. The burned-in timer was redrawn so it counts the delivered file continuously from zero; there are no time jumps in it. The joints are hard cuts. 47 annotated keyframes fell inside removed windows and are not in the case at all, together with 40 auto-labelled ones and 25 transcript segments. Phase pic_scan runs 249 s where the recording had 630 s, and impressions_and_bite 492 s where the recording had 688 s; the other seven phases are untouched.

No face processing has been applied. This is the most important line on this page. The delivered picture has had no blur, no mask and no pixelation. Problem passages were handled by removing them. What remains visible:

• the patient's lower two thirds of the face in individual passages — the eyes are never visible anywhere in this material, but the mouth, chin, lips and the surrounding skin are, at full resolution and in motion;
• skin texture — pores and stubble resolve in the closer passages;
• the dental arch including existing restorations, visible on the maxillary teeth throughout.

The keyframe images are unprocessed originals for the same reason: blurred stills next to an unblurred video would be incoherent.

The audio is not here. The video files carry no audio stream at all, and the sound recording is not distributed in this sample. The surgeon's speech reaches you as transcript text and as verbatim L0 quotes. In the licensed case the audio exists as a separate file, matching the video cut for cut.

The lateral camera stops recording 238 s before the end. From 50:55 to the end at 54:53 the lower panel is black. This is a property of the original recording, not of the edit, and it is why there are no annotated lateral frames after 3031.2 s. A second, shorter passage — 5:55 to 6:40 — has the lower panel deliberately filled black.

Image quality is labelled, not filtered. Of the 1050 images in the case: 913 ok, 96 degraded (motion-blurred but readable), 41 unusable (empty or black). Nothing was discarded. 47 of the 1009 annotated frames are marked as showing no surgical field.

Drill sub-types are only partly resolved. The drill protocol is a confirmed case fact; the per-frame drill diameter is not. The manufacturer's ring colour identifies the implant line, not the diameter of the hole, and no keyframe shows a legible number. One observation is openly inconsistent with the case protocol and is published that way rather than smoothed over.

The transcript is machine-produced. Medical terms are sometimes mangled; known recognition errors are listed in the file. Verbatim L0 quotes were checked against the audio, but the transcript as a whole was not rewritten.

Free-text prose can quote an older clock. Timecodes embedded inside narrative strings were written before the edit and were not rewritten. Every structured time field in every file here is on the delivered video's clock.

This is one case. It is a showcase, not a corpus. Nothing here supports a claim about inter-case variability, surgeon variability or generalisation.

Compliance

• WCG IRB, Protocol / tracking number 20261848, Study #1410330. Approved 2026-06-22, current through 2027-05-29 with annual continuing review.
• Prospective, per-subject written informed consent, covering video, optional audio narration, and de-identified use for research and commercial development and licensing.
• HIPAA Safe Harbor de-identification against all 18 identifiers. Two production slates carrying direct identifiers appear in the source recording; both are blacked out in every derived video, the 26 keyframe images on which those identifiers were readable are on a permanent export blocklist and are not in this repository, and the same identifiers were removed from the annotation free text. File and container metadata are stripped; the images carry no EXIF.
• California Penal Code §632 audio-recording consent is covered by the same consent process.
• Face handling is an open item, not a closed one. The Principal Investigator determined that the patient's mid and lower face is a "comparable image" under Safe Harbor §17, which makes face processing an obligation rather than a precaution. This material has had none applied — see Limitations. Whether the cut-rather-than-mask approach satisfies that determination is a decision that has to be recorded in writing and has not been.

Signed consent documentation is never distributed. A one-page compliance summary, and at deal stage the IRB approval letter, are available on request.

No personal name appears anywhere in this repository — not the patient's, not the staff's, not the investigator's. The approval above can be verified from the protocol and study numbers alone.

Licensing

• This sample: CC BY 4.0. Free, attribution required. Full text in LICENSE.
• The full case — both camera angles at full length, all annotated frames, the complete transcript, the case facts, the audio track, and a measured comparison against a blind vision-only annotation pass of the same images — is available under a negotiated commercial licence and Data Use Agreement. There is no public price list.
• Recording capacity, stated as capacity and not as delivered inventory: the approved protocol supports on the order of 800–1400 hours of new recording per year. This repository is one case.

Contact: andrii@drantipov.com

Citation

No DOI or formal citation has been minted for this dataset. Rather than leave the question open: there is currently nothing to cite beyond this repository's URL. If you need a citable identifier, ask at the contact address above.

Provenance of the numbers on this page

Every figure above was recomputed from the delivered annotation corpus and the delivered video at package build time, not copied from planning documents.

Two counts of the same case exist, and both are correct. The figures on this page are counted over the delivered 54:53 edit: 1009 annotated keyframes, 1070 surgeon statements, 1403 action triplets. Documents describing the working corpus — which spans the full 64:18 recording, before the passages removed for de-identification — count 1056 keyframes, 1095 statements and 1434 triplets. The difference is the 47 annotated keyframes that fall inside removed stretches; they are not in this repository and are not counted here. If you are comparing this card against a proposal or a licence schedule, check which of the two timelines the number is on before treating a mismatch as an error. The planning documents for this project promise instrument bounding boxes, a depth stream and navigation telemetry; this case has none of the three, and the numbers here are the ones to trust.