Dataset Overview
Definition：Million-Scale Neurosurgical MVD Surgery Annotation Datasetis provided byChangsha Langhui Information Technology Co., Ltd.The database is divided by the syntax of ultra-large surgical video created by the DataAssetsAPI platform. A million + frames from the microvascular decompression (MVD) microscope of the cooperative medical facility, modified manually by SOLOv2 pre-mark + neurosurgery specialists, covering 24-type anatomical structures and pixel-class severing of surgical devices, with a multi-level label system of phase 9 + 8.
| Dataset Name | Million-Scale Neurosurgical MVD Surgery Annotation Dataset |
| Total Data Volume | 1 Million+ Surgery video frame (segmentation by frame) |
| AnnotationCategory | 24 Semantics (Type-Anatomy Structure + 12 Operating Devices) |
| Stage Tabs | 9 surgical stages (specific labels per frame), 8 surgical intent (multilabelling) |
| Annotation Format | Pixel Separator Mask (polygon / RLE / binary mask) |
| Annotation Platform | CVAT with SOLOv2 pre-annotation + manual expert correction |
| MetadataField | Certainty score (0-1), occlusion level (0-2), visibility (0-100), quality flag |
| Quality markers | No Abnormality / Blurred / Smoke / Blood Splatter / Underexposed, recorded frame by frame |
| Point Precision | Boundaries must precisely fit the tissue contour; rough edges and approximate polygons are not permitted |
Description category system
Anatomy structure (12 categories)
Operating equipment (12 classes)
Surgery phase and action label
Surgical phase (exclusive label per frame)
Intentions for surgery (multilabelled)
Imaging Data Specifications
| Specification Item | Detailed Description |
|---|---|
| Data Format | Surgeon microscope video frame to support PNG/JPEG sequence export |
| Annotation Format | Pixels Separator Mask, supports pallygon / RLE / Binary Mask |
| Annotation Platform | CVAT with SOLOv2 pre-annotation + manual expert correction |
| MetadataField | Certainty score (0-1), occlusion level (0-2), visibility (0-100), quality flag |
| Quality markers | No Abnormality / Blurred / Smoke / Blood Splatter / Underexposed, recorded frame by frame |
| Point Precision | Boundaries must precisely fit the tissue contour; rough edges and approximate polygons are not permitted |
Annotation Quality System
Video collection and dissemination
Surgery microscope video removes patient information, dates, institutional identification and enters the database through de-sensitive water lines
Key frame extraction
Take critical frames of the operation in accordance with the 1-5fps sampling strategy to ensure coverage of the full process of 9 stages of the operation
SOLOv2 prescript
AI model to generate initial semantic partition mask to provide baseline labels for 24-type anatomical structures+ devices
Expert amendments and cross-checking
Quick-checked by neurosurgery specialist, double-checked, non-conforming frame returned to re-mark
Data Security and De-identification
All data have been thoroughly de-identified, all information that directly or indirectly identifies individuals has been removed and only the autopsy structure and device information within the surgical horizon is maintained.
| Processing Type | Detailed Description |
|---|---|
| De-identified Fields | All personal identification information, including patient name, ID, date of operation, medical facility identification, etc. |
| Reserved Fields | Video frame pixel data, semantic partition mask, phase/action/intentional tag, uncertainty metadata |
Field Data Dimensions
| Field Name | Data Type | Dimension Description |
|---|---|---|
| frame_id | String(32) | Video frame unique sign, associated with the original surgery video time stamp |
| video_id | String(24) | Video number of disembedded surgery, shared with the same operation |
| frame_index | Integer | Video frame serial number, corresponding to the original video location |
| timestamp_sec | Float | Time stamp (sec) in video, calculated from the beginning of the operation |
| surgery_phase | Enum | EXPOSURE | ARACHNOID_OPENING | CEREBELLAR_RETRACTION | VESSEL_ID | CONFLICT_ID | DECOMPRESSION | TEFLON_PLACE | HEMOSTASIS | CLOSURE |
| surgery_intent | Array[Enum] | IMPROVE_VIEW | BLEEDING_CONTROL | ADHESION_RELEASE | sample_IDENTIFY | VESSEL_SEPARATE | DECOMPRESS | NERVE_PROTECT | SPACE_CLEAR |
| anatomy_mask | Polygon/RLE | 12-type anatomical structure pixel semantic partition mask |
| instrument_mask | Polygon/RLE | 12-type surgical devices with semantic pixels partition mask |
| certainty_score | Float | Noted definitive rating 0.0-1.0 below 0.7 subject to review |
| occlusion_level | Enum | 0=NONE | 1=PARTIAL | 2=SEVERE |
| visibility_score | Integer | Key structural visibility 0-100 |
| quality_flag | Enum | NORMAL | BLUR | SMOKE | BLOOD_SPLASH | UNDEREXPOSED |
| annotator_id | String(8) | De-sensitization markers personnel ID |
| review_status | Enum | PENDING | PASSED | REJECTED | ARBITRATED |
| resolution | String | Video frame resolution "1920x1080" or "3840x2160" |
| fps | Integer | Frame sampling rate, usually 1-5fps key frame |
De-identified JSON Data Sample
The structured JSON model for de-identified treatment is presented below, real patient information has been removed and core medical data is maintained.
Other Organiser
"frame id": "MVD-20240815-0128-F03456",
"video id": "MVD-SURG-20240815-0128",
"frame index": 3456,
"timestamp sec": 1847.5,
"surgery case": "DECOMPRESSION",
"surgery intent": ["DECOMPRESS", "NERVE PROTECT", "VESSEL SEPARATE"]
"anatomy mask":
"visible structures": ["CERBELLUM", "sample5 TRIGEMINAL", "SCA", "PETROSAL VEIN", "BRAINSTEM",]
"Mask format": "polygon",
"mask data": "base64 encoded polygon coordinates"
{\cHFFFFFF}{\cH00FFFF}
"instrument mask":
"Visible instruments": ["BIPOLAR FORCEPS", "DSSECTOR", "TEFLON PLEDGET",]
"Mask format": "polygon",
"mask data": "base64 encoded polygon coordinates"
{\cHFFFFFF}{\cH00FFFF}
"Certainty score": 0.92,
"Occession level": "PARTIAL,"
"visibility score": 85,
"Quality frag": "NORMAL",
"Annotator id": "ANN-0472,"
"review status": "PASSED",
"resolution": "1920x1080,"
"fps": 2,
"case description": "The pressure-relief phase: separation of the responsible blood vessels from the trident nerve root with a stripper, preparation for the placement of a Teflon pad."
♪ I'm sorry ♪Data Distribution Statistics
| Dimension | Distribution Description |
|---|---|
| Distribution of surgical sites | The trident neuroses are 68% MVD, thalamus, 22% MVD, thalamus, 6% MVD, 6% other skull neurons, 4% |
| Responsible blood vessels distribution | SCA 47% ⁇ AICA 28% ⁇ vertebrate base artery 12% ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ 5% ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ 5% |
| Video time distribution | 2-4 hours 55% ⁇ 1-2 hours 28% ⁇ 4-6 hours 15% ⁇ 6 hours or more 2% |
| Resolution distribution | 1920x1080 78% | 3840x2160 22% |
| Frame Rate | 2fps Key frame sampling 65% ⁇ 1fps 25% ⁇ 5fps High density sampling 10% |
AI Training Use Cases
Basic model for understanding the surgical scene
Applying model: Mask2Former / MedSAM / Swin-UNETR / YOLOv8-Seg. Multi-syntax split + examples split, training in basic surgical scenario models.
Automatic recognition during surgery.
Applying model: SlowFast / TimeShormer / VidioMAE. Phase 9 classification + 8 is intended to be multiple labels, enabling automatic recording of the surgical process and critical event detection.
Real-time navigation aids in surgery
Applicable model: YOLO-Surgical / DeepLabV3+. Key structure split in real time + device tracking + hazard area alert.
Assessment of surgical skills
Applying model: OSATS automatic / graph neural network. Device track analysis + operational normative + learning curve modelling.
Auto-generated surgical report.
Applying model: Multimodular LLM. Video-text polymodular training, automatically generating structured surgical records.
Surgery teaching and simulation training
Applying model: VR surgical simulation data source/Standard-style reference library/Murder Complication Early Warning Training.
Frequently asked questions
Domain expert network
Covering six major disciplines: mathematics, physics, chemistry, biology, earth sciences, and computer science
Academic credential review, professional competency assessment, data annotation training, and continuous quality monitoring
AI pre-annotation + expert review, 300% efficiency improvement, 99.5% accuracy
Need customised surgical data programs?
Changsha Langhui Information Technology Co., Ltd. The DataAssetsAPI platform is dedicated to providing AI enterprises and research institutions with high-quality, compliant medical data assets.
Data Sample Preview · Customized Data Solutions · Professional Technical Support