EndoSCOP-V

EndoSCOP-V (Endoscopic Screening and Classification of Pathologies — Video) is a video-grounded, multi-turn question-answering benchmark for structured endoscopy reporting and clinical reasoning. It pairs 1,307 real-world upper-GI endoscopy (EGD) and colonoscopy videos with 20,274 question–answer turns. The turns are generated deterministically from clinician-filled structured endoscopy reports. No question or answer is synthetic or rephrased by an LLM.

Accepted at NeurIPS 2026.

• Paper: EndoSCOP-V: A Multi-Turn Video Understanding Evaluation Framework for Multimodal Models in Endoscopy Reporting and Clinical Reasoning, NeurIPS 2026 https://arxiv.org/abs/2611.10626
• Data (canonical record): Harvard Dataverse, https://doi.org/10.7910/DVN/HRHPEX
• Code and evaluation pipeline: https://github.com/abhi25t/EndoSCOP-V
• License: CC BY-NC-SA 4.0
• Contact: Abhishek Tyagi

Dataset summary

Quantity	Value
Cases (videos)	1,307 (1,004 EGD / 303 colonoscopy)
Total video duration	63.2 h; mean clip ~174 s
Resolution	1,269 at 1080p / 38 at 720p (97.1% / 2.9%)
Diseases	54 WEO MST 3.0 diseases + 4 non-disease findings
Question templates	1,033 clinician-reviewed (490 binary, 356 single-choice MCQ, 187 multi-select MCQ)
Question–answer turns	20,274
Turns per case	mean 15.5, median 14, max 73
Multi-disease cases	291 (22.3%), up to six diseases per case
Disease-Rejection (DI-R) questions	3,213 (15.8%)
Clinician baseline (15 endoscopists, ~1,500 questions)	77.9% strict accuracy
Five-rater consensus subset	100 cases, 2,148 questions; Krippendorff's α = 0.77

Files

File	Size	Contents
qa.zip	2.6 MB	1,307 <case_id>_qa.json files: the multi-turn question–answer dialogue for each case
case_reports.zip	946 KB	1,307 <case_id>_case_report.json files: the clinician-filled structured EHR report for each case
disease_ehr_question_templates.zip	147 KB	58 YAML question templates (54 diseases + 4 non-disease findings)
consensus_gt_reannotation.zip	1.4 MB	Five-rater consensus re-annotation of a 100-case subset
sample52.zip	2.4 GB	A 52-case subset covering all 58 diseases and findings, for quick inspection
videos/<case_id>.mp4	282 GB total	The 1,307 procedure videos
ENDOSCOP-V_croissant.json	—	Croissant 1.0 metadata with Responsible-AI fields

Case IDs start with e (EGD) or c (colonoscopy), for example e2493 or c1001.

QA files

Each <case_id>_qa.json holds one multi-turn dialogue:

{
  "case_id": "e2493",
  "video": "e2493.mp4",
  "procedure_type": "endoscopy",
  "diseases_found": ["Schatzki Ring", "Hiatus Hernia"],
  "conversation_model": {
    "total_turns": 22,
    "turn_11_prefix": "This endoscopy video contains Schatzki Ring as seen around frame number 889 ...",
    "...": "..."
  },
  "questions": [
    {
      "question_number": 1,
      "turn": 1,
      "phase": "disease_identification",
      "disease": "Schatzki Ring",
      "task": "DI",
      "subtask": "DI-M",
      "type": "MCQ",
      "multi_select": false,
      "stem": "Given this endoscopy video clip, what condition would an expert endoscopist suspect among the given options?",
      "options": {"A": "Normal", "B": "Esophageal Web", "C": "Schatzki Ring", "D": "Inlet Patch"},
      "correct": "C",
      "answer_schema": {"type": "object", "properties": {"answer": {"type": "string", "enum": ["A", "B", "C", "D"]}}, "required": ["answer"]}
    }
  ]
}

Each question records its task and sub-task, its format (Binary or MCQ, with multi_select for multi-select MCQs), the gold answer, and a JSON schema for the reply. Turn prefixes reveal a disease or a parent answer before the questions that depend on it; this implements the oracle-forcing protocol described below.

Case reports

Each <case_id>_case_report.json contains MetaData (procedure type, patient age and gender, video file name, and the clinician-marked keyframe segmentationFrame). It also contains report: the structured EHR fields filled by the endoscopist, organised by disease → section → attributes. All questions are instantiated from these fields.

Consensus re-annotation

consensus_gt_reannotation.zip contains:

• consolidated_gt.csv: one row per question, with the five raters' answers (rater_original, doctor_1 … doctor_4), n_agree, the consensus label (≥ 3 of 5), original_matches_consensus, and status;
• doctor_responses/: per-rater answer files for each of the 100 cases.

The subset holds 75 cases sampled to match the full benchmark on procedure mix and task share, plus 25 cases over-sampled for the lowest-agreement sub-tasks. 2,044 of the 2,148 questions receive a consensus label; the 104 ties are excluded. The recorded label matches the consensus on 90.7% of questions.

Task taxonomy

Every question belongs to one of 9 tasks and 30 sub-tasks.

Task	Sub-tasks
DI Disease Identification	DI-M MCQ Identification, DI-C Confirmation, DI-R Disease-Rejection, DI-N Normal Assessment
AL Anatomical Localization	AL-O Organ, AL-P Part / Segment, AL-W Wall
MA Morphological Assessment	MA-T Shape & Type, MA-S Surface, MA-C Color, MA-M Margins, MA-SZ Size
PF Pathological Feature Assessment	PF-D Distribution & Extent, PF-M Mucosal Features, PF-U Ulcer / Erosion Features, PF-S Tissue Features, PF-B Contents & Base
QN Quantification	QN-L Lesion Count, QN-S Structure Count
GS Grading & Scoring	GS-MC Morphological Classification, GS-SS Severity Scoring, GS-EC Extent Classification, GS-AG Anatomical Grading
CF Complication Recognition	CF-B Bleeding, CF-S Stricture & Obstruction, CF-A Associated Findings
CR Clinical Reasoning	CR-E Etiology & Context, CR-I Impression & Activity
AR Action Recognition	AR-B Biopsy, AR-V Visualization

Four sub-tasks are video-exclusive: CF-B, CF-S, AR-B and AR-V.

Evaluation protocol

• Input. Models receive uniformly sampled frames (or video, where the model supports it). The clinician-marked keyframe is used only to build the dataset and is never given to a model.
• Oracle forcing. Each turn is graded on its own. Before the next question, the gold answer to a parent question is injected as a prefix. The dialogue history is pruned (Markovian), so an error at one turn does not cascade into later ones.
• Scoring. Binary and single-choice MCQ questions use exact match. Multi-select MCQs are scored both strictly (exact set match) and with partial credit. Strict accuracy is the primary metric. Results are also reported without DI-R items (NDR), because these binary questions are easy for models.

The full pipeline, including providers for open-weight and API models, frame extraction, scoring and Excel reporting, is at https://github.com/abhi25t/EndoSCOP-V.

Baseline results (strict accuracy, %)

	Overall	NDR
Clinician panel (15 endoscopists)	77.9	77.9
Best open-weight model (Qwen3.6-27B)	59.8	53.4
Best frontier proprietary model, spot-check (Gemini 3-flash)	61.4	55.1

The paper gives results for 14 open-weight models, broken down by task, sub-task and disease.

Data collection and annotation

• Source. Archived EGD and colonoscopy procedures at the Asian Institute of Gastroenterology (AIG) Hospital, Hyderabad, India. Collection was approved by the Institutional Ethics Committee (AIG/IEC-Post BH&R 81/03.2026-R1).
• Recording. Procedures were recorded at 1920 × 1080 on Olympus endoscopes (CV-170/190/290, CV-1500) and two pediatric scopes (XP170N, H190DL).
• Sampling. Cases were sampled per disease, targeting about 30 cases for each of the 54 diseases. Because 35 diseases occur only in the upper GI tract, the corpus is weighted towards EGD.
• De-identification. On-screen patient identifiers were cropped from every video before release. Only patient age and gender are retained.
• Annotators. 15 board-certified endoscopists (2 women and 13 men, 2–25 years of endoscopy experience).
  • Each endoscopist filled the hospital's structured EHR report for a disjoint subset of cases, with access to reference material.
  • The same endoscopists reviewed every question template, rewrote question stems and wrote every distractor.
  • Templates were first drafted by an LLM-assisted converter from the EHR field labels; 1,033 of 1,588 drafts were retained after review.

Intended uses and limitations

Intended use. Evaluating multimodal models on structured endoscopy reporting from video, including multi-turn, per-turn diagnostic reasoning. The dataset can also support research on multi-turn evaluation methods.

Not intended for clinical decision-making, or for training or validating a system for clinical deployment without independent validation.

Limitations:

• Single source. All cases come from one institution, and most patients share one ethnicity.
• Single-rater ground truth. Each case's ground truth was filled by a single endoscopist. Only the 100-case consensus subset has five raters.
• Same clinicians throughout. The 15 annotators also provided the clinician baseline. No endoscopist answered a case they had annotated.
• Structured answers only. Free-form narrative reports, calibrated uncertainty and ranked differential diagnoses are not evaluated.
• Clip selection. Clips favour pre-lavage, quick-movement segments. Models may score higher on slow, lavaged footage.

License

CC BY-NC-SA 4.0: non-commercial use, with attribution, and derivatives shared under the same license.

Citation

@inproceedings{tyagi2026endoscopv,
  title     = {{EndoSCOP-V}: A Multi-Turn Video Understanding Evaluation Framework for Multimodal Models in Endoscopy Reporting and Clinical Reasoning},
  author    = {Tyagi, Abhishek and Jain, Tanmay and Biradar, Vilas and Koduri, Krithi Krishna and Masoodi, Misba Hul Bashir and Rapolu, Swetha and Rughwani, Hardik and Kotha, Rachana and Kaipa, Anjan Kumar and Katrevula, Anudeep and Garlapati, Rakesh and Singh, Aniruddha Pratap and Patel, Rajendra and Tyagi, Abhay and Lakhtakia, Sundeep and Ramchandani, Mohan and Kalapala, Rakesh and Reddy, D. Nageshwar},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {39},
  year      = {2026},
}