shahedm2001/Catarct-Irregularity

Cataract Surgery Irregularity Benchmark (CSI-Bench)

CSI-Bench is a standardized, self-contained evaluation benchmark for Vision-Language Models (VLMs) on intraoperative ophthalmic irregularity detection and 3-class taxonomy classification. Derived from the clinical Cataract-1K dataset.

Dataset Overview

• Total Videos: 41 procedures (317.1 minutes of surgical video, 811.0 MB)
• Modality: Pure visual surgical video (all audio tracks completely stripped, silent H.264 @ 10 fps)
• Task: 3-Class Intraoperative Irregularity Classification with Clinical Explanation
• Evaluation Paradigm: 100% Deterministic closed-set MCQ scoring (no LLM judge needed)

Class Distribution

Option	Category	Cases	Percentage	Description
A	Normal	10	24.4%	Standard routine cataract extraction without capsular or pupil irregularity
B	Lens Irregularity	9	22.0%	Capsular tear, zonular laxity/dialysis, subluxation, abnormal lens morphology
C	Pupil Contraction	22	53.7%	Intraoperative miosis, floppy iris syndrome (IFIS), pupil margin irregularities

JSONL Task Envelope Schema

Each .jsonl file corresponds to exactly one .mp4 video and contains a single-line task envelope:

{
  "record_id": "case_XXXX_irregularity_classification",
  "task_category": "mcq",
  "question_type": "cataract_irregularity_classification",
  "reward_type": "deterministic",
  "split": "Evaluation",
  "track": "cataract-1k",
  "video": "case_XXXX.mp4",
  "category": "Normal | Lens Irregularity | Pupil Contraction",
  "prompt": [
    {
      "role": "user",
      "content": [
        {"type": "video", "video": "case_XXXX.mp4"},
        {"type": "text", "text": "You are given a cataract procedure video. Analyze the surgical video carefully, explain the visible intraoperative events, anatomical structures, and any observed irregularities, and then classify the case into one of the following 3 options: A) Normal B) Lens Irregularity C) Pupil Contraction. Respond ONLY with JSON with explanation (3-10 sentences) and answer (A, B, or C)."}
      ]
    }
  ],
  "correct_answer": "A | B | C",
  "reference_reasoning": "...",
  "metadata": {
    "case_id": "case_XXXX",
    "category": "...",
    "source_dataset": "cataract-1k",
    "duration_sec": 233.4,
    "width": 512,
    "height": 384,
    "fps": 30.0
  }
}

Citation

@inproceedings{Cataract-1K,
    author    = {Negin Ghamsarian and
                Yosuf El-Shabrawi and
                Sahar Nasirihaghighi and
                Doris Putzgruber-Adamitsch and
                Martin Zinkernagel and
                Sebastian Wolf and
                Klaus Schoeffmann and
                Raphael Sznitman},
    title     = {Cataract-1K: Cataract Surgery Dataset for Scene Segmentation, Phase Recognition, and Irregularity Detection (to appear)},
    
}