MIMIC-CXR-Diff

Hard-negative VQA pairs mined from MIMIC-Ext-VQA, a large-scale medical VQA dataset built on MIMIC-CXR chest X-rays.

Each row is a pair of visually similar images with the same question type but different answers — designed for evaluating and improving model robustness to subtle visual differences.

Processing Pipeline

1. Normalize the MIMIC-Ext-VQA training split (290K samples → 260K after removing invalid entries).
2. Embed all images using BiomedCLIP.
3. Mine pairs via approximate nearest-neighbor search (FAISS) on image embeddings.
4. Filter pairs:
   • Remove pairs where both sides reference the exact same image (image_similarity = 1.0).
   • Keep only pairs where the question type matches (content_type_a == content_type_b).
   • Keep only pairs where the answers differ (after whitespace stripping).

Result: 1,739,965 hard-negative pairs covering 109,415 unique images.

Schema

Column	Type	Description
pair_id	int	Unique sequential identifier (0-indexed)
image_path_a	string	Relative path to image A within the MIMIC-CXR directory
image_path_b	string	Relative path to image B within the MIMIC-CXR directory
question_a	string	VQA question for image A
question_b	string	VQA question for image B
answer_a	string	Ground-truth answer for image A
answer_b	string	Ground-truth answer for image B
meta	string (JSON)	Nested metadata (see below)

Meta Structure

{
  "question_type": "presence",
  "image_similarity": 0.9833,
  "question_similarity": 0.9118,
  "image_a": {
    "image_id": "d0d24188-dda41b64-...",
    "subject_id": "18386740",
    "study_id": "56503182",
    "semantic_type": "verify",
    "template": "Is the ${object} showing indications of ${attribute}?",
    "template_program": "program_2",
    "template_arguments": {"object": {"0": "right lower lung zone"}, ...},
    "mimic_ext_vqa_idx": 7040
  },
  "image_b": { ... }
}

Meta Field	Description
question_type	Shared question category (e.g., presence, attribute, abnormality)
image_similarity	BiomedCLIP cosine similarity between the two images
question_similarity	Sentence embedding cosine similarity between the two questions
image_a / image_b	Per-image metadata
.image_id	DICOM-derived image identifier
.subject_id	Patient identifier in MIMIC-CXR
.study_id	Study identifier in MIMIC-CXR
.semantic_type	Question semantic type (verify, query, choose)
.template	Question generation template
.template_program	Template program identifier
.template_arguments	Template slot-fill arguments
.mimic_ext_vqa_idx	Index into the original MIMIC-Ext-VQA train.json

Usage

from datasets import load_dataset
import json

ds = load_dataset("mtybilly/MIMIC-CXR-Diff", split="train")

row = ds[0]
meta = json.loads(row["meta"])
print(row["question_a"], "→", row["answer_a"])
print(row["question_b"], "→", row["answer_b"])
print("Image similarity:", meta["image_similarity"])

Note: This dataset contains metadata only. Images must be obtained separately from MIMIC-CXR via PhysioNet (requires credentialed access). Image paths are relative to the MIMIC-CXR root directory.