CausalMedBench

CausalMedBench is an independent downstream benchmark created for this project. It was not proposed or released as part of the MedForge paper or the official MedForge release. CausalMedBench evaluates whether a medical deepfake detector's final decision and explanation are supported by the visual evidence it claims to use.

The upstream authentic images, forged images, manipulation boxes, and real--forged pairing metadata come from MedForge-90K. CausalMedBench introduces the five causal views, fixed splits, evidence-dependency tasks, explanation faithfulness protocol, evaluation metrics, and public scorer. It contains 8,212 source-linked groups (41,060 five-view examples) across brain MRI, chest X-ray, and fundus imaging.

The fixed held-out benchmark contains 1,840 groups (9,200 five-view examples), an ordinary detection set of 6,501 images, and a stratified explanation subset of 300 groups (1,500 examples).

Five causal views

Each group contains:

1. forged_original: the original forged image.
2. authentic_source: the corresponding image before manipulation.
3. local_restored: the forged image with the annotated manipulated box restored from the authentic source; the rest of the forged image is unchanged.
4. local_forged_only: the authentic image with only the annotated forged box inserted; the rest of the image is authentic.
5. authentic_nuisance: the authentic image after a deterministic mild JPEG, blur, denoising, or resize-resampling perturbation that preserves its medical content.

Together, these views test ordinary detection, local evidence sufficiency, global/outside evidence sufficiency, benign-perturbation invariance, five-view consistency, and explanation–evidence alignment.

Load the benchmark

from datasets import load_dataset

five_view = load_dataset("leaf0932/CausalMedBench", "five_view")
ordinary = load_dataset("leaf0932/CausalMedBench", "ordinary", split="test")

print(five_view)
print(five_view["test"][0])

The five_view configuration has train, validation, and test splits. The ordinary configuration contains the fixed 6,501-image detection test split. Images are embedded in Parquet shards, so no local MedForge path is required.

Prediction format

Classification predictions are JSONL records keyed by record_id:

{"record_id":"pair_001174f8d03bfe37::forged_original","verdict":"FAKE"}

For the 1,500 records where is_reasoning_subset is true, structured explanations use:

{"record_id":"pair_001174f8d03bfe37::forged_original","verdict":"FAKE","evidence_type":"mixed","evidence_location":"mixed","evidence":"The localized anatomy is inconsistent and the surrounding texture also differs."}

Allowed evidence types are none, local_medical, local_forensic, local_medical_and_local_forensic, global_forensic, and mixed. Allowed locations are none, local_bbox, global_or_outside, and mixed.

Score predictions

git clone https://huggingface.co/datasets/leaf0932/CausalMedBench
cd CausalMedBench
python -m pip install -r requirements.txt

python evaluation/score.py \
  --ordinary-predictions /path/to/ordinary_predictions.jsonl \
  --five-view-predictions /path/to/five_view_predictions.jsonl \
  --reasoning-predictions /path/to/reasoning_predictions.jsonl \
  --output metrics.json

The scorer reports:

• ordinary Accuracy and Macro-F1, including modality, forgery-type, and generator breakdowns;
• five-view accuracy and all-five-correct consistency;
• local and global evidence sufficiency;
• benign nuisance invariance and authentic recovery;
• explanation verdict, evidence-type, evidence-location, and joint alignment.

Run python evaluation/score.py --help for local dataset and strict-coverage options. A submission template is included in examples/.

Dataset fields

Important fields include record_id, group_id, view_id, target_label, image, bbox_normalized_1000, modality, forgery_type, generator, disease, operation, evidence_available, evidence_removed, and is_reasoning_subset. Bounding boxes use a 0–1000 normalized coordinate system.

The released masks are rectangular box rasterizations, not pixel-precise segmentation masks. The locally composed views use an 8-pixel feathered box boundary. Nuisance operations are deterministic and recorded per row.

Source and license

CausalMedBench uses source data from RichardChenZH/MedForge-90K. The data, including redistributed and derived images, is released under the Computational Use of Data Agreement (C-UDA). Users must comply with C-UDA and preserve the upstream attribution. Benchmark code in evaluation/ and scripts/ is provided under Apache-2.0. See LICENSE and DATA_LICENSE.md for details.

The MedForge citation acknowledges the upstream data source only. It does not attribute CausalMedBench, its five-view construction, its evaluation protocol, or experiments conducted with CausalMedBench to the MedForge authors.

This benchmark is for research and educational use. It is not a clinical diagnostic tool and must not be used to make patient-care decisions.

Citation

Please cite CausalMedBench for the benchmark, five-view construction, evaluation protocol, metrics, and scorer:

@misc{you2026causalmedbench,
  title={CausalMedBench: Evaluating Evidence Dependency and Explanation Faithfulness in Medical Deepfake Reasoning},
  author={You, Maowei},
  year={2026},
  howpublished={Hugging Face dataset},
  url={https://huggingface.co/datasets/leaf0932/CausalMedBench},
  version={1.0.0}
}

The manuscript citation will replace the repository citation after the paper is publicly archived. Because the released images and source annotations originate from MedForge-90K, also cite MedForge as the upstream data source:

@misc{chen2026medforge90k,
  title={MedForge-90K},
  author={Chen, Zhihui and He, Kai and Lei, Qingyuan and Pu, Bin and Zhang, Jian and Xu, Yuling and Feng, Mengling},
  year={2026},
  howpublished={Hugging Face dataset},
  url={https://huggingface.co/datasets/RichardChenZH/MedForge-90K},
  note={Upstream source dataset used to construct CausalMedBench}
}

These two citations serve different purposes: the first credits CausalMedBench; the second preserves the required attribution for the MedForge-90K source data.