PeiyuanHao/MedStyleAudit-Experiments

MedStyleAudit Experiments

This repository contains experimental artifacts, metadata, intermediate results, and evaluation outputs associated with MedStyleAudit, a framework for detecting acquisition-style shortcuts in medical imaging models.

The repository is intended to support the reproducibility and transparency of the experiments reported in the MedStyleAudit study.

Overview

Medical imaging models may exploit acquisition-related, institution-specific, or domain-specific visual characteristics instead of clinically meaningful features.

MedStyleAudit is designed to identify and quantify these shortcut behaviors while preserving clinically relevant regions of interest.

This Hugging Face repository serves as the experimental artifact repository accompanying the MedStyleAudit project.

It may contain:

• experiment metadata;
• dataset split information;
• source and target domain definitions;
• paired or matched sample information;
• model predictions and evaluation outputs;
• intermediate statistical results;
• bootstrap results and confidence intervals;
• aggregated results used to generate tables and figures in the paper.

Repository Purpose

This repository is primarily intended to store experiment-related artifacts rather than redistribute complete medical imaging datasets.

The files are designed to support:

• reproducibility of MedStyleAudit experiments;
• verification of reported experimental results;
• inspection of intermediate outputs;
• comparison between different evaluation settings;
• analysis of acquisition-style and domain-related shortcut behavior.

Repository Structure

The repository is organized according to the datasets and experimental settings used in MedStyleAudit.

Different directories may correspond to:

• different benchmark datasets;
• different hospitals or acquisition domains;
• source-target domain pairs;
• model configurations;
• random seeds;
• evaluation protocols;
• intermediate or final experiment outputs.

The exact structure of each experiment directory follows the corresponding MedStyleAudit experiment configuration.

Experimental Setting

MedStyleAudit evaluates whether medical imaging models rely on acquisition-style or domain-specific visual cues.

Experiments may involve comparisons between different hospitals, acquisition environments, or dataset domains.

Depending on the experiment, the repository may contain results for:

• source-to-target evaluation;
• balanced paired evaluation;
• random paired evaluation;
• common-support evaluation;
• pair-weighted aggregation;
• shortcut-related performance measurements;
• bootstrap uncertainty estimation;
• confidence interval estimation.

These artifacts are intended for experiment verification and reproduction rather than use as a standalone medical imaging benchmark.

Original Medical Data

Original third-party medical imaging datasets are not redistributed in this repository unless their licenses explicitly permit redistribution.

Users who wish to reproduce the complete experiments should obtain the original datasets from their official sources and comply with the corresponding licenses, terms of use, and access requirements.

This repository primarily contains derived experimental artifacts, metadata, evaluation outputs, and statistical results generated for the MedStyleAudit study.

Intended Use

This repository is intended for research purposes, including:

• reproducing MedStyleAudit experiments;
• verifying experimental results;
• studying shortcut learning in medical imaging;
• analyzing domain and acquisition-style effects;
• investigating dataset bias;
• studying model robustness under domain shift.

The contents of this repository are intended for research use only.

They should not be interpreted as clinical predictions and should not be used for diagnosis, treatment, or other clinical decision-making.

Reproducibility

The artifacts in this repository are designed to be used together with the MedStyleAudit source code.

For exact reproduction of an experiment, users should use the same:

• dataset version;
• dataset split;
• source and target domain definitions;
• model configuration;
• random seed;
• preprocessing procedure;
• evaluation protocol.

These factors may affect the final reported results.

Limitations

The results and artifacts in this repository reflect the datasets, models, domains, and experimental settings evaluated in MedStyleAudit.

Observed shortcut behavior may not generalize to:

• other medical imaging modalities;
• other institutions;
• other acquisition devices;
• different preprocessing pipelines;
• different model architectures;
• real-world clinical deployment environments.

The presence or absence of a detected shortcut should therefore be interpreted within the context of the corresponding experimental setup.

Ethical Considerations

This repository is intended to support research on robustness, shortcut learning, and dataset bias in medical imaging.

Any use involving medical data should comply with the licensing, privacy, ethical, and institutional requirements associated with the original datasets.

No result contained in this repository should be treated as a substitute for clinical evaluation.

Citation

If you use this repository or MedStyleAudit in your research, please cite the corresponding MedStyleAudit paper.

Citation information will be updated after the public release of the paper.

Contact

For questions regarding the experiments, repository structure, or reproduction procedure, please contact the authors or open an issue in the corresponding MedStyleAudit source-code repository.