FitzDerm-CF: Controlled Multimodal Dermatology Resource for Skin-Tone Counterfactuals

FitzDerm-CF is a controlled multimodal dermatology resource augmenting images from Fitzpatrick17k and DDI with synthetic clinical notes. The resource, created by Shivam Jangid and hosted on Harvard Dataverse, was last updated on June 5, 2026. It is designed for systematic evaluation of semantic capacity, diagnostic leakage, and skin-tone counterfactual robustness in AI models.

Use Cases

• Benchmarking cross-modal retrieval models based on paired skin images and synthetic clinical notes.
• Probing diagnostic label leakage in multimodal classifiers based on controlled text generation.
• Evaluating counterfactual consistency of models using the skin-type-perturbed text and image splits.
• Developing and auditing dermatology AI systems under controlled distributional conditions.

Strengths

• Provides a controlled resource with synthetic clinical notes generated under strict constraints to prevent explicit disease naming and demographic leakage.
• Includes a counterfactual split constructed by perturbing only the Fitzpatrick skin-type attribute, validated through a four-stage pipeline.
• Notes are generated across multiple temperature settings, enabling studies of text informativeness with different levels of morphological specificity.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Harvard Dataverse, author Shivam Jangid.

Collection Method

Augments images from Fitzpatrick17k and DDI datasets with synthetically generated dermatology-style clinical notes.

Freshness

Last updated 2026-06-05 11:07:10; freshness should be verified.