Corrected Labels for DermaMNIST, HAM10000, and Fitzpatrick17k

Corrected versions of DermaMNIST, HAM10000, and Fitzpatrick17k dermatology datasets address data leakage and duplication issues identified in the original benchmarks. Created by kakumarabhishek and updated in February 2025, this collection provides refined data for skin disease classification. It focuses on improving the integrity of medical image computing benchmarks through rigorous quality analysis.

Use Cases

• Benchmarking melanoma-detection models using deduplicated HAM10000 image sets
• Evaluating skin-disease-classification accuracy using corrected Fitzpatrick17k labels to ensure valid generalization
• Training deep-learning models for computer-aided-diagnosis without the risk of training-to-test set leakage

Strengths

• Removes duplicate entries from the HAM10000 dataset
• Mitigates data leakage issues found in the Fitzpatrick17k dataset
• Released under the Apache-2.0 open-source license

Limitations

• Specific record counts for the corrected subsets are not provided in the primary metadata
• Inherits any underlying sensor noise or bias present in the original source images

Provenance

Source

DermaMNIST (MedMNIST), HAM10000, and Fitzpatrick17k

Collection Method

Data quality analysis and algorithmic correction of existing public dermatology datasets

Freshness

Last updated February 13, 2025.