AI-driven diagnosis of mpox using deep learning models

Area of Science:

• Dermatology
• Computer Science
• Medical Imaging

Background:

• Mpox lesions share visual similarities with other skin conditions, necessitating reliable image-based diagnostic tools.
• Existing studies on mpox image classification are difficult to compare due to variations in dataset creation, augmentation strategies, and evaluation methods.
• Developing a standardized and reproducible benchmark is crucial for advancing automated mpox detection.

Purpose of the Study:

• To establish a leakage-aware benchmark for binary mpox classification.
• To compare the performance of various pretrained deep learning models and a weighted ensemble.
• To provide a trustworthy reference dataset and evaluation methodology for mpox image analysis.

Main Methods:

• A unified dataset was created from MSLD v1.0 and v2.0, ensuring data integrity.
• Seven pretrained deep learning backbones and a weighted ensemble were evaluated using group-stratified five-fold cross-validation.
• Evaluation included original-only testing, validation-based threshold selection, and temperature scaling for model calibration.

Main Results:

• The weighted ensemble model achieved a mean accuracy of 0.8729, an F1-score of 0.8334, and an AUC of 0.9388.
• ConvNeXt-Tiny demonstrated the strongest performance among single models, with an F1-score of 0.8159 and an AUC of 0.9284.
• Results are presented as conservative, trustworthy reference values due to rigorous evaluation protocols.

Conclusions:

• The study provides a transparent and reproducible benchmark for mpox classification, emphasizing dataset curation and grouped evaluation.
• The developed benchmark offers a reliable reference point, though external clinical validation is required for diagnostic deployment.
• Methodological transparency and reproducible dataset curation are highlighted as key limitations and contributions to the field.