Facial-Relevant Dermatology Image Datasets for AI: Biases, Benchmarking, and Readiness for Vision-Language Models
IEEE Access, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Derleme
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3732560
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Anahtar Kelimeler: computer vision, facial dermatology, medical imaging, skin disease datasets, vision language models
- Galatasaray Üniversitesi Adresli: Hayır
Özet
Artificial intelligence (AI) systems for dermatology increasingly rely on heterogeneous image resources, yet facial images combine diagnostic value with biometric identifiability, acquisition variability, and psychosocial sensitivity. Searches through March 2026 identified 60 candidate datasets or resources. After source, relevance, accessibility, and overlap screening, 25 core entries were retained for structured comparison; 2 additional resources were discussed for taxonomy or emerging-method context, and 33 remained supplementary-only. Among the 25 core entries, only 3 are explicitly facial-specific, 12 are broad clinical resources in which facial involvement is plausible but not guaranteed, and 10 are dermoscopic or total-body-photography benchmarks with indirect facial relevance. Fourteen entries have peer-reviewed primary or official challenge evidence, three are official-atlas or emerging resources, and eight are community or derivative sources requiring stronger provenance checks. We compare scale, modality, annotation, metadata, licensing, demographic reporting, task suitability, and benchmarking practice, and assess readiness for vision-language model development. Persistent limitations include under-representation of darker skin tones, incomplete body-site and demographic metadata, inconsistent taxonomies, weak external validation, and inappropriate cross-task metric comparison.We conclude with recommendations for modality-specific benchmarking, clinician review, privacy-preserving governance, and clinically validated multimodal annotation.