TLDR: The eSkinHealth research introduces a new multimodal dermatological dataset collected in Côte d’Ivoire and Ghana, specifically addressing the scarcity of data for Neglected Tropical Skin Diseases (NTDs) in West African populations. Comprising 5,623 images from 1,639 cases across 47 skin diseases, the dataset includes patient metadata, diagnosis labels, semantic lesion masks, visual captions, and clinical concepts. It leverages an innovative AI-expert collaboration paradigm for efficient and high-quality annotation, aiming to foster the development of more equitable, accurate, and interpretable AI diagnostic tools for global dermatology.
Neglected Tropical Diseases (NTDs) of the skin represent a significant health and economic challenge, particularly in impoverished tropical regions. These conditions lead to substantial mortality, morbidity, and profound social consequences, including stigmatization and financial strain on affected communities. Despite the urgent need for efficient diagnostic tools, the development of AI-driven solutions has been hampered by a critical lack of diverse and high-quality data, especially for underrepresented populations and the unique manifestations of NTDs.
Addressing this crucial data gap, researchers from Tulane University, Université Alassane Ouattara, Université de Bouaké, and Bakke Graduate University have introduced eSkinHealth, a groundbreaking multimodal dermatological dataset. This novel resource was meticulously collected on-site in Côte d’Ivoire and Ghana, focusing specifically on skin NTDs and rare conditions prevalent among West African populations.
The eSkinHealth dataset is comprehensive, featuring 5,623 images from 1,639 distinct cases, encompassing 47 different skin diseases. Beyond just images, it includes rich annotations such as patient metadata, confirmed diagnosis labels, semantic lesion masks, instance-specific visual captions, and clinical concepts. This wealth of information is designed to provide a robust foundation for developing advanced AI models.
A key innovation of this work is the proposed AI-expert collaboration paradigm for generating these multimodal annotations. This process leverages the power of foundation models, such as Multimodal Large Language Models (MLLMs) for captions and concepts, and the Segment Anything Model (SAM) for semantic masks. These AI-generated annotations are then rigorously guided and verified by experienced dermatologists, ensuring high quality and clinical accuracy. This synergistic approach significantly enhances the efficiency and scalability of data annotation, a traditionally labor-intensive process in medical imaging.
The data collection itself involved local nurses and health workers, specially trained in dermatological photography, capturing images and clinical information in rural clinics across six health districts known for skin NTDs. Diagnostic accuracy was ensured through independent evaluations by multiple dermatologists, with consensus achieved through discussion, and further confirmed by clinical tests like PCR for Buruli ulcer and DPP rapid tests for yaws in specific cases.
The eSkinHealth dataset is a valuable new resource that aims to catalyze the development of more equitable, accurate, and interpretable AI tools for global dermatology. It provides a unique opportunity for researchers to train and benchmark AI models on real-world data from affected communities, addressing the specific challenges of NTDs. Initial experiments with state-of-the-art models have already highlighted the complex nature of classifying these conditions, underscoring the dataset’s utility in pushing the boundaries of AI diagnostics.
Also Read:
- PlantVillageVQA: A New Benchmark for Agricultural AI
- Advanced Medical AI: Hierarchical Vision-Language Learning for Out-of-Distribution Disease Detection
This work represents a significant step towards democratizing AI in dermatology, with future efforts planned to expand the dataset, explore more sophisticated modeling techniques, and investigate pathways for clinical validation and deployment of AI tools in endemic regions. For more details, you can refer to the full research paper available here.
