BCN20000: Dermoscopic Lesions in the Wild

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• High-quality datasets are crucial for advancing dermatological artificial intelligence.
• Existing datasets often lack the diversity to represent real-world clinical scenarios.
• Unconstrained classification of dermoscopic images presents a significant challenge.

Purpose of the Study:

• To introduce the BCN20000 dataset, a comprehensive collection of dermoscopic images.
• To address limitations in current datasets, including challenging lesion types and locations.
• To facilitate the development and training of robust artificial intelligence models for skin cancer diagnosis.

Main Methods:

• Compiled 18,946 dermoscopic images from 2010-2016 at Hospital Clínic, Barcelona.
• Included diverse lesion types: difficult locations (nails, mucosa), large lesions, and hypo-pigmented lesions.
• Organized images into eight diagnostic categories with a ninth out-of-distribution class for testing.

Main Results:

• The BCN20000 dataset provides a broad spectrum of skin lesions for AI training.
• It encompasses challenging cases often underrepresented in other datasets.
• Baseline classifiers using state-of-the-art neural networks were presented for further research.

Conclusions:

• The BCN20000 dataset bridges the gap between AI training data and clinical practice.
• It supports the development of more accurate and versatile AI tools for dermatology.
• This resource enables further experimentation and advancement in AI-driven skin cancer detection.