[Paper Review] BCN20000: Dermoscopic Lesions in the Wild

Marc Combalia, Noel Codella+9 more|arXiv (Cornell University)|Aug 6, 2019

Cutaneous Melanoma Detection and ManagementMedicine10 references294 citations

TL;DR

The paper introduces the BCN20000 dataset of 19,424 dermoscopic images from 5,583 skin lesions (2010–2016) to study unconstrained skin cancer classification and to support ISIC Challenge 2019.

ABSTRACT

This article summarizes the BCN20000 dataset, composed of 19424 dermoscopic images of skin lesions captured from 2010 to 2016 in the facilities of the Hospital Clínic in Barcelona. With this dataset, we aim to study the problem of unconstrained classification of dermoscopic images of skin cancer, including lesions found in hard-to-diagnose locations (nails and mucosa), large lesions which do not fit in the aperture of the dermoscopy device, and hypo-pigmented lesions. The BCN20000 will be provided to the participants of the ISIC Challenge 2019, where they will be asked to train algorithms to classify dermoscopic images of skin cancer automatically.

Motivation & Objective

• Motivate study of unconstrained dermoscopic image classification beyond well-curated datasets.
• Provide a large, clinically diverse dataset including hard-to-diagnose locations and lesion types.
• Link images with metadata (anatomic location, patient age and sex) to mimic clinical practice.
• Enable evaluation of algorithms on out-of-distribution and challenging scenarios.
• Support the ISIC Challenge 2019 with a robust public dataset linked to diagnoses.

Proposed method

• Assembled a 16-year collection of dermoscopic images from Hospital Clínic de Barcelona (2010–2016).
• Retrieved, organized, and filtered images with computer vision techniques and linked them to diagnoses in a reference database.
• Manually revised diagnoses by multiple readers to ensure plausibility and quality.
• Included metadata on lesion location, age, and sex to reflect clinical workflow.
• Prepared the dataset for ISIC Challenge 2019 and ISIC Archive distribution.

Experimental results

Research questions

• RQ1 How does an unconstrained dermoscopic image dataset, including hard-to-diagnose locations, affect automated skin cancer classification performance?
• RQ2 Can algorithms handle out-of-distribution or non-segmentable/hypopigmented lesions present in the wild?
• RQ3 What is the value of accompanying metadata (location, age, sex) for lesion classification in practice?
• RQ4 How does BCN20000 compare to existing datasets in diversity and diagnostic categories?

Key findings

• The BCN20000 dataset comprises 19,424 dermoscopic high-quality images corresponding to 5,583 skin lesions captured between 2010 and 2016.
• Images cover nine categories: nevus, melanoma, basal cell carcinoma, seborrheic keratosis, actinic keratosis, squamous cell carcinoma, dermatofibroma, vascular lesion, and other.
• Images are accompanied by metadata on anatomical location and patient demographics to reflect clinical practice.
• The dataset is intended for the ISIC Challenge 2019 to assess automatic classification and out-of-distribution detection.
• Ethical approvals were obtained for the data collection and sharing.
• Images were curated and validated through routing from hospital records to ensure plausibility of diagnoses.