Source-linked AI summary
BCN20000: Dermoscopic Lesions in the Wild
Marc Combalia, Noel C. F. Codella, Veronica Rotemberg, Brian Helba, Veronica Vilaplana, Ofer Reiter, Cristina Carrera, Alicia Barreiro, Allan C. Halpern, Susana Puig, Josep Malvehy
TL;DR
The paper addresses unconstrained classification of dermoscopic skin-cancer images, especially difficult lesions encountered outside standard settings. It constructs and curates the BCN20000 dataset from hospital records, providing diagnostic categories and clinical metadata for ISIC 2019 evaluation, including out-of-distribution detection.
Problem
The paper targets unconstrained dermoscopic classification of difficult skin-cancer lesions, including lesions in nails and mucosa, non-segmentable lesions, and hypopigmented lesions.
Method
The authors retrieve, organize, filter, diagnosis-link, and manually review hospital dermoscopic images collected from 2010 to 2016, producing BCN20000 with clinical metadata.
Results
BCN20000 contains 19424 high-quality dermoscopic images corresponding to 5583 skin lesions across multiple diagnostic categories.
Takeaways & Limitations
The dataset supports ISIC 2019 tasks for diagnostic classification and identification of out-of-distribution skin lesions, with access through the ISIC Archive.
Abstract
from arXiv · showhide
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.
1 Background and Summary
The paper motivates dermoscopic image analysis for skin cancer and introduces BCN20000 to support unconstrained classification of lesions encountered in clinical practice, including difficult cases.
- Dermoscopy removes surface reflection and reveals deeper lesion structures, improving dermatologists’ visualization and diagnostic accuracy.
- Deep-learning algorithms for dermoscopic classification have advanced through yearly ISIC challenges, but performance can decline on images from other datasets.
- BCN20000 targets unconstrained classification of difficult lesions, including cases in nails and mucosa, non-segmentable lesions, and hypopigmented lesions.
- Most BCN20000 images are considered hard to diagnose and were excised and histopathologically diagnosed, with anatomical location, patient age, and sex provided as metadata.
2 Methods
BCN20000 was assembled from hospital dermoscopic records captured over several years, then computationally organized, diagnosis-linked, and manually reviewed to produce a curated image database.
- The source images were collected over more than 16 years by the Hospital Clínic de Barcelona Department of Dermatology.
- Figure 1 presents example images spanning nevus, melanoma, basal cell carcinoma, seborrheic keratosis, actinic keratosis, squamos cell carcinoma, dermatofibroma, and vascular lesion.
- Images captured from 2010 until 2016 on three high-resolution cameras were retrieved, organized, and filtered using computer vision algorithms.
- Diagnoses were linked through a reference database and manually reviewed by several readers for plausibility.
- The resulting database contains 19424 high-quality dermoscopic images corresponding to 5583 skin lesions.
- Figure 2 reports image counts by diagnosis confirmation type, including single image expert consensus.
3 Usage Notes
BCN20000 covers multiple diagnostic categories and pairs each image with clinical metadata, while supporting both diagnostic classification and detection of out-of-distribution lesions in the ISIC 2019 Challenge.
- The dataset includes nevus, melanoma, basal cell carcinoma, seborrheic keratosis, actinic keratosis, squamos cell carcinoma, dermatofibroma, vascular lesion, and other lesions.
- Each image is coupled with metadata describing anatomical location and the patient’s age and sex.
- In the ISIC 2019 Challenge, participants classify diagnostic categories and identify out-of-distribution skin lesions.
- The dataset will also be made available through the ISIC Archive.