Skin Lesion and Healthy Skin Dataset for 10-Class Object Detection (YOLO annotation)

This dataset was built to carry out the research described in the paper titled "Enhancing Robustness in Skin Lesion Detection: A Benchmark of 32 Models on a Novel Dataset Including Healthy Skin Images". The full dataset consists of 11,055 images curated from the ISIC archive. It is composed of 10 distinct skin lesion classes (5 benign: Pigmented Benign Keratosis (PBK), Dermatofibroma (D), Solar Lentigo (SL), Seborrheic Keratosis (SK), Nevus (N) ; and 5 malignant: Melanoma Invasive (MI), Basal Cell Carcinoma (BCC), Actinic Keratosis (AK), Melanoma in Situ (MIS), Squamous Cell Carcinoma (SCC) ) and a dedicated 'background' (BGD) class containing 1,000 images of clear, healthy skin. The dataset is balanced, with each of the classes (10 lesions + 1 background) containing between 1,000 and 1,171 images. The dataset was randomly divided into a training, validation, and testing subset (80%, 10%, and 10% of the full set, respectively). The images were annotated using bounding boxes in the YOLO format. The dataset is particularly useful for experiments in automated skin lesion detection and can be used to train and benchmark object detection models (like YOLO or RT-DETR) , with a specific focus on enhancing robustness and reducing false positives on healthy skin.

Using this dataset please cite:

Krukar, N., & Omiotek, Z. (2026). Enchancing Robustness in Skin Lesion Detection: A Benchmark of 32 Models on a Novel Dataset Including Healthy Skin Images. Applied Sciences, 16(1), 99. https://doi.org/10.3390/app16010099