Revolutionizing Skin Cancer Detection with iToBoS Dataset

In the realm of skin cancer diagnosis, artificial intelligence has revolutionized the field by enhancing the detection of malignant lesions with speed and accuracy. While existing datasets have been crucial for developing diagnostic algorithms, they often lack the context of surrounding skin, hindering comprehensive lesion detection. To address this gap, the iToBoS dataset was created, containing 16,954 skin region images from 100 participants captured using 3D total body photography. Each image represents a section of skin with suspicious lesions annotated using bounding boxes, along with metadata like anatomical location and age group.

Skin cancer, including Melanoma, Basal Cell Carcinoma, and Squamous Cell Carcinoma, is a prevalent global issue, underscoring the need for effective diagnostic solutions. Traditionally, dermatologists relied on dermatoscopy for diagnosis, but AI-based tools now play a vital role in differentiating between benign and malignant skin conditions. Dermatoscopic images dominate publicly available datasets, but the reliance on specialized equipment poses challenges for widespread adoption. 3D total-body photography offers a holistic view of the skin, facilitating comprehensive lesion analysis and disease monitoring over time.

The Canfield VECTRA WB360 system, used for 3D total-body photography, automates lesion detection. However, challenges like false positives and over-identification of solar damage lesions persist. To address this, the iToBoS dataset comprises high-resolution images from two clinical sites, providing a diverse set of lesions across various anatomical regions. Each image is annotated with bounding boxes enclosing lesions, alongside metadata like age and sun damage score, enhancing AI model training for lesion detection.

The dataset’s creation involved ethical authorizations, consent acquisition, data collection, annotation, and public subset selection. Annotators meticulously labeled lesions, tattoos, and sun damage levels, with dermatologists ensuring annotation accuracy. The dataset, available under a CC-BY 4.0 license, includes categorized image tiles and annotations in YOLO and COCO formats. It offers a valuable resource for training AI models in lesion detection, with detailed metadata enhancing lesion analysis in clinical and non-clinical settings.

The meticulous validation process by dermatologists ensures the dataset’s precision and reliability. With detailed usage instructions and scripts provided for data processing, the iToBoS dataset stands as a significant advancement in skin lesion detection datasets. Researchers can leverage this resource to develop AI algorithms for early skin cancer detection, ultimately improving patient outcomes in dermatological practice.