ISIC2016

This builder implements ISIC 2016 Task 3: Lesion Classification, preserving the official 900-image training and 379-image test sets. Each row contains image (PIL image), label (integer), and image_id (string). Original JPEG bytes, dimensions, labels, and source splits are retained.

Fig. 6 Three examples per class from the official training set, selected by image ID. Aspect ratios and original labels are preserved; images are not cropped.

Data Structure

Key	Type	Description
image	PIL.Image.Image	RGB PIL image, variable dimensions
label	int	Integer: 0=benign, 1=malignant
image_id	str	Original filename stem, unique within the dataset

Usage

from stable_datasets.images import ISIC2016

datasets = ISIC2016()
train, test = datasets["train"], datasets["test"]
print(train.features["label"].names)  # ["benign", "malignant"]

datasets = ISIC2016(download_dir="/data/downloads",
                    processed_cache_dir="/data/processed")

The official CSV files have no header. Training labels use benign and malignant; test labels use 0.0 and 1.0. Both map to 0=benign and 1=malignant. Labels are joined by image ID, independently of ZIP/CSV order.

Testing

The tests download the original release through the standard dataset API and check split sizes, class counts, image decoding, stable IDs, and cache reopening. Offline parser tests additionally check invalid inputs and label matching.

Source and license

Official task: https://challenge.isic-archive.com/landing/2016/39/

Official data index: https://challenge.isic-archive.com/data/

The 2016 challenge data are listed under CC0. Measured SHA-256 values in SOURCE identify the audited release; they are not publisher signatures.

Citation

@article{gutman2016skin,
    title={Skin Lesion Analysis toward Melanoma Detection: A Challenge at the International
           Symposium on Biomedical Imaging (ISBI) 2016, hosted by the International Skin
           Imaging Collaboration (ISIC)},
    author={Gutman, David and Codella, Noel C. F. and Celebi, Emre and Helba, Brian
            and Marchetti, Michael and Mishra, Nabin and Halpern, Allan},
    journal={arXiv preprint arXiv:1605.01397}, year={2016}
}

Known exact-pixel duplicates are retained as source records. Evaluation code should group duplicate images within partitions and avoid train/test leakage.