flwrlabs/fed-isic2019

Dataset Card for Fed-ISIC-2019

Federated version of ISIC-2019 Datasets (ISIC2019 challenge and the HAM1000 database). This implementation is derived based on the FLamby implementation.

Dataset Details

The dataset contains 23,247 images of skin lesions divided among 6 clients representing different data centers. The number of samples for training/testing per data center is displayed in the table below:

center_id	Train	Test
0	9930	2483
1	3163	791
2	2691	672
3	1807	452
4	655	164
5	351	88

Dataset Sources

• ISIC 2019 Challange website: https://challenge.isic-archive.com/landing/2019/
• HAM1000 database website: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T
• FLamby: https://github.com/owkin/FLamby/tree/main
• FLamby Fed-ISIC-2019 README: https://github.com/owkin/FLamby/blob/main/flamby/datasets/fed_isic2019/README.md
• Fed-ISIC-2019 docs: https://owkin.github.io/FLamby/fed_isic.html

Use in FL

In order to prepare the dataset for the FL settings, we recommend using Flower Dataset (flwr-datasets) for the dataset download and partitioning and Flower (flwr) for conducting FL experiments.

To partition the dataset, do the following.

1. Install the package.

pip install flwr-datasets[vision]

2. Use the HF Dataset under the hood in Flower Datasets.

from flwr_datasets import FederatedDataset
from flwr_datasets.partitioner import NaturalIdPartitioner

fds = FederatedDataset(
    dataset="flwrlabs/fed-isic2019",
    partitioners={"train": NaturalIdPartitioner(partition_by="center"),
                  "test": NaturalIdPartitioner(partition_by="center")}
)
partition_train = fds.load_partition(partition_id=0, split="train")
partition_test = fds.load_partition(partition_id=0, split="test")

# Note: to keep the same results as in FLamby, please apply the following transformation
import albumentations
import random
import numpy as np
import torch


# Train dataset transformations
def apply_train_transforms(image_input):
    print(image_input)
    size = 200
    train_transforms = albumentations.Compose(
        [
            albumentations.RandomScale(0.07),
            albumentations.Rotate(50),
            albumentations.RandomBrightnessContrast(0.15, 0.1),
            albumentations.Flip(p=0.5),
            albumentations.Affine(shear=0.1),
            albumentations.RandomCrop(size, size),
            albumentations.CoarseDropout(random.randint(1, 8), 16, 16),
            albumentations.Normalize(always_apply=True),
        ]
    )
    images = []
    for image in image_input["image"]:
        augmented = train_transforms(image=np.array(image))["image"]
        transposed = np.transpose(augmented, (2, 0, 1)).astype(np.float32)
        images.append(torch.tensor(transposed, dtype=torch.float32))
    image_input["image"] = images
    return image_input


partition_train = partition_train.with_transform(apply_train_transforms,
                                                 columns="image")

# Test dataset transformations
def apply_test_transforms(image_input):
    print(image_input)
    size = 200
    test_transforms = albumentations.Compose(
        [
            albumentations.CenterCrop(size, size),
            albumentations.Normalize(always_apply=True),
        ]
    )
    images = []
    for image in image_input["image"]:
        augmented = test_transforms(image=np.array(image))["image"]
        transposed = np.transpose(augmented, (2, 0, 1)).astype(np.float32)
        images.append(torch.tensor(transposed, dtype=torch.float32))
    image_input["image"] = images
    return image_input


partition_test = partition_test.with_transform(apply_test_transforms,
                                                 columns="image")

Dataset Structure

Data Instances

The first instance of the train split is presented below:

{
'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=224x224>,
 'center': 0,
 'label': 2
}

Data Split

DatasetDict({
    train: Dataset({
        features: ['image', 'center', 'label'],
        num_rows: 18597
    })
    test: Dataset({
        features: ['image', 'center', 'label'],
        num_rows: 4650
    })
})

Citation

When working with the Fed-ISIC-2019 dataset, please cite the original paper. If you're using this dataset with Flower Datasets and Flower, cite Flower.

BibTeX:

FLamby:

@inproceedings{NEURIPS2022_232eee8e,
 author = {Ogier du Terrail, Jean and Ayed, Samy-Safwan and Cyffers, Edwige and Grimberg, Felix and He, Chaoyang and Loeb, Regis and Mangold, Paul and Marchand, Tanguy and Marfoq, Othmane and Mushtaq, Erum and Muzellec, Boris and Philippenko, Constantin and Silva, Santiago and Tele\'{n}czuk, Maria and Albarqouni, Shadi and Avestimehr, Salman and Bellet, Aur\'{e}lien and Dieuleveut, Aymeric and Jaggi, Martin and Karimireddy, Sai Praneeth and Lorenzi, Marco and Neglia, Giovanni and Tommasi, Marc and Andreux, Mathieu},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {S. Koyejo and S. Mohamed and A. Agarwal and D. Belgrave and K. Cho and A. Oh},
 pages = {5315--5334},
 publisher = {Curran Associates, Inc.},
 title = {FLamby: Datasets and Benchmarks for Cross-Silo Federated Learning in Realistic Healthcare Settings},
 url = {https://proceedings.neurips.cc/paper_files/paper/2022/file/232eee8ef411a0a316efa298d7be3c2b-Paper-Datasets_and_Benchmarks.pdf},
 volume = {35},
 year = {2022}
}

Flower:

@article{DBLP:journals/corr/abs-2007-14390,
  author       = {Daniel J. Beutel and
                  Taner Topal and
                  Akhil Mathur and
                  Xinchi Qiu and
                  Titouan Parcollet and
                  Nicholas D. Lane},
  title        = {Flower: {A} Friendly Federated Learning Research Framework},
  journal      = {CoRR},
  volume       = {abs/2007.14390},
  year         = {2020},
  url          = {https://arxiv.org/abs/2007.14390},
  eprinttype    = {arXiv},
  eprint       = {2007.14390},
  timestamp    = {Mon, 03 Aug 2020 14:32:13 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2007-14390.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Other References

The "ISIC 2019: Training" is the aggregate of the following datasets:

BCN_20000 Dataset: (c) Department of Dermatology, Hospital Clínic de Barcelona

HAM10000 Dataset: (c) by ViDIR Group, Department of Dermatology, Medical University of Vienna; HAM10000 dataset

MSK Dataset: (c) Anonymous; challenge 2017; challenge 2018

See below the full citations:

[1] Tschandl P., Rosendahl C. & Kittler H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci. Data 5, 180161 doi.10.1038/sdata.2018.161 (2018).

[2] Noel C. F. Codella, David Gutman, M. Emre Celebi, Brian Helba, Michael A. Marchetti, Stephen W. Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, Allan Halpern: “Skin Lesion Analysis Toward Melanoma Detection: A Challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), Hosted by the International Skin Imaging Collaboration (ISIC)”, 2017; arXiv:1710.05006.

[3] Marc Combalia, Noel C. F. Codella, Veronica Rotemberg, Brian Helba, Veronica Vilaplana, Ofer Reiter, Allan C. Halpern, Susana Puig, Josep Malvehy: “BCN20000: Dermoscopic Lesions in the Wild”, 2019; arXiv:1908.02288.

Dataset Card Contact

If you have any questions about the dataset preprocessing and preparation, please contact Flower Labs.