How to install datasets

We recommend placing all datasets in a single folder. Follow the instructions below to organize the datasets without needing to change the source code. The file structure is as follows:

$DATA/
|–– NCT-CRC-HE-100K/
|–– SICAP_MIL/
|–– LC25000/
|–– skincancer/

Detailed instructions for preparing each dataset are provided below. To ensure reproducibility and enable fair comparisons in future work, we offer fixed train/validation/test splits for all datasets. These fixed splits are either taken from the original datasets (when available) or created by us.

NCT

• Create a folder named NCT-CRC-HE-100K/.
• Download the the train set NCT-CRC-HE-100K.zip and val set CRC-VAL-HE-7K.zip from the official website and extract them under NCT-CRC-HE-100K/. The directory structure should look like:

NCT-CRC-HE-100K/
|–– CRC-VAL-HE-7K/
|–– NCT-CRC-HE-100K/

SICAP_MIL

• Create a folder named SICAP_MIL/.
• Download the the dataset from the official github and extract it under SICAP_MIL/. The directory structure should look like:

SICAP_MIL/
|–– annotation_masks/
|–– dataframes/
|–– patches/
|–– slides/

LC_LUNG

• Create a folder named LC25000/.
• Download the the dataset from the official website and extract it under LC25000/. The directory structure should look like:

LC25000/
|–– colon/
|–– lung/

Here we only use the lung subset.

skincancer

• Create a folder named skincancer/.
• Download the dataset from the official website and extract it under skincancer/. The directory structure should look like:

skincancer/
|–– data/
|   |–– class_dict.json 
|   |–– tiles/
|   |–– tiles-v2.csv