The REAL-Colon dataset is available at this link, and the frame-wise annotations released with this work can be accessed here. Code to automatically download and process the dataset is available at GitHub repository.
Please download the dataset into the following directory: ./data/dataset/RC_dataset
Download the zip file containing all the CSV files for annotations from this link, unzip it, and save everything in ./data/dataset/RC_annotations.
Feature extraction script encodes video frames into their latent representations using a predefined encoder model. It supports augmentation and handles multiple videos in batches.
Usage:
CUDA_VISIBLE_DEVICES=0 python3 feature_extraction.py --config data/feature_extraction_1x_RC.yml
CUDA_VISIBLE_DEVICES=0 python3 feature_extraction.py --config data/feature_extraction_5x_aug_RC.ymlThese splits have been saved at ./data/dataset/RC_lists under the 4_fold and 5_fold directories.
After ensuring that feature extraction was successful, this script checks and creates a dataset for the classification TCN application. It saves a pickle file for every video in the dataset. Each pickle file contains a dictionary where "video_embeddings" is a numpy array of the embedded video features, which can be shaped [1, temporal_size, latent_size] or [n_augmentations, temporal_size, latent_size]; and a list of frame image names at key "image_names".
Usage:
python3 create_embeddings_datasets.py --config data/emb_datasets_v2_mbmmx.yml
Biffi, C., Antonelli, G., Bernhofer, S., Hassan, C., Hirata, D., Iwatate, M., Maieron, A., Salvagnini, P., & Cherubini, A. (2024). REAL-Colon: A dataset for developing real-world AI applications in colonoscopy. Scientific Data, 11(1), 539. DOI:10.1038/s41597-024-03359-0