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 at ./data/dataset/RC_dataset.
Download the zip file containing all the CSV files for annotations from this link and unzip it at ./data/dataset/RC_annotation.
Feature extraction script encodes video frames into their latent representations using a predefined encoder model (ResNet50 pretrained on ImageNet). It supports augmentation and handles multiple videos in batches.
Usage:
CUDA_VISIBLE_DEVICES=0 python3 ./data/feature_extraction/feature_extraction.py --config ./data/feature_extraction/ymls/feature_extraction_1x_RC.yml
CUDA_VISIBLE_DEVICES=0 python3 ./data/feature_extraction/feature_extraction.py --config ./data/feature_extraction/ymls/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