CathAction Dataset

CathAction is large-scale dataset designed for advancing catheterization understanding. CathAction comprises annotated frames focused on catheterization understanding and collision detection, along with groundtruth masks dedicated to catheter and guidewire segmentation.

Please fill out the download form and agree to our license prior to downloading the dataset.

Dataset Structure:

1. Catheterization Action understanding

The CathAction dataset encompasses annotated frames for catheterization action understanding task such as catheterization anticipation and action recognition.

These are five classes: advance catheter, retract catheter, advance guidewire, retract guidewire, and rotate.

The dataset is organized into the following folders and files:

• video_frames/: Contains extracted video frames for each video.
• feature_extractions/: Contains pre-extracted RGB features, extracted using this code.
• training.csv: Groundtruth CSV file for training data.
• validation.csv: Groundtruth CSV file for validation data.

Annotation File Structure

The annotation files (training.csv and validation.csv) contain four columns, with the following structure:

Column Name	Type	Example	Description
video_id	string	video_1	ID of the video where the action segment is located.
start_frame	int	430	Start frame of the action.
stop_frame	int	643	End frame of the action.
all_action_classes	list of int(s)	[1]	List of numeric IDs for all detected action classes in the segment.

The frames and pre-extracted RGB features are located in the video_frames and feature_extractions folders, respectively, and can be generated using this code.

Usage

1. Catheterization Action Recognition and Anticipation Models: Use the start_frame and stop_frame values, along with the ground truth all_action_classes in the CSV file, to train models that recognize action segments and anticipate future catheter actions.

2. Collision Detection

The CathAction dataset is designed for the collision detection task, which involves identifying whether the tip of the catheter or guidewire comes into contact with the blood vessel wall.

The dataset is organized as follows:

• images/: Contains images related to collision and normal events.

• labels/: Contains annotation files for each image, detailing information on bounding boxes and object classes, including collision occurrences and the normal class for the corresponding image

• train_phantom.txt: A text file listing paths to training images and labels for the "phantom" data source in the collision detection task.

• valid_animal.txt: A text file listing paths to validation images and labels for the "animal" source data.

• valid_phantom.txt: A text file listing paths to validation images and labels for the "phantom" source data.

Each .txt file contains a list of image and label paths for its respective category and split (train/validation), enabling easy access and organization for model training and evaluation.

Usage

1. Training: Use train_phantom.txt to load training data for the phantom data source.
2. Validation: Use valid_animal.txt and valid_phantom.txt for validating model performance on different data sources, specifically focusing on the 'animal' and 'phantom' data.

For more information, please visit our webpage.

For inquiries or assistance, please contact the authors at this link.

Best regards,

Authors