Endoscopic Content Area (ECA) dataset

A simple python loader for the Endoscopic Content Area (ECA) dataset. An implementation of the hausdorff distance, optimised for content areas, is also included in the package. The dataset was created and used to test our content area detection package torchcontentarea. Both the dataset and detection algorithm are released alongside our publication:

Rapid and robust endoscopic content area estimation: A lean GPU-based pipeline and curated benchmark dataset
Charlie Budd, Luis C. Garcia-Peraza-Herrera, Martin Huber, Sebastien Ourselin, Tom Vercauteren.
[ arXiv ] [ publication ]

If you make use of this work, please cite the paper.

Installation

To use this dataset, first ensure you have a synapse account, then simply install from pip...

pip install ecadataset

and run the download command...

ecadataset download -d path/to/dataset

You'll be prompted for your synapse credentials and the data will be downloaded. You may also check an existing copy of the dataset with the check command...

ecadataset check -d path/to/dataset

Usage

import matplotlib.pyplot as plt
from ecadataset import ECADataset, DataSource, AnnotationType, content_area_hausdorff

# Create dataset object...
dataset = ECADataset(
  # Path to the directory containing the dataset.
  data_directory="path/to/dataset",
  # Options are: DataSource.CHOLEC, DataSource.ROBUST, and DataSource.BOTH.
  data_source=DataSource.BOTH,
  # Options are: AnnotationType.AREA, AnnotationType.MASK, and AnnotationType.BOTH.
  annotation_type=AnnotationType.BOTH,
  # Whether to use cropping to provide additonal samples without a content area.
  include_cropped=True,
  # Whether to include information about where the frame was taken from.
  include_source_info=True
)

# Iterate through the first 10 samples, slicing is supported...
for image, area, mask, info in dataset[:10]:

    # Circular content area represented as (x, y, r) or None if no area present...
    print("Content area: ", area)
    
    # Origin information in the form (dataset, video, frame)...
    print("Sample source: ", info)
    
    # Image and mask are returned as PIL images...
    plt.subplot(121)
    plt.imshow(image)
    plt.subplot(122)
    plt.imshow(mask)
    plt.show()
    
    # Guessing the content area circle and scoring it against the ground truth...
    width, height = image.size
    area_guess = (width//2, height//2, width//2)
    score, _ = content_area_hausdorff(area_guess, area, (height, width))
    print(score)