ENID: The ITEC Endometrial Implants Dataset

ENID (ENdometrial Implants Dataset) comprises 160 images taken from 100+ gynecologic laparoscopy surgeries and is purposefully created to be utilized for a variety of automatic content analysis problems in the context of Endometriosis recognition. ENID is the result of experiments on reorganizing previously published dataset GLENDA in terms of visual similarity.

Description

Endometriosis

Endometriosis is a benign but potentially painful condition among women in child bearing age involving the growth of uterine-like tissue in locations outside of the uterus. Corresponding lesions can be found in various positions and severities, often in multiple instances per patient requiring a physician to determine its extent. This most frequently is accomplished by calculating its magnitude via utilizing the combination of two popular classification systems, the revised American Society for Reproductive Medicine (rASRM) and the European Enzian scores. Endometriosis can not reliably identified by laymen, therefore, the dataset has been created with the help of medical experts in the field of endometriosis treatment.

Purposes

Overview

The dataset includes region-based annotations of a specific visual endometriosis manifestation: dark endometrial implants. Annotations are created for single video frames, which can contain multiple annotations. Each single annotation is exported as a pixel-based image mask: the annotated regions are colored, while the background is left black. Additionally a particular dataset augmentation is provided: ground truth annotations are tracked over time using the original videos. The tracking results are acquired via kernelized correlation filters (KCF).

Disclaimer

The dataset is exclusively provided for scientific research purposes and as such cannot be used commercially or for any other purpose. If any other purpose is intended, you may directly contact the originator of the videos, Prof. Dr. Jörg Keckstein.

In addition, reference must be made to the following publication when this dataset is used in any academic and research reports:

A. Leibetseder, K. Schoeffmann, J. Keckstein and S. Keckstein. 2021. Endometriosis Detection and Localization in Laparoscopic Gynecology. Accepted: TBA.

BibTeX:

@article{DBLP:journals/mta/LeibetsederSKK22,
    author    = {Andreas Leibetseder and
                 Klaus Schoeffmann and
                 J{\"{o}}rg Keckstein and
                 Simon Keckstein},
    title     = {Endometriosis detection and localization in laparoscopic gynecology},
    journal   = {Multim. Tools Appl.},
    volume    = {81},
    number    = {5},
    pages     = {6191--6215},
    year      = {2022},
    url       = {https://doi.org/10.1007/s11042-021-11730-1},
    doi       = {10.1007/s11042-021-11730-1},
    timestamp = {Thu, 03 Mar 2022 09:23:23 +0100},
    biburl    = {https://dblp.org/rec/journals/mta/LeibetsederSKK22.bib},
    bibsource = {dblp computer science bibliography, https://dblp.org}
}

ENID is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0, Creative Commons License) and is created as well as maintained by Distributed Multimedia Systems Group of the Institute of Information Technology (ITEC) at Alpen-Adria Universität in Klagenfurt, Austria.

This license allows users of this dataset to copy, distribute and transmit the work under the following conditions:

For further legal details, please read the complete license terms.

Dataset

Following table shows dataset statistics and provides preview/download links. Please note that due to the large number of images in the tracked dataset, it can only be previewed once downloaded.

IMPORTANT: By downloading, you agree to and adhere to above conditions.

dataset cases frames annotations size preview download
ENID 108 160 358 12MB
tracked ENID 108 12 533 18 275 973MB -

Splits

Below both of accompanying research's splits for untracked ENID can be downloaded, i.e. 60/20/20 and 80/10/10. For split preview to work, please extract either of archives into the main dataset folder.

IMPORTANT: By downloading, you agree to and adhere to above conditions.

Legend -- c: cases, f: frames, a: annotations
split training validation testing size download
c f a preview c f a preview c f a preview
60/20/20 64 96 240 24 32 61 20 32 57 83KB
80/10/10 83 128 279 12 16 32 13 16 47 81KB

Models

Below table lists the performance of the two best augmented pre-trained Mask R-CNN models of both above splits in comparison to the respective best raw (unaugmented) model. Additionally, a comparison view per split gives insights over the models' qualitative performance. All models have been trained using Detectron2 and the initial model weights are set according to selected models from the Detectron2 Model Zoo (referenced by 'base model' column).

IMPORTANT: By downloading, you agree to and adhere to above conditions.

Legend -- AP: mean average segmentation precision at intersection over union (IoU) overlaps {50-95%, 50%, 75%},
backbone networks: ResNet (R) 50/101 + Feature Pyramid Network (FPN)
split augmentations base model epoch AP AP50 AP75 size download comparison
60/20/20 crop, rotate R101-FPN 29 0.324 0.642 0.255 454MB
rotate R101-FPN 24 0.320 0.640 0.255 454MB
none R50-FPN 39 0.309 0.581 0.325 312MB
80/10/10 crop R101-FPN 44 0.277 0.545 0.258 315MB
crop, rotate R101-FPN 49 0.253 0.472 0.228 451MB
none R50-FPN 39 0.250 0.522 0.195 315MB

Demo

A demo application has been published at CBMI 2021 and can be found at https://github.com/amplejoe/EndometriosisSegmentationTool.