HomeWhy Challenges?All ChallengesCreate Your Own ChallengeContributorsForum
Sign in / Register

Data Description

Overall, the sub-challenge can be divided in two different tasks:

1.) Segmentation: Can you segment the surgical instruments in an endoscopic image? 

2.) Tracking: Can you track the surgical instruments in an endoscopic image sequence? 

Data is provided for rigid  and articulated robotic instruments in laparoscopic surgery. For rigid instrument trakcking, the data reflects typical challenges in endoscopic vision like occlusion, smoke and bleeding. For robotic instrument tracking, the instruments  show typical poses and articulation in robotic surgery and there is some occlusion but no smoke and bleeding in any sequence.

The data is split into a training and test set (see below). Please read the instructions in the  "Readme" file provided with each data set.

Training Data

 

Rigid Instruments

Robotic Instruments
Segmentation

40 2D in-vivo images from 4 laparoscopic colorectal surgeries. Each pixel is labelled as either background, shaft and manipulator (~160 2D images and annotations in total).

Download

4 45-second 2D images sequences of at least one Large Needle Driver instrument in an ex-vivo setup. Each pixel is labelled as either backgroud, shaft, head or clasper. 

Download

Tracking
45 seconds in-vivo video sequences of 4  laparoscopic colorectal surgeries together with pixel-wise labelling of the instrument, (x/y) position of the center point of the end of the shaft and instrument axis in one frame every second as reference (~4500 images and 180 annotated images) 

Download

4 45-second 2D images sequences of at least one Large Needle Driver instrument in an ex-vivo setup. For each frame the associated 2D pose (rotation, translation and articulation of head and claspers) is provided.

Download

 

Test Data

 

Rigid Instruments

Robotic Instruments
Segmentation
  • 10 additional 2D images for each of the 4 recorded laparoscopic surgeries provided for training
  • 2 additional recorded surgeries with 50 2D images

Download

  • 4 15 second sequences corresponding to the 45 second sequences of training data.
  • 2 1-minute 2D images sequence of 2 instruments in an ex-vivo setup (at least 1 will not be a large needle driver). The sequence will contain some small occlusions and articulation. 

Download  

Tracking
  • 15 sec additional video sequences for each of the 4 recorded laparoscopic surgeries provided for training
  • 2 additional recorded surgeries (one minute sequences)

Download

  • 4 15 second sequences corresponding to the 45 second sequences of training data.
  • 2 1-minute 2D images sequence of 2 instruments in an ex-vivo setup (at least 1 will not be a large needle driver). The sequence will contain some small occlusions and articulation. 

Download

 

Reference Standard

For each training and testing image, the reference standard is giving by a pixel-wise labelling where each pixel is either classified as instrument shaft, instrument manipulator or background. 
 
 
 

Rigid Instruments

Robotic Instruments
Segmentation

For evaluating the instrument segmentation, the DICE coefficient between the reference and the submitted result is used. Furthermore, typical classification metrics like precision and recall are calculated.

For evaluating the instrument segmentation, the DICE coefficient between the reference and the submitted result is used. Furthermore, typical classification metrics like precision and recall are calculated.
Tracking
For evaluating the rigid instrument tracking, the reference segmentation is used. As error measure, the Euclidean distance between center point of the end of the shaft and the estimate of this point is used. Furthermore, the angle between the axis of the instrument and the estimated axis is used. 
 
 

For evaluation robotic instrument tracking, the Euclidean distance between the end of the shaft and the estimate of the shaft end is calculated as well as the rotation of the shaft. To assess the articulation estimation accuracy the absolute distance of rotation error is used.

 


Consortium for Open Medical Image Computing © 2012-