-2

I want to know the possible measures that can be used to measure the accuracy of my image segmentation algorithm

I know the following methods:

  • sensitivity
  • specificity
  • Dice index
  • Jaccard index

But I am working with large data set that needs to be tested, So I think measuring accuracy with each pixel value will not be very suitable. So, I am looking for other more convenient measures.

P.S: I have ground truth data

Ahmed Hassaan
  • 99
  • 1
  • 12
  • Do you have ground truth data ? –  Oct 19 '17 at 14:09
  • Yes of course I have my ground truth data – Ahmed Hassaan Oct 20 '17 at 12:13
  • So what aspect of an "accurate" segmentation do you want to promote ? –  Oct 20 '17 at 12:37
  • I want to compare my segmentation to the truth value, and see how much correct is my segmentation. the ground truth is the actual true value that I compare my work with – Ahmed Hassaan Oct 20 '17 at 12:58
  • You don't answer my question. Of course you want to compare to the ground truth. You want to compare what ? What do you call accuracy ? –  Oct 20 '17 at 13:13
  • I have an image from which I want to segment an object. I want to know the accuracy of my segmentation, how many pixels I segmented correctly from the total pixels. is that what you mean ? – Ahmed Hassaan Oct 21 '17 at 09:39
  • Mh, if you define the accuracy as... the accuracy, we can't go very far. What is the purpose of your segmentation ? –  Oct 21 '17 at 09:58
  • I actually don't get your point, the purpose is to segment the lung from CT images – Ahmed Hassaan Oct 21 '17 at 11:54
  • Let us [continue this discussion in chat](http://chat.stackoverflow.com/rooms/157184/discussion-between-ahmed-hassaan-and-yves-daoust). – Ahmed Hassaan Oct 21 '17 at 12:02

1 Answers1

3

The question of "accuracy" in image segmentation is ill-defined, resulting in a number of possible metrics that will not necessarily agree with each other (e.g. segmentation A may be better than segmentation B for one metric and the reverse may be true for a different metric).

As a starting point, this paper discusses 20 different possible image segmentation metrics which could act as a starting point for your request of "possible measures that can be used". Which metric you use will be dependent on what you're segmenting and how you intend to use that segmentation.

JMikes
  • 572
  • 4
  • 12